Professional Documents
Culture Documents
Industrial Electronics
Industrial Electronics
Industrial Electronics
ALARM SYSTEMS
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
ALARM SYSTEMS 393
5. poor location (e.g., alarms not in the operator’s direct to be affecting the system at a given time. The mental repre-
view and not located near associated controls and dis- sentation resulting from situation assessment, called a situa-
plays tion model, is the operator’s understanding of the specific cur-
6. insufficient salience coding (i.e., important alarms fail rent situation and is constantly updated as new information
to draw the operator’s attention) is received. To construct a situation model, operators use
7. inadequate message design (e.g., poor labeling, poor leg- their general knowledge and understanding about the system
ibility, ambiguous messages) and how it functions to interpret what they observe and to
understand its implications. This general knowledge, com-
8. poor acoustic design (e.g., masking of alarms and irri-
monly called the operator’s mental model, constitutes an in-
tating or distracting warnings)
ternal representation of the physical and functional charac-
9. inadequate identification of the absolute or relative im- teristics of the system and its operation.
portance among alarms. Based on the situation model, the operator plans responses
to the disturbance that are designed to bring the system back
To understand the reasons for these problems, it is important to a safe state. The plans are carried out, and the operator
to consider the role of alarms in human information pro- looks for feedback to indicate the success of the plans. Alarms
cessing. play a role throughout this process. When the alarm system
poses problems, operator performance may suffer.
OPERATOR USE OF ALARM INFORMATION The problems encountered by the operator in using alarm
systems are illustrated by examining circumstances when
The primary function of an alarm system is to support the systems deviate from normal operating conditions. The nu-
detection of off-normal situations. Human fault detection can clear power plant incident at Three-Mile Island (TMI) is a
be described in terms of signal detection theory (SDT) (1). good example. The President’s Commission on TMI indicated
Within this framework, the operator and the alarm system that during the first few minutes of the accident, more than
constitute an alerted-monitor system (2,3). The automated 100 alarms were activated in the control room (4). The opera-
monitor is the alarm system, which monitors parameters to tors had no assistance in distinguishing significant alarms
detect off-normal conditions. When a parameter exceeds the (i.e., situations requiring operator attention) from less impor-
criterion of the automated monitor, the human monitor is tant signals (i.e., situations not requiring operator attention).
alerted and then must detect, analyze, and interpret the sig- The Rogovin report on the incident was more specific in iden-
nal as a false alarm or a true indication of an off-normal con- tifying alarm system deficiencies and contributions to safety
dition. The human monitor also assesses system parameters (5). The report indicated that on ‘‘the morning of the accident,
independently of the alarm system. Both the human and au- the alarms were of little use due to the number that were
tomated monitors have their own decision criteria and sensi- flashing and their almost random locations.’’ Some of the im-
tivity. The decision criterion refers to the amount of evidence portant alarms were not located in direct view of the opera-
that is needed before a conclusion is made that a signaled tors. Although auditory signals were associated with these
event is actually present (this is sometimes called response alarms, they could not be distinguished from other alarms be-
bias). Sensitivity is the resolution of the system that deter- cause a single button caused the auditory signals and flashing
mines the ease with which true disturbances are distin- lights to stop for all alarms. Operators indicated that the con-
guished from routine fluctuations. stant buzzing of auditory alarms and flashing lights were dis-
SDT research has many implications for understanding tracting and made their jobs more difficult.
how operators process alarm information. First, the response The types of problems experienced by the operators at TMI
criterion is affected by context, that is, the expected probabil- are typical of the problems faced by operators of many com-
ity that an event will occur and the payoff structure (rewards plex human–machine systems, such as process control facili-
and penalties for making correct and incorrect detections, re- ties, aircraft, and medical systems.
spectively). Significant off-normal events in many systems Alarm overload is a significant challenge to operators.
may have a low probability of occurring. Therefore, operators They often have problems detecting and interpreting alarms.
may be disinclined to decide that they have actually occurred. It has been shown that an operator’s ability to detect off-nor-
There is a conflict between the cost to productivity in re- mal events is reduced as the number of alarms presented in-
sponding (unnecessarily) to a false alarm versus the poten- creases (6,7,8). In fact, it has been generally found that, as
tially significant cost of failing to respond to a true alarm. demands on operators increase, fault detection capability de-
When disturbances have a low probability, operators rely on creases (9). When the number of alarms is large, the opera-
redundant and supplemental information to confirm the tor’s information processing ability becomes overloaded, and
alarmed condition. Upon verification of several confirmatory performance suffers because of high workload (2). This occurs
indicators, the operator accepts the alarm information as in- because the operator’s cognitive resources are in short supply,
dicating an actual off-normal condition (compared with a spu- and the resulting changes in behavior increase the probability
rious condition). of error. First, an operator may adopt inappropriate alarm
Once operators conclude that the alarm information repre- sampling strategies which make the accurate diagnosis of sys-
sents a valid system disturbance, they actively try to con- tem anomalies less likely (10,11). Under normal conditions, a
struct a coherent, logical explanation to account for their ob- sampling strategy based on successive observations of weakly
servations. This cognitive activity is called situation related variables is an appropriate strategy. However, once a
assessment and involves two related concepts: the situation disturbance begins, a more appropriate strategy is to sample
model and the mental model. Operators develop and update correlated variables because this facilitates detecting and rec-
a mental representation of the factors known or hypothesized ognizing a system/component failure. The ‘‘normal’’ sampling
394 ALARM SYSTEMS
strategy plus the operator’s low expectancy of problems can characteristics are alarm selection, alarm states, alarm pri-
delay realization of an event. Alarm system design features, oritization and processing, alarm availability, display, con-
for example, grouping alarms by system and function, help trol, automated and modifiable characteristics, reliability and
minimize this problem. This type of organization enables op- maintenance, location and integration of alarms into the
erators to perceptually integrate the alarms into high-level workplace, and alarm response procedures. The main discus-
information, for example, perceiving several low-level alarms sion pertains to general alarm-design considerations that
as a higher level event, such as ‘‘Pump A is tripped.’’ apply to complex systems, such as process control facilities,
A second result of alarm overload is a shift toward simpler aircraft and surface transportation systems, and medical de-
information-processing strategies. Reason has indicated that vices. Alarm systems for simpler systems may not involve all
operators under high workload conditions employ two prob- of the considerations presented.
lem-solving heuristics: ‘‘similarity matching’’ and ‘‘frequency
gambling’’ (12,13). These strategies give rise to a number of
‘‘basic error tendencies’’ in human performance which account Alarm Selection
for many human errors. Similarity matching reflects the ten- The first consideration in alarm system design is selecting
dency for the operator to attempt to match a perceived infor- conditions to be monitored. Important considerations include
mation pattern (such as a pattern of alarm signals) with (1) the categories (i.e., the events or system states) from
known, remembered patterns. When the perceived informa- which alarms are identified, (2) the criteria to be used in se-
tion partially matches more than one mental representation, lecting alarm parameters to represent the categories, and (3)
the discrepancy is resolved by selecting the one most fre- criteria for determining set points.
quently experienced. This is the ‘‘frequency gambling’’ heuris- The situational categories from which alarms are selected
tic. There are alarm system features that assist the operator are personnel hazards, safety, productivity, and investment
in processing information and thereby help prevent errors. An protection (indicators of damage to important and costly
example of such a feature is the generation of alarms when equipment). Such conditions often require operator attention
(1) ‘‘unexpected’’ alarms (based on the current pattern) occur and action to prevent a system disturbance from resulting in
and (2) an ‘‘expected’’ alarm (based on the current pattern) negative consequences. For purposes of this chapter, a dis-
does not occur. Such alarms call the operator’s attention to tinction is made between alarms and annunciators, although
‘‘outlier’’ alarms which are likely to be missed because of a the two terms are often used interchangeably in the litera-
tendency to focus on indications that confirm the operator’s ture. An annunciator is a status indicator that does not neces-
current representation. The use of alarm overview displays sarily require the operator’s attention. Annunciators should
and hierarchal alarm displays also help minimize these types
not be presented by the alarm system displays because they
of errors.
distract operators from attending to actual alarms.
In addition to alarm overload, nuisance alarms are another
Once the parameters representing these conditions are se-
significant problem with alarm systems. As discussed pre-
lected, set points are determined. A set point is the value of a
viously, fault detection performance is a function of the entire
monitored parameter which defines the boundary between
alerted-monitor system. Optimizing the signal detection pa-
the parameter’s normal range and an alarm condition. An
rameters for one component of the system may not optimize
alarm condition exists when the parameter exceeds the nor-
performance of the entire two-stage system (2). Thus, when
mal range defined by the upper and/or lower set points.
the response criterion of the alarm system is set to maximize
the number of disturbances detected, the number of false Graded alarms may have multiple setpoints outside of the
alarms increases. This problem occurs when alarm paramet- normal range that produce alarms indicating increasing lev-
ric setpoints are too close to the normal operating value or too els of severity of an abnormal condition (e.g., low level, low-
close to the normal value drift. Although this may provide an low level, etc.).
early alert to a potential system disturbance, many false Set points need to be carefully selected to ensure that oper-
alarms are created because of momentary fluctuations in ators monitor and take appropriate action in a timely man-
parametric values. When there are many false alarms, an op- ner. To achieve this, set points are specified at levels that are
erator may lose confidence in the system and adopt a more sufficiently different from the actual limits to allow sufficient
conservative criterion. Such interactions between automated time to respond. However, considering the signal detection is-
and human monitors can result in poor overall performance. sues discussed before, determining alarm set points should
The need to address these deficiencies has led to advances consider the trade-off between timely alerts to off-normal con-
in nearly every aspect of alarm system design. For example, ditions and the creating of nuisance alarms by establishing
alarm processing has been provided to facilitate identification set points so close to the ‘‘normal’’ operating values that occa-
of critical alarms, and alarm displays have been improved to sional excursions of no real consequence are expected.
facilitate association of alarms and other system information.
In addition, many alarm systems are designed with alarm Alarm States
management facilities enabling personnel to sort alarms
along dimensions, such as time and system, and to interro- Alarms have four states: inactive, new, acknowledged, and
gate the alarm system to obtain detailed information about cleared. When the parameter is in the normal range, the
specific alarms of interest. alarm is inactive. A new alarm occurs when a monitored pa-
rameter exceeds a specified limit (set point). The deviation is
ALARM SYSTEM DESIGN evaluated by the processing portion of the alarm system, and
an indication is conveyed to the operator through the display
This section addresses alarm system design characteristics portion of the alarm system. The alarm remains new until it
relative to the alarm functions identified earlier. The specific is acknowledged.
ALARM SYSTEMS 395
An alarm is in the acknowledged state when the operator processing, and alarm generation processing. Each processing
has provided some type of input to the alarm system (such as technique changes the resulting information provided to oper-
pressing a button) to indicate receipt of the alert or message ators.
provided by the alarm system. The act of acknowledging an Nuisance alarm processing includes techniques that seek
alarm typically causes the attention-getting characteristics of to eliminate alarms which have no operational safety impor-
the alarm display to cease or decrease (e.g., the auditory tone tance. Some examples of this class of techniques are time-
stops, and the flashing display changes to a steady illumi- delay and mode-dependent processing. Time-delay processing
nation). applies time averaging and/or time delay to the alarm inputs
An alarm is cleared when the alarmed parameter returns to allow filtering noise signals and to eliminate unneeded mo-
from an alarmed state to its normal range. Some alarm sys- mentary alarms (such as those resulting from instrument
tems provide indications when the parameter enters the nor- noise). Mode-dependence processing evaluates alarm signals
mal range. The operator may be required to acknowledge the for their relevance to the current system mode or configura-
alarm to ‘‘clear’’ it. tion. If a component’s parametric value represents a fault in
some system configurations and not others, it should be
Alarm Prioritization and Processing alarmed only in the appropriate configurations. For example,
the fact that a particular pump has a low discharge pressure
Not all alarms are of equal importance. Therefore, prioritizing may only indicate a fault when the associated fluid system is
alarms is helpful to operators. Alarm priority is typically configured to perform a particular function. Other discharge
based on the immediacy of required operator action and chal- pressures may be appropriate when the fluid system is con-
lenges to safety. The prioritization scheme should be logical figured to perform different functions. Mode processing would
so that those alarms of the highest safety significance receive allow the alarm message to be presented only when it is
the highest priority and so that the prioritization appears rea- meaningful to the fluid system’s current configuration.
sonable to operators. It is best to limit the number of priority Redundant alarm processing includes techniques that ana-
levels to four or less. Prioritization schemes with many levels lyze for alarm conditions that are valid but are considered
require operators to devote excessive attention to the priority less important because they provide information that is re-
level and thus reduce the benefits of prioritization. dundant to other alarms and logically provide no new infor-
Priorities are determined by engineering analyses during mation. For example, a single event (cause) invariably leads
system design. This type of a priori analysis identifies the to subsequent alarmed events that are the direct consequence
static priority of alarmed conditions. Static in this case means of the initial event. Processing is used to eliminate alarms
that the relative importance of alarms is always the same and that follow as logical consequences. Such techniques may,
does not change as a result of real-time analyses during sys- however, minimize information used by the operator for (1)
tem operation. More sophisticated systems use alarm pro- confirmation that the situation represented by the ‘‘true’’
cessing techniques to dynamically prioritize alarms on the ba- alarm has occurred, (2) situation assessment, and (3) deci-
sis of the existing situation and system condition. sion-making. Thus, in addition to quantitatively reducing
Alarm processing techniques were developed to support op- alarms, processing methods may qualitatively affect the infor-
erators in coping with the high volume of alarms that occur mation presented to the operating crew.
during major system disturbances by reducing the number of Significance processing includes techniques that analyze
alarms. Two general classes of alarm processing techniques for alarm conditions that are valid but are considered less
are discussed: signal validation and condition processing. important because of their significance compared with other
When instrumentation failures (such as a failed sensor) occur, alarm conditions. For example, alarms for a condition repre-
biased or false signals are generated. The use of these signals senting a threat to personnel safety may be displayed
by the alarm system may result in the presentation of either whereas a concurrent alarm associated with minor equipment
false or nuisance alarm messages. Such alarm messages are failure is not presented.
misleading and may interfere with the operator’s situation as- Alarm generation processing includes techniques that ana-
sessment or reduce the crew’s confidence in future alarm mes- lyze the existing alarm conditions and, based on the evalua-
sages. Signal validation is a set of techniques by which sig- tion, generate alarm messages which (1) give the operator
nals from redundant or functionally related sensors are higher level or aggregate information, (2) notify the operator
automatically evaluated to determine whether a true alarm when ‘‘unexpected’’ alarm conditions occur, and (3) notify the
condition exists, thereby avoiding presentation of spurious operator when ‘‘expected’’ alarm conditions do not occur. The
alarms to operators. Such techniques include the analysis of generation of alarm conditions and their resulting alarm mes-
normal signal drift and noise signals to eliminate those that sages presents an interesting paradox. Alarm systems should
momentarily exceed the set points but do not indicate a true facilitate the reduction of errors which often reflect the over-
alarm condition. Alarm conditions not eliminated by the loaded operator’s incomplete processing of information. Alarm
alarm signal processing may be evaluated further by alarm generation features may help mitigate these problems by call-
condition processing. ing the operator’s attention to plant conditions that are likely
Alarm condition processing is the rules or algorithms used to be missed. The single most significant problem with alarms
to determine the operational importance and relevance of systems, however, is the high number of alarm messages. Be-
alarm conditions. This is done to determine how the alarm cause alarm generation creates additional messages, it may
messages associated with these conditions should be pre- potentially exacerbate the problem.
sented to the operator. For the purposes of this discussion, There are two additional aspects to alarm processing to be
four classes of processing techniques are defined: nuisance considered in alarm system design: degree of alarm reduction
alarm processing, redundant alarm processing, significance and complexity of processing. The relationship between the
396 ALARM SYSTEMS
degree of alarm reduction resulting from processing and its plastic rectangular ‘‘tile’’ with a label briefly indicating what
impact on operator performance is not fully understood, that the alarm is. The individual alarm tiles are grouped together
is, although we know that most systems have too many into alarm matrices. When an alarm set point is reached, an
alarms, we do not know how much reduction is necessary to auditory signal is the first indication. Upon hearing the audi-
improve the situation meaningfully. tory signal, the operator looks at an alarm panel. At the same
The processing methods applied should not be so complex time as the auditory alert, the specific alarm tile begins to
that operators have difficulty evaluating the meaning or va- flash. So finding the alarm is relatively easy, as long as not
lidity of the resulting alarm messages. To support the under- too many alarms come in within a short time. In some sys-
standability of alarm processing systems, inputs to the alarm tems, alarm tiles are augmented by alarm message printers
processing system (e.g., sensor data) should be available to that print out a chronological list of alarms as they come in.
the operators. This data may be needed, for example, if the In these older systems, the alarm system is typically an infor-
pattern of alarm messages appears to be contradictory or if mation system separate from the other information systems.
operators suspect that there is a problem with the processing The alarm system alerts operators to off-normal conditions.
system, such that the results of alarm processing are in- Then the operators consult other indicators for specific infor-
correct. mation.
There are a number of general types of alarm displays.
Alarm Availability
Based on the types of analyses previously discussed, some • Spatially dedicated continuously visible (SDCV) alarm
alarms are considered more important than others. Alarm displays—A spatially dedicated alarm is always dis-
availability is the method by which the differences in alarm played in the same physical location. Continuously visi-
importance are used to determine which alarms are made ble means that a parallel presentation method is used,
available to operators. Three techniques have been used: fil- that is, the alarm information is always displayed to the
tering, suppression, and priority coding. Note that the terms operator, as opposed to serial presentation methods in
‘‘filtering’’ and ‘‘suppression’’ are often used interchangeably which the operator must select the information to be
in the literature, although different meanings are identified seen. The tile display, described before, is an example of
here. an SDCV alarm. More recently, tile-like video displays
Alarm filtering is the process of eliminating alarms deter- have been used where operators are seated at a worksta-
mined to be less important. Filtered alarms are not available tion.
to operators. Filtering should be employed only where alarms
• Temporary alarm displays—Alarm message lists are a
have no operational significance. Thus, only alarms that have
typical implementation of a temporary alarm display.
no operational significance to operators should be filtered.
Messages appear only when the alarm is activated. De-
In alarm suppression, less important alarms are not pre-
pending on the design, temporary alarms may or may
sented to the operators on the primary alarm displays but
not appear in spatially dedicated locations.
are available on auxiliary displays (which the operators can
retrieve) or are presented automatically when they become • Integrated alarms—Alarm information is presented as
high in importance (e.g., when more important alarm condi- an integral part of other displays, such as process dis-
tions have cleared). plays. For example, if alarms are built into a system
The third option is dynamic priority coding where all mimic display, trouble with a component, such as a
alarms are presented to operators but information about their pump, is depicted by a change in color or flashing of the
priority is given. For example, more important alarms may be pump icon. These types of displays may have fixed or
color coded red whereas less important alarms are white. variable locations.
Although a specific alarm system employs a combination
of these approaches, there are clear trade-offs among them. SDCV alarm displays are superior to temporary alarm dis-
Filtering eliminates the possibility that unimportant alarms plays during high-alarm conditions. The fixed locations are
distract the operators. However, the designer may be remov- often thought to provide perceptual advantages of rapid detec-
ing information that may be used by operators for other pur- tion because operators know alarms by position. Thus when
poses. In addition, the designer must be certain that the pro- an alarm flashes, they often do not have to read its label in
cessing method is adequately validated and functions detail. Further, operators begin to recognize the patterns of
appropriately in all conditions. Suppression provides the po- alarms associated with familiar disturbances. Although mes-
tential benefits of filtering by removing potentially distracting sage lists typically provide more information than alarm tiles,
alarms. But, because the suppressed alarms are accessible, they are problematic because of the workload associated with
they potentially impose an additional workload by requiring reading individual messages. Because spatial cues to the na-
operator action to retrieve them. Priority coding, on the other ture of the alarms are not available, operators must read each
hand, does not conceal any information from operators be- alarm as it comes in to know what it is.
cause all information is presented. However, the operator is In recent years, alarm displays have become considerably
required to search for the important, high-priority alarms more complex. Although audio and visual components are still
among the potentially distracting visual clutter of less impor- important attention-directing features of alarm displays,
tant alarms. alarms are becoming increasingly integrated into normal in-
formation displays used for system monitoring and decision
Alarm Display
making. One of the reasons for this trend is the recognition
In older systems, alarm displays are relatively simple. Each that cognitive processing is facilitated by integrating informa-
alarmed parameter has an indicator, for example, a small tion into a single object (14) or display (15). It is thought that
ALARM SYSTEMS 397
integrated displays enhance the ability of the operators to sented on a video display unit (VDU). A tile-based system nec-
process information in parallel, thus lowering cognitive work- essarily displays less information, requiring that detailed in-
load. Data integration also enables operators to understand formation be provided through other alarm displays, such as
the relationships between display elements better and to as- supplemental VDUs and printers. It is important that the for-
sess the situation more rapidly and accurately. mat of messages on alarms be consistent for all alarms and
Because of their superiority in high-alarm conditions, consistent across types of alarms and between VDU and
SDCV alarm displays should be considered for alarms that printed message displays.
address (1) safety and mission-critical parameters, (2) situa- Alarm title or legend text should be clearly understand-
tions that require short-term response by the operators, (3) able, using standard terminology and addressing conditions
the main alarms used by operators in diagnosing and re- specifically. For example, it is preferable to identify the pa-
sponding to plant upsets, and (4) the principal alarms used by rameters and states (e.g., HIGH PRESSURE) specifically in-
operators to maintain an overview of plant and system status. stead of using one legend for multiple parameters or multiple
Some of the general design considerations for alarm dis- states (e.g., TEMPERATURE-PRESSURE or HIGH-LOW).
plays follow. The discussion is divided into the following top- Operators often want to know the chronology of alarms,
ics: display of alarm states, message design, coding, shared especially for diagnosis. Providing the time the alarm was
alarms, and organization. triggered is beneficial for these types of activities.
The display of important alarms takes precedence over
Display of Alarm States. Each alarm state (i.e., new, ac- lower alarms. For alarms that are not SDCV displays, suffi-
knowledged, cleared, and inactive) should have a unique pre- cient display area is needed for simultaneously viewing all
sentation to support the operator’s ability to rapidly distin- high-priority alarms. For example, operators should not have
guish among them. New alarms are indicated both visually to page or scroll a display to view high-priority alarms. Codes
(e.g., flashing) and audibly. After the operator has acknowl- are used to indicate alarm priority levels. Coding methods in-
edged an alarm (e.g., pressed the acknowledge button), the clude color, position (top to bottom), shape, and symbolic cod-
alarm display changes to a visually distinct acknowledged ing. Color and position are especially effective visual coding
state, and the alerting function (e.g., audible tone) ceases. methods. However, coding by position should not disrupt the
If the operator is required to take action when an alarm functional grouping of alarms. Coding is discussed in greater
clears, the return to normal conditions is also indicated visu- detail later.
ally and by audible means, called ringback. This is not needed If an alarm condition requires verification before action is
for all alarms but is useful when it is important for the opera- taken, then relevant set-point limits should be included in the
tor to know immediately when the deviation has cleared or alarm message. However, one needs to consider whether pro-
when the deviation is not expected to clear for some time. viding set points alone (and not current parametric value)
Techniques that are employed include a special flash rate leads operators to assume that the condition is near the set
(one-half the normal flash rate is preferred to allow discrimi- point when, in fact, it is well above or below it. Whenever this
nation), reduced brightness, or a special color that is consis- is an important consideration, actual parametric values may
tent with the overall control room color-coding scheme. be as important as set points and should be provided.
Cleared alarms should have a dedicated, distinctive audible Where practical, immediate operator actions are presented
signal of finite and relatively short duration. or referenced. For example, immediate operator actions are
Inactive alarms are best indicated by an absence of visual provided in alarm response procedures (discussed later) that
and auditory alarm features. This practice is referred to as are clearly and simply keyed to an alarm tile and located
the dark board (or blackboard) concept of alarm display be- nearby for easy and quick reference.
cause it results in a dark display medium (not illuminated)
when all monitored plant parameters are in the normal Coding. Coding for alarm priority and state was men-
range. Under such circumstances, a new alarm is easily iden- tioned before. Coding is the use of a system of symbols,
tified. shapes, colors or other variable sensory stimuli to represent
specific information. Coding schemes facilitate rapid detection
Message Design. Alarm messages include information, and interpretation of alarms by operators. Each level of a cod-
such as ing dimension must be easily and readily distinguishable
from the other levels. A formal coding scheme is established
• Alarm title and formally documented which encompasses all coding di-
• Time of occurrence mensions (e.g., color, shape, brightness, textures/pattern,
• Alarm source, that is, the particular sensor or group of flashing, and sound) and specifies a hierarchical order of sa-
sensors supplying the signal lience. Then the coding scheme is systematically applied to
alarm information so that the most important alarm informa-
• Alarm priority
tion is associated with the most salient coding. Because cod-
• Set point and parametric values ing adds to display complexity, it is a good practice to mini-
• Required immediate operator actions mize the number of different coding techniques.
• Reference to procedure for more detailed follow-up ac- The primary coding methods used for alarms are visual
tions and auditory. These techniques are discussed in the remain-
der of this section.
The extent to which all of this information is included in an Color is one of the most effective types of coding. Colors
alarm message is constrained by the type of display system should be easily discriminated, each color with a single, pre-
available. More information is provided when alarms are pre- cise meaning consistent with its use in the rest of the system’s
398 ALARM SYSTEMS
operator interfaces. It is important that the use of color is lutely necessary to guarantee effectiveness for alarms indicat-
consistent with and, if possible, takes advantage of operators’ ing extreme danger. Levels this intense are probably appro-
already learned color associations. priate only for situations requiring prompt evacuation of an
A flashing visual signal is typically associated with all im- area. Techniques exist to support the design and production
portant alarms to ensure that the operator’s attention is cap- of signals that more effectively convey alarm-related informa-
tured and directed to these alarms under any and all condi- tion. Recently developed approaches to auditory alarm signal
tions. To achieve this, flash rates from three to five flashes design recommend that the intensities chosen for auditory
per second with approximately equal on and off times are ef- signals take into account the frequency spectrum of the sig-
fective. nals and of the background noise (16,17). Because of the na-
For transilluminated displays, the brightest state should ture of the human auditory system, signals need not be very
be no more than 300% brighter than the inactivated state (but intense to be heard reliably as long as the frequencies at
not annoy operators) and the dim state at least 10% brighter which their energy is concentrated differ from those occupied
than the inactivated state. For VDU displays, the bright state by the background noise. It should be recognized that audi-
should be at least 100% brighter than the inactivated state. tory signals themselves may contribute to the background
VDUs can be used to display two brightness levels, whereas noise, that is, it may be necessary to consider the audibility
transilluminated alarms may display more than two levels of of a signal not just in the presence of ambient noise but also
brightness. It is ineffective to use low-intensity indications in combination with other signals that might plausibly occur
(e.g., dark red) in the periphery of the visual field where color at the same time. To avoid mutual masking, the frequency
coding is used because they may not be readily detected. spectra of auditory signals associated with alarms that may
Under high alarm-volume conditions, consider suppressing be active at the same time should be different. Interference
or delaying the alerting indications (e.g., visual flashing) for among alarm signals is less of a concern when the signals
those alarm conditions that (1) do not require immediate re- consist of a number of widely separated frequency compo-
sponse and (2) do not indicate a challenge to safety. In addi- nents (rather than a pure tone) or of brief groups of pulses
tion, redundant codes (e.g., fast flashing or bright illumina- presented at intervals (rather than a continuous tone).
tion) are used to assist operators in detecting the more It is not good design practice to code auditory signals by
significant alarm messages and to reduce distraction from intensity (loudness). The range of intensities between the
less important alarms. level required to ensure audibility and the level at which sig-
Auditory signals draw attention to alarms even when oper- nals become aversive is relatively narrow. Therefore, the use-
ators are not attending to the visual alarm displays. Conse- fulness of this dimension for coding is limited. If such coding
quently, it is common practice to provide auditory cues for all must be used, use only two levels, and distinguishing the sig-
new alarms under normal operating conditions. The number nals from each other by a minimum of 6 dB(A). Whether this
and placement of sound sources should be such that auditory coding is effective depends on the frequency spectrum of the
signals are free of distortion and are equally audible at any ambient control room noise and the frequency of the signal.
operator work station in the primary operating area. It is best Each auditory signal should be unambiguous and easily
to orient speakers away from surfaces that scatter or diffuse distinguishable from every other tone in the control room. For
the sound. Avoid placing speakers behind structures that example, the auditory signal associated with SDCV alarm dis-
cause distortion, echoes, or sound shadows. When sound local- plays should be easily distinguishable (based on signal char-
ization is used to direct the operator to particular alarm dis- acteristics or sound source) from the auditory signal associ-
play devices, the sound sources should be oriented so that ated with an alarm message displayed by other means (e.g.,
their location is quickly discerned and corresponds to the loca- on a VDU message display). Auditory signals used for new
tion of the intended alarm display device. alarms should be separate and distinct from tones used to
The intensity of auditory signals should be such that they signify clearing alarms. The latter can be momentary or ‘‘self-
are reliably detected under the most adverse, anticipated, silencing.’’ If the tone indicating an unacknowledged alarm
background noise conditions. To guard against the possibility automatically turns off after an interval of time, a reminder
that operators inadvertently reduce the audio level so as to tone can be used to alert the operator to its continued
render the signals inaudible, the systems that generate audi- presence.
tory signals typically do not allow operators to adjust the sig- Current techniques allow designing alarm signals that
nal level. Unfortunately, it is not unusual for signals to be so make better use of the operator’s ability to process audio in-
intense that they irritate or startle the operators. Conse- formation. It is possible to design signals that are more dis-
quently, operators may turn off or disable alarm systems. Im- criminable from one another than conventional signals and
proved approaches to auditory signal design and level selec- also have the potential to carry more information (18). Sets
tion alleviate this problem. of readily distinguishable signals can be designed by varying
A signal level 10 dB(A) above average ambient noise is fundamental frequency, harmonic structure, and temporal
generally considered adequate to ensure that a signal is audi- patterns.
ble. dB(A) refers to decibels as measured using one of three When information is coded by pitch, it is best not to use
standard weighting networks (designated A, B, C) typically more than three frequencies. The frequencies should be
available in sound-level meters. A-weighted sound levels are widely spaced between 500 Hz to 3,000 Hz, although a wider
typically used when the effects of sound on people are of inter- range from 200 Hz to 5,000 Hz may be acceptable. Avoid fre-
est because they correlate well with perceived loudness and quencies in a ratio of 2 : 1 with one another, because it is dif-
speech interference. It has also been recommended that sound ficult to identify pitches an octave apart. Signals with multi-
intensity should be limited to a maximum of 95 dB(A) but ple frequency components (‘‘chords’’ or frequency-modulated
that signal levels of 115 dB(A) may be used if considered abso- tones) are more resistant to masking and more easily localiza-
ALARM SYSTEMS 399
ble than pure tones. (If the location of a source sound is to be • Alarms for several conditions that call for the same cor-
used as a cue, the signal should not be a high-frequency tone, rective action
because such signals are difficult to localize.) No more than • Alarms that summarize single-input alarms elsewhere in
three modulated frequency codes for audible alarms should be the control room
used. Warbling sounds, with frequencies modulating from one
to three times per second, are attention-getting and easily Because they require additional effort by the operator to iden-
recognized, whereas slower modulation rates do not develop tify the specific alarm, it is best to minimize the number of
distinguishable characteristics rapidly enough to be appro- shared alarms, especially when
priate for alarm applications. If modulation of frequency (Hz)
of a signal is used to denote information, the center frequen- • Different actions are to be taken depending on which
cies should be between 500 Hz and 1000 Hz. alarm condition exists and the operator cannot readily
No more than three pulse repetition rates should be used determine which constituent is alarming
for coding purposes. Repetition rates should be between one • The required response must be initiated immediately, so
and eight pulses per second because faster rates may not be that taking time to determine which constituent is
perceived as pulses. It has been recommended that repetition alarming would risk an inadequate response
rates differ by at least a factor of 2 to ensure operator discrim-
• The operator’s understanding is improved by indicating
ination. Recent alarm design research suggests that other-
the conditions separately because of similarity to the lay-
wise distinctive signals having similar temporal patterns are
out of associated controls
readily confused and that this effect may extend to duty cycle.
Accordingly, temporal patterns other than regular on/off cy- • The constituent conditions are not of a similar nature or
cles should be considered. are not of the same order of importance, so that the ac-
If sequences of tones are used to represent information, tion to be taken is very different depending on which con-
the patterns should be easily recognizable. Warning sounds dition is alarming
consisting of ‘‘bursts’’ composed of five or more brief pulses
(about 0.1 s in duration) with inter-pulse intervals of 0.15 s If a new parametric deviation has occurred before a preceding
to 0.3 s are currently recommended (see Ref. 16). The pulses alarm has cleared, the shared alarm should return to the new
may be designed to be distinctive with respect to their onset alarm state (e.g., reflashing). It is also important that the
and offset shaping, fundamental frequency, and harmonic alarm system enable the operator to reactivate the visual and
structure. The bursts may vary as to the number of pulses, audible alert indications for the alarm when subsequent
the tempo at which they are presented, and the rhythmic and alarm conditions occur after the initial alarm condition has
pitch contours. been acknowledged.
It has been recommended that a maximum of nine audi-
tory signals be used when two or more dimensions are used Organization of Alarms. Grouping alarms within a display
in coding. When signals differ in two or more dimensions (e.g., by function, system, or other logical organization facilitates
pitch and temporal pattern), a greater number of signals can the operator’s understanding of the relationships among
be reliably distinguished. This maximum includes auditory alarms and among alarms and system components. Func-
signals used outside the control room (e.g., fire alarm or site tional groups should be visually distinct from one another.
emergency alarm). The number of conditions for which reli- Clear labels and delineations among groups allow the opera-
ably recognizable audio codes can be used is maximized by tors to determine easily which systems have alarms. If alarm
taking advantage of differences in the perceived urgency of displays are organized in matrices, the vertical and horizontal
warning sounds (19) to represent, for example, varying axes of the displays can be labeled with alphanumerics so
alarm priorities. that a coordinate designation is available for any particular
visual element. Coordinate designation is preferred on the left
and top sides of the display. The size of an alarm tile matrix
Shared Alarms. A single ‘‘shared’’ indicator is sometimes should be limited to 50 alarms.
used to represent more than one alarm condition. For exam- Operators use alarm displays more effectively if alarms are
ple, a ‘‘trouble’’ message may combine several potential prob- arranged according to naturally occurring relationships, such
lems associated with a single component, or it may address as those derived from the physical process, for example,
the same problem for a group of similar components (e.g., a
bearing temperature alarm may address bearings from more • Alarms for a given thermodynamic parameter at differ-
than one component). The types of alarms that may be consid- ent points within the system which indicate a progres-
ered for combination include sion (e.g., within a fluid system, a series of pressure
alarms starting with the source tank and ending with
• Alarms for the same condition on redundant components, the system discharge) could be arranged left to right
or logic trains, when each has a separate indicator and • Several alarms for the same variable indicating levels of
the indicators are placed in close proximity on the con- severity (e.g., tank level low and tank level low-low)
sole (e.g., pump A or B trip, logic train A or B actuation) could be arranged in a vertical array
• Alarms for several conditions relating to one component • Alarms related by cause and effect could be adjacent to
or several redundant components, which require the op- one another
erator to obtain further diagnostic information either by
sending an auxiliary operator out to the component(s) or Once an arrangement has been chosen, the arrangement can
checking the computer (e.g., pump A or B trouble) be used consistently within similar systems or alarm groups.
400 ALARM SYSTEMS
Alarm message lists are more effective when segregated by tions for all essential aspects of the alarm system, including
alarm priority with highest priority alarms listed first. In processing logic, audible alarms, and visual alarm indications
addition to priority grouping, provide operators with the (e.g., to detect burnt-out lamps).
capability to regroup alarm messages according to opera- Beyond these silence, acknowledge, reset, and test con-
tionally relevant categories, such as function, chronological trols, computer-based alarm systems may require other con-
order, and status (unacknowledged, acknowledged/active, trols to allow operators to sort the alarms according to time
cleared). For example, it can be useful for diagnosis to or component, to define temporary alarms, to adjust set
arrange alarms in chronological order with the most recent points, and to control processing options.
messages placed at the top of the stack. A separation (blank
row) every four or five alphanumeric messages enhances Automated and Modifiable Characteristics
readability.
In certain situations, such as major system disturbances, it
may be desirable to reduce operator workload by automating
Alarm Controls or modifying some alarm system functions. For example,
Alarm systems typically include controls to: (1) silence the au- lower priority alarms might be automatically silenced, or the
ditory component of the alarm, (2) acknowledge the meaning flashing of unacknowledged alarms temporarily stopped. Sim-
of the alarm, (3) reset the alarm to its monitoring state, and ilarly, automated controls may be implemented to trigger ap-
(4) test the alarm display characteristics. Making these con- propriate displays, such as alarm graphics, data windows, or
trols easily distinguishable from each other by touch and display pages. Other dynamic aspects of the alarm system
sight helps prevent accidental operation of the wrong control. may allow operators to introduce operator-defined character-
Techniques, such as color- or shape-coding of individual con- istics, such as alarm parameters and set points.
trols and color shading or other demarcation of groups of If the alarm system automatically changes operational
alarm controls, can be used. configurations under some situations, an alert is needed to
indicate that the configuration has changed. Alarm systems
Silence. Operators sometimes disable distracting or irritat- may provide the capability for operators to select alternative
ing auditory signals. This can be a dangerous situation. Good functional configurations under some alarm situations, such
auditory signal design, such as described in the previous sec- as automatic silence of auditory alerts for lower priority
tion, helps minimize the need for such action. Another solu- alarms under high-alarm conditions. It is important for the
tion is to provide the capability to silence an auditory signal alarm system to indicate to operators that a requested change
from any set of alarm system controls in the primary op- in system configuration has been successfully achieved. In ad-
erating area. Although manual silence is a generally desirable dition, a prominent display of the present configuration
feature (in that it increases the likelihood that the operator should be available.
has attended to the alarm information), it may become dis- Requiring operator confirmation of any significant change
tracting to silence all alarms manually under high-alarm con- in the alarm system, whether selected by the operator or au-
ditions. tomatically, prevents potential errors resulting from an oper-
Alarm system designs should not allow the operator to de- ator’s misreading of the alarm system’s present configuration.
feat the control. For example, some pushbuttons used for The alarm system may provide temporary, operator-de-
alarm silencing and acknowledgment can be held down by in- fined alarms and operator-defined set points for specific condi-
serting an object in the ring around the pushbutton. tions (e.g., temporary alarms to support increased monitoring
of a problem component or of a parameter trend that is ap-
Acknowledge. Acknowledgment terminates the alarm proaching a limit). A clear indication of operator-defined
flashing and is usually indicated by steady illumination until alarms and set points is needed that is distinct from the
the alarm is cleared. Acknowledgment should be possible only alarm/set points designed into the system. It is also impor-
from locations where the alarm message can be read. If alarm tant that operator-defined alarms and set points not override
information is available at multiple VDUs, then operators or interfere with the existing alarms and set points.
should be able to acknowledge the alarm from the VDU at
which they are working. Reliability and Maintenance
It is important that the hardware and software components
Reset. The reset control places the alarm in an inactive of the alarm system are sufficiently reliable to prevent single
state after an alarm condition has cleared. When it is impor- component failures from causing significant loss of functions
tant to inform operators explicitly of a cleared condition, a or information. For example, redundancy and diversity of the
manual reset is appropriate. An automatic reset option is use- alarm system design can protect against losing alarm indica-
ful when operators have to respond to numerous alarms or tions or generating spurious alarm messages as the result of
when it is essential to reset the system quickly. The reset sensor or signal processing malfunctions.
control should be effective only from locations at which plant Tile-type displays can be designed with dual light bulbs so
personnel know which alarm they are resetting. that a single bulb failure does not interfere with the opera-
tor’s detection of the alarm condition. In case of flasher failure
Test. Given its importance, it is desirable for the alarm of an active alarm element, the element should assume a
system to indicate positively to the operator when alarm sys- highly salient state, such as a high flash rate or a steady on
tem malfunctions occur. By the same token, periodic testing (e.g., illuminated) state rather than a less salient state, such
of the system by operators is good operational practice. Test as off. Although it is preferable in a flasher failure that the
controls should be available to initiate operational test condi- active alarm element remains on (e.g., illuminated) rather
ALARM SYSTEMS 401
than off, a unique and highly salient code is best. In addition, A good ARP contains the following information:
other alerting mechanisms, such as warning messages, may
be used to inform the operator of a malfunction in the alarm • The system/functional group to which the alarm belongs
display system. • The exact alarm text or legend
Where VDUs are the primary means of displaying alarms, • The alarm source (i.e., the sensor(s) sending the signal,
additional reliability is gained by making it possible to access including processing or signal validation and condition-
the alarms from more than one VDU. Alarm printers also ing logic, and the actuating device(s) for the alarm with
back up VDU displays. a reference to a schematic diagram on which such devices
The alarm system should be designed so that maintenance can be found)
can be performed with minimal interference with the activi- • Alarm set points
ties of the operators. Desirable features include built-in test
• Priority
capabilities, modular components that can be rapidly re-
moved and replaced, and rear access panels which prevent • Potential underlying causes for the alarm (e.g., low water
maintenance activities from obstructing the operator’s view of level or feed flow deficient in the long term)
controls and displays. • Required immediate operator actions, including actions
When an alarm is taken out of service for maintenance, its the operator can take to confirm the existence of the
visual and audio signals should be disabled and cues provided alarm condition
to indicate clearly that the alarm is out of service. • Actions which occur automatically when the alarm oc-
curs (and which the operator should verify as having
taken place)
Location and Integration
• Follow-up actions
Locate visible alarm elements within about 60⬚ on either side • Pertinent references
of the direct line of sight of the operator’s normal work posi-
tion. To avoid confusion, do not locate alarms near indicator Just as alarm design conventions should be consistent with
lights that present information about the state of equipment. those applied to other interfaces, the information in the ARPs
Alarm displays and controls should be arranged and located should reflect the same conventions applied elsewhere.
so that the operators who must respond to an alarm can ac-
cess the alarm information quickly enough to respond ade-
quately. Avoid arrangements that require one operator to ALARM SYSTEM EVALUATION
read an alarm message only to recite it to another person. It
This section considers the general evaluation of the alarm
is also important to position alarms near the other displays
system with regard to its functional objectives. It does not
and controls that are required for diagnostic or corrective ac-
address engineering tests of such items as the correctness of
tion in response to them.
wiring or the performance of the circuitry.
Consistency between the alarm system and the other dis-
The objective of the alarm system review is to ensure that
plays and controls in the workplace is important to minimize
the alarm system functionality and its associated controls,
human error. Thus, the same design conventions for ele-
displays, and data processing support safe, efficient, and reli-
ments, such as symbols, icons, acronyms, coding, and mea-
able operator performance. The following design evaluations
surement units that are used in other operator interfaces and
can be performed to help ensure that the alarm system meets
procedures, should also be applied to the alarm system. For
its objectives:
example, if color is used to code priority, the relationship be- Task Support Verification. This evaluation verifies that all
tween the colors used and level of priority should be the same necessary system alarms are provided in the alarm system.
in the alarm system as in the process displays. For example, The method involves comparing a list of required alarms to
if red is used to indicate the highest priority alarm, then red the alarms available in the workplace. A problem is identified
should also be used to indicate high-priority information in if (1) required alarms are not provided by the alarm system
the process displays. or (2) alarms are present in the workplace that have not been
identified as required.
Alarm Response Procedures Human Factors Engineering Design Verification. This eval-
uation verifies that the alarm system design and implementa-
Alarm response procedures (ARPs) provide more detailed in- tion take operator capabilities and limitations into account.
formation concerning the nature of the alarm condition than The method involves evaluating alarm design characteristics
is typically provided in an alarm message. They are especially against human factors engineering (HFE) standards and
important to operators when an unfamiliar alarm is activated guidelines for alarm systems, such as Ref. 20. Problems are
or when an alarm seems inconsistent with the operator’s un- identified if the design is inconsistent with HFE guidelines.
derstanding of the plant state. Integrated System Validation. This evaluation validates
Operators should have immediate access to ARPs from the that the integrated alarm system design supports operator
location at which the alarm messages are read. ARPs may be task performance. This type of evaluation is best performed
hard copy or computer-based. In a tile-based alarm display using an actual system or prototype under realistic opera-
system, the operator’s access to ARPs can be aided by identi- tional conditions. When this is not practical, real-time, dy-
fying and indexing the ARPs consistent with the method of namic simulation of a system may provide an acceptable test
identifying the alarm, for example, by row and column desig- bed. Dynamic performance evaluation addresses both (1) the
nations. operator interfaces associated with operation of the alarm
402 ALGEBRAIC CODING THEORY
system and (2) the quality, accuracy, timing, and usefulness ing design: Relationship between warning sound parameters and
of the information provided by the alarm system to plant per- perceived urgency, Human Factors, 33: 205–231, 1991.
sonnel. Problems are identified if task performance criteria 19. J. Edworthy, Urgency mapping in auditory warning signals, in
are not met or if the alarm system imposes a high workload N. Stanton (ed.), Human Factors in Alarm Design, Bristol, PA:
on plant personnel. Taylor & Francis, 1994, pp. 14–30.
Then problems identified through these evaluation activi- 20. J. O’Hara et al., Human Factors Engineering Guidelines for the
ties can be remedied prior to actual operational use, resulting Review of Advanced Alarm Systems, NUREG/CR-6105, Washing-
in an effective alarm system that helps operators to monitor ton, DC: Nuclear Regulatory Commission, 1994.
the system and to detect disturbances in a timely manner.
Reading List
Annunciator Sequences and Specifications (ANSI/ISA-1979). Research
BIBLIOGRAPHY Triangle Park, NC: Instrumentation Society of America, 1992.
J. Edworthy and A. Adams, Warning Design: A Research Prospective,
1. D. Green and J. Swets, Signal Detection Theory and Psychophys- London: Taylor & Francis, 1996.
ics, New York: Wiley, 1988.
R. Fink, A Procedure for Reviewing and Improving Power Plant Alarm
2. R. Sorkin and D. Woods, Systems with human monitors: A signal Systems, EPRI NP-3448, Palo Alto, CA: Electric Power Research
detection analysis, Human Comput. Interaction, 1: 49–75, 1985. Institute, 1984.
3. R. Sorkin, B. Kantowitz, and S. Kantowitz, Likelihood alarm dis- MPR Associates, Power Plant Alarm Systems: A Survey and Recom-
plays, Human Factors, 30: 445–459, 1988. mended Approach for Evaluating Improvements, EPRI NP-4361,
4. J. G. Kemeny, Report of the President’s Commission on the Acci- Palo Alto, CA: Electric Power Research Institute, 1985.
dent at Three Mile Island, Washington, DC: US Government J. O’Hara et al., Human factors engineering guidelines for the review
Printing Office, 1979. of advanced alarm systems (NUREG/CR-6105), Washington, DC:
5. J. Rogovin, Three Mile Island, A Report to the Commissioners and US Nuclear Regulatory Commission, 1994.
to the Public, Washington, DC: US Nuclear Regulatory Commis- J. O’Hara et al., Human-system interface design review guideline
sion, 1980. (NUREG-0700, Rev. 1), Washington, DC: US Nuclear Regulatory
6. E. Marshall and F. Owre, The experimental evaluation of an ad- Commission, 1996.
vanced alarm system, in Advances in Human Factors in Nuclear R. D. Patterson, Guidelines for Auditory Warning Systems on Civil
Power Systems, La Grange Park, IL: American Nuclear Society, Aircraft, CAA 82017, London: Civil Aviation Authority, 1982.
1986. N. Stanton, Human Factors in Alarm Design, London: Taylor and
7. Y. Fujita and T. Sanquist, Operator cognitive processes under Francis, 1994.
abnormal plant conditions with conventional and advanced con-
trol room designs, 1988 IEEE 4th Conf. Human Factors, New JOHN M. O’HARA
York, 1988.
WILLIAM S. BROWN
8. Y. Fujita, Improved annunciator system for Japanese pressur- Brookhaven National Laboratory
ized-water reactors, Nuclear Safety, 30: 209–221, 1989.
9. A. Ephrath and L. Young, Monitoring vs. man-in-the-loop detec-
tion of aircraft control failures, in J. Rasmussen, and W. Rouse,
(eds.), Human Detection and Diagnosis of System Failures, New
York: Plenum, 1981.
10. N. Moray, The role of attention in the detection of errors and the
diagnosis of failures in man-machine systems, in J. Rasmussen
and W. B. Rouse (eds.), Human Detection and Diagnosis of System
Failures, New York: Plenum, 1981.
11. R. Sorkin, Why are people turning off our alarms?, Human Fac-
tors Soc. Bull., 32 (4): 3–4, 1989.
12. J. Reason, Generic error-modelling systems (GEMS): A cognitive
framework for locating common human error forms, in J. Ras-
mussen, K. Duncan, and J. Leplat (eds.), New Technology and
Human Error, New York: Wiley, 1987.
13. J. Reason, Modelling the basic error tendencies of human opera-
tors, Reliability Eng. and Syst. Safety, 22: 137–153, 1988.
14. D. Kahneman and A. Triesman, Changing views of attention and
automaticity, in R. Parasuraman and R. Davies (eds.), Varieties
of Attention, New York: Academic Press, 1984.
15. K. Bennett and J. Flach, Graphical displays: Implications for di-
vided attention, focused attention, and problem solving, Human
Factors, 34: 513–533, 1992.
16. R. D. Patterson, Guidelines for Auditory Warning Sytems on Civil
Aircraft, CAA 82017, London: Civil Aviation Authority, 1982.
17. C. LaRoche et al., ‘Detectsound’: A computerized model for pre-
dicting the detectability of warning signals in noisy environ-
ments, Appl. Acoust., 33: 193–214, 1991.
18. J. Edworthy, S. Loxley, and I. Dennis, Improving auditory warn-
532 BOUNDARY-SCAN TESTING
BOUNDARY-SCAN TESTING
Boundary-Scan is a collection of design rules which, when im-
plemented in digital integrated circuit (IC) chips, allow the
chips to test faults occurring when these ICs are assembled
on printed circuit boards (PCBs) and in systems. The Bound-
ary-Scan rules are formally presented in a document, IEEE/
ANSI Standard 1149.1a-1993 (1), ‘‘A Standard Test Access
Port and Boundary-Scan Architecture.’’ A Boundary-Scan
compliant IC is one that complies with the design rules in
this document. Boundary-Scan is sometimes referred to as
1149.1 and also as JTAG. JTAG refers to the Joint Test Ac-
tion Group formed in 1985 by major electronics firms around
the world to address a growing problem in board testing. Be-
tween 1986 and 1988, the JTAG Technical Subcommittee de-
veloped many proposals, resulting in approval of the IEEE
1149.1 Standard in 1990. This standard was subsequently
modified in 1993 to address implementation concerns.
It is important to realize that the 1149.1 Standard is only
a document which describes the basic rules to which a Bound-
ary-Scan architecture must comply. While it includes possible
embodiments for various Boundary-Scan components, it does
not define preferred implementations. Specific architecture-
related details are user-defined (2). It is also important to re-
alize that while the 1149.1 Standard is generally identified
with Boundary-Scan, other approaches are possible. This arti-
cle begins by indicating the problem which Boundary-Scan
was designed to address, followed by a description of the basic
Boundary-Scan structure, issues in the application of Bound-
ary-Scan, and emerging extensions of the Boundary-Scan
concept.
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
BOUNDARY-SCAN TESTING 533
fault from primary inputs (the controllability problem) and Figure 1 shows the scan cells of several ICs chained to-
propagate the effect of the fault to the board’s primary out- gether to form a scan path from board scan-in to board scan-
puts (the observability problem). Besides detecting a fault, it out. A scan cell is associated with each input or output pin,
is often very difficult to provide an accurate fault diagnosis. and each can store a bit. That bit can enter the cell by parallel
Conventional, one-sided PCBs were populated with inte- or serial (scan) paths and exit the cell by parallel or serial
grated circuits in dual inline packages (DIPs) having 0.1 paths, depending on how internal switches are set within the
inch ⫽ 100 mil 앒 2.5 mm pin spacings. Several layers of con- cell. Typically, data passes in parallel through the cells in
ducting traces on the PCB were used to interconnect the pins their normal operating mode and passes serially in their test
of ICs and other components. These pins extended through (scan) mode.
the board (through-holes in the PCB) and were soldered on Figure 2 shows the general Boundry-Scan architecture for
the other side of the board. This allowed access to the IC pins an IC as prescribed in the IEEE 1149.1 Standard (1). We dis-
by a bed-of-nails fixture contacting the back of the PCB in an tinguish the following basic elements:
in-circuit tester or by a hand-held probe for fault location.
