Professional Documents
Culture Documents
Creativity and Heuristics in Process Control Engineering
Creativity and Heuristics in Process Control Engineering
a r t i c l e i n f o a b s t r a c t
Article history: We study the relationship between heuristics and standard tool application in the design of
Received 7 March 2015 process control systems. This relationship is illustrated by classical control history exam-
Received in revised form 12 June ples. Features of modern engineering objects are highlighted that challenge classical control
2015 approach and leave substantial space for creative solutions. We focus on inevitable heuristics
Accepted 9 July 2015 in control design for modern complex process systems. The study suggests new systematic
Available online 17 July 2015 approach to heuristic control design. The approach is based on simplification technique
and follow three steps. First, certain assumptions (e.g. negligible nonlinearity or dynamic
Keywords: behavior) help to develop the simplified model and problem setup. Second, standard control
Process control design is performed on this reduced model. Third, the perturbation theory methods help to
Creativity find corrections to the obtained control to cope with the real plant. Special attention is given
Heuristics to the paradigm of concurrent (or integrated) plant and control design. The study suggests
Concurrent design finding a proper compromise between investments in the process and its control. Several
cement manufacturing process examples demonstrate that this method can generate better
effect comparing to autonomous approach.
© 2015 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
∗
Corresponding author. Tel.: +358 504419193.
E-mail address: leonid.chechurin@lut.fi (L. Chechurin).
http://dx.doi.org/10.1016/j.cherd.2015.07.010
0263-8762/© 2015 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
chemical engineering research and design 1 0 3 ( 2 0 1 5 ) 40–49 41
Fig. 4 – Windmill running stone gap adjustment However, the modal reduction method, known from 60s,
mechanism. The gap is proportional to the rotation rate due removed from G(s) components with higher oscillation modes.
to centrifugal forces effect (drawing form http://www.texva. So, the underlying idea had physical meaning. Then the
com/holland/The%20Dutch%20Windmill.htm). paradigm for model reduction became rather mathematical.
First,
the optimal approximation
in respect to quadratic norm
or G (s) − Gred (s) . was introduced in 70s, where
2
done by Maxwell (1868) and Vyshnegradsky (1877). There two
successful analysis used differential equations to model the ∞
1 2
mechanical system and a number of mathematical manipu- G (s) =
2 G (jω) dω.
lations to derive the stability region for system’s parameters. 2
−∞
The same roadmap became standard for the next era of
automation: invention of governing mechanism or device – Then balanced reduction approach came in 80s, with the
(when necessary) stability or performance analysis by math- estimation of approximation in respect on H∞ norm (since the
ematical modeling – device improvement. Here we have to H∞ norm is not smooth, the optimal approximation has not
refer to another famous example: Harrold Black’s invention been found). The H∞ norm of transfer function G(s) is defined
of feedback amplifier (Black, 1932). The invention suggests a by
breakthrough structure for electric signal transmission, where
the feed backing is used to attenuated exogenous noises. But G (s) = sup G (jω) .
∞
ω
the patent reveals just a industry-specific solution only again
(check out terms in the patent “electric wave”, “current”, “volt- This norm, or Hardy space norm, became popular due to its
age”, “amplifier” and not “controller”, “signal” etc.). Similar to use for robust system analysis and design. Finally, optimiza-
steam turbine governor improvement based on mathemat- tion with Hankel norm became the focus of approximations in
ical analysis, it was Black’s colleague, Henry Nyquist who 90s since it provides better H∞ approximation. We emphasis
analytically derived the stability conditions for the feedback that the heuristic design of solutions migrated from physical
amplifier and generalized the conclusions in the form of the- to mathematical meaning.
orem (Nyquist, 1932). Another example of paradigm shifts is robust control
A thoughtful description of this stage of control system/and synthesis. Several concepts were suggested (invented): poly-
theory evolution is given by H. Black himself: “Pertinent to nomial synthesis (frequency domain), state space methods
invention in engineering is the familiar experience that inven- based on Lurie–Riccati equations solving, linear matrix
tion fosters research and vice versa.” (Black, 1958). inequalities methods etc.