This is not practical in modern electronic manufacturing. Sur- 1. Test Access Port (TAP)
face mounted IC packages do not provide access to personal 2. TAP Controller
computer (PC) backs (which may also be populated with com-
3. Instruction Register
ponents). Moreover, the distance between pins is shrinking to
25 mils for surface mount devices (SDMs) and 12 mils for tape 4. Test Data Registers [Boundary Register, Bypass Regis-
automated bonding (TAB), making contacts increasingly dif- ter, Device Identification Register (optional), Design-
ficult. The result is a radical decrease in test point accessibil- Specific Registers (optional)]
ity. This problem is amplified for more recent package tech-
nologies such as Ball Grid Arrays and Chip-Scale Packaging. The TAP provides access to the test support logic built into
Boundary-Scan is a design-for-testability (DFT) philoso- the IC. It consists of three required input connections and
phy. It aims to simplify testing by modifying ICs to improve one output connection. The optional fourth input connection is
the testability of IC interconnections on boards and to provide used to reset the Boundary-Scan test logic. The IEEE 1149.1
access to internal signals. Although the 1149.1 Standard pro- Standard requires that the TAP connections are not used for
vides support for internal DFT and Built-In Self-Test (BIST), any other purpose than testing. The mandatory TAP connec-
internal DFT and BIST are not required for compliance with tions are: (1) Test Data In (TDI), (2) Test Clock (TCK), (3)
the standard. In contract to DFT, BIST allows in-circuit, at- Test Mode Select (TMS), (4) Test Data Out (TDO).
speed testing while minimizing reliance on external testers The scan paths of several ICs are connected by wiring TDO
by adding logic structures such as Linear Feedback Shift Reg- from one chip to TDI of another. The other two reserved pins,
isters (LFSR), signature analyzers, or Built-In Logic Block TCK and TMS, are generally wired in common for all the ICs
on a board so that the states of all Boundary-Scan compo-
Observers (BILBO) (3).
nents are changed similarly and simultaneously. The optional
Scan is a well-established technique for providing access to
Test Reset (TRST) pin, if implemented, allows one to reset
internal test points in large circuits. In scan-based testing, all
test logic on the component independently of the internal sys-
the flip-flops and registers in a circuit are chained together to
tem logic. There is also a synchronizing sequence utilizing
form a path between a primary input and primary output.
TCK/TMS that can be used to reset the logic. Since TRST is
For combinational logic between registers, scan allows test
optional and fairly uncommon, this sequence will be used
vectors to be scanned into the input register and test results
most often for rest. The set of scan cells associated with I/O
to be scanned out of the output register. Boundary-Scan uses
pins can be connected to form a boundary (shift) register.
this technique to test the integrity of board interconnections
The TAP controller is a synchronous state machine that
between IC pins by chaining together scan cells associated
allows the test circuitry to load and execute instructions con-
with every I/O pin of the IC. trolling scan cell operation. The controller responds to
changes in the TMS and TCK signals and generates the clock
and control signals required to shift, capture, or update data
BASIC PRINCIPLES OF BOUNDARY-SCAN through either the instruction or test data registers (bound-
ary register and bypass register). The Bypass Register is a
‘‘Boundary-Scan test is a structured design-for-testability single state shift register which allows data to bypass the
method applicable to digital devices’’ (4). The essential idea scan path on an IC by directly connecting the TDI pin to the
behind Boundary-Scan is described by the term itself. The TDO pin. The instruction register is placed between TDI and
first term, boundary, indicates that the testability circuitry is TDO, just like the boundary register, so that test instructions
placed at the I/O pins of a component, around the system can be shifted in. The test instruction defines the test to be
logic. The second term, scan, suggests that the additional test performed and the test data register to be addressed. Each
circuitry within the component is transformed into a serial Boundary-Scan architecture has to contain the boundary and
shift register during testing. The serial shift register is often bypass test data register. The device identification and de-
referred to as the ‘‘scan path’’ and forms a controllable border sign-specific register are optional. Each test data register can
around the design. Access to the test logic is provided via four be selected, in other words placed between TDI and TDO, by
or five (one optional) dedicated test pins. At the PCB level, scanning the corresponding instruction code into the instruc-
the registers for the individual components can be connected tion register.
in series to form a single path through the entire design. Or The state diagram for the TAP controller has two main
a board design can contain several independent paths. loop sequences, one for handling test data and one for han-
534 BOUNDARY-SCAN TESTING
1 Short 1
1 1
0 1
0 1
System System
logic logic
x
x
IC1 IC2
TDI1 TDO1 TDI2 TDO2
Registers Registers
Control Control
Boundry-scan chain
1
Open
1
1
System System
logic logic
IC1 IC2
TDO4 TDI4 TDO3 TDI3
Registers
Figure 1. Boundary-Scan ICs on a PC board linked together in a Boundary-Scan chain for the
testing of interconnecting wires between the ICs.
dling instructions. Figure 3 shows the loop for handling test nents. It is the key Boundary-Scan instruction. It provides
data as defined by the IEEE Standard. Both loops follow an easy board-level interconnect testing of opens, stuck-at, or
identical sequence, the difference being whether the TAP con- bridging faults, etc. This instruction also facilitates testing
troller is interacting with a test data register or the instruc- of components that do not support the Boundary-Scan test
tion register. themselves but are surrounded by Boundary-Scan compo-
In order to understand how Boundary-Scan operates as a nents. Test values are applied externally to the chip via the
test technique, the key instructions need to be discussed. All boundary register which at the same time isolates the on-chip
instructions are serially loaded into the instruction register system logic from those test signals used to test the integrity
and are decoded to achieve two basic functions: (1) select the of external connections to the IC.
test data register that can operate while the instruction is The SAMPLE/PRELOAD instruction effectively provides
active, (2) control non-selected test data registers so that they two functions. It allows a SAMPLE of the normal operation
do not interefere with the normal on-board operation of the of a component to be taken for examination. Prior to the selec-
particular IC. Several instructions are mandated for Bound- tion of another test operation, PRELOAD loads values into
ary-Scan, and more are optional. Mandated instructions are the boundary register. The BYPASS instruction sets up a sin-
External Test (EXTEST), SAMPLE/PRELOAD, and BYPASS. gle-bit route between the TDI and TDO pin providing a short-
In addition, both optional instructions, such as Internal Test cut route between a component’s TDI and TDO pin for board-
(INTEST), Run Built-In Self-Test (RUNBIST), HIGHZ and level testing.
CLAMP, as well as user-defined instructions, such as ID- INTEST allows static, slow-speed testing of on-chip logic
CODE and USERCODE, are allowed. without affecting the operation of surrounding components on
EXTEST allows for the observation of the input cells and a PCB. As test data has to be applied serially to TDI, the
the control of the output cells of the Boundary-Scan compo- apparent testing rate is greatly reduced. The reduction is di-
BOUNDARY-SCAN TESTING 535
rectly proportional to the length of the boundary register. of its entire Boundary Register. If the optional device identi-
Test data is shifted into the boundary register and conse- fication register is included in the design, an identification
quently applied to the on-system logic. Test results are then code instruction (IDCODE) has to be included as well. It will
captured and shifted out for evaluation. Note that INTEST is provide additional information about the component. If the
an optional instruction. Internal testing of ICs can be facili- component is user-programmable, the user defined code (US-
tated with Boundary-Scan but is not guaranteed in a Bound- ERCODE) instruction has to be incorporated. It contains in-
ary-Scan compliant IC. RUNBIST offers a supplementary formation about the programming of the component.
method of assisting on-chip testing if the IC has BIST capabil- Figure 4 shows a possible implementation of a Boundary-
ity. It causes the execution of a self-contained self-test, with- Scan cell for unidirectional pins (2). Most Boundary-Scan cells
out the need to load complex data patterns or a single step contain a shift stage flip-flop and a parallel hold or update
operation as for INTEST. The optional HIGHZ instruction stage flip-flop. The shift stage flip-flop holds the test data
targets the Bypass Register and allows one to place all out- moving through the boundary register. The update stage is
puts of the Boundary-Scan IC into a high impedance (high Z) updated from the shift stage. For extremely performance-sen-
state. It delivers in-circuit isolation by shutting off the inputs sitive component inputs, the Standard allows a monitor-only
to other ICs on the PCB. Finally, the optional CLAMP in- Boundary-Scan cell which omits the update stage. In this
struction also targets the Bypass Register and forces fixed case, the cell design cannot support INTEST or RUNBIST,
values on an IC’s output pins without incurring the overhead because the system logic cannot be isolated from the effects
Device identification
register
Design-specific register
Bypass register
Exit 1 Exit 1
Pause Pause
Exit 2 Exit 2
of external signals. For Boundary Register support of bidirec- one checks that the boundary register, instruction register,
tional pins, two approaches are available. The first one uses and bypass registers of any Boundary-Scan IC can be loaded
two data register cells: one as an input and one as an output. with combinations of 1s and 0s.
The second one implements a single, reversible cell to perform Once it is established that the Boundary-Scan circuitry is
both the input and output function. In both cases, a control functional, EXTEST can be used to check for interconnect
cell is added that gives the Boundary Register control over opens and shorts. To test the behavior of interconnects con-
the output enables of the drivers. necting the outputs of IC1 to the inputs of IC2 and IC4 (see
Fig. 1), EXTEST is used to control the IC1 outputs (test vec-
tor) and observe IC2 and IC4 inputs (response vector). The
APPLYING BOUNDARY-SCAN TO BOARD TEST interconnects to IC3 are not tested, and therefore, the Bypass
register of IC3 is inserted in the chain. When there are no
Before using Boundary-Scan to check for interconnection faults, the node patterns on the IC1 outputs will match those
faults, one needs to verify that the Boundary-Scan circuitry on the IC2 and IC4 inputs. Identical node values on IC2 or
is present and working. First, one checks that the TDI, TMS, IC4 inputs suggest shorts (bridging faults) between the corre-
TCK, and TDO pins can be driven both high and low. Then, sponding interconnects. Opens will cause IC2 or IC4 inputs
Normal input
Normal output
Figure 4. Example of a unidirectional MUX
Boundary-Scan cell. The pins labeled
‘‘Scan Input’’ and ‘‘Scan Output’’ are the Scan input MUX D D
serial inputs and outputs of the Bound-
ary-Scan register. The regular input/out- Flip-flop Flip-flop
put pins of the cell are labeled ‘‘Normal CLK
Input’’ and ‘‘Normal Output.’’ All other CLK CLK
pins route control signals from the TAP Update Shift stage Update stage
controller to the cell.
BOUNDARY-SCAN TESTING 537
to float high. Not all interconnect tests are as straightforward time for the in-circuit test, compared with 30 minutes for the
as those outlined above. ICs are usually connected in a more ‘‘hot mock-up’’ test.
complicated fashion than described here. Faulty node values In a joint program with Hewlett-Packard (HP), boards
may match expected, correct node values unless both high were selected for analyzing and re-engineering manufactur-
and low values are checked. ing and test procedures. One board contained 1377 nodes,
In practice, detecting and diagnosing faults on board in- 5299 solder joints of which 4593 were testable, 450 discrete
terconnects is usually very complex. Jarwala and Yau (5) analog components, and 96 digital ICs, including 9 ASICS
have summarized the types of faults which can occur on (Application Specific ICs) with Boundary-Scan. A set of seven
boards and the effectiveness of various algorithms in detec- test types was able to achieve 100% fault coverage, even
tion and diagnosis. They consider both multi-net and single- though no single test type was able to achieve 100%. The tests
net faults. Multi-net faults correspond to bridging faults were mixed, including in-circuit probe access as well as
which create a short between two or more nets. A net on a Boundary-Scan. Notable benefits achieved were the elimina-
board corresponds to an equipotential surface which may con- tion of ‘‘hot mock-up’’ tests, the ability to distinguish between
nect multiple input drivers to multiple output buffers. Conse- design faults and manufacturing defects early in the product
quently, multi-net faults include OR-type (driver ‘‘1’’ domi- lifecycle, a 15% reduction in probe count for in-circuit testing,
nates), AND-type (drive ‘‘0’’ dominates), and strong-driver a one-third reduction in test development time, and a three
shorts (one driver dominates). Single-net faults are stuck-at- to four week reduction in board design time. Boundary-Scan
1, stuck-at-0, and open faults on single nets. To clarify the reduced the need for in-circuit test points while improving
issues in fault diagnosis, Jarwala and Yau (5) introduce a set fault coverage. 100% fault coverage guaranteed that incor-
of parallel test vectors applied to all nets in parallel, and the rectly functioning boards were the result of design errors.
corresponding set of sequential test vectors, the vectors ap- Boundary-Scan reduced the time to write ASIC tests because
plied to particular nets over time by a number of parallel test the test process is more automatic. It also reduced design and
vectors. The sequential test vectors for a faulty net provide fixturing time with only a 1 to 5% increase in IC costs to in-
a response which can be used for diagnosis. Effective fault corporate Boundary-Scan.
diagnosis requires identification of which nets are shorted,
open, stuck-at-0, or stuck-at-1. Fault diagnosis is complicated
when the faulty response of one net is the same as the fault- ISSUES IN APPLYING BOUNDARY-SCAN
free response of another net. In any case, interconnect fault
Motorola has reported on the implementation of Boundary-
detection and diagnosis require a careful choice of the test
Scan in the PowerPC RISC microprocessor family (7). The
patterns or vectors to be scanned into the Boundary-Scan
PowerPC 602, 603e, and 604 contain 1.6, 2.6, and 3.6 million
chain.
transistors and are designed for consumer electronics, porta-
Testing is simplified if in-circuit, bed-of-nails testing can
ble, and desktop PCs, respectively. Motorola implemented
be used to supplement Boundary-Scan (2). In particular, this
Boundary-Scan in a customized rather than a standard fash-
can help locate as well as detect faults, which is a great aid
ion. Their PowerPC implementation of Boundary-Scan (IEEE
to the manufacturer in diagnosing the causes of PCB failure.
1149.1) was also used for internal chip testing. This included
This is especially true for PCBs which are only partially popu-
IBM’s Level Sensitive Scan Design (LSSD) internal test meth-
lated with Boundary-Scan ICs. Boundary-Scan allows the
odology, embedded RAM BIST, IDDQ (quiescent CMOS
manufacturer to obtain a higher fault coverage that exceeds
power supply current), emulation, and debug support. Their
that from in-circuit testing. It can provide virtually 100% report focuses on troublesome implementation issues rather
fault coverage for opens and stuck-at pins, while providing a than typical design practice.
high coverage of shorts (bridging faults). A potential problem, Originally, LSSD-based testing was not fully compliant
as with all scan methods, is the time required to scan in and with the 1149.1 standard. In particular, unless the control
scan out long strings of test values. signals for LSSD testing exercised superiority over the 1149.1
Examining recent, representative applications of Bound- boundary register, as discussed in Appendix A of the 1149.1
ary-Scan illustrates many of the practical issues associated standard, it was difficult to maintain LSSD rules. This meant
with use of this technology. Matsushita Electric Industries that several of the LSSD controlling pins were not Boundary-
needed more cost-effective manufacturing and test methods Scan testable, thereby violating compliance (8). This problem
in the face of trends toward significantly more digital compo- was solved by revision 1149.1a-1993 which introduced the
nents, reduced nodal access on PCBs, and an increasingly ‘‘compliance enable’’ concept. Private instructions rather than
competitive marketplace for its diversity of consumer and pro- RUNBIST were used to execute an embedded RAM BIST be-
fessional electronics products (6). Their strategy was to obtain cause the control structures for initialization were not com-
100% fault coverage as early as possible in the manufacturing patible. Private instructions were also used for IDDQ testing,
process. This required tests for opens and shorts on all testa- emulation, and debug support. They note that in PowerPC
ble solder joints, checking the presence and basic function of designs, most implementation difficulties were related to im-
all components, and checking component pins for damage due plementing and verifying SAMPLE rather than EXTEST (5).
to electrostatic discharge. Their assembly yield was about This was a surprise because the opposite had been believed
85%. In-circuit test detected around 70% of the faults, re- during the development of the 1149.1 standard.
sulting in 95% yield after in-circuit test and repair. The re- Designers generally divide device operation into normal
maining faults were not found until a ‘‘hot mock-up’’ test or system and test modes and are aware that SAMPLE must not
after final assembly. The increased cost of low, early fault cov- interfere with normal system operation. However, there is a
erage can be seen in the 5 minute average test and repair third mode, non-normal system operation, which includes fac-
538 BOUNDARY-SCAN TESTING
tory test, debug, power management, and system hard reset. checking compliance and the openness of the 1149.1 standard
Insufficient consideration of non-normal system operation led to user-defined extensions of arbitrary complexity. The do-
to compliance difficulties with the 1149.1 standard. main of possible violations becomes infinite when implement-
For example, they note that the PowerPC 603e includes a ers attempt to map a written-in-English standard to elec-
system hard reset pin (HRESET) and a checkstop input pin tronic circuits.
(CKSTP). When HRESET is asserted, the IC is reset, and all Since IC designers rely heavily on simulation for design
output drivers are placed in a high impedance state. CKSTP verification, the Boundary-Scan Description Language
places all the output drivers in a high impedance state, dis- (BSDL) (8) can be a great help in insuring compliance. BSDL
ables most input receivers, and suspends instruction execu- allows the testability features of components complying with
tion. Directly connecting both HRESET and CKSTP to the the 1149.1 standard to be described in software-readable lan-
driver control logic presented several difficulties when imple- guage. Writing BSDL can uncover compliance errors in the
menting the SAMPLE instruction. One difficulty was that the implementation of Boundary-Scan circuits. For example,
output enable control cells failed to capture the effect of HRE- BSDL cannot describe an illegal configuration in which sys-
SET or CKSTP on the output drivers. The solution was to tem logic is placed between boundary register cells and the
disable the output driver through an output control cell by I/O pins. Programs with different levels of sophistication can
adding CKSTP and HRESET as inputs. be written to check compliance. BSDL was developed as IEEE
Careful design was also required to ensure 1149.1 compli- Standard 1149.1b-1994. Parker (11) notes that, in retrospect,
ance when providing input isolation during power down oper- if the development of BSDL had paralleled the development
ations. SAMPLE captured the state of the Update latch of 1149.1, many of the ambiguities that led to a rewrite of the
rather than that of the input pin. The solution was to disable 1149.1 standard in 1993 might have been avoided.
the Update latch during SAMPLE, forcing the receiver multi- BSDL specifies those parameters which are unique to a
plexer to select the input pin. particular Boundary-Scan implementation; those elements of
Three output pins were initially incorrectly classified as a design which are mandated by 1149.1 are not included in
two-state outputs because the signals were released to high BSDL descriptions. For example, neither the bypass register
impedance only during system reset and checkstop condi- nor the TAP state diagram are described. To assure wide-
tions. They were reclassified as three-state outputs, and an spread use, BSDL has been implemented as a subset with
additional output enable control cell was added for these out- standard practices in VHDL, the VHSIC (very high speed in-
put drivers. They observed that other groups (9) have noted tegrated circuit) Hardware Description Language. As Tege-
incorrect classification, suggesting that this is a common
thoff and Parker note (10), simulation of 1149.1 features with
problem.
verification vectors from automatic test pattern generators
Motorola found that compliance-enable pins and private
(ATPG) can ensure that the BSDL description is correct and
instructions were valuable tools with which to support the
matches the implementation in silicon.
sometimes conflicting demands of chip-level and board-level
ATPG simulation success is a de facto test for compliance
test. Although a considerable amount of internal testing is
and functionality. However, since full compliance is impossi-
incorporated in the PowerPC family, INTEST was not sup-
ble to verify, they believe more robust test vectors are re-
ported. One reason was that the cost of documentation and
quired. For example, robust test vectors are used to check the
consumer support for INTEST appeared to be very high.
timing relationships between TAP signals. In particular, Teg-
Since both private and public test domains are useful, they
ethoff and Parker (10) consider the timing relation between
recommend that the 1149.1 standard should be expanded to
define what constitutes reentry into the public domain from TCK and TDI. ATPG will most likely have TDI and TCK
the private domain. changing on the same vector to minimize the total number of
Recently, Tegethoff and Parker from Hewlett-Packard (10) vectors. All vectors change on the falling edge of TCK, ac-
reviewed the current status of the 1149.1 standard. They ob- cording to the standard, but the proper timing is not checked.
served that this standard had become the methodology of They suggest that a subset of the vectors be generated in
choice for discovering manufacturing defects and enabling which TDI and TMS change a vector before, at, and after the
functional tests. Having long used Boundary-Scan tech- falling edge of TCK to check TDI and TCK timing. This will
niques, their company has converted their internal Boundary- verify that the TAP samples TDI at the correct TCK edge.
Scan protocol to the 1149.1 standard. They present the per- They also raise the question of acceptable violations of the
spectives of a user of the standard and a developer of test 1149.1 standard. They note that while test experts will de-
software supporting the standard. mand full compliance, design experts can often make a case
They note that one problem with verification is fixing re- for a tolerable violation. Of course, chips sold as 1149.1 com-
sponsibility for verifying compliance in an implementation. pliant should be fully compliant. However, they regard viola-
Test experts typically lack the VLSI (very large scale integra- tions in internally-used ICs tolerable if they are transparent
tion) design skills to evaluate the implementation, while de- to any tester or ATPG tool used subsequently and if they will
sign experts have not mastered enough of the subtleties of not hinder use of the standard in the board and system test
1149.1 to assure compliance. They find that most problems strategy. As an example, they consider a glue logic chip de-
associated with 1149.1 in the board test environment come signed in gate array technology for a board with an 1149.1-
from design errors, which are the result of misinterpretation based test strategy. If designers cannot fit the gates necessary
or involuntary violation of the standard. Parker (8) notes that to implement the Boundary-Scan Register (BSR) in the gate
for several reasons the verification of compliance is very diffi- array, they might use external, discrete gates to form the BSR
cult, and a guarantee is virtually impossible. Some reasons cells. An alternative, which might fit all BSR cells on the chip,
he gives are the lack of formalized rules or procedures for is to treat bidirectional I/Os as input-only. Although this vio-
BOUNDARY-SCAN TESTING 539
lates the standard, the loss of coverage on the path from the and provide support for performance/delay testing. P1149.4
driver to the bond pad would be an acceptable risk. allows testability of analog pins and passive components con-
Like all standards documents, 1149.1 is the consensus of a nected to digital pins. P1149.5 provides a protocol to facilitate
committee; revisions and supplements should be expected in operation of a module test and maintaince bus. An important
light of experience. For example, Tegethoff and Parker (10) question for the usefulness of all these standards is whether
discuss subtleties in implementing the SAMPLE instruction. or not they address the needs of a significant constituency in
They suggest that the 1149.1 working group should consider the electronics industry (10).
demoting SAMPLE to an optional instruction, since it is The intent of Standard P1149.2 is to support Boundary-
tricky to implement realistically and may require inordinate Scan for board-level testing as well as internal scan with min-
overhead for its support. This requires that SAMPLE is first imum mandatory features (13). P1149.2 would offer features
divorced from PRELOAD and that PRELOAD continues to similar to those provided by 1149.1, but with two significant
be mandatory. differences. First, the P1149.2 BSR cells can be shared with
They also observe that of all the requirements in the the core logic of the component and are not required to have
1149.1, reset has become a bone of contention for the design separate serial-shift and parallel-update stages. Second,
community. Implementing reset on a Boundary-Scan chip re- P1149.2 uses a direct, parallel-access method to enable the
quires consideration of the reset strategy for the entire sys- different test modes. To allow the use of software tools devel-
tem. It is natural to tie TAP resets to the board power-up oped to support 1149.1, a version of P1149.2, P1149.2T, has
reset. However, if the board tester or one of the chips asserts been proposed which is backward compatible with 1149.1 but
reset during a Boundary-Scan test, the TAPs will also be re- provides additional functions. However, it is likely that
set. Thus, two board resets are required to ensure no sur- P1149.2 may not become a standard.
prises during turn-on. They suggest that one might want to The IEEE P1149.4 working group is defining a mixed-sig-
modify the 1149.1 standard to make the power-up reset re- nal test bus that addresses issues of analog testability (14).
quirement optional. Desirable goals are the ability to stimulate any analog pin
Parker (8) distinguishes non-invasive and pin-permission with an analog source and to monitor any analog pin with an
modes of operation in the 1149.1 standard. The standard analog measurement instrument. A core disconnect capability
specifies a set of Boundary-Scan resources which are guaran- is required to turn off the pins’ core drivers and, perhaps,
teed to be independent of the IC’s internal logic. In non-inva- guard the core from tester-driven stimuli. Since many mea-
sive mode, the Boundary-Scan resources communicate asyn- surement scenarios require a guard point through which to
chronously with external circuits to set up tests or read out route currents, the proposed P1149.4 architecture includes a
results. These activities are invisible to normal IC operation. switch to ground. P1149.4 is an extension of 1149.1 and de-
In pin-permission mode, the I/O pins of the IC are Boundary- pends on the existence of four or five 1149.1 TAP interface
Scan controlled, effectively disconnecting the IC’s internal pins and support circuitry.
logic from external circuits. The P1149.5 standard is for a module test and mainte-
This distinction is important when power is applied to a nance (MTM) bus whose protocol standardizes a method of
board or system to ‘‘bring it to life.’’ Applying power must communicating test and maintenance commands and serial
bring the system to an initial state which is a stable starting data between a subsystem test control module (bus master)
point for future behavior. All Boundary-Scan ICs must ‘‘wake and the other (slave) modules on the bus (15). The MTM bus
up’’ in non-invasive mode. When any one of the Boundary- supports module test, subsystem test, and subsystem diagno-
Scan ICs switches to pin-permission mode, this constitutes sis using observability and controllability techniques such as
‘‘radical surgery,’’ and great care may be needed in ‘‘post-oper- scan and Boundary-Scan. It extends a standard test and
ative recovery.’’ Parker (8) calls this the ‘‘lobotomy problem.’’ maintenance protocol developed in the U.S. Department of
What should an IC do when ‘‘waking up’’ from pin-permission Defense’s VHSIC program. The MTM bus has a multidrop to-
mode to assure that the system doesn’t enter an unsafe state? pology which supports these applications and allows one to
Tegethoff and Parker (10) note that the lobotomy problem is remove a board without breaking the communications link be-
as important and potentially more dangerous to the system tween other modules on the backplane.
than the reset problem. To avoid this problem, Hewlett-Pack-
ard devised the drive inhibit (DRV_ENH) instruction to put BIBLIOGRAPHY
the IC in a safe state with no possible driver fights and the
drive enable (DRV_EN) instruction which restores the IC to 1. IEEE/ANSI Standard 1149.1a, A Standard Test Access Port and
normal bus operations (12). Boundary Scan Architecture, New York: IEEE Standards Board,
1993.
2. B. Geuskens and K. Rose, Implementing a CMOS boundary-scan
EXTENSIONS TO THE 1149.1 BOUNDARY-SCAN STANDARD architecture, Proc. IEEE Int. ASIC Conf., 1994, pp. 392–399.
3. L. Y. Ungar, BIT in a boundary-scan environment, EE-Evalua-
In addition to possible, desirable modifications of the 1149.1 tion Engineering, 29 (11): 141–149, 1990.
standard as discussed above, there is a related 1149.X family 4. GenRad, Inc., Meeting the Challenge of Boundary Scan, Concord,
of standards (11) which extend the Boundary-Scan concept. MA: GenRad, Inc., 1991.
The P prefix denotes a standard which has not completed the 5. N. Jarwala and C. W. Yau, A New Framework for Analyzing Test
acceptance procedure. The 1149.1 standard has been criti- Generation and Diagnosis Algorithms for Wiring Interconnects,
cized as a CMOS (complementary metal oxide semiconductor) Proc. IEEE Int. Test Conf., 1989, pp. 63–70.
compatible standard which is too burdensome for other pro- 6. P. Milo, Success with boundary scan, EE-Evaluation Engineering,
cesses (10). The P1149.2 standard attempts to address this 34 (2): 72–75, 1995.
540 BOUNDARY-VALUE PROBLEMS
BIBICHE GEUSKENS
KENNETH ROSE
Rensselaer Polytechnic Institute
COMPUTER INTEGRATED MANUFACTURING 739
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
740 COMPUTER INTEGRATED MANUFACTURING
be removed to allow for concurrent engineering and multi- and optimize the interactions between tasks and functional
functional teamwork. Information flow from customer order blocks. Functional blocks or modules are machines, worksta-
to engineering to production preparation to manufacture tions, workcells, or software modules where complex opera-
must be streamlined (1). tions, or processes are performed. CIM deals with the effec-
Manufacturing technology started about 200 years ago tive use of computers and automation in manufacturing to
with the English system. There have been six major events in improve production (quality, yield, cycle time) mix of products
the history of manufacturing. Each manufacturing event has running concurrently, and reduce the unit product cost (UPC)
been triggered by the development of new technology that by reducing the recurring and nonrecurring engineering
represented a major milestone in the solution of a particular (NRE) costs. The recurring manufacturing cost of a product
problem. The new technology required changes in the nature consists of the cost of material, labor, and overhead (MLO).
of the organization of manufacturing and in the machines NRE cost is the cost of product design and development and
used to make these changes (2,3). the investment in manufacturing, tooling, machinery, and
The first major manufacturing event was the English sys- processes. It should be mentioned that automation, in and of
tem of manufacturing, originated in the late 1700s with the itself, is not the answer to low cost, and high-flexibility manu-
invention of general-purpose machine tools, such as lathes, facturing. Automation is a tool, and can be used correctly or
that could be used to fabricate a variety of work pieces. This incorrectly. FM is the application of CIM to increase the
development was a breakthrough because it separated the range of products and to decouple UPC from volume (i.e., pro-
product’s function from the process used to make it. As a re- duce products at low cost and low volume). CIM/FM are ap-
sult, improvements in manufacturing were made indepen- propriate manufacturing technologies for cost-effective manu-
dently of the products. Consequently, the manufacturing facturing. Technologies to manufacture affordable devices and
technology became free from the constraints of the product. systems are of significant interest to industry and govern-
The English system emphasized craftsmanship, a one-at-a- ment. With CIM, flexibility in the form of new products and
time method. processes is achievable. CIM provides the capability of per-
The American system of manufacturing emerged in the forming sequential and parallel tasks simultaneously, which
mid-1800s, emphasizing mass production and interchange- increases the versatility and productivity of a CIM factory (1).
ability of parts. The English system emphasized the best pos- Process equipment that is compatible with CIM contains em-
sible fit among components, whereas the American system bedded computers (hardware and software) in the form of
aimed at the greatest possible tolerance without loss of func- micro-controllers, field-programmable-gate-arrays (FPGA),
tionality. The American system is a mass production method digital signal processors, and modems to communicate with
of making many parts of one kind and assembling a product and control other equipment. Process equipment used in con-
from a set of these parts. The manufacture of a rifle is a good ventional factories is not suitable for CIM. From a system
example of the American system of manufacturing. Managing engineering point of view, a CIM factory is network-centric,
variation became the hallmark of the American system. meaning that machines, operators, processes, and operations
The third major manufacturing event is called the period are under computer control through the factory local-area-
of Scientific Management, which began in the late 1800s. Sci- network (LAN).
entific Management was based on the work of Frederick Tay- Over the last four decades, drivers in manufacturing for
lor, an American mechanical engineer. Taylor advocated the competitive success have been:
idea that workers were limiting the speed and efficiency of
machines. Using job analysis and time study, he determined 1. Cost advantage: characterized by highly productive la-
a standard rate of output for each job. The approach of stan- bor, high volume, and low mix of products.
dardization and control of machines narrowed the scope of 2. Quality: emphasizing statistical process control, vari-
work, and left nothing to the worker’s decision. Taylorism ability reduction, and customer satisfaction.
placed the control of manufacturing in the hands of manage-
3. Time-to-market: emphasizing cycle time reduction, inte-
ment, which could monitor a worker’s productivity by compar-
grating product and process development, using produc-
ing it against a standard.
tion and process simulation, and requiring a highly
The fourth period occurred in the mid-twentieth century.
skilled and adaptable work force.
It is based on statistical process control (SPC), which was in-
vented in the United States. In SPC it is assumed that ma- 4. Product variety: characterized by UPC, which is inde-
chines intrinsically produce definable variations. SPC empha- pendent of volume, flexibility, and simulation of product
sis is on out-of-control situations rather than on mean and factory.
performance; therefore, it directs management’s attention 5. Company goodness: emphasizing environmental im-
away from the worker and toward man–machine variations. pact.
The fifth major event in the history of manufacturing is
the introduction of numerical control (NC) in the mid-1940s. Businesses are using several computer-aided techniques in
NC became more mature in the late 1970s with the embedded conjunction with CIM for accelerating the design and produc-
microprocessor-based controller. NC emphasizes monitoring tion of products. Among those that have proven successful
and controlling machines and encourages experimentation are concurrent engineering, just-in-time (JIT) and just-in-
and learning (2,3). sequence (JIS) inventory control in manufacturing and elec-
In the late 1980s, manufacturing entered a new era, com- tronic design automation (EDA) tools (i.e., rapid prototyping
puter integrated manufacturing (CIM) and flexible manufac- and virtual prototyping). Computer simulation of manufac-
turing (FM). CIM is based on the use of information and mod- turing processes has proven to be a very useful tool in comple-
els of functional expertise that make it possible to examine menting the practical know-how of manufacturing planners
COMPUTER INTEGRATED MANUFACTURING 741
and designers. In the late 1980s, the Defense Advanced Re- lows a disciplined series of steps to focus on the objectives and
search Projects Agency (DARPA) and the U.S. Air Force spon- the desired attributes of a product prior to building it. One
sored the Microelectronics Manufacturing Science and Tech- aspect of the system engineering process consists of a logical
nology (MMST) program at Texas Instruments. The program sequence of activities and decisions that transforms a cus-
was a detailed approach to reduce the principles of CIM to tomer specification into a product performance specification,
practice on a factory floor. The objective of the program was a design specification, and a preferred product configuration.
to create a CMOS Integrated Circuits factory with flexibility The design concept aspect is concerned with system design
to process each wafer using different design, with all schedul- and analysis, system integration, and tradeoffs between fac-
ing and processing fully under the control of a centralized tors such as performance, reliability, safety, cost and avail-
system. ability. Another aspect of the system engineering process is
concerned with the system engineering management activi-
ties, which deal with the application of business disciplines
SYSTEM ENGINEERING IN MANUFACTURING such as planning, budgeting, costing, schedule control, policy
direction, acquisition strategy, resources management, and
Recently, the design, development, and manufacture of sys- the evaluation of program performance. A comprehensive sys-
tems has become more complex and has increased the need tem engineering process deals with both aspects (5–7).
for systematic techniques, principles, and methodologies. Sys- During the conceptual phase of system development, a
tem designers and integrators have been faced simultane- methodology must be used to determine the system require-
ously with increasingly complex technical demands under ment, tradeoffs, input/output, and operational environment.
challenging schedule and budget conditions, acutely limited System engineering is the process dealing with all aspects of
resources and test data, and fast-changing technologies. Per- system design. The system engineering process consists of
formance requirements invariably include severe reaction and several phases: (1) requirements/objective determination, (2)
response time constraints that cannot be met without the system performance specification, (3) establishment of system
proper integration and resource allocation of personnel, hard- architecture and tradeoffs analysis, (4) detailed design and
ware, software, and procedures. At the same time, system de- tradeoff of building blocks, (5) integration, costing, testing,
velopment cost reductions, accelerated schedules, and lack of and design specification of the building blocks, and (6) devel-
complete system tests prior to operational usage have com- opment of metrics to evaluate these activities. The develop-
bined to reduce the availability of adequate operational data ment and design of a CIM factory follows a system engi-
bases, either in quality or quantity, and increase the chal- neering approach. The CIM factory is the system, and the
lenge of system integration. The present business environ- building blocks of the factory are those of the system.
ment involves global competition, and the customer is focused After the CIM system architecture is established, the next
on acquiring cost-effective (affordable) systems. Accordingly, step is synthesizing the required functionality to design the
it is necessary to have a comprehensive methodology of sys- building blocks. The design process starts by simulating the
tem design, integration, management, and analysis using all system using the performance specification and generating
available information to pinpoint problem areas, produce several designs with different hardware and software config-
competitive and affordable products, and provide a numerical urations. During the design process, tradeoffs are made be-
estimate of system effectiveness during all phases of the sys- tween performance, cost, safety, reliability, and supportabil-
tem life cycle (4,5). The production of competitive products ity. The end product of the detailed design process is the
requires a system engineering approach to manufacturing. design specification of the hardware and software building
In today’s environment, everyone seems to be working on blocks.
systems or subsystems of some kind. There are communica- System integration is a multidimensional process. At a
tion systems, political systems, transportation systems, eco- given partitioning level, system integration is the process of
nomic systems, radar systems, information systems, manu- blending the hardware and the software building blocks into
facturing systems, and production systems, to mention but a a functioning whole. System integration is the process of com-
few. One might ask the question, Can all these truly be called bining the building blocks, functionally, logically, physically,
systems? An attempt to answer this question requires exam- and socially. The integration of the system with other systems
ining the real world, which is made of material things, organ- and into the operational environment involves supportability,
isms, and ideational constructs that are variously related to interoperability, and readiness activities, such as, repair,
one another, structurally and functionally. Both the func- maintenance, logistics, training, and dealing with interfaces.
tional and structural relationships tend to span all three spa- Vertical system integration is the process of uniting building
tial dimensions and time. The multidimensionality of systems blocks at different levels—system within a system, wheels
leads to a hierarchy of systems. Hierarchy of systems means within wheels. System integration represents a major part of
there is a vertical and horizontal structure of systems. The system engineering (1,5). The design of a CIM factory re-
vertical structure consists of systems embedded within sys- quires a methodology for integrating and controlling the di-
tems. The horizontal structure is characterized by mutual in- verse equipment in a manufacturing facility. The methodol-
terdependence among the units at each tier, where the units ogy must account for the hardware and software interfaces.
are operationally dependent on those of the neighboring tiers. Each manufacturing domain will have a generic interface
Consequently, the term system has broad application. model (GIM) to determine the common interfaces to the di-
System engineering is a multidisciplinary subject; it uses verse equipment, and a framework for the development of ma-
all the engineering disciplines, operation research, manage- chine control software. The basic hardware interfaces deal
ment science, and economics to provide a systematic approach with sensors and actuators. However, there are several cate-
and rationale for building systems. A systematic approach fol- gories of software interfaces called middlewares.
742 COMPUTER INTEGRATED MANUFACTURING
server, workstation, or database or a melting reactor for dop- group tools (groupware) (e.g., e-mail, bulletin boards, fax). In
ing and growing materials. The fourth level in the hardware general, there is a multiplicity of possible architectures. The
hierarchy is made up of function blocks. A grouping of func- determination of the optimal architecture is the subject for
tion blocks makes up a module. A function block consists of further evaluation and tradeoffs (1).
machines, devices, or components that perform specific and At the top level, the system architecture is represented by
simple functions such as drilling, washing, polishing, slicing, a set of interacting subsystems. A logical step in producing a
etching, adding, dividing, and multiplying. The circuit level detailed system architecture is to decompose the subsystems
(fifth level) is characterized by the interconnection of devices into individual hardware and/or software configuration units
that perform simple processes such as amplification, rotation, (CU) that can be developed or acquired. System engineering
linear motion, switching, and controlling. The aggregate of principles are applied to the process of decomposing the sub-
these simple processes contributes toward the realization of systems into configuration units to ensure that the configura-
functions (e.g., circuits that control functions of pressing, tion units are identified in a consistent and rational manner.
heating, or speed control). Also included are logical gates such The application of these system engineering principles to the
as electronics or fluidics gates (e.g., AND, OR, NOT, NAND, decomposition of subsystems:
and NOR gates). Similarly, mechanical or pneumatic gates
controlling valves, pumps, or heat exchangers in a reactor are 1. establishes the initial functional boundaries and inter-
examples of circuit level building blocks. The sixth or the ele- faces consistent with the requirements around the CUs.
ment level in the hierarchy deals with devices, simple pro- 2. allows the identification of CUs to be uniquely assigned
cesses, or components that, when interconnected, perform the to a specific organization for development or acquisition.
circuit-level operations. For example, the heating elements in
3. determines which CUs are to be implemented by hard-
a furnace, the valves of a heat exchanger, the pumps in a
pneumatic circuit, or the electronic devices such as resistors, ware and which are to be implemented through soft-
capacitors, ICs, ASIC (application specific integrated circuit) ware.
chips, microprocessors, motors, and generators, all are exam- 4. identifies which CUs are to be supplied via commercial-
ples of components of a factory hardware hierarchy. off-the-shelf (COTS) products and which are to be de-
The vertical structure of the hierarchy consists of a multi- veloped.
plicity of structures nested within the system. The horizontal
structure of the hierarchy is characterized by mutual support, Applying the same principles to the decomposition of all sub-
interaction, or interdependence among the units at each level. systems will ensure that the resulting CUs are identified in a
Achievement of higher-level objectives depends upon the func- consistent manner.
tioning of the lower-level structure. Software architecture is concerned with the high-level par-
The partitioning of a system into hardware and software titioning of software into major clusters and the specification
is a critical system engineering task because it has a serious of those clusters. It also includes the selection of the environ-
impact on the cost and performance characteristics of the fi- ment for software development (e.g., languages, library struc-
nal CIM factory design. Partitioning decisions performed by a ture, and standards). Software design includes all phases of
designer or by a computer-aided-design (CAD) tool must take software development that precede coding, and as such in-
into account the properties and cost of the resulting hardware cludes software architecture. It also includes the iterative de-
and software building blocks. The tradeoffs in the parti- sign activities of decomposing software modules, leading to a
tioning problem are based on the architectural assumptions, level of detail that allows code to be written in the selected
partitioning objectives, and market constraints. language. Software development and hardware design are
closely coupled. Tight linkages between them are important
especially in software designed to be embedded in hardware
SYSTEM ARCHITECTURE AND DESIGN OF CIM where changes in either one usually require changes in the
other (8–10).
The system architecture gives the top-down description of the The next step in producing detailed CIM system architec-
underlying structure and functions of the system. System ar- ture is to represent the interacting CUs by the function block
chitecture provides the basis for separating the hardware diagram of the system. The function block diagram gives the
from the software associated with the architectural building interconnection and the functionality of the building blocks.
blocks. It gives a combined view of the system operations and In manufacturing, computers are embedded at all levels of
the interaction of the building blocks. An integrated system the factory. They are used in process control in the form of
architecture is a set of interconnected subsystems and their microcontroller, field-programmable-gate array (FPGA), and
characteristics. In terms of design details, there is a hierarchy digital signal processors (DSP). Personal computers (PCs),
of system architecture. The hierarchy of system architecture workstations, databases, and mainframes are used within the
is determined from the ‘‘flow’’ of decisions, information, mate- factory to communicate and to command and control func-
rials, and operations that must be made to accomplish the tions. Computers and the associated software play a critical
objective. The flow is based on the functions to be performed role in the functioning and operation of CIM. Network com-
and the constraints of the environment. The sequence of the puting is used in computer-aided engineering (CAE), com-
flow will determine an architecture of the system. There are puter-aided design (CAD), computer-aided manufacturing
five classes of flow that should be accounted for: (1) decision (CAM), computer-aided testing (CAT), communication, data
flow, (2) information flow (i.e., signal, data, and document), processing, finance, simulation, and virtual prototyping. Vir-
(3) material flow, (4) operations flow (i.e., events, processes, tual prototyping is an integration of data from various
and tasks), and (5) flow of housekeeping tasks such as work- sources that define the total product and its environment. The
744 COMPUTER INTEGRATED MANUFACTURING
Array
design
Customer Cooled
interaction workcells
• Virtual prototype
Customer • Array design
• Electronic
Specs • Subassembly
model
• Spec flow down
Material & Electronic circuits Packaging
detector design design Ship to
• From qualified
suppliers Design Cooled & uncooled Packaging & test customer
workcells workcells workcells
• Wafers
• Substrates
• Ingots
Array
design
Uncooled
workcells
Figure 2. Configuration Units (CUs) of CIM factory to produce cooled and uncooled IRFPA. The
factory architecture is derived from the flow of material and operations.
virtual prototype of a product is built electronically in three- deposition of thin films, photomasking, and mechanical
dimensional solid digital models and evolves throughout the handling (thinning, polishing, dicing).
design process. In CIM virtual prototyping capability allows 3. The CU for Array Design of cooled IRFPA includes
the customer to interact with the factory. The CIM factory capabilities for array layout design, detector size, and
receives customer specification through the internet, intranet, interfaces with the electronics and the cryogenics.
and/or the WWW. Through virtual prototype software the 4. The CU for Array Design of uncooled IRFPA deals with
customer interacts with the factory management and/or engi- array layout design, detector size, thermal isolation,
neering to generate a virtual product. The customer states a and detector interfaces with the electronics.
specification and the virtual prototype evaluates the specifi-
5. The Electronic Circuits CU deals with the design, inte-
cation and responds if the CIM factory can make the product,
gration, and manufacture of input circuits, control cir-
determines the cost of the product, and calculates when it can
cuits, multiplexers, amplifiers, and interfaces with the
deliver the product.
detectors.
The preceding principles are demonstrated in an example
(i.e., the design of a CIM factory to produce infrared (IR) sen- 6. The Packaging and Testing CU deals with the final as-
sors). The following example is based on projects sponsored sembly of the focal plane array, design and manufacture
by DARPA to build a Flexible Manufacturing (FM) factory to of the dewar and cryogenic cooler, and the testing of the
make cooled and uncooled Infrared Focal Plane Arrays (IR- IRFPA (1).
FPA). A focal plane is a two-dimensional integrated circuit
having infrared sensitive elements. The size of the array var- In order to design the CIM factory, the factory software must
ies from 2 ⫻ 40 to 1024 ⫻ 1024 or larger. The IRFPA can account for the software needs within each CU and the con-
operate at different wavelengths. Figure 2 shows the system nectivity with the rest of the factory. Software interfaces of
architecture of a CIM factory. The factory is planned to pro- the CIM factory provide very serious system engineering/inte-
duce cooled and uncooled infrared-focal-plane-array (IRFPA) gration challenges. Figure 3 gives an open system architec-
sensors. The IRFPA factory configuration block diagram con- ture block diagram of the IRFPA CIM factory. The diagram
sists of six CUs. The diagram was derived from the flow of shows where the various applications software are used, such
information, materials, and operations that must occur to as computer-aided-design, computer-aided-manufacturing
make the products (1). (CAM), computer-aided-testing (CAT), and just-in-time sched-
uling. The factory is connected through a local area network
(LAN). The implementation of open system architecture is re-
1. The Customer Interaction CU provides an evaluation of alized by using the client/server configuration. In CIM soft-
customer specification and the design of the IRFPA us- ware interfaces are critical in achieving a seamless and
ing virtual prototype software tools; it matches the cus- smooth-running factory.
tomer specification to the factory capability. Figure 4 shows the factory client/server architecture,
2. The Material and Detector CU deals with all aspects of which uses a two-tier configuration. A client/server architec-
the detector material, such as preparation of substrates, ture is used because it provides: (1) factory flexibility at low
COMPUTER INTEGRATED MANUFACTURING 745
Customer
(RDBMS). The client/server architecture is managed by data-
CAM interaction base connectivity software such as the Digital Equipment
CAT virtual prototype Corporation (DEC) Framework Based Environment (FBE)
Rapid design rules middleware or the Microsoft Object-Linking-and-Embedding
prototype (OLE) middleware. In companies around the world, client/
Inventory server (C/S) systems are helping people work together more
Cooler
packaging/ control efficiently, gain access to vital information, and improve their
testing Scheduling productivity. Client/server architecture requires software de-
JIT
velopment tools to integrate the application within the system
architecture. Visual Basic 4.0/5.0 for Windows or Visual C⫹⫹
Factory are popular 32-bit software development tools for wrapping in
database the client/server architecture. Software wrapping or encapsu-
and toolset LAN lation deals with the software interfaces that are necessary
to make software applications transparent to the operating
system. Visual Basic has a ready-made OLE middleware, and
Material
Array design processing and it can provide software wraps as remote automation objects
cluster detector design on Windows server. Middlewares are programming tools that
provide interoperability and portability within the various
CAM CAM software applications. Middleware gives the system the abil-
Electronic ity to integrate separate hardware platforms, networking pro-
circuit design tocols, legacy applications, LAN environments, and desktop
and layout operating systems (1,11–14).
As manufacturers scale up the CIM architecture to enter-
CAD Workstation prise-wide distributed systems, they are using more complex
three-tier architectures. A three-tier client/server architec-
Figure 3. Block diagram of an open architecture of CIM of a typical
ture provides a more comprehensive functionality spanning
IRFPA factory. The diagram illustrates the information and software
connectivity of the CIM factory.
the enterprise and increases the system flexibility, but it is
more complex. A three-tier architecture decouples the appli-
cation logic from the user interface and the database. A three-
cost, (2) good end-user access to factory data, and (3) excellent tier architecture adds an intermediate layer of servers that
scalability. The two-tier client/server configurations provide a supports application logic and distributed computing services.
simple division of labor. The client application, tier one, per- The intermediate servers facilitate scalability and reusability
forms the user interface, application processing, and business into the client/server environment. Instead of writing the
logic. The server, tier two, performs the database manage- same function or service in every application, application de-
ment, usually by means of a Structured Query Language velopers can write them once and place them on a server,
(SQL)-compliant Relational Database Management System which is accessible by all applications. Effective application
Server 7
Server 1 Server 3 Server 5
Factory Design
Virtual prototype Design Analysis
control tools
and repository service service
service
Figure 4. Two-tier client/server architec-
ture using Framework Based Environ-
Server 2 Server 8 ment (FBE) in factory management. FBE
Server 4 Server 6
Servers Customer Factory is the middleware of the factory software,
Simulation Documentation
requirement control
service service it provides the interface (hooks, protocols,
service service
wraps) between the various applications.
746 COMPUTER INTEGRATED MANUFACTURING
Performance
spec System SW
Software
heirarchy
Subsystem SW
Module SW
Function Enterprise
block SW
Cluster
Circuit SW
Software module
Design Element SW
spec
Application
Figure 8. System Engineering design ap-
proach showing software embedded into Operation
the hardware. The figure shows a snap-
shot of the interaction of the embedded
software throughout the CIM factory. Codeline
may require numerous iterations to finalize the system design and users can participate and collaborate in the various
specifications. Throughout the integration process, decisions stages of the development process. Collaboration in system
must be made as to what should be automated (i.e., hardware design requires tools and applications that provide visualiza-
and software) or what should be done manually by an opera- tion of the system development process. Usually, complex sys-
tor. For example, the objective of CIM is to use computers and tems are specified in a lengthy document that may contain
automation to increase the effectiveness of manufacturing; fuzzy requirements. Collaboration and visualization will clar-
this requires examining the tradeoff of manual operator activ- ify and sharpen the understanding of the requirements.
ities (touch labor) versus automated tasks that can be per- The preceding system engineering principles are used to
formed by machine. Automation is appropriate whenever it is structure graphical simulations to provide meaningful visual-
cost effective and can show payoff considering performance, ization of system specification. The ability to see a visual rep-
reliability, quality, cost, safety, and the like. The information resentation of both the hardware and the software functions
flow of the factory must be analyzed to determine what should removes ambiguity from the specifications. Currently, graphi-
be implemented through the databases and software applica- cal simulation and synthesis tools use a common visual de-
tions or in the client/server terminals. The analysis will pro- sign process called visual collaboration.
vide insights into the tradeoffs between hardware, software, Advances in technology, the emergence of standards, elec-
and manual operator activities. Today’s products are differen- tronic commerce (EC), electronic data interchange (EDI), me-
tiated and upgraded by changing hardware and more often tered software, global competition, and the availability of
changing software. computer hardware and software are (1) providing the oppor-
Embedded software is pervasive; it is built into CIM sys- tunity for automating the business processes and functions of
tems through the microprocessor or microcontroller, field-pro- an enterprise and (2) leading companies around the world to
grammable gate array (FPGA), digital signal processor (DSP), reengineer (restructure) their business practices. Through
and associated memory. Distinction is made between embed- this reconstruction, enterprises are downsizing from main-
ded systems design and design of self-contained computing frame-based systems to distributed and networked client/
systems. Self-contained computing systems are exemplified server implementation architecture (17). These advances in
by computers. The term embedded means being part of a technology are increasing the number of CIM factories and
larger unit and providing a dedicated service to that unit. The improving the productivity of CIM, by lowering the develop-
computing elements in embedded systems are dedicated to ment cost and reducing the cycle time of the product.
performing signal and data processing or control and commu-
nication functions of larger systems. A personal computer can Distributed Computing Environment
be made the embedded control system for manufacturing in
an assembly line by providing dedicated software programs Distributed computing is fundamental to the implementation
and appropriate interfaces to the assembly line environment. of CIM. In manufacturing, distributed computing is consis-
Similarly, a microprocessor can be dedicated to a control func- tent with the Distributed Computing Environment (DCE), as
tion in a computer (e.g., keyboard/mouse input control) and advocated by the Open System Foundation (OSF). In the tra-
can be considered as an embedded controller. Embedded sys- ditional computing factory environment, there is no mecha-
tems differ from traditional computers in the way in which nism to share information with other applications. In other
they are programmed. Programming embedded systems has words, applications provide functions to the end user only and
a serious impact on hardware organization and software com- not to other applications. Each application maintains its own
piler development. database and users must log into a time-sharing system to
The design of complex embedded systems in a CIM factory obtain simultaneous access to shared databases. In the CIM
can be made more effective if system designers, developers, factory we find a collection of computers connected together to
COMPUTER INTEGRATED MANUFACTURING 749
User
Internet/Intranet users. This automation makes it easier for
Processing
users and applications to share services, and it facilitates the
management of the DCE. The distributed file service through
request
the file transfer protocol (FTP) provides access to any file on
any node in the CIM network. It makes a heterogeneous CIM
Client Server network behave like a single system (18).
The distributed time service provides a mechanism for syn-
Result chronizing the clocks on each computer in a DCE network. It
provides time marking, which is used in many transactional
applications. The security service provides authentication and
Figure 9. Distributed computing model using client/server architec- authorization across a heterogeneous network. It validates
ture. The user interacts with the server/application through the cli- that clients (users or programs) are what they say they are
ent. Clients and servers are distributed throughout the CIM factory. and determines whether they have access to the resources
they request. Security permits applications to communicate
with one another. There must be at least one security server
permit concurrent problem solving and sharing of information for each DCE cell. Threads are program elements that enable
(databases). The architecture of distributed computing is multitasking. Threads allow servers to service RPCs and
based on an environment in which applications are divided FTPs concurrently, which improves performance and memory
into clients, which request services from a system or applica- use. In CIM, threads allow the industrial engineer to concur-
tion, and servers, which provide services across the network rently access process recipes, material inventory, status of
on different processes. The DCE requires an overall enter- products, and generate a cost profile of a product. Similarly,
prise management system that uses the client/server archi- threads allow financial brokers to access account information,
tecture running on a connecting network. A DCE using client/ fund information, a portfolio list, and an equity transaction
server architecture provides for the distribution of tasks that list from different servers concurrently. It follows that
used to be done on a single machine. Complexities are added threads speed up the time it takes to create a customer
because the distribution of tasks can occur in a very large profile.
number of ways, since there can be any number of machines Basically two components need to be distributed in a CIM
doing any number of tasks. Figure 9 shows distributed com- factory—data and processing. Encapsulating these compo-
puting using the client/server model (18,19). nents as objects makes them portable, scalable, and sharable
The DCE consists of six services: (1) remote procedure call between applications. Encapsulation is a technique that
(RPC) for communications, (2) cell directory service, (3) dis- allows the software developer to separate the implementation
tributed file service, (4) distributed time service, (5) security of an application from the abstract interface through software
service, and (6) threads package to create multitasking of called wrappers. Users may develop these objects on proprie-
software applications. Together these services act as mid- tary platforms. However, they can aim at architecture that
dleware, the software that supports interactions between cli- will permit them to port objects across a large number of stan-
ents and servers. dards-based platforms in the future. Keeping the logical de-
RPCs are programming calls (messages) that process tasks scription of the application separate from the physical de-
executed by an application. Originally, RPCs were designed scription is very important, especially, when applications are
to connect two computers using synchronous connections. A moved from proprietary platforms, such as Windows, to stan-
protocol is a set of coded characters at the beginning and end dards-based platforms. Functions that previously ran in a
of a message that enables one machine (as a computer) to proprietary environment are simply recompiled as objects in
communicate with another. It is a formal specification con- the industry-based environment. It follows that users save
trolling the meaning of different messages passing between the time invested in developing the application and preserve
sending and receiving computers. There are two dominant the architecture of their applications.
nonproprietary protocols: (1) Open System Interconnection Client/server computing is a combination of technologies
(OSI), (2) Transmission Control Protocol/Internet Protocol that enables personal computers, Macintosh computers, lap-
(TCP/IP). TCP/IP is the primary protocol used on the In- tops, workstations, mainframe, and other devices to interact
ternet. The DCE RPC allows developers to add or replace with computers or software services to access the information
hardware, database, or the Graphical User Interface (GUI) as and functionality of distributed applications. It involves net-
needed. It enables users and applications to communicate worked remote services and applications that work with
with other users and applications throughout the enterprise. shared information across the CIM factory. In client/server
The RPC mechanism provides protocol and network indepen- architecture, processing is logically partitioned between inde-
dence. The cell directory service is a database that contains pendent, but cooperating, client/servers. Clients provide the
the DCE cell resources, such as files, servers, disks, or print interface into the application, and it is the client’s function to
queues. The directory enables users to find these resources present a coherent, usable, and appropriate interface into the
without knowing their locations. Currently, vendors are de- application. The client also sends requests to the server and
veloping advanced directory and security services designed to displays the results back to the user (application). Servers
make it easier for network managers to track (1) end users provide functionality by receiving and processing requests
and (2) network resources across corporate Intranet and the from clients and providing the results. Servers are distributed
Internet. Network managers are looking for a way to consoli- over the network to take advantage of existing hardware and
date electronic mail (e-mail) authentication and security ser- software platforms across the network. The client/server ar-
vices, network operating system directories, and directories of chitecture is becoming the design approach for future soft-
750 COMPUTER INTEGRATED MANUFACTURING
ware applications. The implementation of the client/server The server-centric computing model places the application
architecture is made practical by (1) the availability of low- logic and data on a server and distributes the presentation
cost powerful personal computers and fast networks to link services to clients as shown in Fig. 10(b). The server-centric
them and (2) the need to access databases containing vast architecture duplicates the familiar centralized processing en-
amounts of stored data. Many aspects of C/S application anal- vironment. This is important for companies running mission-
ysis and designs are similar to those of traditional architec- critical applications that need to ensure data integrity and
ture (20). high availability. Collaborative applications such as docu-
The partitioning of the client/server applications is user ment or image management work best with the server-centric
driven and is subject to tradeoffs between the capabilities of architecture because they have heavy processing and storage
the client and server. For example, in a CIM factory a CAD requirements. The servers tier can be a single mainframe,
application may require an extensive amount of data that re- minicomputer, or multiple LAN servers using operating sys-
side on a separate database or mainframe. Therefore, the tems such as Windows, UNIX, OS2 (by IBM), or Netware (by
CAD application is partitioned accordingly. Embedded com- Novell). In any case, adequate bandwidth must be provided
puters into manufacturing are used to perform many impor- to accomplish the transfer of large documents or image files
tant functions, such as communication, supervisory control from servers to clients. The client in the server-centric model
and data acquisition (SCADA), computation, storing data, de- is referred to as the thin client or network computer. In CIM,
sign, and processing and distribution of signals, data, and in- thin clients are commonly used on the shop floor to perform
formation. The optimum integration, distribution, and parti- process and product schedule, tracking, and monitoring.
tioning of these functions within the CIM factory require a The distributed-function model shown in Fig. 10(c) corrects
system engineering methodology to harmonize the operations many of the problems associated with the first two models.
of the factory. How, how much, and where to put the smarts The distributed-function architecture allows developers to di-
of the factory are serious considerations in the design of a vide the application logic between the clients and servers and
CIM factory. Should the factory smarts be distributed within uses RPC messages to join the parts at run time. Using RPCs,
the process equipment, data basis, application servers, or the the client side calls a procedure on a remote server and waits
mainframes? The answer to these questions is dependent on until it receives a response to continue processing. In a CIM
the manufacturing domain and is product-driven. The CIM factory many CAE, CAD, and virtual prototyping software
factory architecture is very sensitive to the partitioning and tools give developers the ability to split application code be-
distribution of embedded computers. There are several appli- tween clients and servers using this model. This partitioning
approach gives developers more flexibility to customize appli-
cation-partitioning models. Each model has its strengths and
cations to satisfy specific business requirements.
weaknesses based on its suitability for different types of busi-
The distributed-services architecture takes client/server
ness and processing requirements. The models are distin-
computing to another level of complexity. The model breaks
guished by the division of the application components be-
the hard-wired connections between clients and servers and
tween clients and servers. The application components
adds support for more flexible forms of distributed computing.
common to all models are (1) presentation, (2) business logic
The architecture divides the presentation, function, and data
or function, and (3) data. The presentation services manage
into autonomous components that interact with each other
the user interface. The logic or function encodes the business
across a network via standard interfaces, as shown in Fig.
rules that the application executes. The data consist of nu-
10(d). The distributed-services model is also called a three-
merical information, text, images, or other types of informa-
tier client/server architecture because the application compo-
tion manipulated by the application. Typically, presentation nents are frequently placed on separate tiers, as shown in
services run on a client, data are kept on a server, and appli- Fig. 5. The three-tier C/S architecture have benefits for large-
cation logic runs on either the client or server or is divided scale Intranet and Internet applications. The distributed-ser-
between the two. Software developers must carefully allocate vice model represents the CIM factory information system as
application components to minimize the impact of network la- discussed in the following section.
tency and outages and to maximize the use of available com-
puting resources. The basic issue is how well each architec-
ture supports the applications given the constraints and CIM Software
requirements of the computing environments. The most com- The CIM software is a distributed computing software of
mon partitioning models are (1) the client centric, (2) the client/server, it is evolving into two groups—highly functional
server centric, (3) the distributed function, and (4) the distrib- but proprietary environment and standards-based open sys-
uted services. A CIM factory will contain all four types of ap- tem environment. The proprietary environment group has
plication partitioning models. evolved out of the PC desktop world, whereas the standards-
The client-centric model is shown in Fig. 10(a). In this based group has grown from the microcomputer/workstation
model the presentation and logic components are placed on world. DCE is implemented using object-oriented technology.
the client, and the data reside on the server. Client-centric In general, CIM software is developed by using Object Ori-
applications allow users to generate SQL queries for trans- ented Technology (OOT). OOT includes OO tools, program-
mission across a network to a database server that processes ming methods, and languages. Combining OO programming
the request and returns a set of results. Most of the pro- with distributed client/server architecture provides the capa-
cessing is made on the client, and the server just manages bility of sharing and controlling CIM information throughout
the data and the access to it by several clients. In CIM, the the factory. Objects are discrete program elements that con-
client-centric model is commonly used in process modeling tain data and the tasks to be performed on that data. An ob-
and industrial engineering. ject is a concept, abstraction, or thing with distinct bound-
COMPUTER INTEGRATED MANUFACTURING 751
Client
Client
Client Presentation services
Presentation services
Presentation services Functional logic
Functional logic
Call
Request response function Result
Request Request One Set of
record set record records
Functional logic Functional logic
Data Data Data
Client Client
Call Call
Result Result
function function
Figure 10. (a) Client-centric model: Presentation and functional logic run on client machines
and data runs on the server. Client-centric model is used in industrial engineering to run CAD
tools and process simulation. (b) Server-centric architecture: The client only handles presentation
services, and the server mimics centralized computing (contains logic and data). The clients in
this model are referred to thin clients or network computers. They are used on the shop floor to
monitor and track factory processes. (c) Distributed-function model: divides the functional logic
of an application into procedures residing on servers available to all clients. In CIM this model
is used in CAE, Auto CAD applications, and simulation. (d) Distributed services model: Client/
server computing components running on a server are deployed as shared services that can be
used by a number of clients. In CIM this model is used in the implementation of the factory
information system.
aries and meanings for the problem at hand. Objects describe stead of the more expensive and limited resources of the per-
elements that are common to a specific environment. For ex- sonal computer. Applets are executable miniapplications. The
ample, objects common to a manufacturing environment Java environment provides a significant step in the process
would include orders, schedules, and inventories. Objects con- of making the Intranet/Internet a truly open and platform-
taining common attributes are grouped into classes. Software independent network. True object-oriented products provide
developers working at the class level can influence the behav- three basic functionality features that allow developers more
ior of all objects within that class at once. Consequently, this flexibility. The first functionality feature is encapsulation,
approach improves productivity. The gains in productivity which refers to wrapping (bundling) data and processing
can be increased by storing individual objects and classes of tasks in a way that hides the individual procedure calls. The
objects in a central repository so that developers within the second feature is inheritance, which is a way for attributes or
enterprise can share and reuse them. The software architec- any changes made to an attribute within one class of objects
ture of the CIM factory must have the flexibility to be useful to be automatically shared with subclasses created from the
in any manufacturing domain, and must be robust enough to parent class. The third functionality feature is polymorphism,
support a variety of pluggable applications. The distributed which supports the application of particular functions on dif-
OO client/server architecture is a unifying force that perme- ferent objects to achieve different results. Microsoft is advo-
ates every aspect of CIM, from its design and implementation cating its ActiveX/Object-Linking-and-Embedding/Distrib-
to its appearance on the screen to the end user. uted Component Object Model (OLE/DCOM) architecture.
The most common object-oriented programming languages The standard-based industry is backing the Common Object
are C⫹⫹, Next Step, Smalltalk, and Java. The Java program- Request Broker Architecture (CORBA), as specified by the
ming language provides the interface that allows developers Object Management Group (OMG) and X/Open Co. Ltd. OLE
to create small applets that use the resources of the Intranet/ is a Windows compound document protocol that allows one
Internet and particularly the World Wide Web (WWW), in- document to be embedded into or linked with another.
752 COMPUTER INTEGRATED MANUFACTURING
CORBA is built using Object Request Broker (ORB), which ages the flow of information and processes between clients
implements the mechanisms required to find an object speci- and servers. For example, the middleware layer uses its own
fied in a client request and to communicate the data making communication code to set up the session with the target en-
up the request (21,22). vironment, initialize protocols, handle error recovery, and re-
A three-tier C/S architecture provides the foundation to turn a delivery confirmation to the workstation. Middleware
‘‘distributed objects applications’’ where all application re- ties the different components of a distributed system together
sources interact as peers. Manufacturing companies use into a single logical environment by providing consistent APIs
three-tier architecture when they create a data warehouse. that are independent of operating systems, networks, and da-
Production data on a host computer is periodically down- tabases. Using middleware enables application developers to
loaded to a warehouse server optimized for user query and write applications independent of operating system interfaces
analysis. Network computing is a natural extension to C/S or specific network transport. Examples of middleware are
environment. Network computing deals with connecting the the DEC Network Application Support (NAS) and Framework
operators, accountants, and other employees to each other Based Environment, Hewlett Packard’s Open View, and Mi-
and the outside world. It is about connecting the enterprise to crosoft OLE (23–26).
the customers and the business partners. The most common There are four types of middleware: (1) distributed data-
vehicle used to implement ‘‘network computing’’ is the base middleware, (2) communication or message-oriented
Intranet/Internet. middleware, (3) transaction-processing (TP) middleware, and
(4) distributed object services middleware. A CIM factory soft-
Middleware/Software Interfaces ware will have all four types of middlewares.
The distributed database or data access middleware makes
Software interfaces are the critical parts of any system inte- databases within the enterprise available for applications. It
gration effort. Interfaces ensure that proper software hooks, provides services that are consistent across diverse databases
protocols, and wraps are used between the various clusters, (e.g., it facilitates communication between client applications
workstations, and workcells of the CIM factory. There are
and databases). The data access middleware provides a com-
six categories of software interfaces in CIM. (1) Protocols for
mon high-level programming interface, such as SQL, that
the CIM system to gain information and control of operations,
packages information so that it can be used by multiple appli-
(2) protocols for CIM to process material on the machine,
cations. The data access middleware is used in the three-tier
(3) interfaces with robots and protocols for moving material
CIM to (1) bring data from databases for analysis and process
and executing transfers, (4) protocols for processing material
control, (2) perform File Transfer Protocol (FTP) and inte-
and data file transfer (5) protocols for recording and accessing
grate legacy and LAN-based data with web servers.
machine performance data, and (6) protocols for preventive
Communication middleware enables developers to parti-
and corrective maintenance scheduling and operations.
tion applications and distribute them across the network. The
Today’s multivendor realities add a subtle need for pro-
gramming tools that will lead to truly successful distributed two main forms of communication middleware are RPC and
computing. In DCE, between the operating system and the messaging. RPCs let application components on different ma-
application lies middleware. Middlewares are programming chines converse synchronously across network. RPCs rely on
tools that provide interoperability and portability within the a simple synchronous call return API. Messaging middleware
various applications. A simple definition of middleware is any enables application components to communicate asynchro-
software layer that resides between an application and its in- nously, meaning that users do not have to wait for a response.
frastructure environment, which isolates the application from Consequently, an application that sends the message is not
the service component by providing an abstraction of them. It blocked and can do other tasks. RPCs are generally consid-
follows that middleware is a layer of software that sits be- ered simpler than messaging, but messaging can be more ef-
tween an operating system and network at one end and the ficient. The messaging middleware adds a layer of software
applications at the other end. These services let developers that becomes a common denominator for all kinds of different
write applications without having to master the details of the technologies. Message-oriented middleware resides between
underlying services that will perform work on behalf of the applications and operating systems to facilitate interaction
application. The application program interface (API) is the with other applications and operating systems. Message-ori-
software specification, and middleware is the software imple- ented middleware stores the messages in a local or remote
mentation of the API. There are three major reasons for using queue until the target is ready to receive it. The receiving
middleware: (1) providing basic transport from data source to target polls the queue until there is a message for it. The
a destination, (2) dealing with differences among platforms, asynchronous communications ensure that an application is
operating systems, database servers, and protocols, and (3) never forced to stop execution while it is waiting for a mes-
protecting the software development team from the network sage. In CIM, the RPC middleware is used to establish chan-
complexity. Therefore, the middleware software program nels of communication between clusters, workstations, and
allows machines to communicate and interoperate by using a workcells. Message queuing is not appropriate for event-
well-defined API. It provides the basic data transport from driven applications, such as GUIs or closely coupled distrib-
source to destination, giving users a single view into a hetero- uted processing applications (14).
geneous environment. The ability of the enterprise to inte- The TP middleware handles the detailed work of managing
grate different multiple hardware platforms, networking pro- traffic among multiple application components and coordi-
tocols, legacy applications, LAN environments, and desktop nates the execution of such tasks as job scheduling, resource
operating systems is realized by the middleware software. location, data updating, and recovery from systems failure. It
Middleware is an information broker that controls and man- is designed to develop and manage online transaction pro-
COMPUTER INTEGRATED MANUFACTURING 753
cessing (OLTP) applications (27). The TP middleware is the 11 shows a seven-layer OSI reference model for computer net-
heart of the virtual prototype of CIM. working and control (10).
The networking breakthroughs such as mobile and wire- The top layer is the application layer, where the user inter-
less networks, LAN switches, Integrated Services Digital Net- acts with the CIM network application. Data are received as
work (ISDN), and Asynchronous Transfer Mode (ATM) are commands from the user. TCP/IP applications communicate
making LAN and WAN indistinguishable. Global business- in client/server pairs at this layer. Application protocols in-
men ventures created virtual workgroups—people working to- clude FTP for file transfers, Telnet P (Telecommunications
gether on common projects that are distributed across the Network Protocol) for remote terminal sessions, SMTP (Sim-
country and around the world. Switchable communication ple Mail Transfer Protocol) for electronic mail and SNMP
protocols such as TCP/IP, shuttled from LAN through Wide- (Simple Network Management Protocol) for network manage-
Area-Network (WAN)-to-LAN by multiprotocol/router prod- ment. At the presentation layer, middleware is used to deter-
ucts, make remote network access transparent to the virtual mine how shared data are formatted for transmission across
workgroup. Businesses are investing many resources into the network. At this layer, data format such as ASCII text
electronic networks that link computers, databases, and other and binary documents are broken into messages. At the ses-
information technologies. The rationale for doing this is that sion layer the DCE interfaces with the operating system of
they must have the capability to (1) track all the components the machines, and the distributed computing facilities act as
and products, (2) synchronize deliveries, (3) keep engineers extensions to the operating system. Each message is serial-
and marketers informed of each other’s plans, (4) alert the ized by adding session layer (operating system) header infor-
research and development organization to the needs of the mation. The serialized session layer elements are tagged with
manufacturing side, and (5) give management a coherent pic- transport layer recovery information so that the communicat-
ture of what is going on. Accomplishing these objectives keep ing parties can request retransmissions of lost or garbled
the businesses competitive (28–31). messages. The transport layer manages the data flow be-
The development of the software for using DCE requires tween two internetwork hosts. The transport layer messages
examining the layers of the software building blocks as they are broken down into routable packets and datagrams whose
relate to the proprietary and standards-based client/server network layer headers identify the sender and receiver by
software. The software layers arise from the presence of dis- their corresponding node numbers. At the network layer, data
tributed control networks within the DCE. The Open Systems are transferred around the internetwork. The Internet Proto-
Interconnection (OSI) of the International Standards Organi- col (IP) is the dominant protocol used for routing packets be-
zation (ISO) provides the most robust standards and models tween hosts and across network links. ICMP (Internet Con-
for distributed data networking. The OSI model is a seven- trol Message Protocol) works with IP to carry control
layer reference model for computer networking, which is ap- messages and status information between networked systems
plicable to a CIM network in which the processing elements with IGMP (Internet Group Management Protocol), which
communicate over a common path. The seven layers are (1) routes data to multiple hosts for multicasting. Both ICMP
application, (2) presentation, (3) session (operating system), and IGMP are carried with IP traffic. A set of methods called
(4) transport, (5) network, (6) data link, and (7) physical. Fig. Internet routing is used to make sure that network traffic is
OSI layers Data format Typical task assigned Control networking requirements Network products
Figure 11. The International Standard Organization (ISO) seven-layer reference model for com-
puter networking structured on the Open System Interconnection (OSI) standard. Software inter-
faces represent the transition from layer to layer.
754 COMPUTER INTEGRATED MANUFACTURING
passed between hosts and networks efficiently. Internet rout- Workstream provides the following functions: (1) Work-In-
ing has its own set of protocols to ensure that the special sys- Process (WIP) tracking, (2) data collection from workcells,
tems that connect networks, called routers, are able to detect labor times, losses, rework, (3) directing data from design
changes in internetwork routes and maintain data flowing cells-to-processes-to-databases, (4) transferring data from
across the Internet. The network layer is very well standard- databases for analysis, and (5) interacting with the process
ized and presents a consistent, transparent interface to the control software that controls equipment. Workstream runs
session layer. The technologies that constitute the network on VAX using UNIX. It forms the backbone of the FIMS and
are understood and well established. The issues network de- serves as the source for the basic manufacturing data.
signers must consider are (1) system protocols, (2) interopera- Cellworks is a COTS product from Fastech, Inc. It runs
bility, (3) command status, (4) security, and (5) topology. The on DEC Alpha server using UNIX. Cellworks automatically
network protocol must be open, to ensure the availability of generates data fields from Workstream. Workstream and
the network to everyone connected to it. The data link layer Cellworks interact to document process recipes, product rout-
is also known as the network interface layer. Data are trans- ing, and other information describing how the product moves
mitted across a network at this layer. Because this layer deals through the factory floor. Cellworks maintains a consistent
with local network transmission exclusively, there is no uni- screen layout throughout the factory. Users connect to Cell-
versal TCP/IP protocol available. Two commonly used proto- works through PCs running Microsoft Windows. When appro-
cols called ARP (Address Resolution Protocol) and RARP (Re- priate, Cellworks launches Windows-based views such as Mi-
verse Address Resolution Protocol) are used on many crosoft Word or Excel or Borland Paradox to assist with data
networks such as Ethernet and Token Ring. Note that these review and entry (1,18).
two protocols are sometimes viewed as being part of the net- The rapid generation and retrieval of information by CAD/
work layer. The physical layer transmits the signal along the CAM/CAE and CIM strain the conventional configuration
various networks in the form of electrical pulses. Finally, management system. The issue is further complicated by hav-
these messages are transmitted through the data link and ing the information stored in different format media and resi-
physical layer protocols of LAN and WAN cables. Presently, dent on different and dispersed computers. The CIM factory
the Internet Protocol is going through several updates in or- requires an engineering data management (EDM) system to
der to meet the security and interoperability requirements. automate the management of product design data. A COTS
IP version 4 (IPv4) currently is widely deployed in routers software called Product Data Management (PDM) from Com-
and desktop software. The next generation of IP, IP version 6 puter Vision is used to automate the factory design data.
(IPv6), includes an improved network addressing method plus PDM is a tool that helps engineers and others manage both
support for moving PCs across subnets without changing data and the product development process. PDM keeps track
their IP addresses. Also it includes optional security features of the masses of data and information required to design,
such as authentication and encryption. manufacture and support a product. The PDM runs on a Sun
Managing the enterprise information is a mission-critical workstation using UNIX.
and challenging task. System engineering principles provide The ability to model processes and factory capabilities
the basic tools for managing the information of the enterprise. leads to improved yields and better matching of product speci-
The following section takes as an example a CIM factory and fications to the factory. Models support the development of
applies the preceding system engineering principle to the de- reliable design rules that enable factory flexibility. The build-
sign of the factory information management system. ing of models consists of identifying the variables in a factory
database as inputs or outputs and programming the computer
to handle the calculation automatically. Process Insights is a
INTEGRATION OF FACTORY INFORMATION commercial software program from Pavilion Technologies
that uses neural network and fuzzy logic technology to model
Computer integrated manufacturing represents our best ap- complex nonlinear processes.
proach to staying competitive in a global economy. System Sound and rapid management decisions require informa-
engineering principles are used to determine the range of tion from the entire factory. The manager, supervisor, and
products to be made, the factory capacity, and the factory engineer need data on cost, schedule, technical specification,
building blocks. The preceding guidelines are used to estab- and resource data. The information exists as data and tools
lish the architecture of the IRFPA CIM factory using commer- distributed throughout the factory. FIMS provides access and
cial-off-the-shelf software applications. The factory is struc- organization to draw resources from across the factory and to
tured into interconnected clusters of workcells. An effective present the results in formats that help managers. Figure 12
CIM factory provides on-demand, correct, and timely informa- illustrates the connectivity of the various software packages
tion to every participant in the integrated team including the in the FIMS architecture.
manager, design engineer, process engineer, floor supervisor, Defects are a very serious manufacturing problem, and re-
operator, and accountant. A factory information management ducing defects will improve the yield and the competitiveness
system (FIMS) is needed to tie the clusters together. Using a of the factory. An important payoff from CIM is the ability
LAN and the FBE middleware, the FIMS connects the partici- to perform on-line statistical quality control (SQC) to reduce
pants to (1) shop floor control, (2) process equipment control, defects in a timely manner. Figure 13 illustrates how CIM
(3) product data management, (4) process modeling, and (5) performs on-line SQC. For a given product, FIMS performs a
support tools. series of tasks: retrieves control limits, locates files, and loads
Workstream from Consilium is a complete factory floor historical data. Cellworks reads current data from the process
control system. It generates and maintains process instruc- equipment, checks it against historical data, and sounds the
tions and handles recipes for automated machine control. alarm when the data exceeds control limits. A similar se-
COMPUTER INTEGRATED MANUFACTURING 755
Workstream
operator Test
data
Finance Design Process & Paradox Word
Industrial
mainframe engineer engineer analysis
engineer
Workstream
execution Document viewing tools
Initiate
work • Specification Recipe creation
requests • Schematics To cellworks
Charge # • Drawings Process specification
Master To Process step Manufacturing
schedule PDM documentation
software
Figure 12. Factory Information Management System (FIMS) gives the architecture of the build-
ing blocks of CIM: clusters, workstations, workcells, and applications are distributed in the fac-
tory. Interfaces/middlewares are represented by arrows and junctions.
On-line SQC
Cellworks
interface
routine Alarm
1 Request 7 Answer
Formatting
FIMS information broker 6
3 2 4 5
Database …
Workstream Cellworks PDM
files
Equipment
Figure 13. On-line statistical quality
Data for Real time File Load control: cellworks reads and checks data
recent lots process data names files against process history.
756 COMPUTER INTEGRATED MANUFACTURING
quence of calls to applications and data files occurs for each minifactories, modules, workcells, operations, and processes.
element of the engineering analysis. Equation (1) gives the relationship between the number of
disjoint regions Rd, the number of terminal nodes Nt, and the
number of factory outcomes.
METRICS: MEASURES OF EFFECTIVENESS
n
Systems are evaluated in terms of their outputs or the out- Rd = Nt = 2n = n
Cm (1)
comes they can provide. The building block model of systems m=0
C
ABC
C
B
B ABC
B C
C ABC
A A
A B
C ABC
C ABC
A B
A ABC A
C
C ABC
B B
B
Figure 15. Network-diagrams: (a) n ⫽ 2,
C Nt ⫽ 4; (b) n ⫽ 3, Nt ⫽ 8; (c) n ⫽ 4, Nt ⫽
ABC C 16. Network diagrams of CIM show the
independent variables explicitly as the
D network is constructed.
ware. Software reliability is the probability that the factory uct is out of the factory; and (3) yield, percent of
software will not cause the failure of the CIM factory, for an product meeting specification (32,33).
interval of time t, under specified conditions. Software relia-
bility is a measure of the frequency of degradation in factory In manufacturing, the aggregation of these factors are
performance that is due to software failure. Software reliabil- grouped into two sets of measures: (1) measures at the factory
ity is different from hardware reliability because software level and (2) measures at the process (operation) or compo-
does not wear out like mechanical or electrical devices. In nent level (1).
physical systems like aircraft, radar, motor, or furnace, relia-
MOEs at the Factory Level
bility is related to wear and tear on the parts of the system.
In software there is no wear or tear. A certain application The objective of CIM is to provide quality products, improve
may be used for years without encountering an error simply yield of factory, reduce cycle time, increase the range of prod-
because an error exists in a rarely traveled path of the soft- ucts, and reduce the unit-product-cost (UPC) by reducing the
ware. Fundamentally, software reliability is a measure of the recurring and nonrecurring cost of producing products. An-
quality of design and implementation of an application. If other attribute of CIM is enhancing the ability of the factory
properly designed, tested, and correctly implemented an ap- to produce at low cost at low volume. The CIM factory flexi-
plication should run without failure for all inputs and uses. bility is based on the ability of the factory to produce low-cost
Consequently, software reliability should increase with use products at low volume. The decoupling of UPC from volume
because the probability of failure decreases with time. How- is illustrated in Fig. 16. The cost-versus-volume curve con-
ever, it is possible for software reliability to initially increase sists of two regions: (1) low rate region characterized by the
as the bugs are removed from an application, then decrease slope m and the knee of the curve and (2) high-production-
as a result of changes in the inputs to the application and/or rate region characterized by the ordinate of the asymptote.
changes in the functionality of the software. This is because Figure 16 shows two measures of low cost at low volume: (1)
these changes lead to new travel paths of the software where the ratio of the knee of the curves (e.g., case 3 is more low
errors reside. In manufacturing, the MOEs of the CIM factory cost at low volume than case 1 by a factor V1 /V3 ⫽ 5) and (2)
are grouped into two classes: the ratio of the slopes m3 /m1. These relationships can be de-
termined by factory simulation of many runs. It is recognized
that productivity of a new factory tends to increase in time.
1. Attributes measures (discrete), such as number of de- The concept of the learning curve has been around for some
fects or nonconformities in parts per million (ppm), cost/ time and has been used as a measure of productivity. Equa-
unit, rolled throughput yield (percent of defect-free tion (2) illustrates the concept:
products), and producibility measures (number rejects/
number produced). Cn = kn−x (2)
2. Variable measures (continuous) can be separated into
where Cn is the unit cost of unit n; k is the cost of first unit
two categories:
produced that met the specification; x is the learning curve
(a) process variation parameters measured by using coefficient, characterized by the cost reduction obtained by
Shewhart charts R, X, . doubling n. That is,
(b) process control capability measures: (1) Cp, Cpk; (2)
cycle time, time from receiving order to time prod- C2n /Cn = (2n)−x /n−x = 2−x = a (3)
758 COMPUTER INTEGRATED MANUFACTURING
steps are not perfect, or do not meet very high standards, the
products will have defects. The concept of parts per million is
C1
used to characterize defects of products and is based on the
C2 area under the curve of the standard normal probability den-
sity function. The standard deviation is used as the describ-
C3
ing parameter.
Consider a product produced under the tolerance specifi-
V3 = V2 = V1 = cation of ⫾3. Where the area under the curve is A ⫽ 0.9973
150 units 500 units 750 units and (1 ⫺ A) ⫽ 0.0027; therefore, out of 1,000,000 units, 2700
Figure 16. Cost versus volume curves illustrate low cost/low volume units will be out of specification. Compare this to the case of
measure of effectiveness. Factory flexibility is demonstrated as the ⫾6 specification, where A ⫽ 0.999999998, which gives 0.002
factory operates at a higher volume level by aggregating several low- of one unit out of 1,000,000 will be out of specification. In the
volume products, as shown in curve #3. real world, there are normal shifts and drifts in the ‘‘mean,’’
which take a toll on product quality. The shift in the mean by
⫾1.5 is the basis for the 6 manufacturing method, which
gives 3.4 ppm defects (32–35).
The factor a is the fractional reduction in cost obtained every Throughput yield is a viable measure of quality and pro-
time the volume is doubled (3). For example, if a ⫽ 0.85, the ductivity. At the factory level, the question arises for a given
learning curve is said to be 85%. Taking the natural loga- number of parts per million, how many of a product could be
rithm of Eq. (3) gives the relationship of x to a as expected to have zero, one, two, three, or four defects? The
answer to this question defines the throughput yield of the
x = − ln(a)/ ln(2) (4) factory. The Poisson distribution is used to calculate the yield.
Assume that the defects are randomly distributed through-
The manufacture of the Infrared Focal Plane Array (IR- out the units of a product. Let p be the occurrence probability
FPA) is a complex, highly technical, and labor-intensive task. of a defect, and let n be the sample size or the number of
A basic benefit of CIM is the ability of the factory to decouple trials. The probability of getting r defects in the sample of
the UPC from volume (i.e., produce products at low cost at size n is Yr:
low volume). Figure 16 shows that the state of decoupling oc-
curs when the factory is operated beyond a threshold volume (np)r e−n p
(i.e., beyond the knee of the curve). This principle is used as Yr = (5)
r!
a guide in setting the capability of the CIM factory. The inter-
est in applying the learning curve equation as a factory met- Expressing Yr in terms of number of defects per unit (d/u)
ric is to get a handle on the cost of producing IRFPAs. gives
Through market research on a potential need for IRFPAs and
the desired capability of the CIM factory, it was determined
(d/u) e−(d/u)
r
that a production rate of n ⫽ 250 IRFPA per month is feasible Yr = (6)
and will allow the factory to operate beyond the knee of the r!
UPC-versus-volume curve. Another benefit of CIM is reducing
the nonrecurring engineering (NRE) cost. Based on current The special case of getting zero defects r ⫽ 0, is called rolled
experience, on the average the cost of the first IRFPA unit yield,
without CIM is k ⫽ $700,000 and with a CIM is k ⫽ $500,000.
Through simulation of a CIM factory operating just below the Y0 = e−(d/u) (7)
knee of the curve, it was determined that the learning curve
coefficient x ⫽ 0.71. Using Eq. (2) and substituting for k ⫽
Rolled throughput yield is given in percent as
$500,000, n ⫽ 250, and x ⫽ 0.71 gives on the average the
projected UPC ⫽ $10,000. The fractional reduction or the
learning curve rate a ⫽ 0.61. This example provides a good Y0 in percent (%) = 100e−(d/u) (8)
rationale and justification for computer integrated manufac-
turing. Equation (8) shows that when d/u ⫽ 1, one defect per unit
Cycle time is a factory MOE. It is estimated by breaking produced Y0 ⫽ 37%. This means that, on the average, 63% of
down the work flow into steps and determining the time it the product will be rejected. The productivity in this case is
takes to perform each step in the manufacturing process. The very poor and is not affordable. Consider a product requiring
proper application of CIM and automation play a key role in 1200 parts and steps. Assume that the product is designed to
reducing cycle time. Through simulation it was determined accept a tolerance ‘‘twice’’ the normal variation of the pro-
COMPUTER INTEGRATED MANUFACTURING 759
2(1.5)σ 3 1
cesses, or ⫾6 with shift of ⫾1.5. This result in 3.4 ppm of k= = = (12)
12σ 12 4
parts or steps variation, and
Cp =
jUSL ; LSLj = 12 = 2 8. G. DeMicheli and R. Gupta, Hardware/software co-design, Proc.
6 6 IEEE, 85: 349–365, 1997.
9. S. Edwards et al., Design of embedded systems: Formal models,
where USL ⫽ Upper Specification Limit, LSL ⫽ Lower Speci- validation, and synthesis, Proc. IEEE, 85 (3): 366–390, 1997.
fication Limit, and u ⫽ Mean. 10. P. E. Green, Computer Network Architectures and Protocols, New
When the process mean is shifted with respect to the de- York and London: Plenum Press, 1983.
sign mean, the capability index is adjusted by a factor k, and 11. W. Eckerson, Client/server architecture, Network World, 1995.
760 COMPUTER INTEGRATED MANUFACTURING
12. N. Engler, Riding the bleeding edge of distributed computing, 26. S. Weissman, Your development tools, Network World, April
Open Computing, 1995. 1995.
13. J. Mullich, The riddle in the middle, Open Computing, August 27. J. M. Willis, TP Software Development for Open VMS, Horsham,
1995. PA: Cardinal Business Media, Inc., 1994.
14. P. Bernstein, Middleware: A model for distributed system ser- 28. R. Whiting, Bridging the CORBA OLE Gap, Danvers, MA: Client/
vices, Commun. ACM, 39 (2): 86–98, 1996. Server Computing Sentry Publishing Co., 1996.
15. T. Williams, Graphical simulation tool for complex systems, Elec- 29. R. S. Raji, Smart networks for control, IEEE Spectrum, 31 (6):
tron. Design, 44 (26): 54–55, December 1996. 49–55, June 1994.
16. T. G. Lewis, CASE: Computer-Aided Software Engineering, New 30. D. Settle, Scalable parallel processing, Comput. Technol. Rev.,
Spring 1995.
York: VanNostrand Reinhold, 1991.
31. D. J. Kaplan, Processing Graph Method Specification Version 1.0,
17. B. T. Harrison, Client/server development, Which way will we
Washington, DC: The Naval Research Laboratory, 1987.
go?, DEC Professional, April 1994.
32. A. A. Afifi and S. P. Azen, Statistical Analysis: A Computer Ori-
18. Open System Foundation DCE, Administration Guide, Core Com- ented Approach, 2nd ed., Reading, MA: Academic Press, 1979.
ponents, Englewood Cliffs, NJ: Prentice-Hall, 1993.
33. E. Grant and R. S. Leavenworth, Statistical Quality Control, 5th
19. Application Portability Profile (APP), Open System Environment, ed., New York: McGraw-Hill, 1980.
Version 1.0, NIST Publication 500-127, April 1991/update June
34. D. H. Evans, Statistical tolerancing: The state of the art, Part I:
1994.
Background and Part II: Shifts and drifts, J. Quality Technol.,
20. A. Cini, Networking for Everyone, Horsham, PA: DEC Profes- 1972.
sional Cardinal Business Media, 1994. 35. B. Smith, Making war on defects, IEEE Spectrum, 30 (9): 43–43,
21. I. Jacobson, Object Oriented Software Engineering, Reading, MA: September 1993.
Addison-Wesley, 1992. 36. V. E. Kane, Process capability indices, J. Quality Technol., 1986.
22. P. VandenHamer and K. LePoeter, Managing design data: The 37. A. Bowker and G. Liberman, Engineering Statistics, Englewood
five dimensions of CAD frameworks, Proc. IEEE, 84: 42–56, 1996. Cliffs, NJ: Prentice-Hall, 1972.
23. D. T. Dewire, Clearing up the Middleware Muddle, Client Server
Computing, Danvers, MA: Sentry Publishing Co., 1995. A. R. HABAYEB
24. A. Radding, Middleware works magic with apps, Information Naval Air Systems Command
Week, June 17, 1996.
25. J. Rumbaugh, Object Oriented Modeling and Design, Englewood
Cliffs, NJ: Prentice Hall, 1991. COMPUTER INTERFACE. See CAMAC.
COMPUTERIZED MONITORING identifying operational problems in rotating machinery. The
main function of soft computing technologies is to extract this
Computerized monitoring is the art of extracting information information and to identify its correlation with the condition
from a system through computer processes for various pur- of the system. Computerized monitoring uses these signals
poses. Computerized monitoring and inspection encompasses and with the aid of soft computing techniques and algorithms
a wide range of applications in various industries. One com- monitors the status of the system or its components.
mon application is to evaluate the condition of a system or to With the improvements in computer technology, modern
inspect the integrity of its components for diagnostic pur- instrumentation systems have the capacity to acquire a prodi-
poses. The concept behind a computerized monitoring system gious amount of data from a wide variety of sensor types,
for diagnosis is to provide information on component faults leaving the software component as the main standard by
by comparison of actual observations with models of normal which to evaluate a computerized monitoring system. In addi-
behavior. The approach is to seek mechanisms and proce- tion, each application would require a software component
dures that can detect deviations from normal operation at an with appropriate customized soft computing techniques and
early stage. Other applications include, but are not limited to, algorithms that are unique to the specific application.
control, industry automation, manufacturing processes, aero- The first basic steps in designing a computerized monitor-
space engineering, laboratory automation, quality control, ing system are to
and robotics.
A computerized monitoring system has two major compo- 1. identify target parameters to be monitored,
nents, hardware and software. The hardware components of
2. design the structure of data acquisition hardware and
a typical computerized monitoring system consist of transduc-
necessary signal preprocessing,
ers (or sensors), signal conditioning, data acquisition hard-
ware interface, and the computer itself (Fig. 1). The basic re- 3. design and develop the algorithm of the soft computing
quirements for the hardware components are reliability, component of the software and design the user interface
accuracy, cost-effectiveness, and speed. and display format, and
The software component of the computerized monitoring 4. reexamine the system for reliability, cost, maintenance,
system should provide effective information display, deter- and updating capability.
mine the status of the system (i.e., pattern recognition and
diagnosis), and, if appropriate, provide decision-making capa- We will demonstrate basic concepts in computerized monitor-
bility. These components are generally designed with the fol- ing with two examples of monitoring and inspection applied
lowing items in mind: reliability, speed, visual effect (i.e., ef- in the nuclear industry. One uses signals from a data acquisi-
fective information display), user friendliness, ease of tion system and the purpose of monitoring is diagnosis and
maintenance, and provisions for upgrading. fault identification using artificial neural networks (3). The
The software component frequently incorporates soft com- other example demonstrates an inspection system for the pur-
puting technologies. Soft computing consists of methodologies pose of quality control using image analysis methodologies
that resemble the real-world model pertaining to imprecision and fuzzy logic algorithms. In addition, a brief description of
and uncertainty. The best example of a real-world model for the concept of an artificial neural network and fuzzy logic,
soft computing is the human mind. Soft computing encom- which are currently the two most popular soft computing
passes technologies including, but not limited to, expert sys- technologies, will be provided.
tems, artificial neural networks, fuzzy logic, genetic algo-
rithms, computer vision and image processing techniques,
data mining techniques, and hypermedia databases. ARTIFICIAL NEURAL NETWORKS
The input variables to a monitoring system could be any-
thing that can bear the information about the monitored ob- Artificial neural networks have become popular tools for pat-
ject. Examples are the current waveform of a circuit, output tern recognition and signal classification (4). They offer great
torque of a motor, image of an object, speed of a car, power of potential for successful application in computerized monitor-
a nuclear reactor, or position of a spacecraft. Transducers ing systems.
sense the monitored variables of the system and produce the Artificial neural networks (ANNs) are information pro-
electrical signal. Typically, these are signals that must be pre- cessing systems motivated by the goal of reproducing the cog-
processed before they are introduced into the monitoring sys- nitive processes and organizational models of neural biologi-
tem software. The preprocessing (i.e., digital signal processing cal systems. The individual computational processor that
[DSP]) may involve filtering, digitizing, sampling, or nor- makes up most artificial neural systems is referred to as a
malizing to a certain maximum value (1,2). Some technologies processing element (PE). Each PE (also called neuron) has
that are applied in DSP include, but are not limited to, filter many inputs, but has only a single output, which can fan out
design, wavelets techniques, fast Fourier transform (FFT), to many other PEs in the network. Each connection to the
time–frequency analysis, and time–scale analysis. Signals ith PE has associated with it a quantity called a weight or
from components of a system carry valuable information re- connection strength. The weight on the connection from the
garding the condition of the components or the system as a jth node to the ith node is denoted Wij. Figure 2 shows a pro-
whole. For example, analysis of vibration data is helpful in cessing element. The specific characteristics of an ANN is a
1
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
2 COMPUTERIZED MONITORING
result of the network paradigm used. The network paradigm connected to the unit whose output is the biggest [Fig. 3(c)].
is specified by the network architecture and neurodynamics. Weights are normalized to avoid increase without upper
bound. Because only one unit becomes active as the winner of
Neurodynamics and Learning Mechanisms the competition, the network is called a winner-take-all
network.
Neurodynamics specifies how the inputs to the PE will be
An adaptive resonance theory (ART) network has the abil-
combined, what type of function or relationship will be used
ity to learn many new things without necessarily forgetting
to develop the output, and how the weights will be modified.
things learned in the past. Patterns of activities that develop
The inputs to the PEs are weighted and are often combined
over the nodes in the two layers of the attentional subsystem
using the summation function. This is called the ‘‘interval ac-
[Fig. 3(d)] are called short-term memory (STM) traces because
tivation.’’ This interval activation is used to generate the out-
they exist only in association with a single application of an
put of the neuron using a continuous or noncontinuous trans-
input vector. The weights associated with the bottom-up and
fer function.
top-down connections between F1 and F2 layers are called
The learning mechanism that handles modifications to the
long-term memory (LTM) trace because they encode informa-
weights and any other organization of the network can be
tion that remains a part of the network for an extended
classified under supervised learning, unsupervised learning,
period.
or self-supervised (reinforcement) learning. Supervised learn-
Among the different rules and procedures developed, the
ing takes place when the network is trained using pairs of
handful mentioned here are accepted by the community:
inputs and desired outputs. In unsupervised learning inputs
backpropagation, counterpropagation, Kohonen feature maps,
are entered and the network is able to organize its own cate-
bidirectional associative memory, neocognitron, Hopfield, and
gories. Self-supervised learning adds the feedback to unsuper-
adaptive resonance theory, including ART2, ART2-A,
vised learning to correct errors in the pattern recognition
FuzzyART, ARTMAP, and FuzzyARTMAP (6,7).
process.
Supervised neural networks do not require a prior fault-
related parameter to be identified and generate their own
Network Architecture
rules by learning from being shown original examples. This
The network architecture defines the arrangement of pro- characteristic of the artificial neural network makes it attrac-
cessing elements and their interconnections. This establishes tive for monitoring purposes and diagnostic applications.
which PEs are interconnected, the inputs to and outputs from
PEs, the group or layers of PEs, and how the information
AN EXAMPLE OF COMPUTERIZED MONITORING
flows in the network. Figure 3 shows several examples of neu-
SYSTEM FOR DIAGNOSIS
ral network architectures.
One of the most popular feedforward neural networks that
Keyvan et al. (3) have developed a prototype of a simple diag-
iteratively determines the weight is the backpropagation Net-
nostic monitoring system using several artificial neural net-
work (BPN) [Fig. 3(a)]. A simple learning algorithm that mod-
works. The system integrates the result of neural network
ifies the weights between output and hidden layers is called
pattern recognition with a preexisting database to classify
a delta rule (5). The backpropagation algorithm is an exten-
faulty signal and through an expert system to identify the
sion of the delta rule that can train the weights, not only be-
fault. The system is developed in an X-windows environment
tween output and hidden layers but also in hidden and input
and uses Motif in a UNIX environment to build the graphical
layers. A sequential network feeds its output back to the in-
user interface (GUI). It is user-friendly and menu-driven,
put units of the network [Fig. 3(b)]. A competitive neural net-
allowing the user to select signals and choose several neural
work is a kind of unsupervised network. It employs a com-
network paradigms including ART2 and ART2-A. The system
petitive learning algorithm that strengthens the weights
provides the status or condition of the signals tested as either
normal or faulty. In the case of faulty status, the system iden-
tifies the fault and indicates the progress of the fault relative
Wij
Ij to normal as well as relative to the previous tests.
W: Weight
Wik I: Input The signals used here are divided into two groups, the ac-
Oi
Ik Σ F O: Output tual collected signal and the simulated signals. The collected
Wil F: Transfer function signal is the pump power signal of the Experimental Breeder
Σ: Summation Reactor-II (EBR-II) nuclear plant; it was collected from the
Il
sensors by the plant data acquisition system on 1/29/91. This
Figure 2. A processing element and its components. Two mathemati- signal is used to simulate and generate faulty signals repre-
cal functions (兺 and F) are applied to the input in order to create the senting several levels of reactor pump shaft degradation. Fig-
output. For a Sigmoid transfer function: F ⫽ (1 ⫹ e⫺兺)⫺1. ure 4 shows the plot of the collected signal data and a faulty
COMPUTERIZED MONITORING 3
Input Hidden
layer layer Output
layer
(a) (b)
Reset
+
+ signal
–
+
+ – +
Gain +
control F1 layer
(c) Figure 3. Examples of neural network architec-
Input vector tures: (a) feedforward neural network, (b) sequen-
tial network, (c) competitive network, (d) ART neu-
(d) ral network.
signal data for a 50 s time period. A comparison of these two monitoring system output showing a typical information
plots reveals the sensitivity that is required of a soft comput- display.
ing algorithm to distinguish these signal patterns.
The neural network runs in the background and classifies
the given input signal into one of three categories : normal, FUZZY LOGIC
faulty, or unknown signal. When the signal is identified as
normal or faulty, a status report is displayed as shown in Fig. Fuzzy logic is often incorporated in a computerized monitor-
5. Note that, in the case of the faulty status report, a sample ing system to better model the causal effect between a system
plot of the normal signal is also shown for comparison. A de- condition and its measurable signal variables.
scription of the fault can be obtained by selecting the Describe In the real world, we often must deal with fuzzy concepts
Fault button shown in Fig. 5. The fault description corre- or variables such as high speed, low temperature, and strong
sponding to the current faulty signal is identified from the signal. Fuzzy logic provides a means to specify fuzzy concepts.
existing fault data base and is displayed at this point. The Fuzzy theory provides a means for representing these uncer-
faults are described as ‘‘Degradation Level 1,’’ ‘‘Degradation tainties and this vagueness. In fuzzy logic, the domain of each
Level 2,’’ etc. When a new fault different from the ones cur- variable is quantified into a finite number of fuzzy concepts.
rently registered in the database is encountered, the fault de- For example, the variable temperature may be fuzzily quanti-
scription will be ‘‘Unidentified fault,’’ as shown in Fig. 6. Fig- fied into low, medium, and high. Application of fuzzy logic is
ures 5 and 6 are the actual computer screen of the diagnostic most suited in (1) very complex models where understanding
Pump power signal
0 50 0 50
Time (s) Time (s)
(a) (b)
Figure 4. (a) Plot of pump #1 collected power signal for a 50 s time period; (b) plot of pump #1
simulated faulty signal for a 50 s time period.
4 COMPUTERIZED MONITORING
Figure 5. User interface showing the status report of a faulty signal. where A1k and A2k are the fuzzy sets representing the kth ante-
By choosing the ‘‘Describe Fault’’ option button, the user is provided cedent pairs, and Bk are the fuzzy sets representing the kth
with another window describing the nature of the fault, i.e., ‘‘Degra- consequence. zk is the fuzzy variable, and Bk is a fuzzy quanti-
dation Level 3’’ in this case. fier for the domain of zk.
The fuzzy system maps an input to an output in three
steps. The first step matches the input to all the IF-part fuzzy
sets in parallel. This step ‘‘fires’’ or ‘‘activates’’ the rules ac-
cording to the degree to which the input belongs to each IF-
Figure 6. User interface describing a faulty signal outside of the database. By selecting the
‘‘yes’’ option, the user is provided with an entry box to describe the new fault.
COMPUTERIZED MONITORING 5
Rule 1
µ µ µ
Low Low Low
1.0 1.0 1.0
Minimum
Rule 2
µ µ µ
High High High
1.0 1.0 1.0
z* Z
Minimum
Figure 7. Graphical (max–min) inference method with crisp input (X, Y). The inference in this
example is done for each rule with a minimum membership function, resulting in the shaded
area of the triangles. The final output is the result of aggregation of the two shaded areas based
on the disjunctive relations between these two rules.
part set. Each input fires a corresponding rule or rules. Then The pictorial representation of the fuzzy mapping process
each fired IF-part set scales its THEN-part set. The second with the input of x and y is shown in Fig. 7. Because the
step adds all scaled THEN-part sets into a final output set. antecedent pairs given in the general rule structure for this
The third step is defuzzification. The system computes the system is connected by a logical and connective, each rule ap-
output as the centroid or center of gravity of this final out- plies the THEN part with a minimum membership grade. The
put set. minimum membership value for the antecedents propagates
Most fuzzy systems involve more than one rule. The pro- through to the consequence and truncates the membership for
cess of obtaining the overall consequence (conclusion) from the consequence of each rule. The inference (process of
the individual consequence contributed by each rule in the applying fuzzy system) is done for each rule. Then the trun-
rule base is known as aggregation of rules. Fuzzy systems cated membership functions for each rule are aggregated. For
differ in how they fire rules and how they combine the fired a set of disjunctive rules, the aggregated output for a max–
rules. Aggregation strategy is based on the two extreme ex- min inference or composition is given by
isting cases—conjunctive system and disconjunctive system.
In the case of a system of rules that must be jointly satis- µ(Z) = max[min[µz1 (input(x)), µz1 (input( y))],
fied (conjunctive), the rules are connected by and connectives. min[µz2 (input(x)), µz2 (input( y))] · · ·
In this case, the aggregated output (consequence) y is found
min[µzr (input(x)), µzr (input( y))]]
by the fuzzy intersection of all individual rule consequent.
The overall output membership function is
where r is the number of the rules that have been activited.
µz ( z) = min[µz1 ( z), µz2 ( z), . . ., µzr ( z)]
An Example of Software Component of Computerized
For the disjunctive system of rules where the satisfaction of Monitoring System for Inspection
at least one rule is required, the rules are connected by the Our second example demonstrates a computerized inspection
or connectives. In this case, the aggregated output is found of quality of a nuclear fuel pellet. Fabricated pellets must be
by the fuzzy union of all rule contributions, as of high quality before being placed into the fuel assemblies
and into service in the core of a nuclear reactor. Computerized
µz ( z) = max[µz1 ( z), µz2 ( z), . . ., µzr ( z)] inspection in this application is expected to increase accuracy
and speed of inspection and will reduce the radiation expo-
Suppose we have the two following fuzzy rules that are acti- sure of the workers. The structure of this computerized in-
vated for input (x, y) in the fuzzy system. spection system is shown in Fig. 8.