A new trend in control theory evolution appeared in the At the present time the amount of mathematically involved
mid of the last century. At one hand, control theory became papers on control theory is greater in powers than the amount
the field of application of mathematics. At the other hand of patents. The efforts of practicing automation engineers
the focus of the research in control theory moved to pure address the adaptation of rigor book solutions to real problems
mathematical problems migrating further and further from like PID/PI controller tuning, approximate solutions to control
the actual problems in industrial practice. The object of con- problems or stability evaluation, complexity reduction etc.
trol became the set of differential equations, the research The interest to systematic creativity or conceptual design
question was how to influence their trajectories by controlled in engineering is gaining momentum as well. Regarding chem-
inputs. As the result we obtained the academic theory of ical engineering design, for example, we can refer to (Ottino,
control, some chapters of which are the monument for math- 2000; Eekels, 2002; Kossack et al., 2008). We strongly rec-
ematical instrumentation. And at the same time, most of ommend an exhaustive review on approaches for chemical
solutions were true but useless. For example, formal meth- process control problems presented by (Sharifzadeh, 2013). It
ods of optimal control with the quadratic performance index contains nearly all possible ideas for process control although
do provide the solution. But the controller is given in the form the heuristics seems to be mostly used to resolve mathe-
of differential equations of high order. The solution is correct matical difficulties. In respect to systematic creativity, one
but unrealistic for practical implementation: high derivatives of the very popular approaches to generate successful inven-
would amplify noises, that always exist. The studies in control tive ideas is TRIZ, the theory for inventive problem solving
chemical engineering research and design 1 0 3 ( 2 0 1 5 ) 40–49 43
raw cement
clay roasting
grinding milling
clinker
cement
gypsum
(Altshuller and Shapiro, 1956; Salamatov, 2005). There are cannot be described by the extensive language of linear analy-
studies on direct adaptation of TRIZ for chemical process sis. Thus, some internal nonlinearities may produce complex
design (Srinivasan and Kraslawski, 2006). At the same time we even unpredictable behavior of system variables, that is hard
focus on creativity in process design and automation issues in to describe in, for example, “classical” frequency domain.
the presented paper. We have to look for heuristic approaches to reduce the
complexity of the problem, to decompose or to reduce the
3. Process engineering control problem dimension, while keeping the same general setup.
setup
4. Heuristic methods of complex control
The manufacturing automation problem can be given a short problem reduction
form. Given the limited amount of resources, we are to maxi-
mize the profit by the material flow intensity and specification There are many approaches to reduce the complexity of the
parameter control subject to the restrictions on the output problem, where complexity means the amount of variables.
quality and quantity are held. We refer again to the review (Sharifzadeh, 2013) that con-
Without the loss of generality, we consider continuous type tains the description of popular methods as well. We cluster
manufacturing process. Any manufacturing is the set of tech- them into two basic strategies: decomposition and aggregation
nological process over the material flows, divided by buffer (Pervozvanskii and Gaitsgori, 1988). The first means break-
tanks. Cement processing scheme is given as the example in ing down the problem into a number of simpler independent
the Fig. 5. ones, although linked by a common goal. Heuristic reason for
The general scheme for manufacturing control is given in decomposition is the weak connection or crosstalk between
Fig. 6. Here the control inputs are the intensities of material some subsystems of lower dimension. Aggregation is the
flows and technological parameters (e.g. fuel consumption, inverse of decomposition: a number of similar variables or
raw mixture consumption, furnace rotation rate and air thrust subsystems are replaced by one macro variable or system,
for roasting process). Process monitoring data and output an aggregate. The dimension is reduced since the amount of
quality specifications are used to develop the control. variables is reduced.
The difficulties of straightforward control problem solving When designing the control system for a complex object,
are originated by the complexity of the problem. As com- the ideas of decomposition and aggregation produce multi-
plexity is known for bearing many meanings, we need to level hierarchical system with block structure. The integrated
provide the details. First, by complexity we mean the practi- automation control scheme depicted in Fig. 7 provides an
cally uncountable amount of possible variants of control that example.
makes the trial and error approach too time demanding even The hierarchy means that every control level (except the
for advanced computational resources. Second, complexity highest, which is driven by demand) obtains orders from
stands for the interconnection of subsystems and system vari- the level above. Each level (except the lowest one) forms
ables. It becomes hard (inaccurate) to single our certain units the order for the lower level in the process of the real-
or a subset of variables to reduce the dimension. Third, the ization of the obtained order. Thus, in similar manner to
complexity reflects the nature of real control processes, that cascade control, systems develop orders to other systems.