First, the input signal (the image of the fuel pellet) is ac-
Rule 1. IF X is low and Y is low, THEN Z is low. quired (using a camera) and converted to the digital signal;
Rule 2. IF X is high and Y is high, THEN Z is high. then, it is converted to an 8-bit gray-scale mode. Next a refer-
6 COMPUTERIZED MONITORING
Machine
vision
ence model is generated to check the presence of a defect on By using this dynamic model search approach, a defect is
the pellet image. For the nuclear fuel pellet, it is challenging enhanced for the next step (i.e., pattern recognition). Next,
to generate a universal model representing a good pellet to be six features are extracted from this enhanced defect informa-
checked against a defective pellet image. This is because the tion pool. Table 1 lists these features and their relations with
gray-scale value of the pixels on the same area for any two the status of the input image, where each possible status of a
good pellets may vary greatly because of the high noise, dif- pellet can be uniquely identified from the corresponding fea-
ferent manufacturing process, and small variations in pellet ture values. Each value of the feature, which is a fuzzy mem-
size. For this reason, a dynamic reference model is generated bership value, encodes the quality criteria of a fuel pellet.
on-line for each pellet individually. The relation between pellet status and feature value,
The shape of gray-scale intensity distribution of the pellet which are fuzzy rules, map the human inspection knowledge.
image reflects the presence of a defect very well. A dynamic The IF-THEN fuzzy rules are:
reference is generated by finding those rows of pixels with a
distribution very close to that of a good fuel pellet. A set of IF Abnormal dark area size is big;
good fuel pellets was selected, and their distribution surface ∧ Abnormal light area size is zero;
was processed to create the reference surface model. Each row ∧ No related closing abnormal dark area and light
of the target image is used to match the distribution of this area;
reference model by using ∧ Shape factor is small;
r (I − I1 )2 + (Ir2 − I2 )2 + · · · + (Im − Im )2
THEN
Banded defect.
S = 1.0 − r1
m
FUTURE OF COMPUTERIZED MONITORING
where AND INSPECTION
S ⫽ the degree of match; Each soft computing technique has unique properties and ad-
m ⫽ the number of total pixels in each row of fuel pellet vantages. Hence, increasing integration of a number of such
image; techniques into a computerized monitoring system is antici-
Ii ⫽ the intensity of ith pixel in the current row; and pated in the near future (9,10). For example, as already men-
Iri ⫽ the intensity of ith pixel in the rth reference model. tioned, neural networks consist of highly interconnected
processing units that can learn and globally estimate input–
The best matching row is selected as the dynamic standard output functions in a parallel-distribution framework. Fuzzy
reference to reexamine the entire pellet image. Those pixels logic systems store and process rules, with output fuzzy sets
whose intensities are below or above the reference value (be- associated with input fuzzy sets in parallel. The similar paral-
yond a preselected tolerance value) are regarded as abnormal lelism properties of neural nets and fuzzy logic systems make
pixels and are classified into two categories—abnormal dark their integration more suitable to the study of the behavior of
and abnormal light. We call those pixels above the reference systems that are imprecisely defined by virtue of their high
distribution abnormal light and those below the reference dis- degree of complexity. Because of their great learning capabil-
tribution abnormal dark. ity, neural networks have been combined with fuzzy logic sys-
tems to form the initial rules of fuzzy systems and tune the
rules and membership functions to manage the fuzzy system
Table 3. A List of Related Resources on Computerized efficiently and accurately. In addition, fuzzy microprocessors,
Monitoring Technology called fuzzy chips, have been successfully applied in control
Web address http://www-dsp.rice.edu and robotics. Hence, it is natural to predict a more intense
Journal Expert Systems with Applications future application of integrated neural networks and fuzzy
Journal of Acoustic Emission logic in computerized monitoring and inspection. In addition,
Artificial Intelligence in Engineering soft computing technologies are the core of computerized mon-
Computers & Industrial Engineering itoring and inspection. Therefore, it is expected that new ad-
Control Engineering Practice vancements in these technologies, such as computer vision
Conference International Conference on Monitoring, Acoustics and data mining techniques, would greatly affect the future
Speech, and Signal Processing
of computerized monitoring and inspection.
International Conference on Robotics and Auto-
mation
International Conference on Intelligent System Ap- RESOURCES AND VENDOR INFORMATION
plication to Power System
Transactions IEEE Transactions on Power Systems Information on several vendors of sensors, transducers, and
IEEE Transactions on Professional Communication data acquisition boards useful in computerized monitoring ap-
IEEE Transactions on Control Systems Technology plications are provided in Table 2. This is not an exhaustive
IEEE Transactions on Instrumentation and Mea-
list of all vendors. Table 3 provides a sample of resources on
surement
IEEE Transactions on Fuzzy Systems
the subject of computerized monitoring and inspection for in-
terested readers.
8 COMPUTERIZED TOMOGRAPHY
BIBLIOGRAPHY
SHAHLA KEYVAN
XIAOLONG SONG
MARK KELLY
University of Missouri-Rolla
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
FAULT LOCATION 277
• Time-varying fault resistance (due to the breakdown of stations to record voltages and currents on the transmission
the insulation, the fault resistance changes during the lines. Again, most of the DFR vendors have implemented a
fault disturbance) fault-location algorithm and provide it as a standard feature
• Changing prefault load conditions (the line may have dis- of their product.
tinctively different load current at a different moment of Even though fault-location implementation can be diverse,
a fault) it should be noted that the accuracy and cost requirements
are always the key consideration. Therefore, it is essential to
Protective Relaying System understand the possible benefits and shortcomings of using
different types of data and system-implementation ap-
The fault-location application requires that full consideration proaches when designing or selecting a fault locator.
is also given to the elements that constitute the relaying sys-
tems: protective relays, instrument transformers, and circuit Cost/Performance Considerations
breakers. Protective relays are supposed to detect the fault
and isolate the line before the system is endangered and fur- The following design considerations directly affect the cost/
ther damage is incurred. The fault clearing time of a typical performance rating of a given fault-location implementation:
transmission-line relay is around four cycles, which should
• One- or two-ended application
provide sufficient measurement time to obtain the waveform
data for the fault-location application. Since the relays give a • Synchronized or unsynchronized data acquisition
determination based on the waveform measurements ob- • Data samples from the adjacent lines
tained by the current transformer and capacitor coupling volt-
age transformers (CT and CCTV, respectively), it is important The least expensive fault-location application is to use a sin-
to understand the errors introduced by the transformers. Typ- gle-terminal measurement of voltages and currents. In this
ical distortion that may affect the current waveform is the case an existing transmission line relay or a DFR can be used.
saturation of the iron core. The CCTV are associated with The main difference between these application approaches is
low-pass filtering characteristics as well as signal oscillations the input data waveform processing requirement. Most of the
in the case of voltage collapse. The instrument-transformer protective relays use a low sampling rate to reconstruct pha-
inaccuracies are very important in determining the overall sors. The DFR sampling is up to 5 kHz and higher and en-
error in the fault-location algorithm. The instrument-trans- ables recovery of other waveform components. The accuracy
former error may significantly affect the fault location error and complexity of the input channels have a bearing on both
causing it to deteriorate for an order of magnitude. Finally, the cost and performance of the fault-location implemen-
the circuit breakers are initiated by the relays to clear the tation.
fault. The phenomena of breaker restrikes and ferroresonance A more expensive but also more accurate solution is a two-
distortion are important when using the waveform data cap- terminal implementation with which the data from the trans-
tured before the breaker opens in calculating the fault lo- mission line ends are collected and brought to a centralized
cation. place where the fault location is calculated. In this case a
communication channel is needed to transfer the required
Implementation Requirements data which increases the cost of the overall solution. A varia-
tion between these solutions is in the way the data sampling
The algorithms for fault location may be implemented using is performed. Most of the implementations do not require that
the data sampling at two ends of the line is synchronized to
• Fault-location devices a common time source, while the most accurate solutions re-
• Protective relays quire the synchronization (2).
• Digital fault recorders Finally, in order to achieve even greater fault-location ac-
curacy, data samples from the lines parallel to the faulted
Stand-alone fault locators are the most flexible option since line, and from all ends of a multiterminal line involved in a
the entire design can be optimized for fault-location applica- fault, can be used. Obviously, more input channels and com-
tion. At the same time, this is the most expensive solution munication facilities are needed in this case, but accuracy can
since the entire device accommodates only one function, be improved significantly (3).
namely, the fault location. Some vendors have opted for such
a solution, justifying an increased cost with a claim that their FAULT-LOCATION ALGORITHM FUNDAMENTALS
fault-location implementation guarantees unsurpassed accu-
racy performance (1). A fault-location algorithm defines the steps needed to obtain
The most common implementation approach is to use the the fault location by using the measurements of voltages and
transmission-line protection relays as the platform for the currents from one or more ends of the line. A set of equations
fault-location implementation. This approach is cost effective representing the mathematical model of the faulted transmis-
since the increment required to accommodate the fault-loca- sion line is needed to define the algorithm. The quantities
tion algorithm is minimal. Almost all of the protective-relay that appear in the equations are (1) voltages and currents, (2)
vendors offer some form of a fault location algorithm as a transmission-line parameters, and (3) fault parameters.
standard feature of their relay designs. The voltage and current in power systems are a combina-
Yet another option is to use a digital fault recorder (DFR) tion of four kinds of signal components: fundamental, higher
design as the platform for fault-location implementation. or lower frequency, transients, and noise. The fundamental
DFRs are commonly used in high-voltage transmission sub- component is a sinusoid having system frequency f 0 that is
FAULT LOCATION 279
In these equations, line parameters l, r, c, and g are induc- • Phasor-based algorithms use the fundamental compo-
tance, resistance, capacitance, and conductance per unit nent of the signals only. The fundamental components
length, respectively; v(x, t) is the voltage and i(x, t) is the cur- then appear as phasors. The line model is usually the
rent. The subscripts x and t denote partial derivatives with lumped-parameter model.
respect to position and time.
• Partial differential equation–based algorithms use tran-
A three-phase transmission line has as a model two matrix
sient components of signals and the distributed-parame-
equation similar to Eqs. (1) and (2). The elements of the volt-
ter model of the line.
age vector are three-phase voltages, and elements of the cur-
rent vector are three line currents. Transmission-line param-
eters are represented by matrices R, L, C, and G and are We will explain the underlying principles of the two groups
composed of self-resistance, mutual resistance, inductance, using their exemplary algorithms.
capacitance, and conductance. The details of this model will
be presented later.
The lumped-parameter model neglects the line conduc- STANDARD APPROACHES: PHASOR-BASED ALGORITHMS
tance g and capacitance c. The partial derivative of the cur-
rent relative to position, in Eq. (2), is equal to zero in this The phasor-based algorithms use a Fourier transform of Eq.
case. Therefore, the current does not change along the line. (3) to model the line. The line is represented by its impedance
The integration along the transmission line from one end (the per unit length Z ⫽ r ⫹ j2앟f 0 l and its length d. Figure 2
sending end) to a point at a distance x from the sending end depicts the circuit model of the faulted line. There are three
produces the following differential equation: groups of quantities in Fig. 2. The phasors of voltages and
currents are known since they may be calculated from the
vx (t) − vs (t) = xri(t) + lx[di(t)/dt] (3) signal samples. The transmission-line impedance Z and its
length are also known from the line construction data. The
In Eq. (3), vs(t) is the voltage at the sending end, vx(t) is the fault position x, the fault impedance ZF, and the fault voltage
voltage at a distance x from the sending end, and i(t) is the VF are not known.
current on the line. In the case of a multiconductor line, the The aim of the algorithm is to find the unknown distance
model is a matrix equation similar in form to Eq. (3). The x to the fault. Two main steps in a phasor-based algorithm
line has a matrix model containing as its elements the self- are (1) calculation of phasors from the signal samples and (2)
resistance, mutual resistance, and inductances. solution of the set of equations for the unknown fault dis-
The Fourier transformation of Eq. (3) can be made if all tance.
the line parameters are constant. Furthermore, if the cur- The phasors are calculated from the corresponding voltage
rents and voltages are the fundamental components, they will and current samples. An arbitrary sinusoid, say voltage v(t),
appear in the equation as phasors. is represented by a phasor V. A phasor is a complex number
Note that due to the linearity of the equations, voltages defined by its real value Re兵V其, its imaginary value Im兵V其, or
and currents in both models may be replaced by their compo- alternatively by its phase and amplitude 兩V兩. The calculation
280 FAULT LOCATION
of the phasor parameters is accomplished using Fourier anal- ing-end currents is obviously equal to zero. Since IF is the sum
ysis. The formulas for real and imaginary part of a phasor are of the sending- end and receiving-end currents, we have
Here VP is the vector having as elements the phasors perti- The one-end algorithms require relatively simple calcula-
nent to phase a, phase b, and phase c. tions, and their implementation is opportune, since the wave-
The matrix A is given by form data are necessary from one side of the line only. They
assume that the fault impedance ZF is a constant during the
1 1 1 fault. Their accuracy depends on the simplifying assumptions.
A= 1 exp( j4π/3) exp( j2π/3) (12) In the case of a high fault impedance the fault current is
1 exp( j2π/3) exp( j4π/3) small; hence the fault components of the sending-end current
are very small. Since the fault current for the sending end is
in the denominator of Eq. (9), the system is ill-defined in this
The equation defining the relation of the phase vector at the
case and errors may be large.
sending end VPS, the phase vector at the fault VPF, the phase
Two-end algorithms require fewer simplifying assumptions
current vector IPS, and the impedance matrix ZP is similar in
and offer potentially more accurate calculations.
form to Eq. (6):
The Two-End Algorithms
V PS = xZ
Z PI PS + V PF (13)
Two-end algorithms fall into two subgroups: algorithms devel-
The impedance matrix ZP has mutual impedances and resis- oped using synchronized samples and those developed using
tances at its off-diagonal terms. When the phasor vectors are nonsynchronized samples. The samples are synchronous if
replaced by the symmetrical component vectors, one gets the two data sampling clocks at the sending and receiving
end ensure that the samples are taken at exactly the same
moments. This may be achieved by global positioning system
V SS = xZ
Z SI SS + V SF (14)
(GPS) of satellites using pulses emitted from a satellite to at-
tune the two GPS receivers that synchronize the sampling
The matrix ZS here is equal to clocks (2). This approach introduces additional cost to provide
GPS receivers and appropriate waveform sampling interfaces.
Z S = A −1Z PA (15) The impact of synchronization will be explained in the follow-
ing paragraphs.
While the matrix ZP has both the diagonal and off-diagonal One must note that phasors are calculated locally. If there
elements, the off-diagonal elements of the matrix ZS are all is a time shift between data acquisition clock pulses at the
equal to zero. Hence the matrix in Eq. (14) may be broken receiving and sending ends, the relative phases of the receiv-
into three independent scalar complex equations: ing and sending end phasors are not the same. The phase
difference between two phasors cannot be calculated by sub-
VSk = xZkk I k + VFk , k = 0, 1, 2 (16) tracting one phase from another. Suppose that the phasor at
the receiving-end voltage is calculated from two sets of sam-
Here Zkk is the corresponding diagonal element of the matrix ples. The first set is taken using the sample clocked by the
ZS. The main advantage of the symmetrical component appli- sending end. The resulting phasor is denoted as VF. The sec-
cation is this decoupling. Each of the decoupled equations de- ond voltage phasor denoted V⬘F is calculated using the receiv-
fines a sequence circuit. They are called the positive-, nega- ing-end clocked samples. If there is a time shift ⌬t between
tive-, and zero-sequence circuits. Since three decoupled the two sets, the phases of two phasors will differ for 웃 ⫽
equations have the same form as Eq. (6), the circuit in Fig. 2 2앟f⌬t. This may be mathematically expressed in the following
may again represent any of the sequence circuits with a suit- way:
able change in notation.
The previously mentioned obstacles of the Takagi et al. VF = VF e jδ (17)
method are eliminated by using the symmetrical components
in the line model (5). In this approach the negative-sequence The phase shift 웃 restates the nonsynchronized phasor (ob-
circuit of the line is used. The decoupling feature of the sym- tained using data from the receiving end) to the frame refer-
metrical components eliminates the mutual inductance in- ence of the sending end. This phase shift is the same for all
fluence. Since the negative-sequence vector is equal to zero in voltages and currents, but it is not known in advance. Note
the prefault condition, the recordings of the prefault current that the phasor in both time references has the same ampli-
tude. The two-end methods consider the phase shift 웃 as an
are not necessary as in the algorithm of Takagi et al. More-
additional unknown and try to solve for the fault distance x
over, according to these authors, the equivalent impedances
by eliminating 웃. Note that the sending-end voltage phasor
of the negative-sequence circuit and the line impedance of the
calculated locally is VS, and the receiving-end voltage and cur-
negative-sequence circuit that make up the current divider
rent phasor calculated locally are V⬘R and I⬘R, respectively.
are more likely to have the same phases than in the case of
An example of an algorithm using non-synchronized sam-
the phase impedances of the line (6). This implies that the
ples is presented in Ref. 5. The line model is constructed us-
assumption that the current distribution factor is a real num-
ing a negative-sequence diagram. By inspecting Fig. 2 and
ber is close to reality. Also, the classification of the fault type
interpreting all phasors as negative-sequence phasors and all
before the calculation is not necessary. However, the exclu-
the impedances as negative-sequence impedances, the appli-
sive use of the negative-sequence representation has a draw-
cation of the Kirchhoff ’s voltage law renders
back. In the (very rare) case of a symmetric fault, the nega-
tive-sequence phasors after the fault remain equal to zero, VF = VS − x ZIS (18)
and the negative-sequence circuit is not suitable for fault lo-
cation. VF = VR − (d − x)ZR IR (19)
282 FAULT LOCATION
Since the absolute value of the fault voltage in both equations Also, the line may be compensated by inserting a series capac-
is the same, one gets the following scalar equation by elimi- itor into the line, or there may be load taps between two line
nating the absolute value of the fault voltage 兩VF兩 from Eqs. ends. In addition, neglecting the line capacitance may intro-
(18) and (19): duce significant errors for a longer transmission line.
However, the most important issue in the phasor-based al-
|VS − x ZIS | = |VR − (d − x) ZIR | (20) gorithms is the need for phasor estimation. Since in reality
there is usually a decaying dc component and noise in the
This is a quadratic equation with respect to x and it may be signal, phasors calculated using the Fourier analysis–based
easily solved. formulas given by Eqs. (4) and (5) will differ from their true
Two-end algorithms using synchronized samples start values.
from the matrix equivalents of Eqs. (18) and (19). Since all The methods based on the distributed line parameters
the phasors are calculated using the samples clocked at the solve some of these problems. Calculation of phasors is not
same time, derived from the same clock, the two equations needed. The line capacitance is included in the model. The
may be combined together. When the fault voltage is elimi- change of the fault impedance is not a problem, and these
nated from these two equations, the following matrix equa- methods work if a series capacitor is inserted into the line.
tion follows:
V S − V R − x ZI S + (d − x) ZI R = 0 (21)
ADVANCED APPROACHES: PARTIAL DIFFERENTIAL
This equation is equivalent to six real scalar equations. Since EQUATION–BASED METHODS
there is only one unknown x, the system is overdetermined.
One alternative in such a situation is to use only a sufficient A solution of a linear partial differential equation may be
number of equations as in Ref. 7. Another option is to use the found using the method of characteristics. The justification
minimum least squares (MLS) technique. The MLS technique for this method may be found, for example, in Ref. 9. The
is often used to identify parameters of a linear system using partial differential equations [Eqs. (1) and (2)] of the trans-
measurements corrupted with Gaussian noise (8). mission-line model have two characteristics: functions of posi-
The basic idea of the MLS method is to compensate for tion and time. The general solution for the voltage and cur-
measurement errors by using more equations than necessary rent along the line is a linear combination of two arbitrary
and thus decreasing the measurement-error effects by averag- functions. Each function has one of the characteristics as its
ing. The solution attained by the MLS method should not ex- argument. The particular value of these functions is set by
actly satisfy any of the equations. When the MLS solution is the boundary conditions. The boundary conditions may be the
put into the equations, the right-hand side of each scalar measured voltage and current signal at the same point of the
equation will not be zero but rather will be equal to a quantity line. Two arbitrary functions are selected so that the general
of the error. The solution offered by the MLS method guaran- solution at this point is equal to the measured values.
tees that the sum of all the squared errors will be the smallest Two approaches based on the partial differential equa-
possible. The matrix Eq. (19) in the MLS technique is repre- tion model have been proposed for the fault location. The
sented as: first method solves partial differential equations using nu-
merical methods with sending-end voltage and current as
Ax + B = E (22) boundary conditions. An inspection of the voltage solution
along the line reveals the fault location. The second method
where vectors A and B are defined as: does not require the solution of partial differential equa-
tions, but instead it exploits a special property of the send-
A = −Z
Z (II S + I R )
(23) ing-end voltage and current and finds the distance by perti-
B = V S − V R + ZI R nent signal processing.
V x = LI t + RI (29)
P One of the
s
characteristics I x = CV x (30)
t +γ x
where the subscripts x and t denote partial derivatives.
The matrices L, C, and R have both diagonal and off-diago-
nal terms. Therefore, the preceding matrix equations cannot
(x0, t0)
be solved using methods described by Kohlas. In addition, the
ρ elements of these matrices depend on the transmission-line
One of the
characteristics geometry and copper resistance only if the ground is not used
Q
t −γ x as a return. However, if the line is grounded, the matrices
depend on the soil conductivity also. This parameter may de-
S x
pend on the weather and type of soil and cannot be easily
Figure 3. Characteristics in the dimensionless position–time plane. determined. To complicate the matter further, as a repercus-
sion, the line parameters then become frequency dependent.
Fortuitously, the two matrix partial differential equations re-
spectively. Along a characteristic, functions u and i are re- duce to three pairs of decoupled partial differential equations
lated by the following two differential equations: similar in form to Eq. (15) by applying modal transformation
as reported in Ref. 11.
du di Modal transformation starts with finding three eigenvec-
−χ = (1 + χ 2 )−0.5 ηi tors of the matrix product LC. These vectors are columns of
ds ds
(27) the transformation matrix M1. The transpose of the matrix
du di
+χ = −(1 + χ 2 )−0.5 ηi M1 is M2. The phasor voltages and currents V and I are trans-
dρ dρ
formed into modal voltages and currents V(m) and I(m) using
the following equations:
These two equations may be solved numerically using the
method of meshes described in Ref. 9. The solution is obtained
V (m) = M −1
1 V (31)
using the sending-end voltage and current as the boundary
(m)
conditions. It is important to note that the value of the volt- I = M −1
2 I (32)
age v(x0, t0) does not depend on all the values of the sending-
end voltages and currents. The voltage depends only on the The matrices R, L, and C are also transformed to modal ma-
boundary conditions in just one segment of time. To find this trices R(m), L(m), and C(m):
segment, it is necessary to identify two characteristics pass-
ing through the point (x0, t0) (see Fig. 3). These two character- R (m) = M −1
1 R M2
istics intersect the t axis at the two points P and Q. Only the
L (m) = M −1
1 L M2 (33)
values of t between these two points affect the value of v(x0,
(m)
t0). This time interval is called the zone of influence. C = M −1
1 C M2
The fault location is found by an inspection of the voltage
along the line by using a property of the voltage. If the fault The particular feature of modal matrices is that their off-diag-
resistance is zero, as in Kohlas’s paper, then the value of the onal terms are equal to zero. Indeed, the modal transforma-
voltage at the fault must be equal to zero. Accordingly, the tion has the same advantage as the symmetrical component
location of the fault is equal to that value of x that annihilates transformation. Actually, if a line is fully transposed, the
the voltage at any time t. When the measurements contain symmetrical component transformation or the Clarke trans-
noise, or when the fault impedance has a low but still nonzero formation will have the same decoupling outcome as the
value, one cannot expect the exact cancellation of the voltage modal transformation. After the application of modal trans-
v(x, t) but rather a minimal value in some sense. Thus, when formation, the transmission-line model consists of three de-
the solution for v(x, t) is found, the next task is to look for the coupled pairs of linear partial differential equations:
value of x at which the voltage is minimal. The problem here
(m) (m)
is that voltage depends both on the distance x and time t. ∂vkk (m)
∂ikk (m) (m)
Instead of inspecting the voltage as a function of time and + lkk = rkk ikk
∂x ∂x (34)
distance, Kohlas proposed to inspect the function of distance (m) (m)
(m)
∂vkk ∂ikk
F(x) that is defined as the square of the voltage averaged in a ckk + =0
∂t ∂x
specific time interval determined by the zone of influence:
T −γ x
Here the subscript k ⫽ 1,2,3 denotes three modes, and super-
F (x) = v (xt) dt
2
(28) scripts x and t denote partial derivatives. One of the modes,
γx known as the aerial mode, has parameters that are least de-
pendent on frequency. Usually, only the aerial mode is consid-
The value of x that minimizes the function F(x) is the esti- ered for the fault location. Once a mode is selected, the proce-
mate of the distance to the fault. The Kohlas idea was subse- dure for the transmission-line model solution is the same as
quently extended and elaborated in detail for the three-phase that for the one-phase transmission line.
284 FAULT LOCATION
where Z0 ⫽ 兹l/c is the surge impedance of the line and 2 ⫽ Figure 4. Lattice diagram.
lc.
The forward and backward traveling waves may be calcu-
flection and the second reflection ⌬t ⫽ t2 ⫺ t1 depends on the
lated from the sending-end voltage v(0, t) ⫽ vS(t) and the
distance to the fault x and the speed of travel:
sending-end current i(0, t) ⫽ iS(t) as follows:
t = 2z × χ (39)
SF (t) = vS (t) + Z0 iS (t) (37)
The idea to use reflections to estimate the fault location ap-
SB (t) = vS (t) − Z0 iS (t) (38) peared in 1930 for the fault location of underground cables. A
cable is energized with a short voltage impulse. The impulse
Fault location uses the transient component of the traveling and its reflection are recorded, and the travel time is found.
waves only. The transient traveling waves appear in the Later, similar devices were used to measure the fault location
transmission line after any abrupt change of its voltages and for transmission lines. These methods are called active
currents. When a fault occurs, the voltage at the fault point methods.
drops. This generates a backward and a forward traveling The calculation of the elapsed time is easy if the inserted
wave at the place of the fault. The backward wave travels to pulse and its reflection have sufficient power. However, trav-
the sending end with a speed ⫺1, and the forward wave eling waves caused by a fault may have a low power, espe-
moves to the receiving end with the same speed. cially if the fault occurs when the instantaneous voltage at
the point of the fault is close to zero. In that case the calcula-
These traveling waves do not change their shape until they
tion of this time requires special signal processing. One of the
reach some discontinuity in the transmission line. The discon-
signal-processing methods most commonly used is the correla-
tinuities are the sending end, the receiving end, and the fault
tion technique (12).
itself. When a traveling wave arrives at a discontinuity, it
The time autocorrelation of the signals x(t) is defined as
ceases to exist in its original form, and two new waves emerge
at the discontinuity. The first is a reflection of the original T
1
wave; it has the shape of the original wave attenuated by a R(τ ) = lim x(t)x(t + τ ) dt (40)
T →∞ 2T −T
reflection coefficient, and it has a reverse direction. That is, a
reflection of the forward wave will be a backward wave. The In real situations, the integration has to start and end with
second wave discussed here, through wave, also has the shape some finite time.
of the original wave attenuated by another coefficient and
continues motion in the same direction as the original wave. T
1
The coefficients affecting magnitudes of both new waves de- R(τ ) = x(t)x(t + τ ) dt (41)
T 0
pend on the type of fault. Low impedance faults have high
coefficients of reflection, and high impedance faults have low For a given signal, autocorrelation is a function of the time
coefficients of reflection. shift . Consider a typical shape of a traveling wave at the
The motion of traveling waves along the transmission line sending end, as shown in the Fig. 5(a) and its time-shifted
and generation of new waves at points of discontinuity are value shown in Fig. 5(b). The autocorrelation is proportional
represented by the lattice diagram in the Fig. 4. The initial to the area of the product of two signals. This area will be
wave arises at the fault point F. The backward wave reaches largest when the first reflection is aligned with the second
the sending end at a time t1. Its reflection moves as a forward reflection as in Fig. 5(c). Then, the time shift is equal to the
wave toward the fault. At the fault, it is reflected again and elapsed time t2 ⫺ t1). Therefore, the elapsed time may be as-
converted to a backward wave. It will arrive at the sending sessed by investigating the maxima of the autocorrelation
end at a time t2. The time that elapses between the first re- function.
FAULT TOLERANT COMPUTING 285
SB(t) the fault distance, (2) the possibility of obtaining a false result
due to the presence of multiple reflections, and (3) a high sam-
pling frequency, increasing the computational burden.
BIBLIOGRAPHY
The accuracy of the fault location is very sensitive to the FAULT TOLERANCE. See GROUP COMMUNICATION;
choice of T and N. If T is too small, the approximation is not
ROUNDOFF ERRORS.
good since an important part of the signal may be missing.
On the other hand, if T is too large, the shape of the forward
wave will contain multiple reflections of both the original
backward and the original forward wave. For example, such
a reflection will appear at time t3 in the lattice diagram. Also,
in nonsymmetrical faults, a fraction of a traveling wave in
one mode may appear in another mode. Therefore, the auto-
correlation will have more maxima, and the identification of
the maxima corresponding to the first reflection and second
reflection will be difficult. In general, the closer the fault to
the sending end, the shorter the window is needed. The other
important factor is the sampling frequency. In general, a very
high sampling frequency (on the order of tens of kilohertz) is
needed to ensure a good approximation of the autocorrela-
tion function.
The limitations of this approach are (1) a lack of firm rules
in the selection of the sample window due to its sensitivity to
FILAMENT LAMPS 435
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
436 FILAMENT LAMPS
40
Based on 60 W, 120 V, and 1000 h life
35
30
Efficacy (lm/W)
25 IR filter
20
Halogen
15
Coiled coils
10 Gas filled
Drawn wire
5 Tungsten
GEM carbon
0
Figure 1. Historical efficacy of incandes- 1880 1900 1920 1940 1960 1980 2000
cent lamps. Year
that perfectly absorbs all radiation incident upon it. By Kirch- visible range. As will shortly be shown mathematically, the
hoff ’s Law, the temperature radiation of a body is at any tem- fraction of radiated power in the middle of the visible region
perature and any frequency the same percentage of blackbody reaches a maximum at around 6500 K. This, of course, is
radiation as the absorbed radiation of the body is of the total about the temperature of the sun, which is also nearly a
impinging radiation. Thus, since a blackbody absorbs per- blackbody radiator.
fectly it also, for equal areas, radiates more total power and Integrating Eq. (1) over all wavelengths, we find that the
more power at any given wavelength than any other source total emittance of a blackbody is proportional to the fourth
operating at the same temperature. The unique feature of the power of the temperature (Stefan–Boltzmann Law):
blackbody is that its radiation characteristics may be speci-
fied solely by the absolute temperature. M eb = σ T 4 W · m−2 (3)
3000 K
1,000 2500 K
where the latest values for the constants are
2000 K
c1 = 2πhc2 = 3.7415 × 10−16 W · m2 (2a)
100
−2
c2 = hc/k = 1.4388 × 10 m-K = 14, 388 µm · K (2b)
In Fig. 2, this implies that the line, not shown, joining the where, as earlier, the b superscript denotes blackbody.
peaks of the emittance for each temperature in the log-log The emissivity of a given surface can vary with the wave-
scale is a straight line, and furthermore that the peak values length, the angle of observation, and the temperature. Thus
increase proportionally to the temperature to the fifth power. the spectral emissivity ⑀ is defined by
It is also clear from Fig. 2 that the emitted power at any
given wavelength below the maximum emitted power in- λ (T ) = Mλ (T )/M λb (T ) (9)
creases more and more rapidly with temperature as the wave-
length decreases. Thus for temperatures between 2500 K and
3000 K, typical of ordinary general lighting lamps, the emit-
100
ted power in the middle of the visible increases at about the
ninth power of the temperature.
By integrating Eq. (1) from 400 nm to 750 nm, the approxi-
mate radiant emittance of a blackbody in the visible region of
Efficacy (lm/W)
The luminous efficacy of a blackbody is equal to the ratio of Figure 4. Lumen efficacy of a blackbody as a function of temper-
its total luminous emittance (flux) to its total radiant emit- ature.
438 FILAMENT LAMPS
Total
where L(, ) is the spectral radiance of the surface element
Emissivity
in the direction (, ), and Lb(, ) is the spectral radiance of
a blackbody at the same temperature. In this context, ⑀ may 0.10
be referred to as hemispherical total emissivity, ⑀ as hemi-
spherical spectral emissivity, and ⑀(, ) as directional total
emissivity. They are all related to the directional spectral
emissivity above by appropriate integrations over wavelength
and/or angles.
If the material is uniformly diffuse, the directional emis- 0.01
100 1,000 10,000
sivity follows that of a blackbody (i.e., ⑀ 앜 cos at all angles Temperature (K)
and wavelengths). If the material surface is gray, the spectral
emissivity is independent of wavelength. Figure 6. Total and average visible emissivity of tungsten versus
The normal spectral emissivity of tungsten, based on data temperature.
from Ref. 5, is shown in Fig. 5. Note that the emissivity
changes quite dramatically with wavelength and also with
temperature. Note, interestingly, that the emissivity is inde- the average visible emissivity obtained similarly to the total
pendent of temperature at around 1300 nm. The directional emissivity, except that the integration is only over the wave-
emissivity of tungsten also deviates from a blackbody such lengths in the visible region.
that the total emissivity is a few percent higher than is calcu- A measure of the tungsten selectivity is the ratio of the
lated based on a diffuse or Lambertian radiator. Specifically, average spectral emissivity in the visible region to the total
note from Fig. 5 that the emissivity of tungsten is higher in hemispherical emissivity. This ratio at any given temperature
the visible portion of the spectrum than in the infrared at all can be deduced from the plots of the average visible and total
temperatures (i.e., tungsten is a selective radiator enhanced hemispherical emissivities plotted in Fig. 6. The spectral se-
for visible radiation). This implies that the efficacy of radia- lectivity is very high at low temperatures; unfortunately this
tion from tungsten is higher than that of a blackbody. is not of much significance because the total lumen output is
The hemispherical or total emissivity of tungsten at a also very low.
given temperature is theoretically obtained by the ratio of the One more comment needs to be made on the shape of the
integral of the product of the spectral emissivity and the spectral emissivity curve in Fig. 5. Note that for every tem-
Planck function over all wavelengths at that temperature di- perature shown, the emissivity is higher in the blue region of
vided by the total radiant emittance, T4. Because spectral the spectrum (around 450 nm) than in the red (around 650
emissivity values are not well known in the infrared nor are nm). This implies that the color or distribution temperature
directional emissivities known in general, the total emissivity of tungsten will be higher than the actual surface tempera-
as a function of temperature is empirically determined. One ture. One way to think of color temperature for an emissive
way to do this is to measure the total power used to re- body is as follows. It is the virtual temperature of a blackbody
sistively heat a long tungsten wire of known surface area up that matches the shape of the emissive body’s emittance
to a known (measured) temperature. Such a measurement curve in the visible region. Thus at a material or true temper-
would be done in vacuum to ensure no cooling and power loss ature of 2800 K, the color temperature of tungsten wire will
to the gas, and with long enough wire so that the end losses be about 60 K higher, even though the absolute emission is
are minimal. The result is shown in Fig. 6. Also shown is higher overall for the blackbody. Coiling the wire tends to
make the radiation emitted more like a blackbody as a result
of internal reflections on the coil, so the difference between
the color and material temperature decreases. For double-
0.5 coiled filaments, which are typical in common household (A-
0.45 2000 K line) lamps, the average color temperature exceeds the true
0.4 2500 K temperature by about 40 K at 2800 K.
2700 K
0.35 Another related concept is that of brightness temperature
3000 K
0.3 at a particular wavelength (generally 665 nm, the value used
Emissivity
Figure 5. Normal spectral emissivity of tungsten for several temper- Tungsten Efficacy. The spectral selectivity of tungsten (i.e.,
ature values. the fact that the emissivity in the visible is higher than that
FILAMENT LAMPS 439
in the infrared) is an important contributor to the efficacy sten atom will escape the wire. In argon at about 1 atm, the
that can be obtained at a given temperature. Adding the spec- number of tungsten atoms that escape is about 1 in every
tral emissivity term to the numerator of Eq. (7) and the total 500 that evaporate from the surface. Obviously, the less likely
emissivity to the denominator, the luminous efficacy is given escape is, the longer it will take to erode the wire, signifi-
by cantly increasing life expectancy. Thus adding a gas to sur-
760 round the filament increases life at a given temperature. Fur-
683 380 λVλ M e,λ
b
dλ thermore, the mass loss process with an inert gas
η = lm/W (11) surrounding the wire is better described as a diffusional flux
σ T 4
of atoms from the cylindrical wire filament in a tungsten
A plot of the radiant efficacy of tungsten compared to that of mass density gradient. The diffusion coefficient is dependent
a blackbody is shown in Fig. 7. Note that at a temperature of on the properties of the inert gas (i.e., gas density, atomic
3000 K the radiant efficacy of tungsten wire is about 27 lm/ mass, and the cross sections for interaction between the inert
W, whereas that of a blackbody is about 21 lm/W, an increase gas and tungsten atoms). Because a gas atmosphere near a
of about 30% resulting from the spectral selectivity of tung- hot wire sets up a convective flow, a major difficulty is en-
sten. Coiling the tungsten, as is the practice in all lamps, re- countered in determining whether, and to what extent, a sta-
duces the selectivity somewhat because the coil tends to be- ble density gradient, which assumes a quiescent gas, can be
have more like a blackbody. assumed.
In addition to reducing the mass loss of tungsten, the
Mass and Heat Loss in a Gas added gas interacts with the filament to remove heat from it.
Thus a sizable fraction of the power input is expended on non-
Operating a tungsten wire at high temperature in a vacuum light-generating processes. For a straight wire, the addition
leads to evaporation of tungsten atoms from the surface. The of an inert gas like argon is more costly in lumen efficacy than
evaporation rate is a very strong function of temperature, be- it is helpful in extending life (i.e., if long life were required, it
ing proportional to about the 36th power at a temperature of
would be better to lower the wire temperature in vacuum to
2800 K. Measurements show (6) that the vapor pressure or
get longer life than adding argon). However, in 1912 Lang-
evaporation rate of tungsten is well described by the Arrhen- muir (7) found that for thin wires, the heat dissipation by free
ius-type equation, where convection increased very slowly with increasing wire diame-
H ◦/RT ter. Moreover, a coil dissipates heat as if it were a wire with
m = 3.8 × 108 exp− g · cm−2 · s−1 (12)
a diameter equal to that of the cylindrical coil. Thus he rea-
soned that by coiling the wire, the heat loss to the gas from a
where ⌬H/R ⫽ 102,300 K.
given length of wire could be greatly reduced because the to-
The evaporation rate changes by about a factor of 3 every tal length of the cylinder is many times shorter than that of
100⬚C when the temperature is between 2700 K and 2900 K. the straight wire. In this way, it was possible to increase the
At a temperature of 2800 K, the evaporation rate, using Eq.
efficacy by operating the coiled wire at a higher temperature
(12) is about 13 ⫻ 10⫺9 g ⭈ cm⫺2 ⭈ s⫺1.
in a gas filled lamp. In 1913 gas filled lamps with coiled fila-
When gas is added to surround the filament in the lamp,
ments were introduced. Later, two coilings (i.e., a coiled-coil
the tungsten evaporating from the surface encounters these filament) improved the situation further. Hence today’s
gas atoms. The resulting collisions of tungsten with the gas household incandescent lamps in the United States use
cause the majority of evaporated tungsten atoms to return to
coiled-coil filaments to take as much advantage of the life in-
the wire. A small fraction escapes (i.e., diffuses away to the crease resulting from the fill gas. This is generally not true in
wall). Thus adding a gas slows the amount of tungsten leav- Europe or other parts of the globe where the mains is 220 V
ing the filament. The higher the gas pressure and the larger to 240 V, where singled coiled filaments are still preferred in
the gas atoms, the less likely it is that an evaporated tung- GLS products. Of course coiling tungsten wire requires a duc-
tile wire. The development of a ductile wire process by Coo-
lidge in 1909 made possible the coiled tungsten filament used
40 in current incandescent lamps.
35 Langmuir explained the reason for the slow increase in
convective heat loss with wire diameter as follows: the veloc-
30 ity distribution around the hot wire of diameter d is such that
Tungsten
Efficacy (lm/W)
gas surrounding the filament. This quiescent gas layer is duced effective surface area. Even though the effect of coiling
called the Langmuir sheath. The difficulty comes in calculat- is beneficial to reducing gas losses, there is a small penalty
ing the thickness of the sheath. See Chapter 2 and Annex A paid in efficacy. This is because light exiting a filament from
of Ref. 8 for how this can be done. the inside of a coil has the opportunity to first undergo one or
more internal reflections. This has the effect of making that
portion of the light more blackbody-like with a lower selectiv-
PRACTICAL FILAMENT LAMPS
ity, albeit higher overall emissivity, than that of a straight
tungsten wire. The effect of coiling on lamp efficacy will be
Why Tungsten?
seen later when the basics of filament coils are introduced.
Tungsten has been the material of choice in incandescent Often, to minimize the thermal gas-loss, it is beneficial to
lamps for about 80 years. What specifically are the advan- double-coil the tungsten wire. That is, the wire is first coiled
tages of tungsten over other incandescent materials? Three into what is called a primary coil. This coil is itself then coiled
major advantages of tungsten over other materials are of pri- into what is referred to as the secondary coil. Some coils are
mary importance. Tungsten has: even triple coiled, although in most cases the benefits are
small to nonexistent.
1. The lowest evaporation rate (vapor pressure) of any Coiling is accomplished in practice using wire mandrels
metal (discussed previously), about which the wire or primary coil is wound. The primary
2. The spectral selectivity which enhances visible light coil is usually made by winding the tungsten wire around a
output (also discussed previously), and molybdenum wire mandrel. This primary mandrel stays in-
side the primary coil while the whole thing is then wound
3. Thermo-mechanical stiffness (non-sag characteristic) at
around a larger retractable mandrel to make the secondary
high temperature.
coil. In the end, filaments are washed in acid to dissolve out
the primary molybdenum coil.
The primary reason tungsten is the preferred material for in-
Coils are referenced with regard to two main parameters,
candescence is that it can be operated at a higher temperature
the pitch ratio Kp and the mandrel ratio Km. Referring to Fig.
than any other metal. The melting temperature is about 3650
8, these coiling parameters are defined as follows:
K. Furthermore, the evaporation rate (vapor pressure) of
tungsten as a function of temperature is lower than any con- h
ductive solid; about a factor of ten lower than tantalum, the pitch ratio: Kp = (12a)
D
next best material in terms of vapor pressure. Everything else m
being equal, the higher the operating temperature the higher mandrel ratio: Km = (12b)
D
the lumen efficacy as demonstrated in Fig. 4 for a blackbody.
As discussed previously, the spectral selectivity of tungsten Here, h is the turn separation of the coil (measured from the
is also an important contributor to the efficacy that can be center of the wire on one turn to the center of the wire on an
obtained at a given temperature. Some ceramic compounds adjacent turn), D is the wire diameter, and m is the inside
such as hafnium nitride or tantalum carbide display an emis- diameter of the coil. This is equivalent to the outer diameter
sivity curve that is even more selective than that of tungsten. of the mandrel wire about which the tungsten wire is wound.
That is, at a given temperature, they give off more radiant The pitch ratio given in Eq. (12a) is actually slightly differ-
energy in the visible region and less in the infrared compared ent than that used by most engineers and coil manufacturers.
to tungsten. These have so far proved unsuitable for incandes- The latter, ‘‘engineer’s’’ pitch ratio is given by
cent lamps because of their extreme brittleness, which causes
an inability to form into suitable filaments and the tendency h
K p = (13)
to disproportionate at high temperatures. Furthermore, de- D
spite higher melting point temperatures, these materials gen-
erally display higher vaporization rates for typical lamp gas
fills than does tungsten (see Ref. 9). D
Finally, the fact that tungsten wire is ductile enough that
it can be drawn from solid tungsten ingots at low tempera-
tures and then transformed into a stiff, rigid body through
doping and recrystallization at high temperatures is a unique m
and important feature of doped tungsten. Without this prop-
erty coiled filaments would not be practical. The capability of
extended operation without sag or distortion at temperatures
greater than 90% of melting is noteworthy; it is a characteris-
tic not found in most metals.
Coiling h
This latter definition is the one generally known because h⬘ Consider the two equations for resistance and power. First,
is easy to measure. It is simply the inverse of the number of the resistance of a wire is a given by the ratio of the wire
turns per unit length (usually written as TPI, for turns per length l to the cross sectional area, A ⫽ 앟d2 /4, times a temper-
inch). The relation between K⬘p and Kp is ature-dependent material property, the resistivity (T). Here,
d is the wire diameter. Thus an initial value for the wire
K2K +(K4(K+ 1)+ 1)
p m length is given by
Kp = 2
(14)
2
p m
RA V 2 πd 2
l= = (15)
ρ(T ) 4Pρ(T )
(Note that the expression for this found on page 146 of Ref. 2
is incorrect.) These coiling parameters are used to relate the Similarly, we can obtain an expression for the wire length l
wire length and diameter to the coil length and diameter, from the equation for power. The power input P to a lamp
which show up in the gas loss equation and a multiplying filament is dissipated by radiative losses Pr, which dominate,
factor 웃 for the efficacy, as will be shown later. as well as heat loss to the gas Pg and conduction to the leads
Pl. Thus
Rudiments of Lamp Design
P = Pr + Pg + Pl = Pr + P (16)
An incandescent lamp is rated by how much power it con-
sumes, how much light it generates, and how long it is ex- where the two nonradiative loss terms have been collected
pected to burn before failure. In addition, the voltage or cur- together in ⌬P.
rent of the power source needs to be specified. These lamp The radiated power from a tungsten wire can be obtained
ratings or operating characteristics are met by choosing a by multiplying the blackbody emittance in Eq. (3) by the total
tungsten wire of a given length and diameter such that, when tungsten emissivity and the wire surface area. The effect of
placed across a given power source, it reaches a temperature reabsorption of radiated power by a coiled wire, called the
that gives the specified power, lumens, and life values. Thus coiling factor 웃, will be described later. The coiling is weakly
incandescent lamp design is principally concerned with simul- dependent on both the wavelength and temperature. For the
taneously matching the conditions of voltage (or current or power, which represents integration over all wavelengths, the
resistance), power consumption, lumen output, and life by ma- expression 웃t will be used. Thus
nipulating the wire length, wire diameter, and filament tem-
perature. Note that because there are only three intrinsic pa- P = π · d · l · δt (T ) · (T ) · σ · T 4 + P (17)
rameters that relate to the filament (i.e., wire length and
diameter and filament temperature), only three of the four Solving Eq. (17) for the wire length l gives
lamp characteristics listed here can be specified indepen- P− P
dently. l= (18)
π · d · δt (T ) · (T ) · σ · T 4
Even though only three filament-related parameters are
considered to be intrinsic (i.e., affect all parameters), other
Equations (15) and (18), both of which give expressions for the
lamp-related parameters such as coiling, gas type and pres-
wire length, can now be solved simultaneously for the wire
sure, and wall temperature have a significant impact on some
diameter d. This gives
of the lamp characteristics. These other parameters, however,
are considered to be supplemental rather than intrinsic pa- 1/3
4Pρ(T )(P − P)
rameters as are the wire dimensions and temperature. A brief d= (19)
V 2 π 2 δt (T )(T )σ T 4
review of the underlying principle physical relationships be-
tween the lamp characteristics and filament parameters fol-
Note that an initial value for ⌬P must be specified. For GLS
lows. These relationships constitute the fundamentals of de- lamps, the lead conduction is small (about 1%) and the gas
sign. Because of the interrelating relationships of the various
loss varies from about 8% to 20%, depending on lamp voltage,
parameters, the approach is by necessity iterative. wattage, and coiling. For automotive lamps or other low-volt-
age lamps, the lead conduction losses and gas convection
Design Procedure losses are about equal, usually around 5% each. A value for
One approach is to start with the lamp voltage V and the ⌬P of about 0.1 ⫻ P is a good place to start the iteration. As
desired power P as being fixed. From Ohm’s Law, this fixes values for d, l, and T become better defined in the calcula-
the filament resistance R ⫽ V 2 /P. As will be shown momen- tion, ⌬P can be better estimated as well.
tarily, the filament wire length and wire diameter may then After an initial estimate for the wire length and diameter
be derived from simultaneous solutions of the resistance and are derived using Eqs. (15) and (19), an initial value for the
power equation, where an initial value of the temperature has total luminous flux from the filament may be derived by mul-
been specified. Given these initial values for wire length, di- tiplying the wire surface area, S ⫽ 앟dl, times the Planck
ameter, and temperature, the luminous flux and life may also blackbody distribution function weighted by the eye sensitiv-
be derived. From the ratio of calculated to specified values for ity curve and the spectral emissivity curve for tungsten in a
either lamp life or the lumen output, the temperature can manner similar to that shown previously. The luminous flux
now calculated. With this new temperature, the wire length (T) is given by
and wire diameter are recalculated. This process is iterated 760
until the solution no longer changes within a specified error. (T ) = S Vλ λ (T )δλ (T )M e,λ
b
(T ) dλ (20)
This process is shown in detail later. 380
442 FILAMENT LAMPS
This equation is solved for the flux value ⌽(T0), in terms of The total tungsten emissivity is also taken from the same
the initial estimate of temperature T0 and given the values source. A second-order fit to the data from 2400 K to 3600 K
for l and d calculated earlier. However, since the lumen value gives the following functional form:
is assumed specified, the ratio of ⌽given to ⌽(T0) can be used to
calculate a new value for the temperature, T1, from the expo- (T ) = − 2.03926 × 10−2 + (1.88601 × 10−4 )T
(27)
nential part of the Me,b term, which is then inserted into Eqs. − (2.35764 × 10−8 )T 2
(15) and (19).
The coiling factor 웃 can be thought of as modifying the The spectral emissivity for tungsten at 2600 K is taken from
emissivity of the coil. For straight wire, 웃 ⫽ 1. For a coil, the Ref. 5. A second-order fit gives the following functional form:
light generated on the inside of the coil has a chance of being
reabsorbed, thereby increasing the filament temperature. For (λ, T = 2600) = 0.55788 − (233904)λ + (4.91792 × 1010 )λ2
a tungsten coiled coil, a typical value at 2700 K is 웃 앑 0.79. (28)
The coiling factor 웃 is a function of temperature and coiling
parameters, and is given, for a single coil, by the following Here the wavelength is measured in meters.
expression (see Ref. 2): For nitrogen, argon, and krypton gas, typical values for
Pg(T), the gas-loss term, may be taken from the following ex-
1
1
1
K −1 pressions, which were derived for a typical 40 W, 120 V fila-
p
δ1 (T1 ) = 1+ + 1− ment lamp with 80 kPa (앑600 torr) cold pressure fill:
2 π (Km + 1) π (Km + 1) Kp − r(T1 )
(21) for N2 : Pg (T ) = −1.2522 + (1.6968 × 10−3 )T
(29)
+ (6.4245 × 10−7 )T 2 W
The coiling parameters Km and Kp are given by Eqs. (12)–(14).
The temperature dependence of 웃 is carried through the re- for Ar : Pg (T ) = −0.8535 + (1.3868 × 10−3 )T
flectivity r(T), which is given by (30)
+ (3.1375 × 10−7 )T 2 W
12
10 N2
Ar
Kr
8
Gas loss (W)
0
2000 2100 2200 2300 2400 2500 2600 2700 2800 2900 3000 3100 3200 3300 3400 Figure 9. Gas loss (watts) calculated for
Temperature (K) different gas fills for 40 W, 120 V lamp.
hotter because the resistance is higher there, and because the The temperature dependencies of the radiated power, lu-
current must be continuous throughout the wire, the thin mens, radiant efficacy, and life for tungsten wire are shown
areas dissipate more power. Thus a positive feedback cycle in Fig. 10. Here the data for each parameter is normalized to
develops, causing the hot spot to increase in temperature one at a temperature of 2800 K to show more clearly the
more and more rapidly until it reaches the melting point of strength of the temperature dependence. Note that because
tungsten and the filament fails (see Ref. 11). It is known that lumens increase about the square of the radiated power, the
hot-spot development is the cause of filament failure, but all efficacy increases at about the same rate as the radiated
the causes of hot-spot development are not known nor is their power. Note further that life, based on the temperature de-
relationship to each other understood well at all. It is proba- pendence of the evaporation rate, decreases most rapidly with
bly because the life-ending mechanism involves hot spots in increasing temperature. Hence, for a given design situation,
which the distribution of life is found to follow a normal curve luminous efficacy increases can be traded against shorter life.
extremely well. Furthermore, the standard distribution of the However, the efficacy increase is only on the order of 10% of
normal life curve is usually quite large being about 20% to the decrease in life. This situation always holds for incandes-
25% of the expected life. cent and halogen lamps. The life at a given temperature can
From the diffusional flow of tungsten atoms through the be increased by increasing the fill gas pressure, going to a
Langmuir sheath, as already discussed, a relationship be- denser rare gas, increasing the wall temperature (all of which
tween expected filament life ⌳ and the filament parameters are done in halogen lamps), and decreasing the voltage that
can be put into the following form: results in larger diameter wire, but the temperature depen-
dence remains as shown in Fig. 10. For further discussion on
C d x the filament lamp design approach used here, see Ref. 12.
= (32)
(T, gas, p, coil, Tw )
1.0E+2
Life
1.0E+1
Normalized performance
1.0E+0
Power
Efficacy
1.0E–1
Lumens
1.0E–2
Figure 10. Temperature dependence of 2000 2200 2400 2600 2800 3000 3200 3400 3600
various parameters for tungsten wire. Temperature (K)
and the flux of evaporating tungsten atoms, although reduced Mechanism of the Halogen Cycle
because of the higher pressure, would nonetheless cause
For the iodine halogen cycle, there is general acceptance that
eventual wall blackening. What is needed is a way to keep
oxygen is a requirement and that the volatile tungsten-iodine-
the wall clean despite the close proximity of the wall to the
containing compound, which returns the tungsten to the coil,
filament.
is WO2I2 (14,15). There is also some agreement that the crude
Halogen lamps are incandescent lamps containing com-
mechanism presented in Reactions 1 through 3 is correct
pounds of one or more of the elements I, Br, Cl, and F. These
(15,16).
compounds act to prevent the tungsten evaporating from the
incandescent coil from depositing on and darkening the lamp
At filament:
wall. The tungsten evaporating from the coil reacts with the
halogen in the cooler regions of the lamp and is converted into W(solid) → W(gas) (R1)
gaseous halogen-containing compounds. These compounds de-
At/near filament:
compose en route to the hotter regions of the lamp, depositing
tungsten metal on the leads or cooler parts of the coil and W(solid, gas) + 2O(gas) → WO2 (gas) (R2)
releasing the active halogen to continue the cycle. For halogen
At/near wall:
cycles used today, the tungsten is not deposited in exactly the
same places on the coil from which it evaporated (13), so the WO2 (gas, solid) + 2I(gas) → WO2 I2 (gas) (R3)
coil in a halogen lamp does not last forever and eventually
fails, generally in the same manner as an ordinary incandes- These reactions are not meant to represent a mechanism in
cent coil. the chemical sense. Not enough is known for that level of de-
With such a wall-cleaning agent, the incandescent lamp tail. In addition to these general reactions, there is a corro-
can be made with a much smaller, and consequently much sion reaction between the tungsten filament and the iodine,
stronger, bulb. This permits the use of higher pressure fill which slightly increases the mass loss rate of the filament.
gases and the economical use of rarer denser inert gases such This is represented by Reaction 4.
as krypton and xenon. With higher pressures and denser
gases, the rate of tungsten mass loss from the coil is de- At filament:
creased, and so life is increased. Denser inert gases also have
lower thermal conductivity, which results in less energy lost W(solid) + n I(gas) → WIn (gas), n = 1, 2 (R4)
by gas conduction. With these changes, the coil may be rede-
signed for increased light output, increased life, or some com- It is likely that an analogous mechanism operates for the bro-
bination thereof. A typical halogen lamp may be 10% brighter mine cycle where the transport species is WO2Br2. There is
and last twice as long as an ordinary incandescent lamp. On also the possibility of an oxygen-free transport cycle for bro-
the negative side, halogen lamps can operate with the lamp mine where the transport species are WBr4 and WBr5. These
walls as hot as 900 K and with internal pressures of 1.5 MPa species are less stable than WO2Br2 and are formed at lower
(15 atm) or more. For safe operation, such lamps must be pre- rates. Wall cleaning and lead corrosion rates are so sensitive
vented from contacting combustible materials and a means of to trace amounts of oxygen that it is difficult to say that, in
containing any fragments from potential wall rupturing must practical bromine cycle lamps, oxygen is not involved. Oxygen
be provided. usually enters the lamp in the form of water or metal oxides.
FILAMENT LAMPS 445
Halogen Lamp Engineering and metal surface oxides. In some cases, there is too much
oxygen contamination, and oxygen getters are required. Many
Halogen Types Used. In an operating lamp, the halogen
getters have been proposed for use with halogen cycle lamps
dose must react quickly enough to convert all evaporating
(18–32). In general, oxygen getters cannot be used in excess,
tungsten to the volatile compound but react slowly enough
as in the case of ordinary incandescent lamps, because this
so that the current-carrying leads and coil supports are not
results in poor lumen maintenance. The sensitivity of the re-
corroded. Ordinarily, to avoid excessive failures from lead cor-
action rates to oxygen, the ubiquitous nature of oxygen, and
rosion and to decrease the extent of coil attack (Reactions 2
the reactivity of oxygen with many common metallic contami-
and 4), halogen lamps are designed to blacken very slowly, a
nants present the main challenge to consistent manufacture
typical maintenance being 85% to 95% of initial lumen output
of high-quality halogen lamps. The fact that there is, in gen-
at 70% of life. The life here is the coil life, which is governed
eral, no easy technique to measure oxygen in all its important
by the hot-spot mechanism described before. This requires not
forms inside a lamp makes the chemical design of halogen
only control of the amount of halogen but also control of the
lamps very difficult. In practice, the halogen dose is deter-
amount of hydrogen and contaminants that can affect the
mined using a design of experiments approach.
amount of oxygen available to the lamp atmosphere.
Control of reaction rates limits use of the elemental forms
to I2, while Br2, Cl2, and F2 are too reactive. The addition of Materials Requirements. To prevent reaction and loss of Cl-,
hydrogen inhibits the rate of lead corrosion and wall cleaning Br-, or I-based halogen doses with the lamp walls, the glass
and permits the use of lamps dosed with HBr, CH2Br2, and used must not contain alkali or alkaline earth elements. This
CH3Br. The addition of oxygen accelerates the rate of lead limits the choice of wall materials to alumino-silicate glasses
corrosion and wall cleaning, permits the use of lamps dosed of negligible alkali content, 96% silica glass (such as Vycor,
with CH3I, and mandates very precise control of the oxygen made by Corning Glass, Inc.) and quartz. At the operating
level in all types of bromine lamps. With the improved manu- temperatures of halogen lamps, Vycor and quartz are both
facturing processes of today, most halogen lamps use bro- permeable to H2 whereas the alumino-silicate glasses are not.
mine doses. Currently, there is no economical lamp wall material that re-
Common bromine doses are HBr, CH2Br2, and CH3Br. sists attack by fluorine-based halogen doses. Nor has any eco-
Doses used in special cases include CH3I, PNBr2, and mix- nomical coating been developed. Consequently, there are no
tures of halogen compounds. HBr, CH2Br2, CH3Br, and CH3I commercially available fluorine cycle lamps. The internal
are gases and are added to the lamp along with the inert fill metal parts of halogen lamps are tungsten or molybdenum
gas. PNBr2 is a solid and is added to the lamp as a solution lead wires and supports and the tungsten coil.
in petroleum ether, with the petroleum ether solvent allowed
to evaporate prior to evacuation and inert gas fill. CH3Br and Temperature Requirements. Vycor and quartz require the
CH2Br2 doses are less corrosive than HBr and have a much use of molybdenum foil hermetic seals for the current-car-
lower tendency to adsorb on the inner surface of the lamp rying leads. This restricts the temperature of the seal area to
exhaust and fill equipment, which increases dose reproduc- less than 620 K if the seal life is to exceed 1000 h, unless
ibility. these seals are not exposed to air. The lamp walls must be
Within seconds of first lighting the coil, CH3Br and hot enough to prevent the volatile tungsten-halogen-con-
CH2Br2 are converted to HBr as shown in Reactions 5 and 6 taining compounds from condensing and removing the halo-
(17). gen from the cycle. For iodine cycle lamps, the coldest permis-
sible wall temperature is about 520 K (33). For bromine cycle
CH3 Br(gas) → HBr(gas) + H2 (gas) + C(solid) (R5) lamps, the coldest permissible wall temperature is about 440
CH2 Br2 (gas) → 2 HBr(gas) + C(solid) (R6) K (33). In general, the upper limit to the lamp wall tempera-
ture is determined by either its rupture strength or the stabil-
If used in excess, the carbon released from Reactions 5 or 6 ity of the tungsten transport species. The latter consideration
can deposit on the bulb wall causing darkening or can be limits the wall temperature to about 1300 K. In practice, how-
transported to localized areas of the coil causing embrittle- ever, the maximum wall temperature is often lower because
ment. The additional hydrogen released from Reaction 5 has of unacceptable water release or impurity element migration
an inhibiting effect both on the kinetics of wall cleaning and to the lamp interior.
on the kinetics of lead corrosion.
The amount of the halogen added typically ranges from Lamp Orientation Requirements. Ordinarily, halogen lamps
0.03% to about 5% (mole or volume percent). The amount of can be operated in any orientation. A special case arises for
halogen added increases with increasing tungsten evapora- the linear halogen lamps [i.e., those with lengths over 5 cm
tion rate, increasing hydrogen to bromine ratio, and decreas- (앑2 inches) having outer lamp wall diameters of about 1.2 cm
ing amounts of oxygen. The hydrogen to bromine ratio is a (0.5 inch)]. To ensure uniform distribution of halogen, these
function of the type of halogen dose, because little additional lamps must be burnt within about 4⬚ of horizontal. When
hydrogen arises from the manufacturing process. burnt off-horizontal, the linear lamps act like thermal separa-
Oxygen is not usually added to bromine lamps because suf- tor columns, and wall blackening will occur in the regions
ficient amounts remain in the lamp as contaminants during where the halogen is depleted (p. 71 in Ref. 15, p. 143 in Ref.
normal manufacture. The most common sources of this oxy- 34) Occasionally, corrosion problems will result for lamps op-
gen include water adsorbed on lamp parts and in the exhaust/ erated in certain orientations when the hot gas flow from the
fill system, water and other oxygen-containing species dis- coil is directed onto metal parts (35,36), but these problems
solved in the lamp walls, oxygen dissolved in the metal parts, are usually resolved in the design phase before production.
446 FILAMENT LAMPS
1.0E-06
O/Inert = 1 ppm
O/Inert = 0.1 ppm
Elemental tungsten gas phase mole fraction
1.0E-08
B
1.0E-09
1.0E-10
formed every 100 K and 0.5 log units in oxygen level, which duced, the minimum wall temperature is increased, and
causes the jagged boundaries between phases. Figure 13 maximum wall temperature is decreased.
depicts the purity and wall temperature required to prevent
The Future
wall deposits in this halogen lamp. Figure 13 shows that
as the oxygen contamination is reduced below 0.1 ppm, the The performance of halogen lamps will continue to improve as
wall temperature range for deposit-free operation is re- designs become more optimized. This will be an incremental
Condensate map
–3.0
yyy
;;;
–5.0 –4.0
Log of O to Ar ratio
;;;; yyy
;;;
–7.0 –6.0
;yy;y;y;
y;
yy;
;;;
–8.0
–9.0
Table 1. Power Balance and Performance of 100 W Incandescent and 90 W Halogen Lamps
Power Efficacy
Type (W) % Visible % Non-Vis. % Non-Rad. (lm/W) Life (h)
Incandescent 100 10 75 15 17.1 750
Halogen 90 11 74 15 17.5 2000
process at best, with the optimum performance limited to that 6. E. R. Plante and A. B. Sessoms, J. Res. NBS, 77A: 237, 1973.
of the basic incandescent lamp. The future of halogen lamps 7. I. Langmuir, Phys. Rev., 34: 40, 1912.
will be brighter when the materials problems associated with 8. W. Elenbaas, Light Sources, Chap. 2, New York: Crane Russak &
using a fluorine-based cycle are solved. This cycle might have Co., 1972.
the promise of increasing life by returning the evaporated 9. L. Bigio et al., Investigation for Incandescent Lamp Efficiency Im-
tungsten to the hot spot (14,15,34). Another, more specula- provement, Final report for New York State Energy Research and
tive, area where halogen or other chemical transport cycles Development Authority contract No. 3034-IABR-BR-94, 1996.
may find use is to stabilize the phases of other materials that 10. R. C. Weast and M. J. Astle (eds.), Handbook of Chemistry and
could be used as incandescent bodies at temperatures ex- Physics, 62nd ed., Boca Raton, FL: CRC Press, 1981–82.
ceeding the melting point of tungsten. 11. H. Hörster, E. Kauer, and W. Lechner, The burnout mechanism
of incandescent lamps, Philips Tech. Rev., 32 (6/7/8): 155–164,
1971.
POWER BALANCE AND PERFORMANCE FOR TYPICAL LAMPS
12. T. D. Page, R. S. Bergman, and M. R. Vukcevich, A computational
first principles approach to incandescent lamp design, J. Illumi-
Table 1 shows the power balance and performance compari- nating Eng. Soc., 18 (2): 109–117, 1989.
son for a typical 100 W incandescent lamp and a 90 W halo- 13. P. Bayle et al., C.R. hebd. Séances Acad. Sci., 258: 4710, 1964.
gen lamp. The percent visible and nonvisible radiation refers
14. J. H. Dettingmeijer et al., Regenerative chemical cycles in tung-
to that which actually exits the lamp. The nonradiative losses sten-halogen lamps, Philips Tech. Rev., 35 (11/12): 302, 1975.
include loss resulting from absorption of the radiation by the
15. A. Rabenau, Chemistry of the incandescent lamp, Angew. Chem.
bulb wall and the base (which is 앑5% for both the 100 W Int. Ed., 6 (1): 68, 1967.
incandescent and 90 W halogen lamps).
16. E. Zubler, Elenbaas Award Lecture, GE Lighting Report 81-LRL-
The performance advantage of the halogen lamp is clear. 1827, 1981.
In general, life and efficacy performance advantages can be
17. L. Bigio and P. J. Codella, FTIR diagnostics of tungsten halogen
traded off for each other. In this example, the main perfor-
lamps, Nicolet FTIR Spectral Lines, 11: 10–12, Spring/Summer
mance advantage of the halogen lamp is taken in life. As al- 1990.
ways, though, there is a price to be paid for the performance
18. K. T. Zarins, Incandescent lamp, US Patent 3,351,407, 1967.
advantage, and this is the extra cost to manufacture halogen
19. J. A. Graves, Getter for incandescent lamps and similar devices,
lamps, which results in higher market prices compared to
US Patent 3,475,072, 1969.
standard incandescent lamps. The higher material costs stem
20. P. D. Johnson, Efficient incandescent light source including light-
mainly from the cost of denser gas types such as krypton and
enhancing metallic iodide vapors, US Patent 3,497,754, 1970.
xenon, which are used in the halogen lamps. The major cost
21. U. Mikami, et al., Electric incandescent lamp, US Patent
addition, though, comes from the fact that an extra filament
3,582,703, 1971.
tube must be made to replace the simple filament in an incan-
22. G. R. T’Jampens, Electric incandescent lamp, US Patent
descent lamp. In addition, the requirements for filling lamps
3,585,435, 1971.
at pressures above 1 bar lead to additional costs, as does that
23. J. R. Coaton and J. M. Rees, A hydrogen-halogen filament lamp
of ensuring that the halogen lamp will contain any particles
with a hydrogen getter flag, US Patent 3,644,773, 1972.
safely in the unlikely event of rupture of the filament tube.
24. D. J. De Fraeye and L. Lens, Halogen incandescent lamp con-
For this reason, little market penetration of the halogen lamp
taining bromine and phosphorus, US Patent 3,800,180, 1974.
in a general lighting household lamp such as the A-line has
25. E. A. Gulbransen and S. A. Jansson, Tungsten halogen incandes-
yet occurred. Thus, the two lamp types can coexist in the mar-
cent lamp with group IVA metal getter and method of manufac-
ket, each filling a particular market niche. ture, US Patent 3,821,585, 1974.
26. G. Prakken and G. R. T’Jampens, Tungsten-bromine cycle lamp,
BIBLIOGRAPHY US Patent 3,829,731, 1974.
27. R. B. Johnston and J. M. Rees, Halogen type filament lamp con-
1. J. M. Anderson and J. S. Saby, The electric lamp: 100 years of taining phosphorus and nitrogen, US Patent 3,898,500, 1975.
applied physics, Phys. Today, 32 (10): 32–40, 1979. 28. W. Kischio, Electric incandescent lamp, US Patent 4,032,808,
2. M R. Vukcevich, The Science of Incandescence, Cleveland: NELA 1977.
Press, 1993. 29. G. R. T’Jampens and H. F. J. Giller, Electric tungsten/bromine
3. L. A. Hawkins, Adventure into the Unknown—The First Fifty cycle lamp and method of manufacturing said lamp, US Patent
Years of the General Electric Research Laboratory, New York: Wil- 4,039,879, 1977.
liam Morrow, 1950. 30. J. E. Oetken and J. P. Buschmann, Tungsten halogen lamp with
4. J. A. Fox, A Century of Light, The Benjamin Company, 1979. coiled getter, US Patent 4,415,833, 1983.
5. W. E. Forsythe and E. Q. Adams, J. Opt. Soc. Amer., 35: 108, 31. R. M. Griffin and R. C. Martin, Long life incandescent tungsten-
1945. halogen lamp, US Patent 4,451,760, 1984.
FILTER APPROXIMATION METHODS 449
32. J. P. Keenan and P. R. Gagnon, Electric lamp including oxygen 54. G. Dittmer and U. Niemann, The influence of solid phases on
getter, US Patent 4,727,286, 1988. transport cycles in halogen-filled incandescent lamps, Philips J.
33. J. R. Coaton, Modern tungsten-halogen-lamp technology, Proc. Res., 42: 41, 1987.
IEE, 117 (10): 1953, 1970. 55. E. Schnedler, Three-dimensional model for the theoretical de-
34. J. R. Coaton and J. R. Fitzpatrick, Tungsten-halogen lamps and scription of tungsten transports in halogen incandescent lamps,
regenerative mechanisms, IEEE PROC, 127A (3): 142, 1980. High Temp. Sci., 19: 237, 1985.
35. T. H. Yu and J. M. Ranish, in The 5th Int. Symp. Sci. Technol. 56. E. Schnedler, in The 4th Int. Symp. Sci. Technol. Light Sources,
Light Sources, York, UK: Univ. York, 1989. Univ. Karlsruhe, Karlsruhe, Germany, 1986, p. 75.
36. J. M. Ranish and L. E. Hoegler, in The 6th Int. Symp. Sci. Technol. 57. S. K. Gupta, Thermodynamic and kinetic aspects of bromine
Light Sources, Technical Univ. Budapest, Budapest, Hungary, lamp chemistry, J. Electrochem. Soc., 125 (12): 2064, 1978.
1992, p. 52.
37. B. Kopelman and K. A. Van Wormer, Jr., Thermodynamic consid- ROLF BERGMAN
erations of tungsten halogen lamps, Illum. Eng., 63: 176, 1968. GE Lighting
38. B. Kopelman and K. A. Van Wormer, Jr., Thermodynamics of LAURENCE BIGIO
tungsten halogen lamps containing chlorine, fluorine, and hydro- GE Corporate Research and
gen, Illum. Eng., 64: 230, 1969. Development Center
39. L. N. Yannopoulos and A. Pebler, Thermochemical calculations JOSEPH RANISH
of tungsten halogen lamps containing bromine, oxygen, hydrogen, Applied Materials
and carbon, J. Appl. Phys., 42 (2): 858, 1971.
40. G. Gottschalk and G. M. Neumann, Simulation of heterogeneous
gas equilibria. I. Metal-halogen systems with examples for tung-
sten-halogen systems, Z. Metallkde., 62: 910, 1971. FILM DEPOSITION BY PLASMA. See PLASMA DEPO-
41. L. N. Yannopoulos and A. Pebler, Direction of chemical transport SITION.
of tungsten in tungsten-bromine lamps, J. Appl. Phys., 43: FILM GROWTH, HTS. See HTS FILM GROWTH.
2435, 1972. FILM THICKNESS MEASUREMENT. See THIN FILM ANA-
42. G. M. Neumann, Halogen incandescent lamps. I. Thermodynamic LYZERS.
analysis of the chemical transport reactions, Tech.-Wiss. Abh. Os- FILTER AMPLIFIERS. See BUTTERWORTH FILTERS.
ram-Ges., 11: 8, 1973.
43. G. M. Neumann, Thermodynamik heterogener Gasgleichge-
wichte. IX Einfluss von Kohlenstoff auf die chemischen Rea-
ktionen und Transportprozesse in den Reaktionssystemen Wol-
fram-Halogen-Sauerstoff-Wasserstoff, Z. Metallkde., 64 (6): 444,
1973.
44. G. M. Neumann, Einfluss von Sauerstoff auf die Transportreak-
tionen beim Wolfram-Brom-Kreisprozess in Halogenglühlampen,
J. Less-Common Met., 35: 51, 1974.
45. J. H. Dettingmeijer, B. Meinders, and L. N. Nijland, Chemical
transport of tungsten in the heterogeneous systems tungsten-
oxygen-hydrogen-iodine/bromine/chlorine, J. Less-Common Met.,
35: 159, 1974.
46. T. Geszti, On the theory of the halogen lamp. I. Radial transport
in the presence of carbon, Acta Technica Academiae Scientiarum
Hungaricae, 78 (3–4): 463, 1974.
47. T. Geszti and I. Gaál, On the theory of the halogen lamp. II. Gas-
controlled axial transport, Acta Tecnica Academiae Scientiarum
Hungaricae, 78 (3–4): 479, 1974.
48. W. J. van den Hoek and G. Rouweler, On thermodynamic calcula-
tions of chemical transport in halogen incandescent lamps, Phil-
ips Res. Repts., 31: 23, 1976.
49. J. R. de Bie and J. C. M. A. Ponsioen, Life and luminous flux
of halogen incandescent lamps related to filament temperature,
pressure, and CH2Br2 content, Lighting Res. Tech., 9 (3): 141,
1977.
50. G. Dittmer and U. Niemann, Heterogeneous reactions and chemi-
cal transport of tungsten with halogens and oxygen under steady-
state conditions of incandescent lamps, Philips J. Res., 36: 89,
1981.
51. E. Schnedler, Description of tungsten transport processes in halo-
gen incandescent lamps, Philips J. Res., 38: 236, 1983.
52. J. A. Sell, Chemical equilibrium calculations of tungsten-halogen
systems, J. Appl. Phys., 54: 4605, 1983.
53. J. A. Sell, Calculations on the role of carbon monoxide in tung-
sten-halogen bulbs containing dibromomethane, J. Appl. Phys.,
54: 6687, 1983.
246 LASER DESKTOP MACHINING
;;
opens up, high-power diodes will become available over a
Light
output Light output
n GaAs
p GaAs
200 µ m
≈ 400 µ m
p – n junction
(a)
5 µm 60°
z
LASER DESKTOP MACHINING 10°
200 µ m
Harnessing a massless beam of light energy for heavy manu- x 5 µm
facturing applications like cutting, drilling, welding, solder-
ing, surface hardening, cladding, sintering metal powder for 1 µm
rapid part fabrication, and so on, is almost like science fiction
becoming reality to engineers (1). There is no need to replace
expensive cutting tools repeatedly. The reality was demon-
(b)
strated in the 1960s with the invention of ruby, Nd–YAG,
and CO2 lasers delivering pulsed high peak power. By the
200 µ m
early 1990s, engineers realized the potential of the desktop Cylindrical lens
manufacturing revolution, with the advent of compact high-
power diode lasers (DL) arrays (2–17).
Of all the lasers, DLs are the most efficient ones. Some of
the commercial ones now reach 40% to 50% electrical to opti-
cal efficiency. They are very compact: a 1 W laser is the size Laser diode
of a grain of table salt, with emission intensity about 1
MW/cm2 to 10 MW/cm2. They operate at about 2 V, and their (c)
pulse rate and shape can be controlled to almost any desired
Figure 1. Semiconductor diode lasers and their coupling to glass fi-
value from direct current (dc) to multigigahertz. DLs have not
ber. (a) An enlarged view of a single microscopic diode depicting the
only helped usher in the Knowledge Age through fiber-optic thin active layer that limits the maximum power and produces highly
communication network, but they are also beginning to com- divergent and asymmetric beam. (b) An example of a bar of a mono-
pete in the manufacturing applications when a larger number lithic incoherent array of diode lasers. (c) The cross-sectional view of
of them are combined to produce tens to thousands of watts coupling of such a bar of laser array to a multimode optical fiber using
of continuous-wave (CW) or pulsed power. DLs are compara- a cylindrical lens.
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
LASER DESKTOP MACHINING 247
106
109
104
Marking
102
Intensity (W/cm2)
Drilling
107 100 Deep
penetration
Sp Glazing welding
ec Cutting
ifi
c
en
er Welding
105 gy
(J
/c Lam
m2 Currently
)
accessible
Hardening Figure 2. Intensities and interaction
with laser
Brazing diodes times required for various laser material
processes. (Courtesy of Lawrence Liv-
103 ermore National Laboratory and Fraun-
10–8 10–6 10–4 10–2 10–0
hofer Resource Center at Ann Arbor; see
Interaction time (s) also Ref. 10.)
much wider range of wavelengths from deep blue (0.35 애m) culty, along with the limited market acceptance, has kept the
to near-infrared (5 애m). cost of the current commercial system relatively high. A diode
coupled fiber bundle as shown in Fig. 1(c) is capable of provid-
ing CW power exceeding 100 W or more. Such devices are
commercially available from several international suppliers.
HIGH-POWER DIODE LASERS Figure 3 shows a different geometry to achieve intensity
reaching 50 kW/cm2 to 100 kW/cm2. This is achieved by fold-
Diode lasers (Fabry–Perot type) are produced by cleaving epi- ing the beams from a bar of incoherent laser array into a
taxially grown p–n junction semiconductor materials GaAlAs closely packed vertical stack to achieve higher brightness.
(0.8 애m), GaAsSb (0.98 애m), InGaAsP (1.55 애m), and so on, Much higher total power in the kilowatt domain is com-
and then electrically pumping across the p–n junction orthog- mercially available with a geometry called ‘‘rack-n-stack.’’ In
onal to the microscopic waveguide cavity (5) (500 애m to 1000 this geometry a two-dimensional dense stack of incoherent di-
애m long) [see Fig. 1(a)]. Single-mode waveguide facets are ode bars are stacked on very special thin cooling plates. The
about 1 애m ⫻ 3 애m, through which a 10⬚ ⫻ 60⬚ divergent average intensity at the surface of the stack can reach as high
beam is emitted. as 2 kW/cm2. By separately collimating each horizontal bar
A monolithic array (6–9) of such lasers can easily emit 1 by miniature cylindrical lenses (monolithic or discrete), fol-
W to 20 W as shown in Fig. 1(b). The light is generally inco- lowed by a large focusing lens, one can obtain an intensity
herent from element to element, but a cylindrical lens can
collect it into a glass fiber of approximately 200 애m core diam-
eter (as an example), matching the width of one monolithic
diode array [Fig. 1(c)]. Such a system, with 80% coupling effi-
ciency, can easily provide an output working intensity ex-
ceeding 50 kW/cm2.
Let us focus briefly on the intensity (or power density) re-
quirements for various laser materials processing functions
(10,11) as shown in Fig. 2. The figure does not explicitly iden-
tify processes like metal powder sintering for rapid prototyp-
ing, soldering, cellulose cutting, vapor phase deposition, and
so on, that require power density in the low kilowatt domain
with total power from 10 W to 100 W only. Figure 2 does
show that most of the traditional laser material processing
functions can (eventually) be carried out by DLs since their
intrinsic emission intensity is about 107 W/cm2 albeit at low
individual device power of about 1 W or less. Thus, heavy
manufacturing requires collecting laser light from a large
number of devices into a small spot to achieve high total Figure 3. High brightness diode laser spot from an incoherent bar
power at high intensity. Technically, this is called high after vertically stacking horizontal source array by a pair of tilted
brightness requirement; and the associated engineering diffi- and translated mirrors. (Courtesy of Opto Power Corporation.)
248 LASER DESKTOP MACHINING
5 cm
Diode
array Collimating
lens array
f/1 coupling Optical fiber
lens 0.5 mm diameter
Embedded Bragg grating
(a)
Figure 5. Schematic diagram of a next-generation high-power DL
using broad area embedded gratings. Second-order distributed Bragg
reflector is part of the laser resonator and output coupler through
the surface.
2000
1800
1600 Slope efficiency = 42.9 W/A
1400 Threshold lasing current = 14.8 A
Figure 6. Diode laser marking by a two-dimensional array of lasers
1200 on a plastic plate.
Watts
1000
800
600
400
200
0
0 10 20 30 40 50 60 70
Amps
(c)
Figure 8. Diode laser soldering of electronic components in (a) and melting of solder in (b).
250 LASER DESKTOP MACHINING
exceeding 105 W/cm2. Figure 4(a) gives the schematic diagram tion can be very wide—for example, 15 애m ⫻ 200 애m instead
of such a rack system, and Fig. 4(b) is a photo of such a sys- of 1 애m ⫻ 3 애m as in edge-emitting Fabry–Perot laser stripe.
tem showing an unfocused laser beam. Figure 4(c) shows the This broader coherent source size provides a much higher
optical output power curve against the dc pump current. brightness source due to lower divergence and is capable of
Figure 5 shows the geometry of the next generation DL giving an intensity of 1 MW/cm2 or higher at a focused
that emits the light vertically through the wafer surface by spot. A two-dimensional array of such lasers will eventually
virtue of second-order Bragg gratings. The emission cross sec- replace most of the current lasers for laser material pro-
Figure 9. Diode laser welding, cutting, and cladding. (a) Welding of a razor blade to a steel saw
blade. (b) Cutting a razor blade. (c) Cladding a valve by stellite powder. (d) Laser-assisted
(heated) machining of SiN cylinder. [Courtesy of (a) Zediker of Nuvonyx, (b) SDL Inc., (c) Fraun-
hofer IWS, (d) Fraunhofer Institute of Production Technology.]
LASER DESKTOP MACHINING 251
Computer Laser
CAD/CAM Diode
Fiber output
Optic system
Working piece
X N2
He
O2
Y
Z Z
Vertical axis
Powder Atmospheric
X–Y stage handling controlled
controller system chamber
Figure 10. Diode laser surface hardening of a steel tape. (a) Micro-
graph of pearlite phase before laser hardening. (b) Micrograph of the
held, independently addressable, high-peak power diode
martensite phase after hardening by diode laser.
array can carry out a number of the commercial jobs. Figure
6 shows an elementary demonstration of marking (engraving)
plastic (Acrylonitrite butadiene styrene) with a fiber-coupled
cessing. Commercialization of such lasers is expected in the
laser array. The laser wavelength was 810 nm; power density
near future.
was 앑100 W/cm2.
(a)
(a)
(a)
Wavelength Effect. Absorptivity of metal powder increases Part Density with Particle Size. The density and the hard-
with shorter laser wavelengths. This is verified by using two ness of sintered parts are of critical importance if this method
different wavelengths at 810 nm and 980 nm. The two pic- is to become a commercially acceptable process for rapid pro-
tures in Fig. 14 compare the differences in sintering effects. totyping. Figure 15 shows the intuitively obvious assumption
The same 44 애m Fe powder was used in both cases under an that finer particles make denser and harder parts. Two parts
inert Ar atmosphere using the same total power of 15 W in a were made using identical computer-controlled processes
focused spot of 750 애m at a scan speed of 1 mm/s. The test from Fe-bronze powders of 44 애m and 150 애m particle sizes.
showed that, while 980 nm just sinters the powder, 810 nm The part density made of 44 애m powder is 80% [Fig. 15(b)]
sintered and also partially melted some particles due to with partial melting, while that made out of 150 애m powder
higher absorption. is only 50% [Fig. 15(b)] and sintered only. Laser power was
256 LASER DESKTOP MACHINING
15 CW (at 980 nm) focused to a spot size of 0.8 mm and scan Because of compactness and wavelength diversity, complex
rate of 1 mm/s. Crystollographic analysis of the microstruc- three-dimensional optoelectronic and other devices can be
ture showed that the miscibility of particles, of different met- fabricated on a single substrate by changing the gas composi-
als, in the sintered part is significantly better for 44 애m pow- tion inside a small chamber. Spatially distributed individual
der than for 150 애m powder. laser beams with different optical frequencies and beam ener-
gies can simultaneously or sequentially develop complex
Part Density with N2 and Ar Atmosphere. An appropriate gas multilayer optoelectronic devices comprising dielectric, metal-
inside the sintering chamber can be utilized to control the lic, and semiconductor materials.
part density. This was verified by fabricating sintered parts
under N2 atmosphere for Fe powder in contrast to the inert ACKNOWLEDGMENTS
Ar gas. Sintered parts grown under otherwise identical condi-
tions showed 80% density for N2 gas, due to nitride formation, The author would like to acknowledge Weiq-un Chen and
in contrast to 70% for inert Ar gas. The micrographs are Tariq Manzur for carrying out most of the experiments dem-
shown in Fig. 16(a) and Fig. 16(b), respectively. Low-melting onstrated in our institution. Acknowledgments for experimen-
Pb powder is another possibility of obtaining higher density tal results from outside institutions have been explicitly men-
and strength for laser prototype parts. We verified the antici- tioned in the figure captions. Dong-Ik Lee and Doug Bradway
pated result; but because of environmental risks, we would have put the manuscript together. Support from Connecticut
not recommend it for manufacturing. Innovations, Inc. of Connecticut is gratefully acknowledged.
The SiC growth experiment was carried out under DARPA
SOLID FREE FORM FROM GAS PHASE support received by Harris Marcus of the Institute of Material
(CHEMICAL VAPOR DEPOSITION) Science, University of Connecticut.
CHANDRASEKHAR ROYCHOUDHURI
University of Connecticut
Lamp voltage properly sized capacitor can be placed across the lamp so that
Mains voltage
Ballast voltage by resonance (with the inductive ballast) the voltage across
100 the lamp builds up to the ignition level. Once discharge is
established the lamp (arc) resistance loads the resonant cir-
cuit such that the voltage reaches the operating voltage of
the lamp.
50
0 0.2 0.4 0.6 A ballast that starts a fluorescent lamp in this manner de-
ILamp scribed is known as an instant-start circuit. The cathodes
(filaments) are not heated and the voltage required to ignite
Figure 1. Graph illustrating the need for a ballast to operate a dis-
the lamp is quite high because there is no thermionic emis-
charge lamp. The V–I curve of a compact fluorescent lamp (solid line)
sion from the cathode to help initiate the discharge. Such a
is overlaid with the load lines of the mains supply only (dotted line)
and mains supply with resistive ballast (dashed line). The lamp oper- starting circuit is detrimental to lamp life as the cathodes
ates at the stable intersection of its V–I characteristic and the bal- gradually lose the emission mix that lowers their work func-
lasted mains load line. If operated off the mains with no ballast there tion at every start. Among the starting methods that apply
is no intersection and the current would increase till either the lamp cathode heat prior to raising the lamp voltage to starting lev-
fails or the mains relays trip or fuses blow. els the main ones are known as rapid start and switch start
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
LIGHTING CONTROL 339
(1,3). In a switch-start implementation, a switch closes a cir- last, the efficacy increase provided an early impetus for elec-
cuit consisting of the two cathodes in series with the bal- tronic ballasts.
lasting inductor and the voltage supply. The current estab- In spite of these advantages, the penetration of electronic
lished in the circuit heats the cathodes. The switch is then ballasts had been slow principally because of increased costs
opened which results in the voltage across the inductor build- relative to a conventional ballast. Recently however, lamp
ing up until the lamp ignites. The starting voltage required is systems have been developed where electronic ballasts pro-
lower than in the instant-start case. vide capabilities that cannot be matched by electromagnetic
In a rapid-start implementation circuit the cathodes are ballasts. Some examples of these capabilities are arc straight-
preheated for about a second before ignition is initiated. The ening in HID lamp (11–14), dimming in fluorescent and HID
cathode heating current can be applied using additional wind- lamps (15–17), high-pressure sodium (HPS) lamp color im-
ings that are placed either on the isolation transformer of the provement by pulsing (18–20) and lightweight, small sized
ballast or on the ballasting inductor itself. The ballast is de- fluorescent lamp replacements for incandescent lamps (21–
signed so that the open circuit voltage (applied during the 23). In addition to the increased capability offered by electron-
cathode heating period) is insufficient for starting with a cold ics there are some systems that cannot be practically operated
cathode. However, once the cathode is heated the voltage is at the mains frequencies. Examples of these systems are the
sufficient to ignite the lamp. electrodeless fluorescent lamp (24–27), the electrodeless HID
Another commonly used approach for preheating the cath- lamp (24,28), and the microwave sulfur lamp (29). The geome-
odes for rapid-start lamps is placing a positive temperature try of these systems and the efficiency of coupling the electric
coefficient resistor (PTC) in series with the cathodes. At start energy into the discharge dictates the frequency of operation.
the impedance of the PTC is low, so a relatively large current Figure 2 shows the block diagram of an electronic ballast
flows through the cathodes and the PTC. This heats up the for a discharge lamp. The principal blocks will be found in
cathodes as well as the PTC, which becomes high resistance one form or another in every electronic ballast. The first block
after about 1 s. During this time the cathodes are heated to consists of an ac-to-dc rectifier which converts the line fre-
the thermionic emission temperature thus allowing for a quency voltage to a dc voltage, and also always includes an
smooth discharge to be set up. Note, even though this is a electromagnetic interference (EMI) filter. The EMI filter pre-
simple start scheme the PTC does consume some power which vents high-frequency noise generated in the ballast from be-
affects the overall efficiency by 3% to 4%. ing conducted back into the mains, in compliance with regula-
High-intensity discharge (HID) lamps go through several tory controls on emissions [e.g., Federal Communications
phases during starting: initial breakdown, glow discharge, Commission (FCC) in the United States and International
glow-to-arc transition, and thermionic discharge (3–7). Each Electrotechnical Commission (IEC) internationally]. In addi-
of these phases has its specific requirements that must be met tion to the basic ac-to-dc conversion functions, this block could
by the ballast in order to successfully start the lamp. The bal- include a power factor correction circuit or a dc bus voltage
last designer must understand the behavior of the lamp and control circuit. In the case of a microwave system this block
its impact on the circuit during each of the phases. For exam- would be the power supply for the magnetron.
ple, a ballast that does not supply sufficient current during The second block is a dc-to-ac high-frequency converter. In
the glow discharge phase may cause the lamp to remain in circuits operating at frequencies from the tens of kilohertz to
that phase and not transition to the arc mode. Operation in the tens of megahertz this usually consists of one or more
this mode is detrimental to lamp life as it causes the cathode switching power devices [bipolar transistors, metal-oxide-
to sputter off material which blackens the lamp walls and semiconductor field effect transistors (MOSFETs), insulated
eventually causes lamp failure. Starting can be enhanced by gate bipolar transistors (IGBTs), etc.] that chop the dc voltage
the application of starting pulses or the enhancement of UV into high-frequency pulses. This block includes the ballasting
radiation around the lamp by additional electrodes to promote element (inductor or capacitor) and auxiliary components for
the initial ionization for starting (7). starting. Starting components can be as simple as a resonant
capacitor, or as complicated as a separate high voltage sup-
ply. As in the rectifier block, control functions and feedback
ELECTRONIC BALLASTS FOR DISCHARGE LAMPS loops may also be included in this block. In a microwave sys-
tem this block would represent the magnetron.
The same principles and requirements applied to electromag- The third block is the coupling block that applies the high-
netic ballasts are applied to electronic ballasts. In essence, an frequency signal to the lamp. Typically with electroded lamps
electronic ballast simply consists of a frequency converter this is simply wires connecting the ballast to the lamp. How-
that takes the mains (line) frequency (50 or 60 Hz) and con- ever, in electrodeless lamps the coupling component is a very
verts it to a much higher frequency (tens of kHz to GHz). As important one, worthy of being defined as a separate block.
the impedance of an inductor is proportional to the frequency, Some examples of these are the coils used in the electrodeless
the inductance needed for ballasting is reduced accordingly. fluorescent lamps recently introduced by Phillips (26) and GE
A line-frequency electromagnetic ballast is therefore much (27), the drive coil for the electrodeless HID lamp described
larger and heavier than a high-frequency electronic ballast. by GE (28), and the microwave cavity used in the microwave
In recent years electronic ballasts have begun penetrating the sulfur lamp introduced by Fusion Systems (29).
lighting field at an increasing rate. It was found early on that The design of any ballast requires that the electrical engi-
operation of linear fluorescent lamps at frequencies above 10 neer understand the principal characteristics of the lamp as
kHz resulted in an increase in lamp efficacy by 10% to 15% a load as well as the requirements for properly starting and
(8–10). Coupled with a higher electronic ballast efficiency, its running a lamp. Traditionally, the lamp was treated as a
smaller size and weight relative to the electromagnetic bal- black box where only the terminal characteristics were impor-
340 LIGHTING CONTROL
v v v v
t t t t
Figure 2. Block diagram of a generic electronic ballast, showing the different voltage waveforms
at the inputs of the different blocks. All electronic ballasts contain these building blocks in one
form or another.
tant. However, the design of the coupling blocks in the new types). However, such a system is not feasible if the electron-
electrodeless systems requires a more intimate cooperation ics needed to operate the lamp at 13.56 MHz is either unavail-
between the lamp and ballast designers. The lamp/ballast able or too inefficient.
combination becomes more of an integral system that needs The field of power electronics, of which ballast engineering
to be designed as a single unit. In fact, decisions made by the is a specialty, has been advancing in the direction of increas-
lamp designer can have a significant impact on the ballast ing the efficiency of systems operated at ever higher frequen-
and vice versa. A process whereby lamp performance is opti- cies. Whereas 10 years ago operation at 50 kHz was consid-
mized without any regard to system impact may result in a ered an upper limit for higher efficiency, now systems are
system that is far from optimum. being built that operate at 2.5 MHz and higher without com-
promising the efficiency.
Component Technology for Electronic Ballasts
These advances have come as a result of both improve-
In general standard ballasts operate with efficiencies in the ments in the efficiency of components as well as advances in
85% to 93% range. Therefore, very little room exists for fur- circuit techniques. The improvements in components have
ther improvements in system efficacy to be obtained from bal- come in every area. In semiconductors, power MOSFETs have
lasts. However, insofar as the use of electronics enables types increasingly replaced bipolar transistors as MOSFET channel
of lamps to exist that would normally be impractical or impos- on-resistances have decreased and switching speeds in-
sible to make and allows sophisticated light control schemes creased. Similar improvements have occurred in power diodes
to be implemented, they can have an indirect positive impact with lower forward drops and reverse recovery times.
on the overall power consumption for lighting application. For Magnetic materials used in ballasts such as ferrites have
example, the electrodeless HID system (28) achieves lamp also improved over the years (31). Ferrites are now able to
and coupling system efficacies of 140 lm/W (for the purpose run at higher flux densities, higher temperatures, and higher
of comparison, Table 1 lists efficacy of other standard lamp frequencies. However, the improvements there have not been
as impressive as with the semiconductor devices. In fact as
Table 1. Typical Efficacy of Standard Lamp Types (30) the operating frequency gets up in the MHz range powdered
iron cores are used, and finally as the frequency gets higher
Wattage Efficacy
air core inductors gain the upper hand.
Lamp Type (W) Lumens (lm/W)
Ceramic capacitors have allowed operation at higher fre-
Incandescent 60 to 100 870 to 1750 14.5 to 17.5 quency with their low loss and extremely small size (32).
Fluorescent However, one component that has seen very little improve-
Warm White 40 3200 80 ment over the years has been the electrolytic capacitor. Elec-
Deluxe Warm White 40 2200 55
trolytic capacitors have, by far, the highest energy density per
Warm White 32 2800 82.35
(Watt Miser)
unit volume of all capacitors. They are used as the energy
Deluxe Warm White 32 1925 56.60 storage capacitor in the rectifier to allow the circuit to operate
(Watt Miser) during the zero voltage crossings of the mains voltage.
HID-Mercury Vapour However, electrolytic capacitors suffer from a very low
Deluxe White 400 22,500 56.25 temperature range of operation. Although rated for up to
HID-Metal Halide 105⬚C their life at this temperature is only 2000 h. The life of
Clear, ‘‘any position’’ 400 36,000 90 an electrolytic capacitor doubles for every 10⬚C temperature
HID-High Pressure reduction. If we take into account that a metal halide lamp
Sodium (Lucalox)
lasts 20,000 h, and that a ballast should last as long as two
Clear, mogul base 150 to 250 15,000 to 100 to 110
27,500
lamps, the electrolytic temperature must then not exceed
65⬚C. Given that electronic ballasts operate in ambient tem-
LIGHTING CONTROL 341
High frequency
Vd inverter
Rectifier
Figure 3. Circuit schematic of a conven-
EMI Filter Blocking tional electronic ballast (low power factor)
Control for fluorescent lamps. Each functional
capacitor Ballast
logic inductor Lamp Cathode heat block of Fig. 2 is illustrated. The EMI fil-
Main and vc
Cs circuit ter, rectifier and line filter capacitor form
voltage gate
Line filter drive Cd L Starter the ac-to-dc block. The control logic, high
Cp
electrolytic Resonant capacitor frequency inverter, load network (Cd, L,
capacitor C
capacitor Positive and C) and cathode heat circuit form the
R thermal dc-to-ac block. In this case, the ballast to
coefficient lamp coupling block consists simply of the
resistor wires connecting the ballast to the lamp.
342 LIGHTING CONTROL
Circuit layout and device packaging are circuit areas in which symmetrical waveform switching between ⫺Vd /2 and ⫹Vd /2
we can expect rapid improvements in the near future. The to the series resonant circuit consisting of the ballast induc-
use of finite element modeling to help with designing circuit tor, L, and resonant capacitor, C. The lamp, which is the load,
boards and device packages for minimal parasitics is begin- is connected across the resonant capacitor. The gating or
ning to have an impact on the frequency of operation of power switching frequency of the MOSFETs is the control parameter
circuits. Whereas a few years ago high-efficiency operation at for resonant ballasts to control starting or ignition of the
hundreds of kilohertz was the state of the art, converters op- lamp, to control lamp current crest factor against modulation
erating at tens of megahertz are being reported in the liter- caused by the dc bus voltage ripple and against component
ature. variations with temperature and life, and to maintain lamp
It would then seem that for ballasts operating at frequen- power regulation against changes in the mains voltage. Typi-
cies from the tens of kilohertz to the tens of megahertz we cally, lamp current feedback is used as a control signal to a
can only expect small incremental improvements in efficiency. voltage controlled oscillator (VCO) to vary the switching fre-
Most of the impact in years to come will be in the area of quency. All feedback, control, and drive electronics are repre-
reduction in cost and size coupled with an increase in the bal- sented as a functional block in Fig. 3.
last functionality as smart controllers using ICs become more Starting scenarios for rapid-start ballasts have been dis-
prevalent. Dimming functions, light control functions, and en- cussed earlier. Two typical implementations for preheating
ergy saving functions will be more easily implemented. Im- the filaments prior to ignition are as follows: (a) using a posi-
provements in circuits will be mostly directed toward raising tive temperature coefficient (PTC) resistor in series with the
the frequency at which circuits can operate with high effi- filaments of the lamp (Fig. 3), and (b) using auxiliary wind-
ciency. Ballasts operating at 100 MHz to 200 MHz will be- ings on the ballast inductor to provide a preheating cathode
come as common as ones operating at 2.5 MHz. voltage in the range of 3 V to 5 V as specified by ANSI.
With the circuit topology selected, design of the ballast in-
Design Methodology volves the selection of the control circuit and power devices,
the design of the resonant load network, and the ballast in-
The primary focus of this section is the operation and design ductor. The key inputs to the design process are: the input
methodology of the dc to high frequency ac inverter and load voltage, the output power, the lamp impedance (or VI charac-
network stage that interfaces with the discharge lamp, typical teristics), the starting voltage, and the input current require-
of most electronic ballasts available in the market today. Fur- ments. Electronic ballasts are always competing with the
thermore, there are two broad categories of electronic bal- electromagnetic ballasts which places a severe cost constraint
lasts—low-power factor and high-power factor—the differ- on the electronic ballast. Hence the cost of the ballast is also
ence being in the front-end line frequency ac to dc conversion a significant input and it drives the selection process. Fur-
stage. In the low-power factor front-end, the electrolytic ca- thermore, selection of the PTC for filament heating and its
pacitor (required for energy storage and mains filtering) is influence on starting will also be incorporated in the design
directly connected to the dc side of the mains voltage diode methodology.
bridge rectifier. The current drawn from the mains is zero To simplify the analysis of the ballast the following as-
except for the narrow spikes required near the crest of the sumptions are made (37,38):
mains voltage to refurbish the energy in the capacitor that is
subsequently consumed by the ballast and the lamp. This pro- 1. The MOSFET switches are ideal with zero on resis-
cess repeats once every half cycle of the mains frequency. This tance, infinite off resistance, and negligible output ca-
highly discontinuous and peaky mains current has a high pacitances.
harmonic content resulting in the low power factor, typically 2. The loaded quality factor of the resonant circuit is high
in the range of 0.6 to 0.7. Various high-power factor front- enough so that the currents through inductance, L, ca-
ends, discussed in detail in a later section, differ in their ac- pacitance, C, and load resistance, R, are sinusoidal.
tive mains current wave shaping schemes. 3. Operating frequency is fixed at the undamped natural
Figure 3 shows the circuit schematic of a typical high-fre- frequency of the resonant circuit.
quency electronic ballast, shown with a low-power factor
front-end for the sake of simplicity. The dc to high frequency The resonant circuit in Fig. 3 is a second-order low-pass filter
ac stage consists of a pair of n-channel power MOSFETs con- and can be described by the following parameters:
nected in a totem-pole configuration across the dc bus, Vd. The
electrolytic capacitor across the dc bus is assumed to be large 1 L R
f0 = √ Z0 = Q=
enough such that Vd has a relatively small ripple at twice the 2π LC C Z0
line frequency. Hence, Vd is approximately constant and equal
to the peak of the line voltage. where, f 0 is the undamped natural frequency, Z0 is the charac-
The MOSFET pair is gated on and off at a duty cycle teristic impedance, and Q is the loaded quality factor at f 0.
slightly less than 50% in a complementary manner to gener- Note that R represents the effective resistance seen by the
ate a trapezoidal wave across the lower MOSFET that resonant capacitor which is different under starting and run-
switches between zero and Vd. The less than 50% duty cycle ning conditions. During starting, R ⫽ RP, where, RP is the
allows a small dead time (approximately 1 애s to 2 애s) be- resistance of the PTC which increases as the PTC gets hotter.
tween the turn-off of one switch and the turn-on of the other Although the PTC is in series with the filaments, the filament
which is necessary for zero voltage switching and prevention resistance is small compared to the PTC hot resistance. Also,
of short-circuit across the dc bus. Cb is a dc blocking capacitor the lamp discharge resistance which appears in parallel with
that removes the dc bias on the trapezoidal wave to apply a the PTC resistor is near infinite before ignition.
LIGHTING CONTROL 343
Under steady-state discharge conditions, R ⫽ RL, where, Note, that the peak inductor currents during starting and
RL is the lamp resistance under nominal discharge conditions. running conditions are given by using the respective values
This is a ratio of the nominal voltage and current specified for 兩Z兩.
for the lamp. It is assumed here that the hot PTC resistance Given the value of the ballast inductor, L (calculated from
is 3 to 4 times bigger than the lamp discharge resistance, thus resonant circuit design methodology outlined previously),
having little influence on the effective load resistance, R. nominal operating frequency, f 0, starting and running peak
The magnitude of the fundamental component of the trape- inductor currents a simple iterative design process for the in-
zoidal voltage, vc is: ductor can be formulated. First, a suitable core magnetic ma-
terial (ferrites) and geometry (e.g. E-E, E-I) are selected. The
2 sin(φd ) following three equations are iterated to satisfy the peak flux
Vf = Vd (2) density specified by core manufacturer while arriving at a
πφd
reasonable number of turns, Nt, and the air gap, lg where Ac
is the core cross sectional area and IL (start) and IL (run) are
The following relevant quantities can be derived for the de-
the starting and running peak inductor currents, respectively.
sign and selection of the resonant components. All these
quantities are valid only at f ⫽ f 0, which is assumed to be the
fixed operating frequency. By neglecting the impedance of the The number of turns is:
dc-blocking capacitor Cb and using the previously defined pa-
rameters, the magnitude and phase of the input impedance of Llg
Nt = 104 (9)
the resonant circuit, respectively, are given as 0.4πAc
1 The starting and running peak flux density (in tesla) are:
|Z| = Z0 (3)
1 + Q2
0.4πNt IL (start) −4
1 Bm (start) = 10 (10)
φL = arctan (4) lg
Q
0.4πNt IL (run) −4
Bm (run) = 10 (11)
As stated earlier, for high-efficiency operation requiring ZVS lg
the circuit must operate in a lagging (inductive) mode. Opera-
tion at the undamped natural frequency ensures that the bal-
High-Power Factor Ballasts
last will always run in the lagging mode.
The magnitude of the voltage across the load (or resonant In the US market, fluorescent lamp ballasts have for years
capacitor, C) is been marketed as high-power factor ballasts. This standard
has not been enforced by any regulatory agency, but has be-
Vo = Vf Q (5) come a de-facto standard. The power factor definition in this
case is accepted to be greater than 0.9. In Europe and other
countries, the IEC (International Electrotechnical Commis-
where, Vo is the peak voltage at either starting or discharge
sion) has established a standard applicable to lamp ballasts
conditions depending on the value of R as defined. Hence,
known as IEC 61000-2-3 (39). This standard controls the
given the starting conditions, and using the definition for Q
power factor by specifying the maximum relative amplitude
and the load voltage Eq. (5),
of harmonics of the mains current up to the fortieth harmonic
(Table 2). The standard requires that any lamp ballast con-
Vf
Z0 = RP(hot) (6) suming more than 25 W has to be a high-power factor ballast.