Direct changes of material flows and regime parameters hap-
pen at the lowest hierarchical level. But if the whole system
flow
intensity is well designed, control goals at each level are harmonized
with the main strategic goal: economic optimization of the
Manufacturing output
enterprise (Yakovis, 2010).
technology
parameters
Measurement Control
year, quarter, ERP
output planning month (Enterprise Resource
Planning)
week, APS
scheduling (Advanced Planning &
day Scheduling)
MOM,
operative control day, shift,
MES,
hour LIMS, EAM
DSC,
distributed control of TP shift, hour, SCADA,
unit 1 … unit N minute APC
…
Manufacturing
unit 1 … unit N
…
Fig. 7 – The scheme of multi-level hierarchical system for manufacturing control (MOM—Manufacturing Operations
Management, MES—Manufacturing Execution Systems, LIMS—Laboratory Information Management Systems,
EAM—Enterprise Asset Management, DCS–Distributed Control System, SCADA—Supervisory Control and Data Acquisition,
APC—Advanced Process Control).
dimension of initial problem. For example, the whole manu- of intermediate and final products. These results in “bound-
facturing is considered for material volume planning at the ary” control regimes. But if the upper control level should not
highest level of control. This does not deal with “frequent” make effective the “boundary” technological process orders
control designs (material dosing in real time). It analysis to the lower control level since there are disturbances always.
annual amounts of materials used in manufacturing only. Since these disturbances are stochastic and therefore unpre-
That makes the problem still difficult but solvable. dictable, the lower level is not able to observe the upper level
Lower level control directly manages material flows and control specification. Therefore, the process parameters that
processing regime parameters. As the properties of processed were supposed to be within specification boundaries, would
materials vary stochastically, the control is to be fast to react exceed them in 50% cases approximately. It will lead to quality
them. That means the high frequency control signals. But degradation, high reject rate, claims etc., in other words to seri-
the technological processes are slow and inertial, therefore ous economic losses. Thus the upper level control should take
the future responses of any control input is to be taken into into account certain parameter instability. It is done by setting
account. In contrast to planning that is based on static mate- certain “emergency” margins, in other words by stepping back
rial balance, the control design is to be based on dynamic from the dangerous boundaries inside the admissible set. The
physical–chemical processes and dosing system models. It described situation is shown in Fig. 8. The setpoint Ỹ of index
complicates the control design a lot if we try to approach it Y changes due to its instability.
directly. Instead, placing intermediate buffer tanks between The figure explains the economic meaning of stabilization:
process stages makes them independent. In general, manu- the more index stability is attained at the lower control level
facturing site with shop structure reflects the heuristic idea of the closer the upper level controlled technological parame-
decomposition, where each shop can be considered indepen- ter to the most profitable value Ȳ. It is worth mentioning
dently. that incorrect “emergency” margin evaluation leads to eco-
Even after decomposition or problems of lower control level nomic losses regardless its sign. Indeed, the underestimation
studied above, the control problem is still difficult since the of the instability leads to quality drops while overestimation
large amount of variables and dynamic properties of the sub- to unnecessary deviation from the most profitable regimes.
systems. We are going to suggest a heuristic idea of further If there many indexes (that is general case) to be optimized
manufacturing control decomposition by two-level control and kept at the required level than two level control becomes
design for certain processes (Yakovis, 2008). more complex but the effects of narrowing the admissible
Let us assume that the optimization problem is solved at regime parameter sets are the same for parameter instability
the upper control level while the lower level stabilizes the pro- analysis. To illustrate this statement we consider the problem
cess parameters within the given specification obtained from with two indexes Y, that has to belong to a polygonal depicted
above. in Fig. 9. Some index values and the index gradient are also
It is worth mentioning that in cost-sensitive optimization shown. Obviously, the optimum of Y is attained at the least left
problems the solution generally belongs to the boundaries point of the admissible set Ỹ 1 . But it assumes no disturbances
of the available set. For example, economy of cement batch that bring about index Y instability, that produce an elliptic set
manufacturing tells to increase the share of slag usage (that around the setpoint Ỹ. Therefore, we have to choose much less
is cheaper than clinker) and to decrease the roasting tem- economic regime Ỹ 3 . Instead, having optimized the regime in
perature (to save fuel). The same situation happens in beer respect to stabilizing (stabilization reduced the size of the devi-
brewing, where water share in the product is to be maximized ation ellipse), we obtain the setpoint Ỹ 2 , corresponding to the
for maximum profit. But theses trivial means are restricted correct interaction between upper and lower control levels.
by quality specification. The latter is given the form of tech- In addition to “emergency” margins evaluation we also
nological tolerances on various physical–chemical properties have to define proper priorities for stabilization at the lower
chemical engineering research and design 1 0 3 ( 2 0 1 5 ) 40–49 45
control into two levels where the upper one ensures the opti-
mal regime while lower one ensures stabilization. The scheme
of system interactions for this two level structure can be sug-
gested as in Fig. 11.
There are two information flows in the scheme. One down-
stream of program controls as well as index function forming
request from upper level to lower level. As the result, the lower
level system operates with the common to the whole sys-
tem optimization criteria, as usual. But the difference in the
scheme in Fig. 11 is that flows from lower level to upper one
are shown. In particular, these flows are used to evaluate those
emergency margins that define the restrictions of upper level
problem. This is typical situation for any neighboring control
levels in the multilevel hierarchical control system.
· · ·
Fig. 10 – The effect of non-uniform resource assignment of stabilization system.
46 chemical engineering research and design 1 0 3 ( 2 0 1 5 ) 40–49
Measurement
system
Controlled
process
partial differential equations for modeling or transport theory methods. Then the obtained solution is heuristically
delays, (inventively) modified and the closed loop performance is eval-
• multivariablity of the dynamic objects, that can reflect the uated by simulations.
continuous nature of the controlled plant again. Partial The question of the choice of the successful (neither over-
differential equations may require many variables for finite- simplified nor lost in details) is also of heuristic nature.
element approximation for reasonable accuracy, However we formulate here typical patterns:
• subsystem’s interconnection, any of control input influence
every output,
• complex and time-variant output dependence from control • Systems with weak nonlinearity. The “weakness of nonlinear-
and disturbance inputs, ity” does not seem to have been given a rigor definition but
• stochastic and time-variant uncontrollable external distur- the main idea is the following. Let an operator A over x
bances, be nonlinear A(x) = y. Let the operator A has a linear com-
• tough boundary restrictions on output and control vari- ponent such A0 such as y0 = A0 x. If ||y − y0 || = ||A(x) − A0 x||
ables, is small then “the main part of A(x)” is linear, so we
• significant measurement noise, and can call this nonlinearity weak. Thus, the basic model is
• “nonstandard” control quality criteria. chosen to be linear (A0 ) and the control is designed by
standard linear control design procedures. Then simula-
tions are used to check the performance of the control on
The most powerful MIMO control theory synthesis meth- actual nonlinear plant with A(x). If necessary, the linear
ods use state space approach (Åström, 2012). But this control solution can be corrected: the first approximation
representation is not comfortable to cope with delays and is evaluated with neglected nonlinearities and the per-
boundary restrictions. Moreover, it has been already men- turbed problem is solved again by linear theory methods
tioned that they lead to controllers of high orders in general. (Pervozvanskii and Gaitsgori, 1988).
Heuristics and simulations can help an engineer in this sit- • Systems under weak disturbances. The basic model assumes
uation. Control design (that ideally is be done simultaneously zero disturbances. Then program control methods can be
with the plan design, we return to this point later) should be used and, in particular, static optimization for stationary
assisted by modelling as well as control itself can use sim- systems. The upper level control provides the optimal pro-
ulations in its algorithms (Perelman, 1978). The modelling gram or regime setpoints at the second stage. At the same
and simulation nowadays are well automated and does not time the setpoint deviations are stabilized by additional
require intensive mathematical knowledge (see, for example feedback controller, that attenuates relatively weak distur-
SIMULINK and MATLAB). Visualization of these modelling and bances. Linear dynamic control synthesis methods can be
simulation tools is similar to typical SCADA interfaces, that are used for that based on linearized object description. This
known by automation engineers. A study of the role of simu- heuristic idea was in fact implemented for two level process
lation and visualization in modern digital manufacturing can control scheme described above.
be found in (Mourtzis et al., 2015). • Systems with weak dynamic properties. The impulse func-
Once the idea of control algorithm is found, the simulation tion of stable dynamic objects is known to vanish as time
tools can numerically help to tune its parameters, to anal- increases ad infimum. It means that the dependence of future
yses stability and robustness, to evaluate the performance outputs from past inputs decreases with time difference.