Vs Electronic ballast engineers have devised many circuits to
meet power quality requirements. Some have been adapted
where, Vs is the specified lamp starting voltage and RP(hot) is from other power electronics disciplines while some others
the PTC resistance after 1-s preheat. For the given lamp char- have been specifically devised for lighting applications. The
acteristics and a selected PTC resistor Z0 can be calculated. need to balance the cost, input current, starting, running, and
Note, the selected PTC resistor must have a cold value low current crest factor requirements for the ballast has resulted
enough to allow sufficient filament current for proper pre- in the development of some very innovative circuits.
heating and prevent lamp ignition before thermionic emission
sets in, for rapid-start lamps. Selecting f 0 and knowing Z0, the Valley Fill Power Factor Correction Circuits. The first type of
resonant components L and C are calculated from the defini- power factor correction circuit is used primarily in the US
tions given previously. where the cost requirements are stringent and the power fac-
The magnitude of the ballast inductor current and load tor requirements are vague. These circuits are known as the
current, respectively, are given by valley fill circuits (40,41). The name refers to the dip that
occurs around the zero crossings of the voltage at the output
Vf 2Vd sin(φd ) of an unfiltered mains rectifier. As described earlier, the nar-
IL = = (7)
|Z| πφd |Z| row spiky nature of the mains current drawn by a low-power
factor electronic ballast has a high harmonic content. The val-
2Vd sin(φd )
IR = (8) ley fill circuits attempt to increase the conduction angle of the
φd Z 0 input current as a way of improving the input power factor
344 LIGHTING CONTROL
Table 2. IEC 61000-2-3 specifications and the performance of amount of lamp current feedback which serves to improve the
three high-power factor circuits. The modified valley fill and current crest factor. As the lamp current decreases in ampli-
multiresonant boost values are measured, while the tude the switching frequency also goes down which counter-
integrated boost values are calculated. acts the reduction in current amplitude. However, the gain of
IEC Modified these feedback circuits is typically very low. The self-oscillat-
61000-2-3 Valley Integrated Multiresonant ing circuit helps reduce the lamp current crest factor, but it
Harmonic Specification Fill Boost Boost does not eliminate the problem entirely.
fund 100 100 100 100 The power factor of the valley fill circuits is marginal. In
3 28 17 13 0.83 order to increase the power factor the conduction angle must
5 10 6.7 0.13 0.14 be increased. The modified valley fill circuit [Fig. 5(b)]
7 7 16 0.21 0.39 achieves this by charging the electrolytic capacitor to a value
9 5 15 0.14 0.45 lower than half the line voltage amplitude. The circuit is a
11 3 8.9 0.056 0.48 constant power circuit, thus the current is an inverse sine
13 3 3.1 0.36 0.49 wave. The power factor of such a circuit reaches a maximum
15 3 5.2 0.03 0.44
value of 0.961. This value occurs at a conduction angle of 135⬚
17 3 7.7 0.065 0.37
19 3 6.7 0.42 0.4
(i.e. the conduction occurs between 22.5⬚ and 157.5⬚ relative
21 3 3.6 0.033 0.34 to the line zero voltage crossing).
23 3 2.2 0.0064 0.34 None of the valley fill circuits described meets the require-
25 3 4.5 0.18 0.28 ments of IEC 61000-2-3. Table 2 shows the harmonic compo-
27 3 n/a 0.065 0.15 nents of the input current for the valley fill circuits compared
29 3 n/a 5.7 0.17 to the IEC requirements.
31 3 n/a 0.049 0.19
33 3 n/a 0.053 0.23
Boost Power Factor Correction Circuits. One of the circuits
35 3 n/a 5.7 0.25
that meets the IEC requirements has a discontinuous current
37 3 n/a 0.065 0.2
39 3 n/a 0.18 0.19 boost power factor correction circuit as the input circuit [Fig.
6(a)]. The average value of the high-frequency discontinuous
boost inductor current is proportional to the mains voltage.
The circuit senses the mains voltage and the inductor current
by charging the rectifier filter capacitor to some value smaller to ensure that the mains current mirrors the mains voltage.
than the peak mains voltage. The storage capacitor, Cs, voltage is higher than the peak of
The simplest of these circuit uses two capacitors and two the mains voltage. The circuit has been used in ballasts for
diodes configured in such a way that the capacitors are conventional fluorescent lamps in the US market (Motorola)
charged in series but discharged in parallel [Fig. 5(a)]. Thus, and may be used for high-intensity discharge lamp ballasts
the mains voltage supplies the current to the ballast for the which typically cost more than fluorescent lamp ballasts. For
duration when the mains voltage amplitude is greater than compact fluorescent lamp (CFL) this ballast is not typically
half the peak amplitude. During the rest of the period the used for the cost reason but also because of size restriction.
capacitors provide the voltage to the ballast. This results in a The circuit comprises two fully functional power electronic
power factor that is somewhat over 0.9. The variation in circuits that operate in series which makes it hard to fit in an
mains voltage causes the lamp current crest factor to be large. CFL ballast housing.
Lamp current feedback is used in these circuits to mini- Integrated Boost Circuit. The boost power factor correction
mize the crest factor. Self-oscillating circuits have a certain circuit can be modified by merging the boost inductor, Lb, into
Resonant
(a) Valley fill capacitor
circuit
(b)
Figure 5. Schematics of two simple high-power factor electronic ballasts with (a) two capacitor
valley fill power factor correction circuit, and (b) modified valley fill power factor correction cir-
cuit. These circuits are relatively easy to implement and are commonly used in the US, but do
not meet IEC standards for high-power factor.
LIGHTING CONTROL 345
Lb Db Vd Lb Db Vd
To To
L Cd L Cd
Boost lamp vd lamp
vi control vd vi Cs
Cs
circuit
C C
PFC current
sense resistor
(a) (b)
Figure 6. Schematics of electronics ballasts with (a) conventional boost power factor correction
circuit and (b) integrated boost power factor correction circuit. The first circuit can meet all high-
power factor requirements, but is considered too expensive and large for compact fluorescent
lamps. The second is a simplified lower cost version which has very good power factor perfor-
mance for the US market but fails to meet IEC standards for the European market.
the ballast circuit. In that circuit [Fig. 6(b)], the boost induc- The input power, Pi, to the ballast is equal to the lamp output
tor is connected to the midpoint between the two devices via power and the ballast losses. The equation for the input
a diode. Thus, the boost circuit power device and control cir- power to the ballast is combined with Eq. (14) to give the
cuit are all eliminated. The boost inductor has a 50% duty value of the boost inductor:
cycle applied to it at the switching frequency of the ballast.
The boost inductor current during one half cycle will flow 1
Lb = vi ii dt (15)
through one of the power devices (in the case shown it flows 4 f s Pi
in the top one). In the second half cycle the current flows
through the diode, Db, and into the capacitor, Cs. The current The design of the inductor follows a procedure similar to the
rises during this period with a slope equal to vi(t)/Lb, where one described for the ballasting inductor.
However, this circuit though attractive in terms of the cost
√
vi (t) = 2Vi sin(ωt) (12) and simplicity fails to meet the IEC requirements (Table 2).
Another problem is that the dc link voltage is twice the line
voltage. Since the rms line voltage in many countries that
and 웆 is the line angular frequency. The inductor is selected
follow the IEC requirements is 230 V, the dc link is at 650 V
to be at the critical discontinuous conduction level at the peak
which would require the use of power MOSFETs with a volt-
mains voltage, which means that when the mains voltage is
age rating of at least 800 V. These devices may be cost effec-
at its peak the inductor discharges exactly in the next half
tive for stand-alone, multilamp ballasts, but they are too ex-
switching cycle. As the peak current reached in the inductor
pensive for compact fluorescent ballasts. The integrated boost
during the charging half of the switching cycle is proportional
circuits are commonly used in the US because they have a
to the mains voltage during that time, this peak current will
significantly higher power factor and lower crest factor than
have a sinusoidal envelope. The volt-second balance on the
the valley fill circuits.
inductor operated in this manner causes the voltage across
Multiresonant Boost Circuit. In recent years, circuits ad-
the electrolytic capacitor to be equal to twice the peak line
dressing the particular needs of compact fluorescent lamp bal-
voltage. The discharge duration varies during mains cycle
lasts and the IEC specifications have been developed. One of
and is given by
these circuits is known as the multiresonant boost high-power
sin(ωt) factor circuit (42). The circuit consists of essentially a low
td (t) = (13) power factor circuit to which a boost capacitor and two diodes
2 f s [2 − sin(ωt)]
are added (Fig. 7). The dead time between the switching de-
vices is used to charge the boost capacitor from the mains and
The voltage across the electrolytic can be assumed for simplic- to discharge it into the electrolytic capacitor. The voltage to
ity to be constant at the peak value. As the mains voltage which the electrolytic capacitor is charged is determined by
varies, the peak current reached during each switching cycle the amount of energy pumped in during each cycle of the
will be equal to vi(t)/2Lb f s. Since the inductor is fixed and the switching frequency. If the voltage across the electrolytic ca-
switching frequency is fixed (to first order), the peak current pacitor is maintained at slightly above the mains voltage then
is proportional to the input line voltage. The filter at the input all the input current is drawn by the power factor correction
to the ballast averages out that current so that the line cur- (boost) capacitor. The charge pumped into Cb during each
rent is equal to the average of the boost inductor current. This switching cycle is proportional to the instantaneous mains
average value can be calculated as follows: voltage. Therefore, the filtered mains current is also propor-
tional to the mains voltage. Thus, the power factor will be
Vi sin(ωt)
ii (t) = (14) very high and the harmonic content will be very low. This
4 f s Lb 2 − sin(ωt) circuit is capable of meeting IEC 61000-2-3 requirements.
346 LIGHTING CONTROL
Cb Vc L
Because of zero-voltage switching the losses in the circuit are
vi(t)
conduction losses in the power devices and in the components.
Pout to
Cd IL lamp
Assuming an overall efficiency, , the total power is given by:
Cs C
Po
Pi = (21)
η
Alternately, the dimmer phase could be set so that the high- trodeless lamps. These lamps use inductive or capacitive cou-
est voltage applied to the ballast is much lower than the peak pling at radio frequencies (rf) and cavities at microwave
mains voltage. The ballast then operates at a dc bus voltage frequencies to apply the power to the plasma. The lamps can
that is much lower than the design limits. The lamp may then then be built without electrodes which results in several fea-
flicker severely and the restarts may stress the ballast compo- tures and advantages. First, new light-emitting materials can
nents to the point of catastrophic failure. be used as one does not have to be limited to chemicals that
Because of their sinusoidal input current, the high-power are compatible with the electrodes. Second, as a general rule,
factor circuits with high-quality input currents will not cause the electrode life and hence the lamp life is limited by the
the dimmer to fail. A carefully designed high-power factor cir- number of ignition cycles. Electrodeless lamps have been
cuit with sophisticated control circuitry can operate on a demonstrated to sustain a very large number of starts with-
phase control dimmer. However, no one has come up with a out failure. Finally, lamp designers can design lamps with
viable practical solution to this problem. All manufacturers of shapes that would not have been possible with electrodes. The
compact fluorescent lamps print warnings on their products basic limitation currently facing electrodeless lamps is the
against operating the lamps in sockets that are connected to cost of the electronic ballasts.
phase control dimmers.
Commercial users are interested in the conservation of en-
ergy as a means of cutting their electrical bills. Most commer- Microwave Lamps
cial locations already use high-efficiency lighting in the form Fusion Systems recently introduced a microwave lamp that
of fluorescent lamps in offices, high-intensity discharge lamps uses optical emissions from sulfur in the plasma to produce
in retail stores, warehouses, and factories, but they still need
light. The microwave system uses a device known as a mag-
further reductions in their lighting bill. Dimming controls,
netron to generate the microwave frequencies (43). The effi-
motion sensors, and architectural modifications in the work-
ciency of magnetrons in converting dc power to microwaves is
place are some of the ways of achieving these cost reductions.
in the range of 30% to 70%. Combining this with the conver-
Laptop computer screens have backlit LCD screens. The back-
sion efficiency of the ac-to-dc supply of the magnetron (85%
lighting is provided by cold-cathode fluorescent lamps. The
to 95%) results in at best a 67% efficiency from line cord to
ability to control the brightness of the screen by dimming the
microwave. Even if we assume the microwave cavity coupling
lamps is a standard feature. Therefore, demand exists for the
the energy into the lamp to be lossless, the relatively low effi-
ability to control the lighting level, especially in offices. For
ciency of the electronics has a detrimental impact on the over-
that purpose electronic dimming ballasts have been developed
all efficacy of a microwave system. Therefore, one would hope
for controlling fluorescent lamps.
The problem with designing a ballast for operation over a that improvements in the efficiency of these devices are forth-
wide range of output powers is mostly caused by the lamp’s coming, although the technology is quite mature. One possible
impedance characteristics. Electronic ballasts mostly use the approach for obtaining the same performance from the micro-
switching frequency as a means of controlling the output wave operated lamps while raising the system efficacy would
power. The frequency is raised to dim the lamp. The ballast be to operate the same lamps at frequencies in the tens of
load curve must remain above the lamp impedance curve or megahertz where high-efficiency ballasts have already been
the lamp will extinguish. Figure 8(a) shows the load curves demonstrated.
for a standard ballast with an LC circuit. At some frequency
the ballast load line falls below the lamp’s, which means it is Electrodeless Lamp Ballasts
extinguished. The available dimming range for this load cir-
cuit is given in Fig. 8(a). The addition of a series capacitor to In recent years, several lamp companies have introduced or
the lamp increases the dimming range significantly. Figure described electrodeless lamps operating at RF. In this section,
8(b) shows the load curves and the dimming range for that the design of the ballast for an inductively coupled elec-
circuit. trodeless lamp is described. In such a lamp the arc forms a
At very low dimming level, the discharge may begin to single turn secondary in a very loosely coupled transformer
show some instability, striations show up, and the light level (44–46). The arc has resistance and inductance, Ra and La,
may flicker. The application of a very low level dc current to respectively, as well as a coupling coefficient to the primary,
the lamp has been shown to reduce the instability of the or drive coil. In order to simplify the notations we will use the
plasma. Whereas, under normal operation the lamp current arc Q defined as: Qa ⫽ 웆La /Ra (웆 is the switching angular
is sufficient to maintain the temperature of the cathode, at frequency here) to represent the arc impedance. The reflected
very low dimming ranges this is not true anymore. In these resistance of the arc into the primary circuit is given by:
cases, additional cathode heat is applied to the lamp by an
auxiliary circuit as the lamp is dimmed. Thus, dimming Qa
ranges as large as 1000 : 1 have been achieved in practical Rr = k2 ωLc (24)
1 + Q2a
circuits (15).
V0(V) P0(W)
200 62.5 kHz 20
Io 65 kHz
Cd L
72.5 kHz 18
Lamp 16
100
Vo Lamp
C 14
12
0 0.2 0.4 60 65 70
I0(A) f(kHz)
(a)
V0(V) P0(W)
Io 200 62.5 kHz 30
L 65 kHz
72.5 kHz
Cs 20
100 Lamp
Cp V Lamp
o 10
0
0 0.2 0.4 60 65 70
I0(A) f(kHz)
(b)
Figure 8. The choice of load network can have a significant impact on the dimming range avail-
able for a given lamp. The dimming characteristics of two ballast load circuits showing the cir-
cuit, a chart of the lamp load line and the ballast load lines at 62.5 (solid), 65 (dashed), and 72.5
kHz (large dashes), and a chart of the dimming range. (a) A standard ballast circuit with a
dimming range of 18 W to 14 W. The lamp extinguishes above 70 kHz (the 72.5 kHz load line is
below the lamp load line). (b) A ballast circuit with the added series capacitor. The dimming
range is from 30 W to 0.2 W.
The loaded impedance of the drive coil is: pling efficiency is the frequency of operation. This frequency
is thus selected with a view to optimizing the coupling effi-
ZL = Rc + Rr + j(ωLc + Xr ) (26) ciency, while meeting regulatory requirements on electromag-
netic interference (EMI).
where RC is the series loss resistance of the drive coil. An electrodeless fluorescent lamp using a ferrite core drive
As the reflected resistance of the arc appears in series with coil reaches acceptable efficiency levels at frequencies above 1
the loss resistance of the drive coil, the same current flows in MHz. For these lamps, the frequency of operation is chosen
the two resistors. The power dissipated in the reflected resis- to be in the middle of a relaxed emission level area in the
tance represents the arc power, and that dissipated in the CISPR 15 regulations between 2.2 MHz and 3 MHz. There-
series coil resistance represents coil losses. Therefore, the fore operating in the middle of that frequency window makes
coupling efficiency is a function of the ratio of the two resis- it possible for the system to meet the EMI regulations in a
tances: the larger Rr is relative to RC the higher the efficiency. cost-effective manner. In the case of an air-core-driven elec-
The coil resistance, Rc, is minimized during the design of the trodeless HID lamp the optimum frequency for efficiency and
drive coil as allowed by constraints on the coil (e.g., the maxi- EMI consideration is 13.56 MHz which is an allowed fre-
mum outer diameter) placed by the requirements of the lamp. quency for industrial, scientific, and medical applications.
Therefore, Rr must also be maximized. Examining the expres- The circuit for an electrodeless lamp ballast is very similar
sion for Rr we see that it is a function of the coupling coeffi- to a standard ballast (Fig. 9). The higher frequency of opera-
cient, the coil inductance, the frequency, and a factor depen- tion makes lossless switching a key requirement. The shorter
dent only on the arc impedance. The coupling coefficient and switching period requires the designer to consider ZVS from
the coil inductance are primarily dependent on the geometry the very beginning. The current amplitude and phase must
of the system. Therefore, decisions made by the lamp designer meet the power requirement of the lamp, Po, and the losses in
have a significant impact on the coil designer and vice versa. the ballast (including the coil losses), Pl, as follows:
The lamp/coil system has to be optimized for maximum effi-
2(Po + Pl )
cacy as a unit, and cannot be done separately. IL = (27)
The least-restricted parameter in the expression for cou- Vf cos(φL )
LIGHTING CONTROL 349
Vd
Inductive
coupling
k Figure 9. Electrodeless discharge lamps are among the ex-
citing new developments in lighting technology. The ballast
C Rc
for an inductively coupled electrodeless fluorescent lamp is
vi shown. The ac-to-dc and dc-to-ac blocks are similar to the
Drive vc
circuit IL generic ballast of Fig. 3. The load network consists of the
Lc La Ra resonant capacitor and the drive coil. The ballast to lamp
coupling block also includes the drive coil (with an air core
or ferrite core) which acts as the primary of a loosely cou-
Arc pled transformer and the discharge which is ring shaped
and acts as a single turn secondary. The equivalent circuit
Drive of the arc is shown in the figure. The drive coil also acts as
coil the starter for the lamp.
19. D. E. Brabham, Improved color HPS lamps: Systems considera- 43. G. K. Farney, Crossed field tubes, in D. G. Fink and D. Christian-
tions and a comparison of commercial lamps, J. Illum. Eng. Soc., sen (eds.), Electronics Engineers’ Handbook, McGraw-Hill, 1989,
19 (1): 110–117, 1990. pp. 9-52–9-66.
20. D. M. Rutan, L. Nerone, and D. Kachmarik, Practical application 44. L. Nerone and A. H. Qureshi, Mathematical modeling and optimi-
of pulsed high pressure sodium lamp technology, J. Illum. Eng. zation of the electrodeless, low-pressure, discharge system, Proc.
Soc., 21 (1): 60–68, 1992. IEEE Power Electron. Specialists Conf., 1993, pp. 509–514.
21. A. Bouwknegt, Compact fluorescent lamps, J. Illum. Eng. Soc., 45. R. B. Piejak, V. A. Godyak, and B. M. Alexandrovich, A simple
11 (2): 204–211, 1982. analysis of an inductive rf discharge, Plasma Sources Sci. Tech-
22. E. E. Hammer and L. Nerone, Performance characteristics of an nol., 1 (3): 179–186, 1992.
integrally ballasted 20-W fluorescent quad lamp, J. Illum. Eng. 46. D. O. Wharmby and S. A. El-Hamamsy, Low power compact elec-
Soc., 22 (2): 183–190, 1993. trodeless lamps, Proc. 7th Int. Symp. Sci. Technol. Light Sources,
23. J. J. de Groot and H. Houkes, Triple-U electronic compact fluo- Kyoto, Japan, August 1995, pp. 27–36.
rescent lamps with reduced harmonics, J. Illum. Eng. Soc., 23
(1): 45–51, 1994. S. A. EL-HAMAMSY
24. D. Wharmby, Electrodeless lamps for lighting: A review, IEE M. H. KHERALUWALA
Proc.-A, 140 (6): 465–473, 1993. General Electric
25. K. Shinomiya et al., Development of the electrodeless fluorescent
lamp, J. Illum. Eng. Soc., 20 (1): 44–49, 1991.
26. J. Schlejen, Inductively coupled fluorescent lamps: The QL light- LIGHTING, INDUSTRIAL. See INDUSTRIAL LIGHTING.
ing system, in L. Bartha and F. J. Kedves (eds.): Proc. 6th Int.
Symp. Sci. Technol. Light Sources, Budapest: Tech. Univ. Buda-
LIGHTING TESTING. See IMPULSE TESTING.
pest, 1992. LIGHT METERS. See PHOTOMETERS.
27. Genura娃 R80 Lamp, GE Brochure, April 1994. LIGHTNING. See LIGHTNING, LIGHTNING PROTECTION AND
28. C. H. Stewart et al., Inductively coupled HID lighting system, in TEST STANDARDS.
L. Bartha and F. J. Kedves (eds.), Proc. 6th Int. Symp. Sci. Tech- LIGHTNING GROUNDING. See GROUNDING.
nol. Light Sources, Budapest: Tech. Univ. Budapest, 1992.
29. R. Dolan, M. G. Ury, and C. H. Wood, A novel high efficacy micro-
wave powered light source, in L. Bartha and F. J. Kedves (eds.),
Proc. 6th Int. Symp. Sci. Technol. Light Sources, Budapest: Tech.
Univ. Budapest, 1992.
30. D. R.. Lide (ed.), CRC Handbook of Chemistry and Physics, 72nd
ed., Boca Raton, FL: CRC Press, 1991–1992, pp. 10-286–10-287.
31. H. Domingos, Capacitors, in C. Belove (ed.), Handbook of Modern
Electronics and Electrical Engineering, New York: Wiley-Intersci-
ence, 1986, pp. 228–244.
32. A. Goldman, Magnetic materials, in C. Belove (ed.), Handbook of
Modern Electronics and Electrical Engineering, New York: Wiley-
Interscience, 1986, pp. 151–162.
33. M. Jordan and J. A. O’Connor, Resonant fluorescent lamp con-
verter provides efficient and compact solution, Proc. IEEE Appl.
Power Electron. Conf., pp. 424–431, 1993.
34. P. Wood, Electronic ballasts using the cost-saving IR2155 driver,
Application Note No. AN-995, International Rectifier.
35. S. J. Freeland, An introduction to the principles and features of
resonant power conversion, in K. K. Sum (ed.), Recent Develop-
ments in Resonant Power Conversion, Intertec Communications
Inc., 1988, pp. 20–43.
36. S. A. El-Hamamsy, Design of high-efficiency RF class-D power
amplifier, IEEE Trans. Power Electron., 9: 297–308, 1994.
37. R. L. Steigerwald, A comparison of half-bridge resonant converter
topologies, IEEE Trans. Ind. Electron., 35: 174–182, 1988.
38. M. K. Kazimierczuk and W. Szaraniec, Electronic ballast for flu-
orescent lamps, IEEE Trans. Power Electron., 8: 386–395, 1993.
39. R. Redl, P. Tenti, and J. D. Van Wyk, Power electronics’ polluting
effects, IEEE Spectrum, 34 (5): 32–39, 1997.
40. R. R. Verderber, O. C. Morse, and W. R. Alling, Harmonics from
compact fluorescent lamps, IEEE Trans. Ind. Appl., 29: 670–
674, 1993.
41. M. H. Kheraluwala and S. A. El-Hamamsy, Modified valley fill
high power factor circuit, Proc. IEEE Power Electron. Specialists
Conf., 1995, pp. 10–14.
42. M. H. Kheraluwala and S. A. El-Hamamsy, Multi-resonant boost
high power factor ballast, Proc. IEEE Power Electron. Specialists
Conf., 1997, pp. 60–65.
MANUFACTURING PROCESSES 371
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
372 MANUFACTURING PROCESSES
Controlling factor
Design
specifications
and tolerance limits
Manufacturing process
Precision
grinding
Figure 1. Representation of a manufac- machine
turing operation.
flows from cathode to anode and this effect is advantageously as shown in Fig. 3. In a conventional operation, the motion of
utilized to remove burrs or to drill a hole. It is estimated that grinding grits is tangential to the work surface. This produces
10,000 A of current remove steel at the rate of 1 in.3 per a cutting type of miniature chips. Sometimes this process is
minute. also called ultrasonic grinding (USG). Ultrasonic welding
(USW) is another manufacturing process that utilizes the
same principle for welding.
Ultrasonic Machining
In ultrasonic machining (USM) a transducer is used in con- Powder Metallurgy
junction with a tool that vibrates at low amplitudes but at A crude form of powder metallurgy may have existed in Egypt
high frequencies (25 kHz, for example). The principle is to as early as 3000 B.C. This process gained popularity during
remove material by microchipping and erosion. This is facili- the late nineteenth century. Powder metallurgy is a manufac-
tated by an abrasive slurry contained between the workpiece turing process wherein finely-powdered materials are blended
and tool. The vibration of the tip of the tool results in im- and pressed into a desired shape by a process known as com-
parting a very high velocity to the fine abrasive grains. Thus, pacting. Then the compacted mass is heated at a controlled
the motion of the grinding grits is normal to the work surface, temperature to bond the contacting surfaces of the particles,
;;;
a process called sintering. Thus the final product is manufac-
tured in the shape required. In addition, it also possesses the
desired properties and characteristics. Further, the product
often needs no machining or finishing. There is almost no
Servo-
motor wastage of material. Porosity and permeability of the product
or are easily controlled. To be cost effective, powder metallurgy
;;;
control (abbreviated P/M) requires high production volume. Quality-
Moveable grade metal powders, precision punches, compacting dies, and
electrode specialized sintering equipment are all very expensive, and
R
E C Tool
negative
;;;
Power
Transducer supply
Figure 2. In the EDM process a powerful spark erodes the workpiece Figure 3. In ultrasonic grinding, motion of the grinding grits is per-
to a desired configuration. pendicular to the surface of the workpiece.
MANUFACTURING PROCESSES 373
these contribute to high costs. Nevertheless, casting by tradi- the workpiece and generate heat which accomplishes the de-
tional means may not be feasible for some high temperature sired manufacturing operation, say, for example, drilling a
alloys. Forging or hot extrusion, on the other hand, may re- hole or cutting a pattern in a precious metal. Caution must be
sult in poor tolerances and cause unnecessary die wear. Ma- exercised while using EBM because electrons, high voltages,
chining obviously generates waste material during pro- vacuum, and metal surface all combine to generate hazardous
cessing. Powder metallurgy produces a wide range and X rays. Higher material removal rates (compared with EDM
variety of goods of diverse shape and size with good and ac- or LBM) are achieved by plasma-arc cutting (PAC). In this
ceptable tolerance limits. Modern manufacturing methods case ionized gases (plasma beams) are used which are partic-
utilizing highly automated equipment produce several types ularly useful when cutting materials like stainless steel, and
of consumer items, toys, and automotive parts by the million. where very high temperatures (17,000⬚F) are needed.
Because labor cost per part is low, powder metallurgy offers
a viable alternative and is often preferred. Powder metallurgy Injection Molding
is also preferred in cases where parts are produced in small
quantities. Stringent specifications, tolerances, and desired This type of fabrication is very popular for manufacturing
metallurgical properties sometimes dictate the use of powder complex-shaped plastic components. This method is similar to
metallurgy. Small parts made from nickel-based super alloys, ‘‘die casting.’’ In both, molten thermoplastic resin or some low
beryllium processing, certain types of self-lubricating bear- melting point alloy is injected into a die. Then it is allowed to
ings, and metallic filters are some examples where powder cool and harden. Modern day industry uses a wide variety of
metallurgy is applied. plastics and polymers. The idea implemented is to convert the
plastic raw material directly into a finished product in a sin-
Laser Beam Machining (LBM), Welding, and Cutting gle operation. The process selected for fabricating these mod-
ern day plastics depends mainly on one criterion, whether the
The principle used here is focusing the high-density energy of polymer is thermoplastic or thermosetting. Thermoplastic
a laser (Light Amplification by Stimulated Emission of Radia- resins and polymers can be heated to a fluid state, so that
tion) beam to melt and evaporate portions of the workpiece in they can be poured into a die or injected into a mold. In the
a controlled manner. A schematic representation is shown in case of thermosetting polymers, the polymerization process
Fig. 4. Extreme caution has to be exercised while using lasers and the shape-forming process are achieved simultaneously
because they can cause permanent retinal damage to the because, once the polymerization has taken place, no further
eyes. LBM is widely used in electronics and automotive parts deformation is possible. Some of the methods available are
manufacturing where precision drilling of holes (0.005 mm or casting, extrusion, thermoforming, etc. In addition, a variety
0.0002 in.) is required. Reflectivity and thermal conductivity of molding techniques are extensively used with plastics and
of the workpiece surface play a major role in LBM. Excimer polymers. Injection molding, reaction injection molding
lasers are very popular for drilling holes and marking plastics (RIM), rotational molding, foam molding, transfer molding,
and ceramics. Pulsed carbon dioxide lasers are commonly cold molding, compression molding, and hot-compression
used for cutting ceramics and metals. Neodymium : Yttrium- molding are some of the manufacturing processes most com-
aluminum-garnet (Nd : YAG) lasers and ruby lasers are used monly used by the plastics industry.
for welding metals.
Unlike laser beam machining, electron beam machining re- Space age technologies and the computer revolution have re-
quires a vacuum. Dc voltages as high as 200 kV are used to quired the manufacturing industry to produce prototypes of
accelerate electrons to speeds comparable to the speed of parts and components economically at a faster pace. This has
light. These high-speed electrons impinge on the surface of resulted in the development of rapid prototyping techniques,
also called desktop manufacturing or free-form fabrication.
The idea is to manufacture an initial full-scale model of a
product. The part is made directly from a three-dimensional
CAD drawing. One of the methods is called stereolithography,
a process based on curing and hardening a photocurable liq-
Laser uid polymer to the desired shape, using an ultraviolet laser
source. Some of this equipment costs as much as half-a-mil-
lion dollars. However, in many instances, this method is
Lens much cheaper than conventional prototyping, and the manu-
facturing industry has quickly recognized the importance and
Laser beam economic impact of these new technologies. Some of the other
focuses optical
energy on the surface techniques used are selective laser sintering, three-dimen-
sional printing, ballistic particle manufacturing, photochemi-
Workpiece cal etching, and laminated object manufacturing. Almost all
of these methods use CAD systems, and many cost in the re-
gion of hundreds of thousands of dollars. Some metals are
Figure 4. The principle behind the laser-beam-machining (LBM) used, but rapid prototyping with plastics and polymers, such
process. The reflectivity and thermal conductivity of the workpiece as polystyrene, epoxy, polyester, PVC, and nylon, is more
influence LBM effectiveness. common.
374 MANUFACTURING PROCESSES
Tool holder a wide selection depending on the application and needs. Un-
der the category of welding one can list shielded metal arc
Workpiece welding, submerged arc welding, gas metal arc welding, flux-
Tool
Work table
cored arc welding, electrogas welding, electrosag welding, gas
Drive motor tungsten arc welding, plasma-arc welding, laser-beam weld-
;;;;;;;;;;;;;;;;;;;;;;;;; ing, electron-beam welding, inertia friction welding, linear
;;;;;;;;;;;;;;;;;;;;;;;;; friction welding, resistance spot welding, resistance seam
Lead screw Feedback
welding, resistance projection welding, flash butt welding,
sensor
stud arc welding, percussion welding, explosion welding, and
Signal to Feedback diffusion welding.
Machine
drive motor path signal
control unit, In many cases welding may not be the proper choice. For
computer example, alloys containing zinc or copper are considered un-
weldable. Aluminum alloys are weldable only at a very high
Figure 5. Improved quality and reduced manufacturing time are temperature. Brazing and soldering processes use much lower
some of the advantages of using computers and closed-loop feedback
temperatures compared to welding. Further, soldering tem-
systems.
peratures are lower than those used for brazing. Brazing is a
joining operation wherein a filler material is placed between
the surfaces to be joined and the temperature is raised to melt
Assembly or Joining and Fastening Processes the filler material but not the workpieces. As such, a brazed
In most cases, this joining process is inevitable because the joint possesses higher strength. It is believed that brazing
product cannot be manufactured in one single piece or one dates as far back as 3000 B.C. Brazing methods are identified
single operation. For example, a pressure cooker lid may be by the various heating methods employed. Torch brazing, fur-
made from aluminum, but it has a plastic handle. In addition, nace brazing, induction brazing, resistance brazing, dip braz-
the replaceable sealing ring is made from rubber. In other ing, infrared brazing, and diffusion brazing are noteworthy.
words, selected products may have to be replaced frequently, Brazing is conducted at relatively high temperatures. For ex-
according to a routine maintenance schedule. In some cases ample, stainless steel and nickel-copper alloys need high
it might be more economical to manufacture, transport, and brazing temperatures on the order of 1120⬚C. At the other
assemble individual components at the customer’s site. The extreme, titanium can be brazed at 730⬚C, using silver alloys.
functionality of different components may dictate that the de- Soldering is similar to brazing but requires lower tempera-
sired properties be different. Besides traditional mechanical tures. In this case, the filler material melts below 450⬚C and
fastening, a variety of joining and fastening processes are again, as in brazing, the base metal does not melt. A general
available. If the material is ‘‘weldable,’’ then the engineer has purpose soldering alloy widely used in electronics assembly
ning is labor-intensive and time-consuming because the ‘‘pro- an efficient and clean cutting operation. Pressures ranging
cess planner’’ has to selectively choose the methods and se- between 500 and 1200 MPa (1 Pascal ⫽ 1 Newton/meter2 and
quences required for the production and assembly operations. 1 pound per square inch ⫽ 6891 Pa) are used to direct a jet
The planner also selects the necessary machine tools, fixtures, of water to act like a saw. Water-jet machining, which is also
and dies. This tedious task is made simple by using computer- called hydrodynamic machinging, can be very conveniently
aided process planning (CAPP), a powerful tool that views the used to effectively cut plastics and composites. The food pro-
complete manufacturing operation as one integrated system. cessing industry uses. WJM for slicing a variety of food prod-
There are two types of CAPP systems, the derivative system ucts. Whether it is a strong and solid material like brick or
(wherein the idea is to follow a standard process stored in the wood, or a soft and flexible material such as vinyl or foam,
computer files) and the generative system (wherein the process hydrodynamic machining offers the engineer an advanta-
is automatically generated based on some sort of ‘‘logic’’). geous choice for the selected manufacturing operation, be-
CAPP obviously requires expensive, sophisticated software cause WJM eliminates the need for certain requirements,
that works appropriately with CAD/CAM systems. Some of such as, for example, pre-drilled holes. A water-jet cutting
the benefits include reduced planning costs, decreased ‘‘lead machine can be seen in Fig. 6.
times,’’ and improved product quality. Computers have helped Robots have made a significant impact on the manufactur-
in inventory management and other areas. Group technology ing shop floor, relieving humans from dull, dirty, and danger-
(GT), cellular material-requirements planning (MRP), manu- ous environments. They have been manufacturing high qual-
facturing resource planning (MRP-II) are some of the areas ity goods with minimal waste and at reduced costs. Robots
destined to gain wider acceptance and usage during the are continuing to play a dominant role in streamlining several
twenty-first century. Coordinate measuring machines (CMM), manufacturing processes. An example of a gantry robot in-
lasers, vision systems, ultrasonics, and other noncontact mea- stallation is shown in Fig. 7.
surement techniques are helping to streamline inspection and Robots have helped the electronics manufacturing industry
quality control. in a variety of ways. An example of a Robot being used in a
Programmable automation has several advantages. Some semiconductor manufacturing processes is shown in Fig. 8.
are listed here:
BIBLIOGRAPHY
MYSORE NARAYANAN
Miami University
POSTAL SERVICES 559
POSTAL SERVICES
On July 26, 1775, members of the Second Continental Con-
gress, meeting in Philadelphia, agreed . . . that a Postmas-
ter General be appointed for the United States, who shall
hold his office at Philadelphia, and shall be allowed a salary
of 1,000 dollars per annum . . ..
The final two digits designated small post offices or postal mail in walk sequence for the mail person to effect efficient
zones in larger zoned cities. Thus a ZIP code of 48323 would delivery.
identify the West Bloomfield Township in the state of Michi- Although MLOCRs have been deployed throughout the
gan. In fact the first two digits (48) would identify Michigan. United States, some formidable challenges remain. For exam-
An important milestone occurred in November 1965 when ple, the same MLOCR machine that is deployed across the
the Postal Service installed a high-speed optical character nation has a performance that ranges from a low 35% ZIP⫹4
reader (OCR) in the Detroit Post Office. This first-generation encoding for places like Queens, NY (which has addresses
machine read the city/state/ZIP code line of typed addresses with hyphens and numeral street names), to a high 65% in
to sort letters to one of the 277 pockets. This automation led places like San Diego (which mainly has long street names
to increased productivity. In order to offset rising costs associ- and a limited number of high-rise buildings). Today, the main
ated with growing mail volume and to reduce the number of hurdle in total automation is the inability of MLOCRs to han-
mail piece processing, the Postal Service developed an ex- dle the handwritten addresses that constitute nearly 15% of
panded ZIP code in 1978. all letter pieces (amounting to several million pieces) handled
by USPS.
ZIPⴙ4. The ZIP⫹4 code, which was introduced in 1983, The USPS address database has evolved over the years
added a hyphen and four additional digits to the existing five- and has advanced technologies in areas of data gathering,
digit ZIP code. The first five numbers continued to identify an data storage, data validation, and database maintenance.
area of the country and delivery office to which mail is di- Driven by the need for automation, the level of details found
rected. The sixth and seventh numbers denote a delivery sec- in the database started with a few delivery points in a city to
tor, which may be several blocks, a group of streets, a group every single delivery point in the United States. That is quite
of post office boxes, several office buildings, a single high-rise an achievement.
office building, a large apartment building, or a small geo-
graphic area. The last two numbers denote a delivery seg-
ment, which might be one floor of an office building, one side MACHINE RECOGNITION OF HANDWRITTEN ADDRESSES
of a street between intersecting streets, specific departments
in a firm, or a group of post office boxes. The process by which people recognize handwritten charac-
The age of automation was ushered in in September 1982 ters, words, and documents has been the subject of intense
when the first computer-driven single-line OCR was intalled interest and investigation by researchers from very diverse
in Los Angeles. The equipment utilized OCR to read the letter fields. A good understanding of the mechanism of human rec-
and print a barcode on the envelope. At the destination post ognition of handwritten documents will have a significant im-
office, a less expensive barcode sorter (BCS) sorted the mail pact on the development of machines capable of recognition
by reading its barcode. By the end of 1984, 252 OCRs capable and interpretation of handwritten documents. However, the
of processing 24,000 pieces of mail per hour were installed in human recognition process is quite complex, and it incorpo-
118 major processing centers across the country with an aver- rates information extracted at different levels: characters,
age productivity rate of 6200 pieces per work hour. This was whole words, key words, and contextual processing. The effi-
a substantial increase when compared to the 1750 pieces per ciency of human recognition of handwritting can be attributed
work hour processed previously. Currently, USPS has de- to the effective integration of multiple cues and exploitation
ployed more than 800 multiline optical character readers of redundancies contained in most documents. However, if the
(MLOCRs) that can read addresses at 40,000 letter pieces per goal of this study is to develop machines that are capable of
hour and assign corresponding ZIP⫹4 codes. Of the letters automatic transcription of handwritten documents, then one
currently fed to the MLOCRs, 15% have handwritten ad- must recognize the immense difficulty of adopting the human
dresses and these are mostly rejected by the reader. Today, a recognition process.
new generation of equipment is changing the way mail flows In this article the primary focus will be on the development
and improving productivity. MLOCRs read the entire address of practical approaches to handwriting recognition. The word
on an envelope, print a barcode on the envelope, and then sort ‘‘document’’ is used in a very general sense. Thus, a document
it at the rate of more than nine per second. Wide area barcode will include characters, words, phrases, sentences, and whole
readers can read a barcode located anywhere on a letter. Ad- paragraphs. There are two main approaches to handwriting
vanced facer-canceler systems face, cancel, and sort mail. The recognition: (1) techniques based on holistic approaches
remote barcoding system (RBCS) provides barcodes for hand- whereby an entire word or a character string is recognized as
written script mail or mail that cannot be read by OCRS. a unit and (2) techniques based on extraction and recognition
The ZIP⫹4 code has reduced the number of times that a of characters (also referred to as segmentation-recognition ap-
piece of mail needs to be handled and has shortened the time proach) contained in a word or a string. Due to the focus on
carriers spend casing their mail (placing it in order of deliv- practical approaches, this article will present an in-depth
ery). The delivery point barcode, which represents an 11-digit overview of recognition techniques based on segmentation-
ZIP code, will eliminate the need for carriers to sort mail be- recognition.
cause mail will arrive in trays at the delivery post office This article will be organized as follows:
sorted in ‘‘walk sequence.’’ The MLOCR reads the barcode
and address, then imprints a unique 11-digit delivery point 1. Address interpretation system—an overview
barcode using the Postal Service’s National Directory and the
2. Taxonomy
last two digits of the street address. The 11-digit code consists
of ZIP⫹4 code with two additional digits that uniquely identi- 3. Image normalization processes
fies the addressee. The barcode sorters essentially sort the 4. Image presegmentation
POSTAL SERVICES 561
5. Context-free recognition of primitives and concatena- point code determination. Figure 2(a) illustrates the proce-
tion of primitives dure for address interpretation, and Fig. 2(b) shows a typical
6. Lexicon-driven recognition based on word matching result. Address interpretation requires the determination of
the correct ZIP⫹4 code by analyzing the image of a given ad-
7. Case studies
dress image. Several preprocessing steps are implemented
prior to the critical task of address recognition. The first task
is the determination of the destination address block (AB). In
ADDRESS INTERPRETATION SYSTEM—AN OVERVIEW this process, one utilizes the fact that destination addresses
are generally found in the southeast section of a flat mail
In this section, the authors present an overview of an inte- piece. By analyzing this region of the image, one may reliably
grated handwritten address recognition system that requires extract the destination AB. The second preprocessing subsys-
detection and recognition of ZIP code field, city/state field, tem applies tilt correction, line segmentation, slant correc-
street number field, street name/PO box field, and finally the tion, and word presegmentation.
correct nine-digit ZIP code. The interesting feature of this The first step in address interpretation is the detection and
study is the lack of any a priori information about the nature recognition of the ZIP code field. Again, one uses the common
of the address. Addresses may contain a PO box and/or street practice of writing the city, state, and the ZIP code in the last
number/name fields. The integrated system is required to de- line of the address. The ZIP code line recognition subsystem
termine the type of field present and determine the nine-digit generates several ranked ZIP code candidates.
ZIP code. Figure 1 illustrates images of handwritten ad- The street line recognition subsystem generates several
dresses that contain street number and name and/or PO box ranked pairs of street numbers and street names for given
designation. five-digit ZIP code. If the top candidate pair is accepted with
Three ZIP code directories are used to generate lists of sufficient confidence, it is sent to the DPC determination sub-
cities, states, and streets. The five-digit ZIP code directory system together with the five-digit ZIP code. The PO box line
consists of about 100,000 records containing information recognition subsystem generates several ranked PO box num-
about the five-digit ZIP codes of all cities in the United States. bers for given five-digit ZIP code. If the top candidate is ac-
The ZIP⫹4 directory consists of about 26 million records con- cepted with sufficient confidence, it is sent to the DPC deter-
taining information about every street, street number range, mination subsystem with the five-digit ZIP code.
and PO box numbers in the United States. The third direc- If the top candidate is a five-digit unique ZIP code with
tory, called the delivery point code (DPC) directory, has over sufficient confidence, it is encoded directly in the DPC deter-
100 million records that virtually locates any valid address in mination subsystem. If the top candidate is a nine-digit ZIP
the United States. Efficient use of these directories is very code on a mail piece with sufficient confidence, it is also di-
crucial to successful encoding of mail pieces. rectly encoded to DPC in the DPC determination subsystem.
One basic problem in address matching remains, that is, The DPC determination subsystem encodes given information
the address in the database is a USPS standard address. Pa- from each subsystem to a DPC. If no valid DPC is obtained
trons often do not actually write addresses on mail pieces us- and the five-digit ZIP code has sufficiently high confidence, it
ing USPS format. is accepted. Otherwise it is rejected. Each subsystem is de-
The handwritten address interpretation system consists of scribed in subsequent sections.
subsystems for preprocessing, ZIP code line recognition,
street line recognition, PO box line recognition, and delivery ZIP Code Line Recognition Subsystem
The ZIP code is first assumed to be at the last field of the last
line. If the likelihood of the detected ZIP code is less than a
threshold, up to two preceding lines are assumed successively
to be the ZIP code line until a ZIP code with sufficient likeli-
hood is detected. In actual presegmented images, ZIP code
fields are often split and divided into several pieces, which
have to be merged again into a field. This problem is resolved
through multiple use of the word recognition algorithm to a
set of successive presegments. The word recognition algo-
rithm employs a lexicon free word matching described in the
section entitled ‘‘Word-Matching Algorithm.’’
Preprocessing
Encode
DPC
PO box line recognition
Accept Accept
PO Box 5-digit
Reject
Accept
5-digit
Accept
9-digit
(a)
Original image:
Word segmentation
Barnett Bank for field extraction
150 Tequesta Drive
33469
Street name
recognition Top choice: Tequesta Drive
Rank-orderd
street name list
Lexicon generated from ZIP+4
database, using ZIP code and
street number as indices.
(b)
Figure 2. (a) Block diagram for handwritten address interpretation. (b) Result of processing an
address image.
POSTAL SERVICES 563
3. The third part deals with the recognition of street name 4. Overlapped Lines. This is a serious problem in hand-
through the use of the word-matching algorithm de- writing recognition. Due to limited spacing and the
scribed earlier. generally poor writing habits of humans, words from
one line intrude into adjacent lines, often intersecting
Street Number Location and Recognition. The street number with words in the lines above and below.
is assumed to be the first field of the street line. If ZIP code 5. Discrete Handwriting. Most people write in a hybrid
line includes only the ZIP code, the second preceding line is (neither purely cursive nor discrete) format. A typical
first assumed to be the street line; otherwise the immediate handwritten word may have discrete components (sin-
preceding line is assumed to be the street line. The rest of the gle isolated characters) as well as cursive components
street number location works in the same way as in the ZIP consisting of several characters. Often these compo-
code location. The word recognition algorithm employs a lexi- nents of a single word are spatially separated. The
con free word matching. main problem with discrete writing is the confusion
with regard to the location of precise word boundaries
Street Name Recognition in a line of handwritten text.
A street line recognition system is composed ot the ZIP/street 6. Imprecise Punctuation. In handwritten documents, it
number recognition system and the lexicon directed word rec- is often not easy to recognize punctuation marks be-
ognition algorithm. The lexicon is generated through the cause they are not precisely rendered during writing.
ZIP⫹4 directory search for a given pair of ZIP code and street Thus a comma may be mistaken for a character owing
number. The street name recognition is performed in the long to both its size and location.
word lexicon scheme (i.e., the predirectional, the street name, 7. Broken Characters. This is a very common occurrence
and the suffix are concatenated in a word) and is dealt as a in handwritten documents. Thus the horizontal bar of
single word. The word images in a street line except the street T may be physically disconnected from the vertical
number image are supplied as a single word image to the limb. This also occurs for characters such as A, B, D,
word recognition algorithm. The word recognition algorithm H, R, and so on.
employs a lexicon-directed word matching described in the
8. Similarly Shaped Words. In handwritten words, it
section entitled ‘‘Word-Matching Algorithm.’’
would often be very difficult to distinguish the word
‘‘clean’’ from ‘‘dean,’’ when written in a cursive mode.
TAXONOMY Contextual interpretation would be needed to resolve
this type of confusion.
An important goal of this article is to provide a taxonomy of 9. Ligatures. Again due to the diverse writing styles of
handwriting recognition. In its most general form, handwrit- people, long and sometimes unusual ligatures connect-
ing recognition is the transcription of a handwritten docu- ing adjacent characters often add confusion to the
ment into machine written text. The process starts at the doc- identity of the word. This is especially problematic
ument level and goes through the following steps: (1) with words in which one encounters w, u, v, and so on.
extraction of lines from the document, (2) extraction of words Other examples include unwanted connections such as
from the line, (3) holistic or character-based recognition of in tt and ff.
words, including punctuation marks such as periods, commas,
10. Overlapping Characters. Another common occurrence
colons, semicolons, apostrophes, and so on, and (4) a postpro-
is the overlapping of character fields within a word.
cessing step for integration of contextual and a priori knowl-
Thus a t or a T with a long horizontal limb often over-
edge to improve and enhance the recognition process. Figure
laps with characters in adjacent positions, making it
1 displays samples of handwritten documents with the char-
quite difficult to extract characters. Such overlaps also
acteristics described above.
occur with such combinations as gh, gy, and so on, es-
Techniques for the extraction of lines and extraction of
pecially when the bottom loop of g or y is quite wide.
words would be properly classified as preprocessing steps. In
general, a handwritten document can have one or more of the
following characteristics: It is in the context of the above observations that one must
approach the goal of developing machines capable of reading
1. Slant of the Writing. It is very common to find a dis- handwritten documents. It is important to note that even
tinct slant in the writing habits of most humans. Slant without most of the problems cited above, handwriting recog-
is measured as an angle with respect to a vertical nition would still be a formidable challenge.
frame of reference.
2. Skew or Tilt. A skew occurs when the lines of words Recognition Strategies
are at an angle to the horizontal frame. This often oc- Word recognition algorithms may be classified into the follow-
curs when the writer introduces a skew owing to his/ ing categories: (1) holistic approach or (2) character extrac-
her inability to write on a reference line that is often tion approach.
not physically present. The holistic approach generally utilizes shape features ex-
3. Underlines. It is very common to encounter under- tracted from the word image and attempts to recognize the
lines in many handwritten documents. The underlines entire word from these features. The character extraction ap-
which are often undulating and not straight are in- proach segments the word image into primitive components
tended to emphasize some key features of the docu- (typically characters). Character segmentation prior to recog-
ment. nition is called external character segmentation, while con-
564 POSTAL SERVICES
current segmentation and recognition is called internal char- It is also necessary to recognize that segmentation is not a
acter segmentation. local process; rather it is dependent on both the previous ex-
tracted character (and its identity) and the likely character
Holistic Approach that follows the current character. Contextually, it is clear
In the holistic approach a word is recognized as a unit, and that if the current character is q (this character could easily
techniques are derived from recognition of the entire word have been recognized as g), then the previous character is
without attempting to analyze the letter content of the word. most likely a vowel and the following character is, with a
A set of features (strokes, holes, arcs, ascenders, descenders, probability of 1, the character u. However, if the previous
etc.) characterizing the entire word is derived and used in character were g, then the next character would be any legiti-
recognition. It is generally accepted that holistic methods are mate letter. Di-gram and/or tri-gram analysis would be
feasible only when the number of words to be recognized is needed to eliminate ambiguous letter strings. This article will
small (typically less than 100). One of the earliest papers in describe in detail the word recognition strategies based on
holistic recognition is the classic work of Frishkopf and Har- segmentation-recognition.
mon of Bell Laboratories (1). In this approach a word is repre-
sented in terms of its horizontal and vertical extremes. In this CONCURRENT SEGMENTATION-RECOGNITION
context, an extreme is defined as a point at which one finds a OF WORD IMAGES
horizontal (vertical) maximum or minimum. Thus a word is
represented as an ordered list of extremes. Recognition is In this section, word recognition algorithms based on segmen-
based on the best match between test features and features tation of a word image into primitive components (characters
derived from dictionary words. Although the test inputs were and subcharacters) and concurrent concatenation and recog-
obtained in an on-line mode, recognition was essentially off- nition of these primitive components will be described. There
line. Another early paper in holistic recognition was from Ear- are two approaches to word recognition:
nest (2), who used off-line recognition strategies to on-line
test data. Earnest used features extracted from the middle 1. Context-free recognition where the recognition system
zone of the words, ascenders, and descenders. yields an optimum letter string and a lexicon is used as
As noted earlier, holistic approaches can be used in two a postprocessor to select the best match.
principal environments: (1) when the words to be recognized
are prespecified and when the number of words is small and 2. Lexicon-directed recognition, where a presegmented
word image is matched against all the words of the lexi-
(2) when the main objective is the reduction of lexicon size by
eliminating obvious mismatches, thereby facilitating a more con to obtain the best match.
accurate, but computationally more intensive, technique to be
used for final word recognition. While word recognition may be based on context-free or
lexicon-directed techniques, numeral string recognition such
Character Extraction Approach as ZIP code recognition or street number recognition in an
address is predominantly based on context-free techniques.
In this approach, which is also described as a character-based The recognition of words in a document follows a hierarchical
approach, algorithms are derived to extract and recognize the scheme as described below:
individual characters of the word. There are three principal
issues that need to be considered:
1. Remove tilt (skew) of the document.
1. In cursive writing, it is very difficult, if not impossible, 2. Extract lines of words from document.
to extract the characters of the word. If one considers 3. Remove slant from each line.
words containing the characters w, m, and d, and letter 4. Extract words from each line.
pairs rn, nr, un, iv, and so on, it is evident that many 5. Presegmentation of each word into primitive compo-
segmentations leading to identifiable characters are nents (characters and subcharacters).
possible. Also in cursive writing, it is often difficult to
6. Concatenation of components followed by character rec-
distinguish the letter o from a especially when ligatures
ognition to recognize each word.
are involved.
2. Erroneous recognition of characters extracted from the
In the section entitled ‘‘Recognition Algorith,’’ lexicon-di-
word image can lead to incorrect word recognition. It
rected word recognition will be discussed in detail. The exten-
is more typical to encounter letter strings that do not
sion to context-free recognition will be illustrated.
constitute a legitimate word. In such cases, it would be
necessary to incorporate a postprocessing stage to select
the closest words from a lexicon, using expression- Preprocessing (Line/Word Segmentation)
matching techniques. In this section we discuss segmentation of lines in a docu-
3. The availability of a lexicon of words that also contains ment, segmentation of words in a line, and related techniques
the true word is crucial to developing efficient tech- such as tilt and slant correction. Line segmentation is defined
niques for word recognition. Fortunately in certain ap- as the process of extracting the individual lines of words from
plications such as check processing and address recogni- a document. Word segmentation is defined as the process of
tion, a suitable lexicon can be generated. However, in extracting words from a given line. Character segmentation is
the case of numeral strings, a lexicon is not usually defined as the process of extracting the individual characters
available. that constitute the word unit.
POSTAL SERVICES 565
Horizontal projection of a document image is most com- cient, unless the down-sampled image itself is needed in suc-
monly employed to extract the lines from the document. If ceeding processes.
the lines are well-separated and are not tilted, the horizontal During the process of tilt estimation, several other charac-
projection will have well-separated peaks and valleys. These teristics of the document image (e.g., the number of lines, the
valleys are easily detected and used to determine the location interval, and the average height of characters) can be esti-
of boundaries between lines. This simple strategy fails if the mated, which are of great use in subsequent operations. Fig-
lines are tilted with respect to the horizontal axis of the image ure 3 shows an example of tilt correction. In this example, tilt
plane. The peaks and valleys would not be distinctive for is estimated as the direction which maximizes the variance of
tilted document images, and the text lines cannot be sepa- crossing counts was selected over a range varying from ⫺8⬚
rated by horizontal boundary lines. There are two approaches to ⫹8⬚.
to handle the problem of tilt. In the first approach the tilt is Hough Transformation. Hough transformation is one of the
estimated and the document image is corrected for tilt prior most common technques for detection of line segments in an
to line segmentation. In the seond approach, tilt correction is image. It maps the original image to a – parameter plane,
applied to each line after line segmentation has been per- and a line in the original forms a cluster in the parameter
formed. plane. Once the location of the clusters are determined, the
Vertical projection of a line image is employed in the ex-
traction of words in a line, analogous to the process of line
segmentation. Again this simple strategy fails if the words Input_word_image
have a slant with respect to the vertical axis of the image
plane. To resolve the problem, it is necessary to perform slant
estimation and correction. Slant correction is effective and
widely used as a preprocessing operation for character seg-
mentation in a word. Slant estimation and correction are also
useful in detection and recognition of printed italic letters.
Some documents are handwritten along preprinted or
hand-drawn underlines, which interfere with the segmenta-
tion and recognition of words and characters. In such cases, it
is necessary to detect and eliminate the underlines. A simple
underline elimination algorithm is also described in this
section.
Tilt Correction
Tilt Correction Before Line Segmentation Crossing_counts
Zero Crossing Method (3). Tilt correction is generally per-
formed in two steps. In the first step, the tilt of the document
is estimated and in the second step shear transformation is
applied to remove the tilt. Tilt or skew is estimated by finding
the direction in which the projection of the document has min-
imum separability with regard to the peaks and valleys in the
projection profile. In the zero crossing method, only crossing
points are counted to obtain the projection. The use of the
zero crossing points rather than the entire foreground pixels
is advantageous both in improving the separability of the pro-
jection and in saving computation time. The variance of the
number of zero crossing points is used as a simple measure of
separability. To find the direction which maximizes the sepa- –8 –6 –4 –2 0 +2 +4 +6 +8
rability measure, multiple projections in different directions
differing by one or two degrees are calculated within the Tilt-corrected_image
range of the expected tilt. Once the tilt of the image is esti-
mated, tilt correction is implemented as a shear transfor-
mation.
It is worth observing that this straightforward enumera-
tive search for maximum separability is more efficient than
expected, if it is implemented properly: All the multiple pro-
jections are calculated in a single raster scan. Only the zero
crossing points are projected in multiple directions. The map-
ping is performed incrementally without any multiplication.
If the document image is quite large and has sufficient resolu-
tion, the raster scan can be performed for every two pixels on
alternate scan lines. This interleaving is equivalent to pro-
cessing a down sampled image, and the process is more effi- Figure 3. Example of tilt correction.
566 POSTAL SERVICES
tilt of each line and the average tilt is easily estimated. The span more than one text line are split into two or more compo-
Hough transformation is very attractive and useful because it nents and placed in separate text lines.
is available to detect not only solid lines but also broken lines Morphological Method (6).. In the morphological method,
and even text lines consisting of characters and words. This core regions of the image are generated using a morphological
generality, however, sacrifices the processing efficiency; also, operation. These core regions generally fill in the body of each
the implementation has to be performed carefully, utilizing word, but eliminate ascender and descender strokes. For the
domain specific knowledge: To reduce the processing time, core images, a technique similar to the zone method is ap-
only border points or crossing points are required to be plied. Because of the preceding morphological operation, less
mapped to the parameter plane instead of the entire fore- heuristics are required to split or merge blocks (to form text
ground pixels. Further reduction is achieved by restricting lines) than for the zone method. Zones in the morphological
the range of in the parameter plane. It should be noted that method are assumed to have overlapped areas, while zones in
if the transformation is performed only for specified (e.g., the zone method are mutually disjoint.
every 2⬚ from ⫺8⬚ to ⫹8⬚), the Hough transformation ap-
proach is equivalent to the crossing point method described Line Segmentation After Tilt Correction
above. A section of the parameter space at a specified is Projection Method. As described earlier, the horizontal pro-
simply the pixel projection in direction. jection of a document image is calculated, and the valley
points are detected and used to determine the location of
Tilt Correction After Line Segmentation. Techniques for tilt boundaries between lines. Some valley points may be merged
estimation in an entire document such as an address block or removed based on heuristics such as the expected interval
generally yields an average value; individual lines in the doc- of lines. If two valley points are closer than a given threshold,
ument may still exhibit residual tilt that needs to be removed they are merged. The advantage of the projection method is
for more accurate word recognition. A different approach is its robustness in dealing with documents containing con-
needed to estimate tilt in a single line of words. It is also nected lines due to extenders. The disadvantage is the under-
observed that the method described in the previous section lying assumption that line boundaries are horizontal.
may yield incorrect tilt estimates, when the number of lines Component Clustering Method (7). Line segmentation can
in the document image is less than two. A common approach be considered as a problem of clustering for the connected
to estimate tilt in a single-line image is the use of a least- components. Each connected component is mapped into a two-
squared error line fit to the bottom profile of the line image dimensional space, in terms of its vertical extents (ymin, ymax)
(ignoring any descenders in profile derivation). The slope of [Fig. 4(c)].
the resulting line yields a good estimate of the tilt in a single The clusters are detected using typical clustering algo-
line image. rithms such as the K-means clustering (8). The K-means clus-
tering initially requires K clusters or K centers of clusters.
These initial values can be obtained in the process of tilt cor-
Underline Elimination
rection. The projection method may be employed to obtain
Projection Method. Many techniques have been proposed these initial values, if necessary. To suppress the undesirable
for the detection and elimination of underlines. These include influence of small noise components or large components in-
methods based on Hough transform and morphological opera- cluding multiple lines connected by extenders, the extended
tions with suitable structuring elements. An approach based version of the K-means clustering (called the weighted K-
on a line fit to the bottom profile of a single line image has means clustering) is known to be useful. In the weighted K-
also been used to eliminate underlines characterized by low means clustering, the center of a cluster is the weighted mean
curvature segments. In a handwritten document consisting of of the samples. The weight of each connected component is
several lines of data, vertical extents of underlines are simply defined so that the closer the height of the component is to
estimated in the horizontal projection of the tilt corrected doc- the estimated character height, the larger is its weight.
ument image. Within the vertical extents, short vertical runs If the number of clusters is not uniquely estimated, but
which are isolated in the extent are removed. A novel algo- within a range of 6 ⫾ 1, the K-means clustering is applied for
rithm using a morphological approach has been proposed by all possible values of K. Among the results, clusters which
Liang et al. (4) for removing interference strokes, including have poor separability are discarded. Clusters that do not sat-
underlines from word images. This method is capable of re- isfy spatial constraints required by valid document lines are
moving hand- and machine-drawn underlines, even when also discarded. The remaining clusters yield the number and
these underlines cut across the characters of the word image. the position of the lines, and the components are assigned to
these clusters (lines). Multiple line components occupying a
Line Segmentation vertical extent of more than two lines are split at the line
boundary [Fig. 4(b)]. The advantage of the component cluster-
Line Segmentation Before Tilt Correction ing method is its ability to construct complex line boundaries.
Zone Method (5). The algorithm divides an input image
into vertical zones. For each zone, a horizontal projection is
Slant Correction
computed. A vertical extent of the projection with nonzero
values form a block. Blocks which are horizontally adjacent There are two principal approaches for estimating the slant
and vertically overlapping are connected to form text lines. of a word. These include the projection method and the chain
Heuristics are used to split and/or join blocks. Connected code method. A brief description is provided below.
components that are located entirely in blocks from a single Projection Method. The average slant of characters in a
line are assigned to that line. Connected components that word or in a line is estimated by the analysis of slanted verti-
POSTAL SERVICES 567
Word Segmentation
Word Boundary Analysis (7). Words are assumed to be sepa-
rated by a space, a comma, or a period. The space detection
algorithm detects the spaces by classifying each gap between
the character segments as ‘‘between words gap’’ or ‘‘within
word gap,’’ respectively. If the gap is wider than a threshold,
the gap is classified as ‘‘between words gap’’, otherwise as
‘‘within word gap.’’ The threshold, based on the distribution
(b)
of the gap width for text lines is found by applying a standard
technique such as Otsu’s method (11).
Exact segmentation of a handwritten line field is very dif-
ficult unless it is integrated with the word recognition pro-
cess. It is interesting to observe that human beings usually
employ this approach very efficiently for word recognition. In
a typical integrated word segmentation process, the word seg-
mentation is assumed to yield oversegmented word images,
where some words can be split and divided into subwords.
These (pre-) segments are merge again into a whole word
through multiple application of the word recognition algo-
rithm to a set of successive segments. To obtain an overseg-
mented word image, the line is subject to further segmenta-
tion, if undersegmentation is anticipated. If the number of
words in a line is too few or the estimated length of a word is
too long, the word is divided at the maximum within-field
gap. This procedure is repeated until no further subdivision
is necessary.
Convex Hull (12). Different metrics can be employed to
measure the spatial gaps with varying degrees of accuracy
(c) (12). Convex-hull metric requires computation of convex hulls
for each of the components in a line. The distance between
(c)
(a) (c)
cal projections (histograms) at various angles (9). The average
slant is found by looking for the greatest positive derivative
in all of the slanted projections.
Chain Code Method (10). In contrast with the tilt estima-
tion, the average slant of characters in a word or in a line is (d)
easily estimated using the chain code of the border pixels. The
(b)
average slope (tangent of the slant angle) is given by
n1 + n2 + n3
m= (1)
n1 − n3
(e)
where ni is the number of chain elements at an angle of i Figure 5. Different stages of processing for word recognition: (a)
times 45⬚ (/ or 兩 or ). Shear transformation is then applied to Original image, (b) extracted chain code, (c) image slant corrected, (d)
remove the slant. It is interesting to see why this simple ex- oversegmented image, (e) final segmentation-recognition.
568 POSTAL SERVICES
the convex hulls of the components along the line joining their
centers of gravity is used as a distance measure.
RECOGNITION ALGORITHM
Character Segmentation
Contour Analysis (13, 14). The contour analysis method is
suitable to obtain oversegmented character images. Possible
segmentation points are detected through local extrema anal-
ysis of the upper contour of the word image. Among the local (c)
minima, those that are not deep enough from the adjacent
local maxima are sequentially removed. In order to obtain Figure 7. (a) Example of presegmentation points, (b) disjoint box
segmentation, and (c) optimum character segmentation.
characters separated by vertical lines, segmentation points
determined in the previous step are often shifted horizontally
to the right or the left as follows: Disjoint Box Segmentation (13,14). Characters and charac-
If the minimal point is not open vertically upward, the ter segments are more easily handled if their bounding boxes
point is shifted to the right or to the left of this point de- are mutually disjoint. If oversegmentation is permitted, hori-
pending on where the number of runs and the total length of zontal overlapping of character segments is resolved by a sim-
runs is minimal. Figure 7(a) illustrates this process. ple algorithm. A word image is split vertically at each preseg-
Profile Analysis (15). Instead of the upper contour of a word, mentation point and is separated into horizontally
the upper profile is analyzed. Here the upper profile is defined nonoverlapping zones. A connected component analysis is ap-
to be the set of the topmost foreground pixels in each column. plied to the split image to detect the boxes enclosing each
Postprocessing is required to find some segmentation points connected component. These boxes are usually disjoint and do
overshadowed by character segments such as the long ‘‘t’’ bar. not include parts of other connected components [Fig. 7(b)].
Run Length Analysis (16). Ligatures between ‘‘or’’, ‘‘on’’, etc., The disjointedness of the bounding boxes is necessary for
occasionally do not have a valley point in the upper contour. high-speed feature extraction.
To detect and split these ligatures, a single-run stretch is de-
tected and split at the middle point. The run is vertical
streaks of one or more black pixels, and the single-run is the WORD-MATCHING ALGORITHM
unique run on a single vertical line. The single-run stretch is
a horizontal stretch of single-runs shorter than a threshold Lexicon-Directed Algorithm (13, 14)
determined depending on the average stroke width. Figure 5 The number of boxes (or segments) obtained by the disjoint
illustrates this analysis. Among these single-run stretches, box segmentation is generally greater than the number of
those which have well-defined peaks and valley points in the characters in the word image. In order to merge these seg-
upper contour are removed. The remaining single-run ments into characters so that the final character segmenta-
stretches are split at the middle point. Figure 7 shows an ex- tion is optimal, dynamic programming (DP) is applied, using
ample when the ligature connecting ‘‘r’’ and ‘‘n’’ of California the total likelihood of characters as the objective function.
had to be cut. The likelihood of each character is given by a discriminant
function. To apply the DP technique, the boxes are sorted left
to right according to the location of their centroids. If two or
1
more boxes have the same x coordinates at the centroids, they
are sorted top to bottom. Numbers above the boxes in Fig.
4 5
2 3 7(b) show the order of the sorted boxes. It is worth observing
that the disjoint box segmentation and the box sorting process
reduce the segmentation problem to a simple Markov process,
in most cases. For example, in Fig. 7(b), boxes 1 to 4 corre-
Figure 6. Segmentation-recognition with DP. spond to the letter ‘‘C’’ of California, box 5 corresponds to ‘‘a’’,
POSTAL SERVICES 569
4 — — 4.87 4.57
2 1.65 3.11 — —
1 1.90 — — —
k씮 1 2 3 4
Letter F o u r
box 6 corresponds to ‘‘l’’, . . ., and so on. This assignment of where L(k, j(k)), the maximum likelihood of partial solutions
boxes to letters can be represented as given j(k) for the kth letter, is defined and calculated recur-
sively by
i 씮 1 2 3 4 5 6 7 8 9 10
Ai C a l i f o r n i a
i(i) 씮 4 5 6 7 8 9 10 11 13 15 16
k
L(k, j(k)) = Max (Ai , j(i − 1) + 1, j(i))
j(1), j(2),..., j(k−1)
where i denotes the letter number, and j(i) denotes the num- i=1
ber of the last box corresponding to the ith letter. Note that
the number of the first box corresponding to the ith letter is This can be written in a recursive form as shown below:
j(i⫺ 1)⫹1. Given [j(i), i ⫽ 1, 2,. . .; n] the total likelihood of
the character is represented by = Max [(Ak , j(k − 1) + 1, j(k)) + L(k − 1, j(k − 1))]
j(k−1)
n
L= (Ai , j(i − 1) + 1, j(i)) (2) with L(0, j(0)) = 0, j(0) = 1, 2, . . ., m (4)
i=1
Starting from Eq. (5), all L(k, j(k))’s are calculated for k ⫽
where ᐉ(Ai, j(i ⫺ 1) ⫹ 1, j(i)) is the likelihood for the ith letter. 1, 2, . . ., n using Eq. (4) to find j(n)* using Eq. (3). The rest
In the lexicon-directed algorithm, an ASCII lexicon of pos- of j(k)*’s (k ⫽ n ⫺ 1, n ⫺ 2, . . ., 1) are found by back-tracking
sible words is provided and the optimal character segmenta- a pointer array representing the optimal j(k ⫺ 1)*’s which
tion is found for each lexicon word. All lexicon words are then maximizes L(k, j(k)) in Eq. (4).
ranked regarding the optimal likelihood per character (L*/n) The example illustrated in Fig. 6, Table 1, and Table 2
to select the best candidate word. The optimal assignment shows the values of L and j(k ⫺ 1)* for given k and j(k). In
(the optimal segmentation) which maximizes the total likeli- this example, L(n, j(n))* ⫽ L(4, 5) ⫽ 6.71 and j(n)* ⫽ 5. Suc-
hood is found by using the dynamic programming technique
ceeding j(k)’s are 4, 3, 2, 0 respectively. The box ‘‘0’’ corre-
described below. The optimal assignment j(n)* for the nth let-
sponds to a virtual box standing for the last box of the letter
ter is the one such that
preceding the first letter ‘‘C’’. Figure 6 illustrates the word-
L∗ = L(n, j(n)∗ ) = Max L(n, j(n)) (3) matching procedure based on segmentation-recognition
j(n) with DP.
4 — — 3 3
2 0 1 — — k and j(k)
1 0 — — —
1 2 3 4 Letter k씮
1 2 4 5 j(k)
Letter F o u r
570 POSTAL SERVICES
n
L= (Ai , j(i − 1) + 1, j(i)) tion to determine the best match between the lexicon words
i=1 and the optimum letter string derived in a context-free recog-
n nition mode.
= Max{(Ai , j(i − 1) + 1, j(i))} (5)
Ai
i=1
PERFORMANCE EVALUATION—CASE STUDIES
instead of Eq. (2). The character likelihood for all letters are
calculated and the maximum value and the associated letter Performance Evaluation of Word Recognition
Ai* are determined. The word matching process is applied A total of 2998 word images extracted from the ‘‘bha’’ data-
only once for an input word, and the recognition result is base (handwritten address block data collected at Buffalo,
given by A*1A*2 . . ., A*n. When the word length n is un- New York) were used for this test. The test images included
known, an upper bound is estimated and used as the value of city, street, state, personal, and business names, including
n. It is very convenient that the likelihood array L(k, j(k)) and abbreviated forms. The style of writers ranged from strictly
the pointer array for a word of length n includes all the en- printed to strictly cursive, and from upright to very slanted.
tries for shorter words. The lexicon free algorithm is required The distribution of writing styles reflected the corresponding
when no lexicon is available for numeral string recognition. distribution for the mail stream.
It is also efficient and suitable if the input word is written Each word image in the test database was associated with
neatly and segmentation and recognition of characters are three separate ASCII lexicons of size 10, 100, and 1000, re-
relatively accurate. spectively. The lexicons were generated through a random se-
K-Best Path Algorithm (17). The lexicon free algorithm can lection process from a statistically significant sample con-
be extended so that it generates not only the optimal word taining city, street, state, and personal and business name
interpretation but also the K best interpretations for a speci- words. Each of the three lexicons included the ASCII repre-
fied K. Among the K interpretations, invalid interpretations sentation of the associated word image, spelled exactly as it
are removed through lexicon search and the best valid inter- appeared in the image. Thus if the word Illinois was wrongly
pretation is selected. spelled as Illionis, then the lexicon contained the word Illi-
Expression Matching (18,19). Expression matching is the nois. Table 3 shows the cumulative correct recognition rate.
process of matching a character string against words in a lexi-
con and computing a measure (called ‘‘edit distance’’) indica-
tive of the degree of match between the given string and the
words in the lexicon. Spell-checking operations use expression
matching to determine alternative words to correct misspelled
words. This process uses three operations:
Table 4. Correct Rates of ZIP Code, Street Number, Street Name, and PO Box Recognition
ZIP Code Street Number Street Name PO Box Number
Correct Correct Correct Correct
Recognition Recognition Recognition Recognition
Top N (%) Top N (%) Top N (%) Top N (%)
1 2910 (82%) 1 2093 (77%) 1 422 (58%) 1 540 (71%)
2 3021 (85%) 2 2119 (78%) 2 434 (60%) 2 556 (73%)
3 3045 (86%) 3 2134 (79%) 3 438 (61%) 3 566 (74%)
4 3056 (86%) 4 2141 (79%) 4 440 (61%) 4 572 (75%)
5 3066 (86%) 5 2148 (79%) 5 441 (61%) 5 575 (76%)
6 3070 (86%) 6 2150 (79%) 6 443 (62%) 6 577 (76%)
Rest 449 (12%) Rest 551 (20%) Rest 231 (32%) Rest 182 (24%)
Reject 21 (0.59%) Reject 7 (0.26%) Reject 46 (6%) Reject 0 (0%)
Total 3540 Total 2708 Total 720 Total 759
The character classifier was designed (trained) using the (2) lexicon truncation to achieve low error rate and high pro-
character samples extracted from state name and city name cessing speed.
word images in the ‘‘Bd’’ database. The number of characters A word recognition algorithm using the segmentation-rec-
used for classifier design was 22606 (435 per character in av- ognition approach is shown to be robust, accurate, and com-
erage) and the correct character recognition rate was about mercially feasible. Context-free recognition is shown to be fea-
74.2% for the design samples. The top correct recognition rate sible for numeral string recognition, while a lexicon-directed
was 98.01%, 95.46%, and 91.49% for lexicons of size 10, 100, approach is recommended for word recognition.
and 1000, respectively. In conclusion, it can be stated that handwriting recogni-
Figure 8 shows examples of correctly recognized words. tion is a feasible technology and can be used with advantage
The speed of word recognition was 2.0, 2.5, and 3.5 s/word for in many commercial applications such as address recognition,
each lexicon on a SUN SPARC Station 2. The integrated ad- forms processing, check processing, and so on. As this article
dress interpretation system was designed to determine the is being concluded, the USPS has announced that the Hand-
nine-digit ZIP code by locating and recognizing the ZIP code, written Address Interpretation system is currently being de-
the street number, and/or the PO box fields. ployed in some postal sorting centers on the east coast. Pre-
The USPS five-digit city/state/ZIP directory consisting of liminary indications are that the systems are performing
100,000 records was used to generate a lexicon of city names satisfactorily. Advances in this field can be directly credited
for city name recognition. The USPS ZIP⫹4 address directory to the USPS, which initiated and supported basic research
consisting of 26 million records was used to generate a lexicon in this field through grants and contracts to industries and
of street names for street name recognition. The performance educational institutions.
of the integrated system was evaluated using ‘‘bha’’ test sam-
ples. All the samples from bha_6000 to bha_7603 were used
for this test. BIBLIOGRAPHY
Tables 4 and 5 summarize the performance at different op-
erating points specified in column 1 of the table. The error 1. L. S. Frishkoff and L. D. Harmon, Machine Reading of Cursive
rate for one set of operating points was 1.12% with 50.19% Script, in C. Cherry (ed.), Information Processing, London: But-
encode rate. With a different set of operating points, and error terworth, 1961, pp. 300–315.
rate of 0.87% with 43.12% encode rate was obtained. In other 2. L. D. Earnest, Machine Recognition of Cursive Writing, in C.
words the system could be tuned to achieve a specified error Cherry (ed.), Information Processing, London: Butterworth, 1961,
rate. pp. 462–466.
3. Y. Ishitani, Document skew detection based on local region com-
plexity, Proc. 2nd ICDAR, 1993, pp. 49–52.
CONCLUSION
4. S. Liang, M. Ahmadi, and M. Shridhar, Segmentation of interfer-
ence strokes using morphological approach, Proc. 3rd Int. Conf.
The performance of the integrated system developed for the
Document Anal. Recognition, Montreal, Canada, 1995, pp. 1042–
US Postal Service exceeded the performance specifications set 1046.
by USPS for processing handwritten addresses. The inte-
5. V. Govindavaju, A. Shekhawat, and S. N. Srihari, Interpretation
grated recognition system incorporates several novel features
of handwritten addresses in US mail stream, Proc. 3rd IWFHR,
such as (1) tunability for adjusting error-rejection rates and 1993, pp. 197–206.
6. M. J. Ganzberger et al., A system for handwritten address inter-
pretation, Proc. 5th Adv. Technol. Conf., 1991, pp. 337–351.
Table 5. Error Versus Encode Rate 7. F. Kimura, Y. Miyake, and M. Shridhar, Zip code recognition us-
tt1 tt2 Encode rate Error Correct ing lexicon free word recognition algorithm, Proc. 3rd ICDAR,
1995, pp. 906–910.
20.0 5.0 50.19 (803) 1.12 (9) 98.88 (794)
40.0 7.0 43.12 (690) 0.87 (6) 99.13 (684) 8. A. Devijver and J. Kittler, Pattern Recognition, London: Prentice-
Hall International, 1982, pp. 409–410.
572 POTENTIAL TRANSFORMERS
9. D. Guillevic and C. Y. Suen, Cursive script recognition: A sen- T. Wakabayashi et al., Accuracy improvement through increased fea-
tence level recognition scheme, Proc. 4th IWFHR, 1994, pp. ture size in handwritten numeral recognition, Syst. Comput. Jpn.,
216–223. 26 (8): 35–44, 1995.
10. D. Lee and S .N. Srihari, Handprinted digit recognition: A com- S. Watanabe and N. Pakvasa, Subspace method of pattern recogni-
parison of algorithms, Proc. 3rd IWFHR, 1993, pp. 153–162. tion, Proc. 1st Int. Joint. Conf. Pattern Recognition, 1973, pp.
11. L. R. Rabiner and B. H. Juang, An introduction to hidden Markov 25–32.
models, IEEE Acoust. Speech Signal Process. Mag., 30 (1): 4–16,
1986. M. SHRIDHAR
12. S N. Srihari, V. Govindaraju, and R. K. Srihari, Handwritten text University of Michigan–Dearborn
recognition, Proc. 4th IWFHR, 1994, pp. 265–274. GILLES HOULE
13. F. Kimura, M. Sridhar, and Z. Chen, Improvements of a lexicon TRW Enterprise Solutions
directed algorithm for recognition of unconstrained handwritten
words, Proc. 2nd ICDAR, 1993, pp. 18–22.
14. F. Kimura et al., Context directed handwritten word recognition
for postal service applications, Proc. 5th Adv. Technol. Conf., 1992,
pp. 199–213.
15. E. Lecolinet and J. Crettez, A grapheme-based segmentation
technique for cursive script recognition, Proc. 1st ICDAR, 1991,
pp. 740–748.
16. R. M. Bozinovic and S. N. Srihari, Off-line cursive script word
recognition, IEEE Trans. Pattern Anal. Mach. Intell., 11: 68–83,
1989.
17. C. R. Nohl, C. J. Burges, and J. I. Ben, Character-based hand-
written address word recognition with lexicon, Proc. 5th Adv.
Technol. Conf., 1992, pp. 167–180.
18. F. Kimura, M. Shridhar, and N. Narasimhamurthi, Lexicon di-
rected segmentation-recognition procedure for unconstrained
handwritten words, Proc. 3rd Int. Conf. Frontiers Handwriting
Recognition, Buffalo, 1993, pp. 122–131.
19. H. Bunke, A fast algorithm for finding the nearest neighbor of a
word in a dictionary, Report of Institut für Informatik und An-
gewandte Mathematik, Universität Bern, Switzerland, 1993.
Reading List
M. Chen et al., Variable duration Hidden Markov Model and morpho-
logical segmentation for handwritten word recognition, IEEE
Trans. Image Process., 4: 1995, 1675–1688.
K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd ed.,
New York: Academic Press, 1990, pp. 76–78.
N. Gorsky, Off-Line Recognition of Bad Quality Handwritten Words
Using Prototypes, in S. Impedovo (ed.), Fundamentals in Hand-
writing Recognition, NATO ASI Series F: Computer and Systems
Sciences, Vol. 124, Berlin: Springer-Verlag, 1994, pp. 199–217.
F. Kimura, Y. Miyake, and M. Shridhar, Relationship among qua-
dratic discriminant functions for pattern recognition, Proc. 4th
IWFHR, 1994, pp. 418–422.
A. Kundu, Yang He, and P. Barl, Recognition of handwritten word:
First and second order Hidden Markov Model based approach,
Pattern Recognition, 22 (3): 283–297, 1989.
M. Leroux, J. C. Salome, and J. Badard, Recognition of cursive script
words in a small lexicon, Proc. 1st Int. Conf. Document Anal. Recog-
nition, St. Malo, France, 1991, pp. 774–775.
E. Oja, Subspace Methods of Pattern Recognition, Oxford, UK: Re-
search Studies Press, 1983.
N. Otsu, A threshold selection method from gray-level histograms,
IEEE Trans. Syst. Man Cybern., SMC-9: 62–66, 1979.
T. Paquet and Y. Lecourtier, Handwriting recognition: Application on
bank cheques, Proc. 1st Int. Conf. Document Anal. Recognition, St.
Malo, France, 1991, pp. 749–750.
B. Plessis et al., Isolated handwritten word recognition for contextual
address reading, Proc. USPS 5th Adv. Technol. Conf., 1992, pp.
579–580.
Threshold of Electric Shock. The level of electric shock dif-
fers from one individual to the other and depends on sex, age,
weight, and chemical balance (a function of the physical con-
dition of the person). The effect of a shock largely depends on
the frequency, duration of contact, and the amount of electric
current passed through the body, rather than the voltage. For
instance, a shock from a 100 V source can be as deadly as
that from a 1000 V source. Low frequency currents from 60
Hz (power line frequency) down to direct current cause more
severe shock because they penetrate the skin more deeply and
quickly and burn the flesh much faster. Currents of higher
SAFETY SYSTEMS frequencies change direction several times per second, at a
rate much faster than the rate of a normal heart beat. Thus,
Some fundamental objectives of a society are the safety, pros- high-frequency currents have less tendency to initiate fibril-
perity, happiness, and well-being of its people. Safety is of lation of the heart than low-frequency currents. Currents as
paramount importance because it is the brickwork on which low as 1 mA can be perceived as an electric shock. The ac-
the other aspirations of the society must stand upon. Several cepted maximum harmless current intensity is 5 mA, and this
factors affect safety, and they are the focus in this article. is also the maximum current allowed to leak from home appli-
ances and still pass Underwriters Laboratories (UL) specifi-
cations (3). A current of 100 mA or above will cause ventricu-
ELECTRICAL SAFETY lar fibrillation, which prevents the heart from pumping blood;
death may result.
Many people are killed every year, and many more are in-
jured, while in contact with electrical energy. Of the 3740
work-related deaths reported by the Bureau of Labor Statis- Estimation of Electric Shock Current. An electric shock re-
tics for 1984 in the United States (1), 10% of the fatalities, or sults when an electric current passes through the body caus-
about 370, were the direct result of electrocution at work. Un- ing physical stimulation of the body. The magnitude of the
like other industrial accidents, electrical accidents often hap- current I in amperes (A), obeys Ohm’s law,
pen to knowledgeable workers and professionals. For in-
stance, in Great Britain, of the 805 accidents reported in V
factories in a typical year, 47% were electrical accidents in- I= (1)
R
volving skilled workers (see Table 1) (2). Many of these deaths
and injuries could have been prevented by the use of appro-
priate safety equipment and techniques. Table 2 shows that where V is the applied voltage in volts (V), and R is the total
ignorance, negligence, and forgetfulness account for most of resistance in ohms (⍀) of the current path, which may include
the electrical accidents recorded in a typical year. the ground on which the person is standing, the boot being
worn, and the human body. When these resistances are ac-
counted for, Eq. (1) is modified to
Electric Shock
When the human body forms a conduction path for electric V
current, the effect it causes on the body is called electric I= (2)
RA + RB
shock. The three main hazards of electricity are shock, arc,
and blast. Since the low resistance of the body diminishes its
ability to withstand the passage of electric current, most elec- where RA is the resistance of ground plus boot, and RB is the
trical systems can be hazardous. Even a minor electric shock resistance of the body. For example, if the resistance of shoes
can create a serious injury due to a fall as a result of reflex to ground of a man holding a pair of pliers is RA ⫽ 1000 ⍀
action. Electric shock is a safety hazard in most laboratory and the resistance of the body is RB ⫽ 5000 ⍀, then for a 110
environments, and it can be caused by improper use or han- V fault voltage, the total current through the man’s body from
dling of electrical appliances or equipment. It can also come Eq. (2) is
from faulty equipment: equipment with a factory defect that
causes it to malfunction, or equipment failure as a result of 110 V
fatigue or aging. Shorted cables or worn electrical conductors I= = 18.33 mA (3)
(1000 + 5000)
can leak electric current away from its desired path and cause
electric shock. The severity of the effect on the body may in-
clude tingling, a burn on the skin at the point of contact, mus- The current of 18.33 mA is over the paralysis threshold as
cular contraction, inability to control the muscles, and loss of shown in Table 3 (4). The victim cannot release the pliers,
grip on the electrical conductor or equipment at the inception and the result may be fatal. The numbers used in this exam-
of the electric shock. ple are only approximate. The actual values will depend on
629
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
630 SAFETY SYSTEMS
Table 1. Electrical Accidents Analyzed by Occupation Table 2. Conditions Leading Up to Accidents in One Year
Occupation Fatal Total Cause Fatal Total
Skilled Failure or lack of testing 5 91
Testing 5 87
Supervisory staff 2 37
Ignorance, negligence, forgetfulness, and inadver- 24 354
Switchboard substation attendants — 2
tence
Testing staff — 18
Accidents resulting from fault of persons other 18 160
Electrical tradesmen and their mates 7 278
than injured person
Engineering apprentices (under 18) — 7
Working on live gear deliberately 3 108
Engineering apprentices (over 18) 1 17
Misunderstood instructions or failure of permit-to- 1 16
work system
Unskilled
All men not included in the above 16 396
All women not included in the above — 50
Total 26 (3.2%) 805 (100%) vicinity of explosive mixtures or in an explosion-prone envi-
ronment, as for instance in a gunpowder manufacturing
plant. In this category also are the accidents in a manufactur-
ing plant or assembly line and nuclear accidents in a nuclear
several factors, which include the contact resistance between power plant, including nuclear-propelled vehicles, subma-
man and metal and his weight and physical condition. The rines, and aircraft carriers, as well as hazards inherent in the
nominal resistances of various parts of the human body are restarting of nuclear plants. Nuclear war and nuclear-weapon
given (5) in Table 4, and the resistances of various materials accidents are technological hazards as well. An example of
are given in Table 5. hazard due to sabotage is the explosion of an airplane due to
Investigators have also established that the resistance of a terrorist bomb.
the body from hands to feet depends on the area of contact Hazards can also be broadly classified as environmental,
and on whether the skin is wet, moist, or dry. These values physical, chemical, and biological. Typical environmental haz-
range from 1000 ⍀ and 10,000 ⍀. An empirical formula is ards include falling objects, improperly enclosed workplaces,
given by and improperly lighted shop floors. Physical hazards include
lifting heavy objects, being exposed to heat, bright lights, ex-
cessive noise or vibration, and mild doses of radiation, and
RV k = C (4)
being shocked by improperly grounded equipment or under-
sized power cables. Irrespective of the cause and type of haz-
where R is the resistance in ohms, V is the voltage in volts, ard, hazardous conditions should be avoided or eliminated
and k ⫽ 0.83 and C are constants. Table 6 (6) shows the elec- where possible by redesigning the workplace and wearing the
trical characteristics of the human body at 50 Hz in dry condi- appropriate protective clothing and equipment. These and
tion computed with Eq. (4). other hazards result in injuries and loss or damage of prop-
erty. When hazards are not contained, they cause work-re-
lated accidents which include minor to severe burns, physical
HAZARDS and bodily injuries, back pains due to physical exertion, loss
of hearing or vision, or death.
Hazards can in general be natural, technological, or caused
by an act of sabotage. Examples of natural hazards are earth- Sources of Hazards
quakes, floods, chemical spills during transportation, hurri- Hazardous conditions may be unnoticed or ignored. Exposed
canes, and lightning. Examples of technological hazards are conveyor belts in a workplace can catch a finger or entangle
automobile, marine, and airplane failure as well as fire and clothing. The light from an arc-welding machine can be too
explosion in mines. Electrostatic hazard due to capacitive dis- bright unless for the naked eye. The noise level from a grind-
charges can be a source of ignition in the presence of or in the ing machine or metal cutting tool may be too intense if protec-
Table 4. Nominal Resistance Values for Various Parts of the Table 6. Calculated Electrical Characteristics of Human
Human Body Body at 50 Hz in Dry Conditions
Resistance V (V) R (⍀) I ⫽ V/R (mA)
Condition (area to suite) Dry Wet 12.5 16,500 0.8
31.3 11,000 2.84
Finger touch 40 k⍀–1 M⍀ 4–15 k⍀
62.5 6,240 10.0
Hand holding wire 10–50 k⍀ 3–6 k⍀
125 3,530 35.2
Finger–thumb grasp 10–30 k⍀ 2–5 k⍀
250 2,000 125
Hand holding pliers 5–10 k⍀ 1–3 k⍀
500 1,130 443
Palm touch 3–8 k⍀ 1–2 k⍀
1000 640 1560
Half around 1.5 in. pipe (or drill
2000 362 5540
handle) 1–3 k⍀ 0.5–1.5 k⍀
Two hands around 1.5 in. pipe 0.5–1.5 k⍀ 250–750 k⍀
Hand immersed — 200–500 ⍀
Foot immersed — 100–300 ⍀ Food processing plants in dirty or poorly maintained envi-
Human body, internal, excluding ronment
skin — 2–1000 ⍀ Exertion of physical strength far in excess of the individual
capability, such as lifting heavy loads
Mental and psychological stress as a result of long hours
tive equipment is not worn. Workers neglect to wear helmets of work at a computer
in situations where falling objects are commonplace. Inade- Fatigue due to continuous hours of work without break, or
quate boots are worn in areas where heavy tools can drop from the neglect of break times stipulated by manage-
and hit a toe. Rubber gloves are not worn around chemicals. ment
Thermally insulated gloves are not used to handle hot objects. Financial pressures dictating long and continuous shift
Work paths clearly marked are not followed. Restricted areas hours beyond the safe limit of a worker
are not observed. Thus, the following conditions constitute
hazards: Electrical Hazards and Their Prevention
Unmarked walk paths to guide movement of workers Various electrical hazards are considered and safety practices
within a plant are suggested:
Heavy tools dropping on the floor from the workbench
Any electrical installation of 600 V and above should be
Dropping objects around construction sites guarded with a physical barrier to keep out unqualified
Very high noise levels around workshops persons or unskilled personnel.
Extremely bright lights from welding machine A second person capable of helping in case of emergency
Exposed conveyor belts must be present when one person is working on a live
Objects falling from overhead crane line.
Absence of proper warning against radiation-intensive Lines and electric equipment are assumed live, unless pos-
area itively established as deenergized.
Existence of radiation sources Operating voltages of all equipment must be known before
attempting to work on them.
Lack of proper skill required to operate heavy road con-
struction equipment If the nature of the system requires it to remain energized,
as when troubleshooting a circuit or if deenergizing in-
Lack of adequate training necessary to properly carry out
terferes with the operation or proper function of a safety
a prescribed function in a factory or workshop
system, then the system can remain energized, so long
Toxic, odorless fumes in accessible areas as the personnel are competent and aware that it is live.
Absence of or inactive fire extinguisher Only qualified personnel should be allowed to switch on or
Absence of clearly marked and visible exit signs switch off any live system.
Viruses from sick or dirty animals in animal clinics
Bacteria from improperly disinfected or nonsterilized nee- The primary safety procedure is to deenergize the parts of
dles and syringes the system exposed to the worker. This eliminates the hazard
Spread of bacteria from contaminated foods of shock, arc, and blast. Thus, before working on lines,
grounding jumpers such as are shown in Fig. 1 (7) should be
used to bridge a deenergized line. These conductors, also
Table 5. Nominal Resistance Values for Various Materials called safety ground, ensure that an accidental reenergizing
of the system will not cause injury, short-circuiting and
Material Resistance
grounding of conductors throughout the service.
Rubber gloves or soles ⬎20 M⍀ Figure 2 (8), in which Rj is the resistance of the ground
Dry concrete above grade 1–5 M⍀ wire (typically 0.001 ⍀), shows two applications of ground
Dry concrete on grade 0.2–1 M⍀ wires. The short-circuit bridge intentionally placed across
Leather sole, dry, including foot 0.1–0.5 M⍀
lines for safety must be removed before switching on the lines.
Leather sole, damp, including foot 5–20 k⍀
A second opinion for line clearance should be obtained before
Wet concrete on grade 1–5 k⍀
energizing. Protective devices and specially designed flame
632 SAFETY SYSTEMS
Tower Rg
ground
(b)
Rectifier assembly
Grid lights 5 10 5 15 20 10 15 20 20 30 5 5
Resistance heater
E-1 E-2 D-1 D-2 B-1 F-1 L-2 X-1 X-2 X-3 W/R-1
C-101 tube
on pump 1A
C-101 tube
on pump 1B
Misc. load system
Indicates auxiliary
circuit interconnection
440 BUS 5
Contact load dispatcher
* before operating
BUS 5 Circuit breaker
Figure 4. Safety electrical one-line diagram (SEOLD). (Courtesy Cadick Professional Services.)
first in Great Britain and then in the United States and other Computer-Aided Manufacturing
countries. Fire insurance and accident insurance companies
Programs, which are stored in microprocessors, are mounted
made efforts to enforce safety rules and to educate the public.
on machine tools and used to perform drilling, boring, tap-
Factory inspectors and inspectors from fire insurance and
ping, reaming, and other metal-cutting functions without the
casualty insurance companies carried on a campaign against
need of a human operator. CAM became popular with its ap-
unsafe conditions and actions, and at the beginning of the
plication in the fabrication of large-scale integrated circuits.
20th century a new branch of engineering developed, devoted
to finding and eliminating such hazards. Laws concerning
workers’ compensation were passed in Germany in 1884, Group Technology
Great Britain in 1897, and the United States in 1908. By plac- This is a modern approach to manufacturing technology
ing the financial burden of caring for injured workers on the which increases productivity by reducing delays in material
employer, such laws created an incentive for providing safe handling process, including the waiting time (delay) suffered
machinery and working conditions and for improved selection by a job in order to go through a specific manufacturing
and training of employees. In the United States, the National phase.
Safety Council was formed in 1913. This noncommercial orga-
nization has since been a leader in accident-prevention activi-
ties, especially in the publication of educational literature; the Flexible Manufacturing
compilation of statistics; and the coordination of safety in This is the ability to manufacture different finished products
schools, clubs, industrial organizations, and state and munici- and adjust to changes in product design in response to mar-
pal agencies. ket needs.
which can be used in areas such as arc welding, spot welding, MOTOR-OPERATED VALVES
and paint spraying.
Modern nuclear power stations are very complex. By compari-
son, approximately 40,000 valves are needed in a single US
Advanced Manufacturing System nuclear plant, while an oil- or coal-fired plant of a similar
The need to mass-produce products of uniform quality was size requires only 4000 valves (22). The motor operated valve
the drive behind the move by industries to automate manu- (MOV), actuated by a valve actuator motor (VAM), is one of
facturing processes; automated manufacturing is based on such valves. They play a very indispensable role in nuclear
machines designed to perform predetermined manufacturing power plants. They are called upon to perform safety-related
functions. Originally, these machines were not flexible and services in case of nuclear accidents. Hence, all efforts must
were not easily adaptable to accommodate changes in product be made to ensure that after their installation they remain in
design. Their high cost and the need for a more flexible sys- good working condition, ready to function when called upon
tem led to the use of robots. Robots are capable of performing to do so. Since 1973, failures in nuclear power plants have
a variety of manufacturing functions at a lower production been dominated by valve failures, 34% (23); followed by in-
cost because they can be reprogrammed. Basically, an indus- strumentation, 16%; pumps, 8%; control rods, 8%; and diesel
trial robot is a general-purpose, computer-controlled manipu- generators, 7%. Miscellaneous and human failures formed the
remaining contribution. It is the combination of valve failure
lator consisting of several rigid links connected in series by
and the occurrence of a nuclear accident that spells a nuclear
joints. Each joint–link pair produces one degree of freedom.
disaster. Thus, the MOV plays a critical role in the safe opera-
Motion of the joints results in the relative movement of the
tion of nuclear power plants.
links. The assembly rests on a base and tools are attached to
In general, MOVs are not continuously rated machines be-
the free end. The robot uses the tools to perform assembly
cause they do not operate continuously. Rather, they are
tasks. In essence, it is composed of an arm, a wrist, plus a
short-time-rated, high-torque machines, which operate for
tool designed to reach jobs within its area of operation. The
short periods of time when they are engaged to open or close
wrist unit consists of three rotary motions namely pitch, yaw,
a valve. They are not National Electric Manufacturer’s Asso-
and roll. It is the combined effect of these motions that en-
ciation (NEMA) designed motors, which are rated in horse-
ables the robot to orient the tools attached to its arm to suit
power, thus implying continuous operation. Instead, they are
the configuration and placement of the job. An industrial ro- torque-rated, in pound-foot, for a given duration usually in
bot typically has six joints, which provide six degrees of free- minutes.
dom as illustrated by the Cincinnati Mailacron T3, and the
Unimation PUMA 560 shown in Fig. 5 (20).
In a broader sense, a robot has three components: a me- Special Properties of Valve Actuator Motors
chanical unit comprising rigid bodies or links connected by VAMs are not ordinary motors. They are usually manufac-
joints, a power supply, and a controller. To pick up an object, tured to specifications of the user. They are used for valve
a robot arm responds to the force of its actuators. Servomech- controls and are usually furnished in weatherproof, explosion
anisms are used to exercise control via continuous feedback proof, or submersible enclosures. They may run for only 30 s,
on the actuators. Vision with TV cameras and hearing with just for the time required to stroke the valve. They are widely
microphones gives robots some artificial intelligence, because used in nuclear power plants. Their primary function is to
they are now equipped with eyes and ears. The most powerful open valves for water intake in case of emergency or a nuclear
robot sensory capability is vision, which is commonly referred accident. Being high-speed high-torque machines, VAMs are
to as machine or computer vision. It can be subdivided into made either from an induction motor with special design or
the six principal areas (21): sensing, preprocessing, segmenta- from a dc compound wound motor. These two types of motor
tion, description, recognition, and interpretation. Each of possess high starting torque capabilities because of their de-
these areas is an expert area of study. sign. Special rotor construction is needed in order for the in-
Yaw
duction motor to be suitable as a VAM. In VAMs the locked 3. Overall or Unit Ratio: The unit or overall ratio (OA) is
rotor torque may be as high as 3 to 5 times that of an equiva- used to compute the motor torque so as to select motor
lent NEMA design B motor of the same nominal rating (24). size. The unit OA can only be computed after the stem
Such high levels of torque are achieved while keeping the torque determined from Step 2 is used to select the
physical size of the motor small in order to reduce its inertia proper HMB- or SMB-type Limitorqe valve controls. By
by the following means: using the stem torque and SEL-9, the appropriate value
control is picked:
1. Increasing the flux level in the motor
motor design (or rated) speed
2. By using special materials, such as magnesium which OA = (8)
has a high resistivity acutator or stem rpm
3. A combination of 1 and 2 stem speed
actuator or stem rpm = (9)
stem lead
Some of these ac motors are manufactured by Reliance Elec-
tric, Cleveland, OH. Similarly, Peerless Electric Division, of 4. Calculated motor torque (MCT) is needed in order to
H.K. Porter Company, Inc., Warren, OH, manufactures dc select the proper motor size. This is given by:
motors. Both companies provide motor performance curves.
These curves are used by Limitorque Corporation to properly ST
MCT = (10)
match their actuators to the proper motor. The company does OA × EFF × AF
not manufacture the VAMs. Both the motor and actuator are
assembled into a homogeneous unit. The projected life of a where ST is the stem thrust from Step 1, and EFF is
Limitorque actuator is 40 years if operated at ambient tem- the pullout efficiency, which is determined by the manu-
perature. All motors are furnished with ball bearings and pro- facturer through calculation and confirmation by test. A
vided with grease seals. No lubrication of these motors is nec- typical value of unit pullout efficiency is 0.4 (obtained
essary, since they are lubricated at the factory for lifetime from actuator manufacturer). AF is the application fac-
operation. All three phases of ac motors are of squirrel cage tor and it is used to give a conservative estimate of the
design, and dc motors are compound-wound. MOV torque. It is also used to apply a reduced voltage
for some motor calculations. A sample value for, AF is
Selection of a VAM 0.9.
5. Selection of motor: The calculated motor torque from
The correct selection of a valve actuator motor is critical to Step 4 is a guide used to select the next available motor
its ability to seat and unseat a valve in order to perform the from SEL-9. Thus, conservatively, the motor to be se-
safety-related function in case of a nuclear accident. Since the lected is of the next higher torque size. For example, if
failure of a VAM is a hazard that may be attributed to an MCT ⫽ 51 ft ⭈ lbs, then choose the next higher motor
improper choice of actuator motor, a descriptive outline of the size, which is the 60 ft ⭈ lb, 1800 rpm motor from SEL-
procedure, which will lead to a correct choice, is given. The 9. Similarly if MCT ⫽ 30 ft ⭈ lb, then choose the 40 ft ⭈ lb
five major types of valves commonly used are the gate valve, torque motor.
the glove valve, the plug valve, the ball valve, and the butter-
fly. Only the gate and glove valves are considered since the Safety Guide
selection of other types of valves follow a similar procedure.
Typically, all standard gate and globe valves require maxi- Although the motors are lubricated for lifetime operation, it
mum stem thrust/torque to seat and unseat the valve against is advisable to check the lubricant every six months. It is also
a given differential pressure. By using the Limitorque selec- important to see that the commutator brush on dc motors is
tion guide (SEL) (25), the steps are as follows: clean and operates freely. It cannot be over-emphasized that
proper size of wires should be used to insure against a large
1. Maximum thrust, or stem thrust (ST). The manufac- voltage drop at the terminals when the motor starts.
turer usually provides ST. If not, this is given by:
Standards
ST = TDP + SP + SL (5) There are no unified standards by an industrial body, like
TDP = DP × A × FV (6) NEMA, or from an educational and professional organization,
like IEEE, to regulate the world of MOVs. However, there
where TDP is the thrust due to differential pressure, A are some guides that have been developed by the American
is the seating area, DP is differential pressure, FV is National Standard Institute in conjunction with the IEEE
the value factor given by SEL-3, SL is the stem load or (26) and the US Nuclear Regulatory Commission (27–29) on
piston effect, and SP is the stem packing or stuffing box the protection of MOVs used in nuclear generating stations.
load obtained from SEL-3. Because of the absence of a unified standard, the design, man-
2. Stem torque: The next step is to determine the stem ufacture, and operation of MOVs are left to the judgement of
torque from ST obtained from Step 1. The stem torque the engineer and end user. One direct result of this is a varia-
is given by: tion in standard and application of MOVs from one nuclear
power plant to the other. For example, in some cases, the
stem torque = ST × FS (7) thermal overload relay (TOL) is bypassed in the protection
of MOVs, while in some others they are not. In fact, many
where FS is the value stem factor. investigators (30–32) have been alarmed by this lack of uni-
SAFETY SYSTEMS 637
fied standards and have pointed out the dangers presented by points accurately and correctly. Another was the difficulty of
such practices. providing proper thermal overload relay protection for the in-
termittent duty actuator motor.
Constraints for Starting and Operation The most significant electrical cause of MOV failure is the
overheating in the stator or rotor of the valve actuator motor
MOVs are high-torque high-current starting and short-time-
during locked rotor condition (35). It accounts for 90% of the
run motors. The usual requirements of maximum torque at
reported failures. The overheating caused by the stalled mo-
starting (zero speed) with reduced starting current that ap-
tor condition results in a high rate of temperature rise in the
plies to all types of motors are also desired for MOVs. Nearly
stator as well as in the squirrel cage rotor. It was found that
all squirrel-cage induction motors are capable of starting at
squirrel cage rotors made of magnesium alloy had a signifi-
full rated voltage without being damaged. However, their
cantly larger failure rate than those with aluminum rotors.
starting current is so high and the power factor so low that
The reason is that although the use of magnesium rotors pro-
the power supply may be adversely affected by an excessive
vided the much-needed high starting torque, they are suscep-
voltage dip when the motor is started. For MOVs, the dc mo-
tible to cracking when overheated by sustained locked rotor
tor can be started at 70% rated voltage, while the ac motor
current. Also, they are susceptible to corrosion when installed
will start at 80% rated voltage. While the 70% figure is well
in a hot, humid environment. This weakness is due to the
accepted for the dc motor, the 80% figure is not yet well estab-
brittle nature of the magnesium alloy with its relatively low
lished for the ac motor. At the upper limit, both types of mo-
melting temperature and large galvanic potential between
tors may operate satisfactorily at voltages 10% above their
magnesium and steel used in the rotor lamination.
rated values.
Protection of MOV
Failure of MOVs
Measures that can limit or reduce the failure of MOVs include
In nuclear power facilities, the concerns for radiological expo-
the following:
sure, fast emergency system operation, and the requirement
to bring the plant to a quick and safe shutdown point to the • Operator training
indispensable and critical role that the MOV must play. Al-
• Correct operational procedures
though everything possible is done to minimize the failure of
MOVs, an Institute of Nuclear Power Operations (INOP) • Preventive maintenance
Study (33), which investigated 644 different Licensee Event • Valve/actuator matching
Reports (LERs) submitted to the Nuclear Regulatory Commit- • Periodic inspection of magnesium rotors
tee (NRC) between 1974 and 1982, showed that failure of
MOVs has indeed consistently occurred. The failures are at- Any protection strategy for MOVs must be primarily tempera-
tributed to mechanical, electromechanical, electrical, and mo- ture sensitive because excessive heat is the major cause of
tor control circuits (MCC) causes. Their analysis, shown in motor burnout. Since MOVs are specialized intermittent duty,
Table 7 indicates that 32% of the documented failures were high-torque, high-slip motors, wide fluctuations in current are
due to electromechanical torque switches and limit switches. their characteristics, and they are designed to withstand
Since the summary also shows that 22% of the failures were them. Hence, protectors designed to be current sensitive are
mechanical in nature, then 54% of the total MOV failures an- likely to stop a motor that is functioning normally or fail to
alyzed were due to electromechanical components within the stop a motor that is overheating. Basically, three types of pro-
valve actuator. tectors are available for overload protection. They are:
Further evidence showed that the documented failures
could be attributed largely to MOV hardware, equipment de- 1. Internal devices located on the stator windings: These
sign, application, operation, and maintenance practices. How- are internal temperature sensors like thermistors or
ever, the biggest cause of failure was the torque switches and thermally actuated contacts located on the stator wind-
the position limit switches. For this reason, Electric Power ing. They are good for motors that are stator tempera-
Research Institute (EPRI) funded EPRI NP-4254, Project ture limited, but are ineffective for motors that are rotor
2233-2 (34) titled ‘‘Improvements in Motor-Operated Valves’’ temperature limited. MOV motors are classified as ei-
in order to find remedies for the shortcomings in the design ther rotor limited or stator limited. If the temperature
of the MOV. Other investigations on causes of MOV failures of the rotor reaches its allowable limit before the stator
across nuclear power plants showed that one major difficulty reaches its own allowable limit, then the motor is re-
was the problem of setting up the torque and limit switch set ferred to as rotor limited. However, if the temperature
of the stator reaches its allowable limit before the rotor
reaches its, then the motor is said to be stator limited.
Table 7. MOV Failure Analysis—INPO Report 83-037 (12) They are vulnerable to failure due to heavy vibration
found in some actuator applications and are not easily
Type Percent accessible, thus creating maintenance difficulties.
Mechanical (failure to operate, bent stems, damage to 22 2. External devices actuated by motor currents: These are
valve seats, gear binding and damage) bimetallic or eutectic thermal overload relays in the mo-
Electromechanical (torque switch failure, torque switch 32 tor–starter circuit and the thermal–magnetic trip ele-
adjustment, limit switch adjustment)
ment in the circuit breaker. They provide protection for
Electrical (motor, contacts, MCC and others) 27
overload and locked rotor conditions. They usually con-
All others (vibration, wear, other) 19
sist of the current-carrying portion, which produces the
638 SAFETY SYSTEMS
heat, and the tripping mechanism, which is actuated as with specialized rotor design or from a compound-wound dc
a result of the heat. motor. Since it is required to drive a valve to a fully close or
3. Combination of internal and external devices. fully open position, the MOV is fitted with position limiting
switches to stop the unit at the full open and full close posi-
More protection is provided by a combination of 1 and 2. tions. In addition to position limiting switches, valve motor
Whichever protection mechanism is selected, it must pro- actuators are equipped with a torque-limiting device. This
tect the motor from the following: torque switch is designed to limit the output torque of the
actuator to the peak torque required by the valve usually to
• Motor overheating due to locked rotor conditions seat or unseat it.
• Motor overheating due to anticipated overloads While some manufacturers provide internal temperature
• Nuisance trips during acceleration sensors located on the stator windings, the most common
mode of protection is through current-sensing devices located
• Nuisance trips due to anticipated overloads
external to the device. The industrial or utility approach is to
• Nuisance trips during operation within the duty cycle of use thermal overload relays (TORs) as part of the combina-
the valve tion starters provided in the MCC.
It is most essential to reduce the failure rate of MOVs be-
Maintenance of MOV
cause of the importance of their function in nuclear power
It is known that in many cases, MOV failures result from plants, especially in case of nuclear accident. To this end, the
inadequate training of personnel and the failure to implement Licensee Event Reports (LER) between 1974 and 1982 were
existing maintenance schedules. Therefore, preventive reviewed (33) in order to identify the causes of the docu-
inspection/maintenance based on time or cycle provided by mented MOV failures. It was found that up to 54% of the
the valve manufacturer should not be ignored or overlooked. failures were electromechanical in nature, while 27% were
The manufacturer’s recommendations should be followed. Use electrical. The report also indicated that a significant number
of lubricants specified by the manufacturer and careful atten- of MOV failures were due to improper operator training.
tion to quantity, quality, and consistency must be adhered to. In order to ensure that MOVs remain in a sound functional
The Limitorque Corporation recommends the following main- condition, they must be routinely maintained as recom-
tenance practices for their actuators: mended by the manufacturers. Some of the maintenance tips
include lubrication of the main gearbox and geared limit
Lubricate main gearbox and geared limit switch. switch and inspection of the stem and stem nut system and
Do not fill the actuator to 100% capacity. Leave an air lubrication as necessary. Note that for lubrication purposes,
space in the actuator to allow for thermal expansion of the quantity and quality of grease recommended by the man-
the lubricant. ufacturer must be adhered to. Finally, operator training is
Check shaft penetrations for seal leakage. Note that some necessary in torque switch and limit switch setting, as well
oil leakage is expected and acceptable. Replace seals if as in the protection and control of MOVs in order to reduce
abnormal grease leakage occurs. human error in their utilization.
Remove moisture if found in the limit switch compartment.
Ensure cleanliness of electrical contacts and check ter- REGULATIONS AND SAFETY STANDARDS
minal connections for tightness.
Inspect stem and stem nut. Internal and external wiring Regulations have been established by government agencies to
inside connection compartment should be checked for guide the safety practices of its people. These days everyone
damaging abrasion cuts or distortion of conductor insu- must conduct his or her work in accordance with this statu-
lation. tory legislation. Some of these standards organizations and
function are:
Summary
MOVs are commonly used in direct gear-driven valve actuator National Electric Code. National Electric Code (NEC) is
assemblies in nuclear power plants. They are short-time one of the oldest, and was first developed in 1897. It
rated, high-torque, high-slip motors. They are rated in torque sets standards that cover industrial, commercial, and
(i.e., in lbs. ⭈ ft.) and not horsepower (hp) as in conventional residential electric systems to help minimize the possi-
motors. For normal operation of the valve actuator, the motor bility of electrical fires.
may run up and down the torque scale from light load to stall American Society for Testing and Materials Standards. The
torque. The corresponding horsepower follows this same pat- American Society for Testing and Materials (ASTM)
tern. Thus, the horsepower fluctuates accordingly, making it Standards produces standards for safety equipment de-
useless as a means to rate the MOV. sign, usage, and testing.
Industry or professional society standards do not cover mo- Occupational Safety and Health Administration.
tors employed as direct gear-driven valve actuator motors. Occupational Safety and Health Administration
For this reason, there are no universally accepted methods of (OSHA), an agency of the US Department of Labor, was
protecting the motor. Therefore, the design, manufacture, and established by an act of Congress in 1970. Its main re-
utilization of the MOV are left to engineers, manufacturers, sponsibilities are to provide for occupational safety by
and end-users. reducing hazards in the workplace and enforcing man-
Because MOVs require a very high starting torque, they datory job safety standards and to implement and im-
are manufactured from either a squirrel cage induction motor prove health programs for workers. OSHA regulations
SAFETY SYSTEMS 639
CLARENCE N. OBIOZOR
University of North Florida
ISDN 721
ISDN
WHAT IS ISDN?
than today’s fastest analog modems. ISDN uses the existing ISDN line can transmit information almost 10 times faster
phone line and provides up to 128 kbyte/s of bandwidth for than a standard 14.4 kbyte/s modem. BRI provides service to
information transfer. Compared with a 14.4 kbyte/s analog Centrex multiple customers.
modem, it becomes clear how much easier ISDN can provide
for bandwith-intensive applications—such as uploading or BRI Applications. Some of the typical ISDN BRI-based ap-
downloading large files, video conferencing, and surfing the plications include:
World Wide Web. ISDN uses the digital binary language of
computers, ones and zeros, to transmit voice, data, and video
• High-speed access to the Internet or online services
over existing copper wire telephone lines. By replacing much
larger and slower analog signals with faster integrated digital • Telecommuting/work at home
packets, ISDN squeezes much more potential out of the com- • Desktop video conferencing
puter and phone line. ISDN also allows us to use a single • Businesses needing the voice capabilities of Centrex and
phone line for data transfer as well as voice or fax calls at the the speed and bandwidth of ISDN
same time. For example, we can talk on the phone and send
a fax simultaneously at speeds of up to 64 kbyte/s. If we are Technical Capabilities. The technical capabilities of ISDN
not engaged in a voice call, however, the entire bandwidth BRI are:
(128 kbyte/s) can be dedicated to a high-speed data connec-
tion for video conferencing or transfer of files to and from a
• End-to-end digital transmission over ISDN-compatible
remote database. ISDN uses the same standard phone jack
network facilities
as current technology. It is an inexpensive, convenient, reli-
able way to speed data communications and provide high- • Standard interface open systems interconnection (OSI
speed access to the Internet. model levels 1–3)
D-Channel (OSI L1-L3)
FUNDAMENTALS OF ISDN TECHNOLOGY B-Channel (OSI L1)
Supports up to eight ISDN terminals on a single ISDN
Announced in the early 1980s, ISDN progressed from stan- BRI line
dards committees to technology trials to successful applica- • Circuit-switched B-channel data at speeds up to 64
tions. Using a standard interface, ISDN provides high-speed kbyte/s
simultaneous digital transmission of voice, data, and video,
• Standard rate adaption (V.120) for sub-64 kbyte/s speeds
which allows for universal connectivity to the public tele-
phone network. • Packet-switched data over B- or D-channels
Each ISDN line is made up of 64 kbyte/s separate ‘‘chan- • D-channel at 9.6 kbyte/s
nels’’ that send and receive calls. In addition, it provides a • X.25 packet multiplexing over B-channel, for example,
channel that is used primarily for signaling. Standard ISDN multiple virtual D packet sessions on a single B-channel
basic rate interface (BRI) lines have 2 B-channels, and more • Calling party number delivery on incoming calls
powerful ISDN PRI lines have 23 B-channels. Each channel
can be used separately for any communications task, includ- • Dynamic service selection over B-channels
ing voice calls, faxes, and data transmission. The channels • Out-of-band D-channel signaling
can also be combined for information-intensive applications
like video conferencing. There are two basic types of ISDN Primary Rate Interface (PRI) Technology
service:
ISDN PRI service is well suited to meet the needs of small
• Basic rate interface (BRI)—(2 B-channels) and large companies. PRI has 23 B-channels with a capacity
of 64 kbyte/s, each for simultaneous transmission of voice and
• Primary rate interface (PRI)—(23 B-channels)
data, and one 64 kbyte/s channel that is used only for net-
work management functions.
Basic Rate Interface (BRI) Technology
The 23 B-channels can be used for any mix of voice, virtual
ISDN BRI service is the easiest and most inexpensive way to private lines, packet-switched data, circuit-switched data, or
use superfast technology to dramatically increase data trans- video. B-channels can be combined for increased bandwidth
mission capabilities and be able to transmit data or faxes and when higher speeds are required.
talk on the phone at the same time. ISDN BRI, also called ISDN PRI uses four-wire T1 technology to provide reliable,
2B⫹D, provides simultaneous integration of voice, data, and high-speed switched connections to the public network. PRI
video over ordinary twisted-pair telephone wiring. BRI sup- allows call-by-call selection for access to a variety of commu-
ports two bearer (B) channels at 64 kbyte/s, each for trans- nication facilities, potentially reducing trunking requirements
mission of voice and data. These B-channels can be used indi- by as much as 33%. Used as a tie line, PRI provides feature
vidually; two voice or data calls can be made or received transparency between like-vendor Centrexes/private brands
simultaneously. The B-channels can be combined into one su- exchanges (PBXs).
perfast channel for high bandwidth applications like video Signaling System 7 (SS7) is the standard that provides the
conferencing. BRI also has a D-channel for signaling between virtual network services for ISDN PRI. It delivers the infor-
the ISDN equipment and the phone company at 16 kbyte/s. mation used with customized call identification and selection
The D-channel can also be used for packet switched data. features as well as high performance call-setup features. SS7
With a total communications capacity of 144 kbyte/s, a single is associated with ISDN because it plays an increasingly im-
ISDN 723
portant role in inter- and intracarrier communications. The wire S/T interface. The S/T interface supports multiple de-
portability of 800 numbers among service carriers, for exam- vices (up to seven devices can be placed on the S/T bus) while
ple, is an SS7 application. it is still a full-duplex interface; there is now a pair of wires
to receive data, and another to transmit data. Today, many
PRI Applications. Some of the typical ISDN PRI-based ap- devices have NT1s built into their design. This design con-
plications include: tains inexpensive and easier-to-install devices, but often re-
duces flexibility by preventing additional devices from being
• Data, video, and voice network integration connected.
• PBX-to-PBX connectivity for ISDN feature commonality Technically, ISDN devices must go through a termination
• PBX-to-Central Office (CO) connectivity for trunking 2 (NT2) device, which converts the T interface into the S in-
• Inverse multiplexer-to-inverse multiplexer connectivity terface. (Note: The S and T interfaces are electrically equiva-
for video conferencing lent.) Virtually all ISDN devices include an NT2 in their de-
sign. The NT2 communicates with terminal equipment and
• LAN-to-LAN bridging or routing
handles the layer 2 and 3 ISDN protocols. Devices most com-
• Imaging, for both client/server and computer aided de- monly expect either a U interface connection (these have a
sign (CAD) application built-in NT1) or an S/T interface connection. Devices that
• Video conferencing connect to S/T (or S) interface include ISDN capable tele-
• Connecting PBX systems to the public switched tele- phones and fax machines, video teleconferencing equipment,
phone network (PSTN) for voice transmission bridges/routers, and terminal adapters. All devices that are
• Connecting cluster controllers to PSTN for data trans- designed for ISDN are designated terminal equipment 1
mission (TE1). All other communications devices that are not ISDN
capable, but have a standard telephone interface (also called
• Connecting LANs and hosts
the R interface), including ordinary analog telephones, fax
• Connecting multiplexers to PSTN machines, and modems, are designated terminal equipment 2
Technical Capabilities. The technical capabilities of ISDN (TE2). A terminal adapter (TA) connects a TE2 to an ISDN
PRI are: S/T bus.
Going one step in the opposite direction takes us inside the
• End-to-end digital transmission over ISDN-compatible telephone switch. Remember that the U interface connects
network facilities the switch to the customer premises equipment. This local
• Standard interface (OSI model levels 1–3) loop connection is called line termination (LT function). The
connection to other switches within the phone network is
D-Channel (OSI L1-L3)
called exchange termination (ET function). The LT function
B-Channel (OSI L1) communicates via the V interface.
Supports up to eight ISDN terminals on a single BRI
line Layer 1—Physical Layer
• Circuit-switched B-channel data at speeds up to 64 The U interface for BRI is a two-wire, 160 kbyte/s digital con-
kbyte/s nection. Echo cancellation is used to reduce noise, and data
• Standard rate adaption (V.120) for sub-64 kbyte/s speeds encoding schemes (2B1Q in North America, 4B3T in Europe)
• Packet-switched data over B- or D-channels permit this relatively high data rate over ordinary single pair
• D-channel at 9.6 kbyte/s local loops.
• X.25 packet multiplexing over B-channel, for example,
2B1Q. 2B1Q (2 binary 1 quaternary) is the most common
multiple virtual D packet sessions on a single B-channel
signaling method on U interfaces. This protocol is defined in
• Calling party number delivery on incoming calls detail in 1988 ANSI spec T1.601. In summary, 2B1Q pro-
• Dynamic service selection over B-channels vides:
• Out-of-band D-channel signaling
• A very flexible mechanism for call control, network man- • Two bits per baud
agement, security, and so on • Transfer rate of 160 kbyte/s
• Baud rate of 80 kbaud/s
INTERFACE
Bits Quaternary Symbol Voltage Level
In the United States, the telephone company will be providing 00 ⫺3 ⫺2.5
its BRI customers with a U interface. The U interface is a 01 ⫺1 ⫺0.833
two-wire (single pair) interface from the phone switch. It sup- 10 ⫹3 ⫹2.5
ports full-duplex data transfer over a single pair of wires; 11 ⫹1 ⫹0.833
therefore, only a single device can be connected to a U inter-
face. This device is called a network termination 1 (NT1). The This means that the input voltage level can be one of four
situation is different elsewhere in the world, where the phone levels. (Note: Zero volts is not a valid voltage under this
company is allowed to supply the NT1, and thereby the cus- scheme.) These levels are called quaternaries. Each quater-
tomer is provided an S/T interface. The NT1 is a relatively nary represents 2 data bits, since there are four possible ways
simple device that converts the two-wire U interface into four- to present 2 bits, as in the table above.
724 ISDN
Frame Format. Each U interface frame is 240 bits long. At Information—layer 3 protocol information and user data.
the prescribed data rate of 160 kbyte/s, each frame is there- CRC (2 octets)—cyclic redundancy check is a low-level test
fore 1.5 ms long. Each frame consists of: for bit errors on the user data.
Flag (1 octet)—this is always 7E16 (0111 11102).
• Frame overhead—16 kbyte/s
• D-channel—16 kbyte/s SAPI. Service access point identifier (SAPI) is a 6 bit field
• 2 B-channels at 64 kbyte/s—128 kbyte/s that identifies the point where layer 2 provides service to
layer 3.
Sync 12* (B1 ⴙ B2 ⴙ D) Maintenance
SAPI0 Description
18 bits 216 bits 6 bits
0 Call control procedures
• The sync field consists of nine quaternaries (2 bits each) 1 Packet mode using Q.931 call procedure
in the pattern ⫹3 ⫹3 ⫺3 ⫺3 ⫺3 ⫹3 ⫺3 ⫹3 ⫺3. 16 Packet mode communications procedures
32–47 Reserved for national use
• (B1 ⫹ B2 ⫹ D) is 18 bits of data consisting of 8 bits from
63 Management procedures
the first B-channel, 8 bits from the second B-channel,
127 Reserved for future use
and 2 bits of D-channel data.
• The maintenance field contains cyclic redundancy check TEIs. Terminated endpoint identifiers (TEIs) are unique
(CRC) information, block error detection flags, and ‘‘em- IDs given to each (TE) on an ISDN S/T bus. This identifier
bedded operator commands’’ used for loopback testing can be dynamic; the value may be assigned statistically when
without disrupting user data. the TE is installed or dynamically when activated.
line, and a suffix that is sometimes used to identify features Dynamic Allocation of B-Channels in a PRI. For practical pur-
on the line. If an ISDN line requires a SPID, but is not cor- poses, combining multiple channels in a PRI for large video-
rectly supplied, then layer 2 initialization will take place but conferences, data transfers, and the like is most often pro-
layer 3 will not, and the device will not be able to place or grammed into the digital switch serving the location.
accept calls. However, new bandwidth-on-demand controllers have begun
to enable a network manager to combine larger bandwidths
Information Field Structure. The information field is a vari- in real time to meet specific needs. They can also monitor
able length field that contains the Q.931 protocol data. quality and traffic on both corporate leased-lines and ISDN
networks and perform dynamic allocation of B-channels to re-
Information field lieve bottlenecks or backup error-prone or damaged lines.
1 2 3 4 5 6 7 8
Protocol discriminator CONNECTING TO ISDN
0 0 0 0 Length of CRV There are three ways ISDN can be ‘‘delivered’’ from an ISDN-
0 Message Type ready digital switch:
Mandatory and optional information elements (variable)
Through a Direct BRI Connection from an ISDN Switch.
One or more standard BRI (2B⫹D) connections can be
These are the fields in a Q.931 header: used to link a company directly to an ISDN-ready
switch in a central office. These lines can connect
Protocol discriminator (1 octet) identifies the layer 3 directly to ISDN equipment in a small office or resi-
protocol. If this is a Q.931 header, this value is always dence or can be connected through an outside connec-
0816. tion.
Length (1 octet) indicates the length of the next field, the Through ISDN Centrex Service. One or more BRIs can also
CRV. be linked to ISDN Centrex service. This arrangement
Call reference value (CRV) (1 or 2 octets) is used to offers several advantages for an individual or company.
uniquely identify each call on the user–network inter- Since the ISDN switch functions as the switching sys-
face. This value is assigned at the beginning of a call, tem, the company does not have to own or maintain a
and this value becomes available for another call when PBX or key system. It also offers a low-cost, virtually
the call is cleared. unlimited growth path.
Message type (1 octet) identifies the message type (e.g., Through a PRI Connection. A PRI delivers 23 B-channels
SETUP, CONNECT). This determines what additional plus one D-channel from the telephone company to the
information is required and allowed. PBX or other control device, which then distributes the
B-channels as needed throughout an organization. How
Mandatory and optional information elements (vari-
able length) are options that are set depending on the this configuration is set up can vary greatly. Users with
message type. heavy data traffic, for example, might configure the con-
nection through an ISDN router, multiplexer, or con-
troller rather than PBX, reducing the chance of conges-
Layer 3 Call Setup. These are the steps that occur when an
tion through the switch.
ISDN call is established. In the following example, there are
three points where messages are sent and received: (1) the
caller, (2) the ISDN switch, and (3) the receiver. ISDN AS A NETWORK CONTROLLER
1. Caller sends a SETUP to the switch. ISDN offers a command-and-command structure that can ac-
2. If the SETUP is OK, the switch sends a CALL PRO- tually mold the public phone network to individual need. In
Ceeding to the caller, and then a SETUP to the receiver. effect, the D-channel command language can transform the
3. The receiver gets the SETUP. If it is OK, it then rings network itself into:
the phone and sends an ALERTING message to the
switch. 1. A virtual PBX that has no boundaries, creating what is
truly a virtual office
4. The switch forwards the ALERTING message to the
caller. 2. An automatic call distributor, or more accurately a call
optimizer, that logically directs incoming calls to agents
5. When the receiver answers the call, it sends a CON-
worldwide
NECT message to the switch.
3. A user-controlled link between computers and voice ap-
6. The switch forwards the CONNECT message to the
plications anywhere
caller.
7. The caller sends a CONNECT ACKnowledge message There are several companies today offering computer-based
to the switch. systems that allow users to communicate with the telephone
8. The switch forwards the CONNECT ACKnowledge mes- network in the actual command language used by the net-
sage to the receiver. work itself. The practical result is that companies with many
9. Complete. The connection is now up. offices statewide or even nationwide can link them all into a
726 ISDN
single, responsive telephone center. Calls can be routed to the Network Congestion
best person, regardless of location. At the same time, calls to
When a user dials in the Internet or Intranet via an analog
a local office where everyone is busy can be seamlessly for-
modem, a set of telephone network resources are reserved for
warded to the nearest site with someone available, or even to
that user’s exclusive use during the call, whether or not any
agents working at home. A distributed call center gives users
data are being transmitted. Those facilities include talk paths
an enormous ability to shape and manage the network itself
within the telephone switches at the originating and termi-
through the D-channel. A regular customer, for example, can
nating ends of the call, trunks that interconnect the telephone
be routed to a specific sales agent, while someone needing
switches, a port at the ISP or corporate location that the user
specialized assistance, as entered in initial prompts, can be has called, and, of course, the user’s own phone line.
directed to an appropriate specialist. Distributed calls centers With ISDN and Always On the packet network maintains
can also balance work loads, centralize after-hour calling, and a set of pointers directing the flow of packets to and from the
work around service interruptions at any site. A range of cus- user. However, no other network resources are used except
tomer options can program call redirection. when data are actually being transmitted; thus, the network
can easily support many simultaneous users with the same
facilities. The user can actually use two B-channels for other
THE POWER OF PACKET SWITCHING telephone calls while the Internet or Intranet connection is
maintained over the D-channel. Many Always On users can
Many users need to be continually connected when they are be simultaneously funneled into the same port at the ISP or
working from home. They require e-mail notification so they corporate location.
can quickly respond. They need regular updates of changing
information such as stock quotes or news headlines. Or, they
Standards
want to be available for an on-line conference that might be
initiated by co-workers at another location. Always On/Dy- The D-channel packet capability is defined in the interna-
namic ISDN (AO/DI) satisfies these requirements by provid- tional standards governing ISDN. Packets are formatted ac-
ing a continual connection to the corporate network (Intranet) cording to X.25, another international standard. The point-to-
or Internet for telecommuters, remote workers, and indepen- point protocol (PPP) and the multilink point-to-point protocol
dent professionals who require Always On connectivity for e- (ML-PPP), which are in widespread use today for Internet ac-
mail and data. Building on the feature-rich ISDN platform, cess and remote dial-up, are also used with Always On. More-
which permits simultaneous voice, data, video, and e-mail on over, the newer bandwidth allocation control protocol (BACP)
a single ISDN connection, D-channel Always On offers a cost- will also work with Always On. In addition, Internet protocol
effective way of maintaining a real-time link without having (IP), IPX, and NetBEUI can also operate with Always On.
a ‘‘dial-up’’ connection to the corporate network or Internet The user’s ISDN terminal adapter must have an X.25 packet
service provider (ISP). Always On/Dynamic ISDN offers the capability and the Internet/Intranet location into which the
best of all worlds—an Always On digital platform with scal- user is dialing must have a network connection to the packet
able connections that is paid for only as it is used. network in order for the user to establish the Always On link.
The availability of local ISDN BRI and PRI services is conserve on electricity bills and equipment wear. Most ven-
growing rapidly throughout the nation. Full digitization of dors offer their equipment with NT1 units, so the user has
many regional networks has been scheduled, and ISDN inter- a choice.
connections to the backbone networks of the IECs are now
proceeding at a rapid pace. It is also possible today to estab- ISDN TAs and NICs
lish ISDN-compatible data links to out-of-state locations not
served by ISDN, by using Switched 56 services. The major ISDN TAs work essentially the same as standard modems.
interexchange carriers, all regional operating companies, and They look the same physically and have the same applications
many specialized international carriers offer Switched 56 ser- for data transfer. Like standard modems, there are all types
vices. These are single-channel, dialed, data-only connections, of TAs: internal, external, and PC cards (PCMCIA).
capable of speeds up to 56 kbyte/s. Virtually all are connected One thing, however, must be considered when deciding to
to the interexchange carriers nationwide digital networks. go with an external TA, even though it is not as much of a
problem as it used to be. The user equipment must have the
serial chip set that can handle the higher speeds of ISDN.
ISDN EQUIPMENT
Nothing less than a 16550UART will do for full ISDN speeds
via external TAs. If possible, having an advanced serial card
Now that we have looked at some of the more popular applica-
and driver is ideal to eliminate any problems and take full
tions of ISDN, we will look at what types of equipment best
advantage of the ISDN speeds. However, most PC 486, Mac
implements these applications, including NT1s, TAs, NICs,
68040, and later models will handle the serial speeds without
bridges, and routers.
any problems.
When choosing a modem, it is important to make certain
NT1s
that the modem supports 2 B-channel operation and the ven-
ISDN abides by a set of standards that define the layers of dor supplies drivers that will allocate the B-channels. It is
contact between the central office and the user’s equipment. imperative to allocate the channels so that the user can take
Each point of contact has a specific function and communi- advantage of ISDN’s channel allocation features. Some older
cates using a different layer of protocols according to the modems do not allow use of the second B-channel.
needs of the user’s equipment. The different points are as fol- ISDN NICs carry many of the same attributes as TAs but
lows: NT1, NT2, TE1, and TE2. NT1 represents the actual function as regular LAN cards would on a 10BASE-T net-
termination of the ISDN circuit to the site and handles the work. Just as a TA appears as another modem to the applica-
communication to the telephone switch for the devices that it tion, an ISDN NIC fools its application into believing that
connects via TDM. That is why a device accordingly called it is talking to a regular NIC. The vendor usually provides
a network termination (NT1) must reside between the special NDIS or ODI drivers for the ISDN NIC that replace
switch interface (called the U interface) and each ISDN- traditional ones used for Ethernet. These drivers handle
compatible equipment interface. This interface is called an the ISDN functions of the card as well as the translations
S interface for equipment such as TAs and/or a T interface from the network to the next layer of protocols such as IPX
for other switching devices such as PBXs; often both are and IP. ISDN NICs are essentially an interface to another
implemented NT1 units, hence the S/T label of the inter- network, ISDN.
face. Other devices that do not abide by the ISDN stan- Which one is better? That naturally depends on user needs
dards, such as analog phones, can be accounted for by and implementation. TAs can be made to perform the same
another interface (R interface) that is not usually imple- functions through software and can serve many other simple
mented into simple NT1 units but is commonly incorporated roles such as plain terminal connections. NICs are often, but
into small office/home office (SOHO) oriented ISDN devices not always, more transparent to the user when initiating and
or more complex NT1 devices. terminating connections and can function as a NIC. TAs are
NT1 units typically connect ISDN-compatible devices to typically more popular than ISDN NICs.
the ISDN line. More complex NT1 units connect noncompati-
ble devices such as analog phones and fax machines as well
ISDN Bridges
as ISDN-compatible devices. NT1 units can be purchased as
external devices or can be purchased as built-in components ISDN bridges do just that. They bridge data from LAN to re-
to ISDN equipment being connected. Every ISDN component mote LAN transparently to the user, just as if the data were
needs an NT1 to operate. There are advantages and a disad- on the destination LAN. Like TAs, ISDN bridges come in in-
vantage to having a built-in NT1 unit. The built-in NT1s are ternal and external varieties and offer built-in NT1 unit op-
usually less expensive than external ones and the user does tions. They are essentially TAs or NICs with bridging firm-
not have to deal with finding a place to put it. The simplest ware that will transfer LAN traffic regardless of protocol
external NT1s are usually about two-thirds the size of a stan- between remote sites. Bridges are a good choice for sites that
dard modem, and they cost around $100 to $150. The disad- run many protocols and operate over relatively short dis-
vantage of internal NT1s is that the user must leave the tances. Bridges offer some packet filtering to help reduce traf-
equipment powered up to use the other devices connected to fic but do not filter out most of the packets generated by
the NT1 device. This often involves leaving the entire com- routers and servers that could keep an expensive long dis-
puter running, so it can power the TA card, and the TA card tance link up.
can power its NT1, so the phone connected to the NT1 can Users must plan carefully before deciding to implement a
operate. This is often unacceptable to companies trying to bridged solution as opposed to routing and choosing which
ISDN 729
product to use. Many bridges take full advantage of stan- network terminating device (NT1) to which the line is con-
dard ISDN channel allocation protocols and compression nected and which must be powered at the customer site. This
algorithms to help manage the connection and minimize means that if the NT1 loses power, the ISDN line will not be
the usage of the ISDN line, reducing the cost of operation. usable. Traditional analog telephone service does not require
It is important to find out which product supports the fil- any terminating equipment other than the phone, which
tering and line management features that are needed for usually does not require commercial power and therefore
the network. functions even when commercial power is absent. For this
reason it is strongly recommended that ISDN be installed
ISDN Routers as a separate line unless customers agree to be responsible
for providing their own power supply in the event of pro-
Among the more complex of the ISDN devices, routers allow
longed outage.
remote sites to take full advantage of ISDN link and traffic
moving between the various locations. Complex LAN config-
The Spectrum of Digital Technologies
urations can be made simple and effective using ISDN
routers. Packets can be routed and filtered according to proto- ISDN is, in fact, one of a number of emerging technologies
col and packet type and can be intelligently routed quickly to designed to exploit the power of the worldwide digital commu-
various sites using either a single BRI connection, multiple nications network. These technologies include asynchronous
BRIs, or PRI connections. transfer mode (ATM), switched multimegabit data service
The router vendors have been hard-pressed over the past (SMDS), and frame relay (all designed for high-speed network
few years by standards for remote access and LAN-to-LAN access through dedicated, leased lines) as well as ISDN and
routing connections and link management. Standards such as Switched 56 services (which offer both dedicated and dialed
multilink PPP and STAC compression are among the most access).
popular of router solutions. These standards define common
protocols for ISDN devices to establish an ISDN link, route The Unique Value of ISDN
specific protocols, allocate bandwidth on demand, and com-
To many users, especially individuals and those in smaller
press data between different vendors’ equipment. ISDN inter-
companies, ISDN is by far the most important of these tech-
operability among various vendors has proved to be a night-
nologies. To literally millions of users, it offers inexpensive
mare in the past. This is important when scalability is a
dialed service, high-speed data transmission, and the ability
factor in choosing a routing solution or when third parties are
to send and receive voice, data, and moving images through
to be integrated into the network solution such as ISPs.
the same fully digital connections. Communications to most
Many of the routers can link remote dial-in users as well
of the major business centers of Europe and the Pacific Rim
as linking LANs to LANs. Many also offer bridging capability
can now be set up with a simple dialed call.
for the protocols that they do not route. There are standards
for each. As a result of the standards effort, many ISDN
ISDN Closes Digital Loop
routers now offer the most cost-effective solution to linking
branch offices that need the bandwidth and low cost. Good The real promise of ISDN, as noted earlier, is in replacing
ISDN routers will tend to cost more than any of the other the dialed analog collection of the past with the dialed digital
ISDN devices already mentioned but will soon pay for them- connections of the future. The ramifications for those who use
selves in cost of operation when linking central LANs and computers are many.
remote LANs.
Aside from single BRI routers, higher-end solutions are • Digital-to-analog conversions will go away. Modems to
available that work with these lower-end BRI routers. Some covert the digital signals of PCs, LANs, and other devices
vendors offer routers with multiple BRIs with the option to to the analog signals needed for the analog network will
add more. Users must weigh the cost of multiple BRI connec- be replaced. Instead, high-speed pulses will flow directly
tions against a PRI and related equipment. Large companies from one digital device to another through a totally digi-
with many remote users and ISPs typically use PRI circuits tal network.
and routers to best manage many users dialing in from many • Many digital connections will become dialed, not dedi-
destinations. PRI is very expensive, but ideal in these situa- cated. It will no longer be necessary to lease expensive
tions. For the rest, BRI can be scalable and cost effective if dedicated digital lines simply to link a high-speed data
planned right. device to the network. The deciding factor will be the eco-
nomics of how fast the line must be and how frequently
ISDN AS THE ONLY LINE OR AS A SECOND LINE it is used.
• Endless connectivity possibilities will open. It will be-
While ISDN is specifically designed to deliver digital connec- come possible to link individuals, networks, and systems
tions through existing copper twisted-pair lines, many that are neither economical nor practical to link today.
smaller locations (and many larger ones as well) will have
to decide whether or not to use ISDN as the only telephone Some obvious examples of these new applications are telecom-
connection, or to install it as separate, second line. muting, LAN-to-LAN or LAN-to-host interconnection, high-
It is important to know that in the United States, the Fed- speed (greater than 28.8 kbyte/s modems) Internet access,
eral Communications Commission (FCC) chose to implement remote image sharing and retrieval; inexpensive teleconfer-
ISDN in a way that requires users to supply at their site a encing, combined voice and document or image collaborations,
730 ISDN
high-speed access to remote files or databases, and acceler- Note that often these dialed connections can be quite brief,
ated links to remote libraries and research systems. which means that many at-home workers can share the same
While most of these applications are technically possible ISDN channels installed at a network or host, keeping up-
today without ISDN, very few are practical, either because front costs reasonable. Moreover, efficient LAN or host access
costs are too high or transmission speeds are too slow. The can also be important to someone who is traveling, who is ill
purpose of this article, in fact, is to look at these innovative and at home, or who is spending the day at another location.
uses. Many are important breakthroughs, while others are With ISDN, they have not only better telephone contact with
merely helpful conveniences. But all in concert show beyond co-workers and messaging systems, but efficient laptop access
a doubt that an important new technology has begun to take to their LAN or file server.
its place on the stage of teleconferencing.
LAN-to-LAN Links
The Increase of User’s Data Speeds
A typical LAN-to-LAN configuration is used, and charges are
Logarithmic views show how transmission speeds available incurred only when data are sent. One of the most immedi-
to individual users have accelerated in the past two decades. ately accepted and widely used applications of ISDN is in
Digital speeds should continue to grow, while analog speeds linking LANs to each other and to the outside world. Informa-
begin to peak. tion in most local area networks travels through dedicated
fiber-optic or coaxial cables at speeds of 10 Mbyte/s to 30
Mbyte/s, which means that, contrary to a common misconcep-
WORKGROUP CONNECTIVITY
tion, ISDN was not designated to replace these LANs or
bridge them into larger local or wide area networks. Rather,
Today, both large and small organizations are choosing ISDN
it is ideal for the cost-effective, temporary linking of LANs to
to bring people, offices, and locations together in effective,
each other, to remote hosts, or to individual non-LAN users
easy-to-use, communications systems. ISDN, with end-to-end
or locations for the timely transfer of specific information or
digital connections, offers giant steps forward in speeds up
files. In fact, the growing popularity of these applications has
to 128 kbyte/s today, with the quantum leaps of digital data
spurred manufacturers to offer comprehensive lines of ISDN
compression still to come. But that, in fact, its just the begin-
LAN-bridging and file-transfer equipment.
ning. As more and more ultra-high-speed channels begin to
Are data rates fast enough? The answer today is a solid
bind the network together, and the broadband cables that will
yes. Current data transfer rates are typically characterized
carry high-definition television begin to reach into more and
by users as acceptable; a rating that should improve quickly
more areas, the growth of dialed digital transmission speeds
as B-channel speeds increase in the near future. Higher LAN-
available to ISDN users should continue unchecked well into
to-LAN speeds are also possible now by inverse multiplexing
the future.
multiple B-channels into bandwidth connections.
Telecommuting
A Typical LAN-to-Host Arrangement
Across the nation, legislative mandates are being put in place
Dialed connections offer enormous savings compared with
to reduce the number of cars commuting to and from an office
dedicated links, ensuring fast, efficient file transfers. Most
each day. For example, the Clean Air Act affects all large and
transfers of large files today take place through dedicated
mid-sized companies, and limits the number of cars in a com-
broad digital connections or on magnetic media physically
pany parking lot based on number of employees. One immedi-
transported by messengers and overnight couriers. The flexi-
ate answer to this problem is telecommuting.
bility of ISDN PRI connections, in contrast, is that it expands
a system manager’s options. It enables the same dialed B-
A Basic ISDN Setup for Home or Office
channels used for individual connections during the day to be
An important fact to remember is that an ISDN line does not combined into higher-speed links for after-hours transfers
provide its own power. Therefore, an independent power sup- and from multiple points. Similarly, many remote sites that
ply is always required. currently have only analog access content themselves with
The idea of ISDN telecommuting is simple: to ‘‘transport’’ weekly revisions of pricing and inventory updates, manufac-
as much of the functionality of the office as possible to a re- turing schedules, and the like. The time required to transfer
mote site through a single ISDN BRI connection. This func- these files from a central system to multiple satellite locations
tionality includes: through analog connections simply takes too long to make
more frequent transmissions practical. Similar limitations ex-
• Acceptably high-speed access to the user’s LAN and file ist in many industries. Equity pricing services, for example,
servers often use disks and tapes, messengers, and overnight couriers
• Full access to mailboxes, and the ability to send, receive, for weekly price updates to the thousands of trust depart-
forward, and annotate both voice and written messages ments, pension funds, and other money managers they serve.
• Reasonably fast interconnections to other company LANs
or hosts, remote systems, and other networks, such as Remote Telemetry and Security
the Internet ISDN connections can also let machines talk to each other. At
• Teleconferenced meetings, or the full-color images of co- a leading California biotechnology company, an ISDN system
workers through a range of rapidly emerging PC video initially installed for LAN-to-LAN connectivity, telecommut-
technologies ing, and other uses now links computers to a growing number
ISDN 731
of monitoring devices for room and liquid temperature con- Easy-to-Use Call Management
trol, fluid disbursements, animal feeding, and the like. For
ISDN also offers better call management. Not only is the
each, the system helps maintain virtually flawless control.
range of features extended, but the telephone’s liquid-crystal
Several other companies are also using ISDN for physical se-
display (LCD) is linked to the ISDN D-channel so that useful
curity in warehouses, receiving docks, and other similarly
telephone functions are no longer buried in a complicated sys-
vulnerable sites. Both B- and D-channels can be used as inex-
tem of double presses and two-second holds.
pensive conduits for remote TV cameras and to monitor locks,
Rather, they become as quickly available as when pressing
alarms, and strategically placed sound, movement, heat, and
a button on an automatically displayed menu. Normal ISDN
other sensors.
BRI links include the most widely used features:
Two Conversations on the Same Line 1. Hold. For anyone who has inadvertently disconnected
a caller rather than put them on hold, the easy use of
As we know, ISDN enables two separate voice conversations
this capability will make it the important tool it should
to take place, at the same time, through the same single
be.
twisted-pair telephone wire that traditionally, in an analog
environment, carried only one conversation. What is more, 2. Three-Way Conference. Adding another voice to the con-
since ISDN delivers two separate channels through the wire, versation is a powerful feature, yet one rarely used.
it is also possible to conduct a conversation on one channel 3. Call Transfer. An important feature frequently used by
and simultaneously use the other for a data device such as a receptionists and operators, who use it enough to re-
PC or facsimile machine. These multiple conversations could member, but rarely by anyone else. ISDN makes it al-
also take place while the speaker or someone else uses the D- most automatic.
channel for a third simultaneous transmission.
Extended Call Management
The Advantages of ISDN Voice For those who take advantage of an ISDN Centrex service,
or those in a large organization or campus served by their
It is also possible to attach up to eight devices—telephones,
own comparably equipped digital switch, the range of call
computers, faxes, and more—on every BRI connection and to
management features can be greatly extended. Some of
give these devices as many as 64 call appearances of the same these include:
telephone number, or virtually any combination of different
numbers. This feature alone offers enormous advantages. A
1. Call Forwarding. Forwards calls to a preselected num-
few examples will show why:
ber when the called number is busy, after a preset num-
ber of rings, or permanently.
1. In a busy sales location, several representatives who 2. Call Pickup. Allows an incoming call to be picked up at
are frequently out of the office might share a series of another station where the line is busy, after a preset
numbers on a single ISDN line. Each incoming call can number of rings, or permanently.
be answered correctly (‘‘Jane Green’s office’’ or ‘‘Dan
3. Directed Call Pickup. Allows calls to a specific line, to
Brown’s line’’); theoretically as many as 64 representa-
be answered only by another specified line.
tives served by up to eight telephones, faxes, or PCs re-
quire only one ISDN line, not the many lines currently 4. Message-Waiting Indicator. Shows with a light or lamp
needed. on the set that a voice message has been received.
5. Direct Inward Dial to Direct Outward Dial (DID/DOD)
2. A busy individual might have two or more appearances
Transfer. Allows a call answered at the Centrex loca-
of a single number on an ISDN telephone, as well as
tion to be transferred to any other number with the
different numbers linked to a PC and/or a fax, so that
caller staying connected as though the call were only
the appropriate device can answer incoming calls. At
being transferred across an office.
any time, a voice conversation on one call appearance
can be put on hold, and another voice call made or re-
ceived on another call appearance. These calls can even ISDN Telephones
be conferenced into a three-way conversation. Yet all of Softkeys offer quick, effortless access to features and func-
this takes place through a single ISDN, not the three or tions. The ‘‘inspect’’ button displays additional features.
four phone lines required today. Sometimes called a ‘‘virtual key system,’’ Centrex ISDN fea-
tures make a formidable competitor to premises-based PBX
The Economics of ISDN or key systems. The centrex switch becomes, in effect, the
PBX serving a location. It offers extended functionality and
In an office equipped with ISDN, two conversations, or PC or the potential for unlimited growth and enhancement, with
fax transmissions, or any combination of these could take much reduced capital investments and little risk of obsoles-
place at the same time, and all devices share the same line. cence.
Furthermore, multiple data devices such as credit units, PCs, When Centrex locations are scattered yet still served by
remote sensors, and the like can all contend for and use the the same central office (a requirement on all Centrex sys-
packet-switched D-channel at the same time that two conver- tems), the unifying features of the Centrex combine with the
sations take place on the B-channels. ISDN capabilities to create a powerful and versatile system.
732 ISDN
In this case all calls between Centrex stations (voice or data, Examples of ISDN applications that can be effectively ex-
even if they are located at different addresses) are free, being tended through Switched 56 services are those that:
‘‘in system’’ or intercom calls. For telecommuters located near
their offices, or businesses with multiple locations served by • Do not depend on voice transmission
the same telephone company’s central office, this capability, • Do not use several channels simultaneously for multi-
which extends big system features to the remote user, is un- media transmission
matched.
• Do not need data speeds greater than 56 kbyte/s
• Do not depend on the call-by-call intermixing of circuit-
Hi-Fi on the Digital Highway switched data
Throughout the world radio stations now use ISDN for clear, • Do not depend on out-of-band D-channel signaling
quiet sound transmissions from baseball, basketball, and
other games, concerts, news conferences, political conven-
NETWORK SECURITY CONCEPTS
tions, and similar events. At most locations, temporary ISDN
lines are installed, although at more venues, permanent
ISDN is a public network and, as such, data security mecha-
ISDN lines have been put in place by both broadcasters and
nisms need to be implemented to make the network secure
entrepreneurs who leased them to others for major functions.
for the exchange of confidential information. Also, test and
Modern sound studios are also using ISDN for remote re-
evaluation methodologies need to be implemented to properly
cording of announcers’ voices, live music, or other components
characterize the level of security that is being implemented.
of a firm’s video, advertising, or audiovisual presentations.
For that reason, high data security applications such as
The latest digitizing equipment can compress the highest-
transfer of confidential medical, industrial, commercial, and
quality audio signals (CD quality stereo), which would ordi-
banking information require private, dedicated network con-
narily require 1,411,200 bits/s to transmit into bit streams of
figurations. Due to the large number of applications and lack
56, 64, and 128 kbyte/s. Many studios also download stereo
of standardization, it is very expensive and difficult to imple-
tracks recorded elsewhere for mixing and enhancement on
ment standard tests and maintenance procedures for private
their more sophisticated equipment, or for incorporation into
networks. An alternative, less expensive, and more appro-
a film or television presentation. Many voiceover and auto-
priate solution would be to use a public network such as
matic dialog replacement (ADR) sessions are now done with
ISDN, with standard security mechanisms and encryption
talent and the mixers and directors all in different locations.
schemes to make it equivalent to a private network. However,
Today, many voice samplings for dubbing into foreign lan-
end users such as hospitals, laboratories, and other organiza-
guage films are recorded and then transmitted over ISDN.
tions would be ready to accept the public network as secure
ISDN lines and state-of-the-art compression equipment
as the private network only if they can test, evaluate, and
have been used to link singers in San Francisco, Hollywood,
characterize those secure schemes themselves. Unfortunately,
and Hawaii into a simple real-time concert. It is believed that
those standards are not available for ISDN-based data en-
the technology opens many possibilities, from remote con-
cryption schemes.
certs, talent auditions, and collaborative jam sessions, to tele-
jukeboxes in which customers link their stereo systems to a
Sources of Threats to the Public Switched Network
compact disk library to hear and perhaps even buy new re-
leases or hard-to-find classics. Recently, several government documents have reported the
growing vulnerability of the public switched network (PSN)
and the government’s concern about maintaining the integ-
ISDN in Marketing and Service
rity of the PSN against intruders. Specifically, a report by the
Telemarketing call centers and telephone service centers are President’s National Security Telecommunications Advisory
voice installations that also give representatives access to Committee (NSTAC) concluded that ‘‘until there is confidence
data stored in a computer. These sites have flourished in the that strong, comprehensive security programs are in place,
past decade, especially for answering incoming sales calls the industry should assume that a motivated and resourceful
from advertisements or catalog service calls for parts, repair, adversary, in one concerted manipulation of the network soft-
or assistance. ware, could degrade at least portions of the PSN and monitor
or disrupt the telecommunications serving [government] us-
Basic Telemarketing ers.’’ Unfortunately, users of expensive leased lines pay for a
false sense of security, since these lines are sometimes subject
The PBX prompts for an account number and delivers both to the same threats as the PSN. Furthermore, the new open
the call and customer records to the next available agent. Us- systems telecommunications environment has been charac-
ing a range of computers, from small PCs to mainframes, terized as one with: a large number of features; multimedia,
these systems automatically answer each call, prompt the multiparty services; partial knowledge of the feature set by
caller for an account number (or read the incoming caller-ID service designers; lower skill and knowledge levels of some
number when available), and simultaneously deliver both the service creators; multiple execution environments from differ-
call and the caller’s computerized record to the next available ent vendors; and distributed intelligence. A Bellcore investi-
agent. ISDN adds a new dimension to these facilities, because gation reports that ‘‘while the advent of open systems inter-
through a single telephone connection, ISDN offers not only a faces has assisted the acceptance and international
voice connection to the caller but a simultaneous high-speed deployment of networking technology, it has also seen a down
data connection to a computer. side in that it has become easier to intrude on networks de-
ISDN 733
signed with such open features.’’ The Telecommunications Se- licious hacker listening on the D-channel. Furthermore,
curity Guideline ranks the most significant threats to the idle times on the lines could be padded in such a way
PSN as listed in Table 1. that it would be impossible to discern when the chan-
In this research, we are interested, in particular, in how to nels are not used. As it is now, there is always some
mitigate the impact of malicious hackers and sabotagers on activity (handshaking between the switch and the CPE)
the public ISDN. Some of the safeguards against malicious on the D-channel even when the corresponding B-chan-
hackers are: nels are idle; therefore the required changes are techni-
cally feasible.
1. Install mechanisms that will validate network software
and check for and remove malicious software . . .. En- One advantage the ISDN has over the PSN is out-of-band
crypt sensitive data or protocol information that may be signaling versus in-band signaling for network control. In in-
transmitted by wireless or other unprotected channels. band signaling, the tones to signal switches could be intro-
2. Use security-oriented access technologies such as dial- duced by users through their telephone instruments to de-
back modems, dial-in passwords, and validation of in- fraud network service providers. ISDN uses out-of-band
coming calling number against an authorization data- signaling through the separate D-channel for network control.
base. But, out-of-band signaling in itself does not make ISDN pub-
3. Develop and implement an overall network security ar- lic networks secure from fraud.
chitecture. Computers should mediate access to net- A detailed guideline for the eight high-level security re-
work software through appropriately applied user iden- quirements includes:
tification and authentication mechanisms. Biometric,
token-based, and third-party authentication systems 1. Identification. Pertains to the process whereby the ap-
should be used. plication system recognizes a user’s unique and
auditable, but not confidential, identity, such as the
Also, threats to the PSN that could be mitigated by system user ID.
(or customer) based software include: 2. Authentication. Refers to the process of verifying the
identity claimed by the user. Authentication can be pro-
1. Masquerade. Refers to a user posing as another author- vided by a password or ‘‘smart card’’ and must be kept
ized user. A user will effectively masquerade as another confidential.
user through replay of data or insertion of data, which 3. System Access Control. Refers to allowing access to the
appears genuine, into the communication path. This applications only to those users who have been identi-
threat can come from outside users accessing the appli- fied and authenticated.
cation from the PSN, locally connected users, or system
administrators through direct-connected or dial-up con- 4. Resource Access Control. Pertains to guaranteeing users
only the least privileges required to perform their job
trol.
function.
2. Disclosure of Information. Refers to data disclosed with-
out authorization. 5. Data and System Integrity. Refers to the reliability of
the application and its resources.
3. Unauthorized System Access. Refers to a user accessing
system software information. An example of this threat 6. Audit. Refers to a trail for investigating security-rele-
is a user accessing and modifying the password table. vant events.
4. Denial of Service. Refers to a degraded performance of 7. Security Administration. Refers to the tools for manag-
the application. An example of this threat is a malicious ing security-relevant tasks.
hacker trying to repeatedly log into the application, 8. Documentation. Describes how the security features of
which might prevent an authorized user from accessing the application should be provided.
it.
5. Traffic Analysis.. Refers to a hacker observing traffic in ISDN is a public network and, as such, data security mech-
the PSN and makes an inference using the source and anisms need to be implemented to make the network secure
destination addresses. This concern can only be ad- for the exchange of confidential information. Also, test and
dressed by a switch software modification, perhaps evaluation methodologies need to be implemented to properly
through the encryption of the D-channel information. characterize the level of security that is being implemented.
This, if done properly, would maintain the required High data security applications such as transfer of confiden-
throughput but render this information useless to a ma- tial medical, industrial, commercial, and banking information
require private, dedicated network configurations. Due to the
large number of applications and lack of standardization, it
Table 1. Most Significant Threats to the Public is very expensive and difficult to propose standard tests and
Switched Network maintenance procedures for private networks. An alternative,
Threat Likelihood Principal Impact on Network less expensive, and more appropriate solution would be to use
a public network such as ISDN, with standard security mech-
Employees 50% Availability and integrity anisms and encryption schemes to make it equivalent to a
Natural disasters 20% Availability
private network. However, end users such as hospitals, labo-
Hackers 15% Availability, integrity, privacy
ratories, and other organizations would be ready to accept the
Sabotage 15% Availability
public network as secure as the private network only if they
734 ISOLATION
BIBLIOGRAPHY
http://www.bellcore.com/
BC.dynjava?ADSIVMWISvcsPDGeneralProductDescription
http://www.bellcore.com/NIC/
http://www.bellcore.com/CILS/
http://www.bellcore.com/ISDN/
http://www.bellcore.com/ISDN/digest.htm
http://www.bellcore.com/ISDN/dcm-1.htm
http://www.bellcore.com/NIC/links.htm
http://www.bellcore.com/
BC.dynjava?BTECCatalogCHPGeneralContentHomePage
IRMA BECERRA-FERNANDEZ
SUBBARAO V. WUNNAVA
Florida International University