indexes etc. The focus of the problem so far is how to find the Thus, the basic model can be one with the “truncated”
conceptual idea for the control. And modern control theory dynamic memory. Then the obtained solution is upgraded
can help provide those principles for control concept design. with respect to truncated parts. For example, if the con-
Namely, the basic approach can be derived by disturbed trol inputs intervals in the basic model are longer than the
system analysis methods. First, we consider idealized (sim- “truncated” memory depth, we arrive at simper problem of
plified) problem set up. Since many complicating details are static object control. Now, the control is to be upgraded by
released, this generating problem is solved by classical control lower level control design that moves the operation regime
chemical engineering research and design 1 0 3 ( 2 0 1 5 ) 40–49 47
to the upper (static) level control setpoints as fast as possi- We consider an example of cement manufacturing plant
ble. (Fig. 5) as an illustration. More precisely, we focus on the raw
• Systems with weak control. Let us consider manufacturing material mixing followed by the roasting. The first heuristic
systems that consist of a sequence of technological pro- idea is that the roasting in the rotating oven is the process of
cesses with material flow between them. Since the volume very complex physical and chemical nature. The development
of buffer tanks is limited we face the complex problem of of reliable dynamic model of such a process is very difficult
simultaneous control design with one or several criteria. time demanding effort. Therefore, the problem of quality con-
These criteria describe the manufacturing and economic trol of roasting process is difficult too. The less stable the
indexes of the whole plant. Some of these controls (let us chemical stability of the raw material inflow the more difficult
call them “of first type”) are material consumption per time the roasting quality control. But the mixing process is easy to
unit. Others (let us call them “of the second type”) are regime control as it is described by universal material balance equa-
parameters for every technological process. It is reasonable tions. Thus, maximum efforts should be allocated to chemical
to assume that if all the controls belong to the admissible contents stabilization system design. Having reduced the vari-
set than the first type control inputs influence the material ations of main external disturbances for the roasting oven,
output flow for some processes. Therefore we can simplify we essentially simplify the control design for this sensitive
the problem to so called operating control problem, in other process.
words, the problem of routing of material flows through the Let us turn to the chemical contents stabilization problem
technological processes. To solve it the regime parameters for raw materials to be mixed. We face chemical nonuniform-
are fixed first at certain nominal value, for example in the ity of the materials that come from a quarry. Two approaches
median of admissible interval. Second, the initial manu- can help to achieve the contents stability: “by control” and “by
facturing and economic criteria and the obtained routing technology”. “By control” means careful share control for the
problem solution are taken into account: every local tech- materials to be dispatched to the mill. It will obviously main-
nological process is given a performance index. The best tain the required chemical contents. “By technology” means
regime parameters are defined as local optimization prob- just dosing the raw materials in constant specified proportions
lem solution. The operating control problem solution can be and averaging the milled mixture in huge homogenizing tanks
corrected at this stage since local regime parameters vary with mixing facilities.
the process productivity (Gaytsgori et al., 1986). To control the mixed materials share we need frequent
contents measurements and precise dosing equipment. Sys-
tems for frequent control of bulk materials are expensive and
not precise. Therefore chemical contents sensors for milled
7. Systematic design of automated mixture are engaged. But it generates substantial transport
technological complexes delay. The control system attenuates low frequency distur-
bances only due to this delay. On the other hand, we need
The abovementioned features of technological processes seri- giant averaging tanks to reduce low frequency disturbances,
ously complicate the control and force the control design that are originated by raw material chemical contents vari-
engineers to engage sophisticated and not always reliable con- ation. Capital costs as well as operation costs are too high
trol algorithms. A natural question appears: is it possible to for these big volume averaging tanks. But they filter high
address control difficulties at the stage technology equipment frequency contents fluctuations efficiently. These heuristic
design? In other words, what if it is allowed to change the analysis result in the conclusion that an optimal design would
plant to be controlled? These inventive plant (re)design, or combine the control system as well as averaging tanks of rela-
concurrent (simultaneous) plant and control system design, or tively small volume. With this conceptual idea we can turn
“design-for-control”, or integrated process and control design to quantitative design by approximate analytical stochastic
would provide more radical means to increase control qual- control methods as well as numerical simulation.
ity. Regarding the process control it would mean systematic The general scheme of this quantitative evaluation is the
design of automated equipment or, in more general sense, following. First, the achievable (under control) output prod-
automated processing complexes (APC). uct stability is estimated for each set of selected technological
The systematic design of automated processing complexes parameters. Then comparing the derived estimation with
is becoming more and more popular topic among tech- the specification we define admissible solutions. Finally, the
nology and automation designers both (Ambartsumian and cheapest design is selected over all admissible designs. This
Kazansky, 2008). The general goal is to maximize the economic man-machine approach was implemented in practical design
output by concurrent choice of the structure and parameters of averaging-mixing complexes or cement manufacturing
of technological complex (TC) as well as control system (CS). plant (Gorenko et al., 1987; Doroganitch et al., 1989). MATLAB-
There are studies on concurrent automated process design, Simulink software toolbox was developed for analysis and
the problems of buffer tank volume choice is attacked in optimization of averaging-mixing automated complexes. The
Gel’fand et al. (1980) under condition of operating control and toolbox perform multi-scenario simulations for design and
maintenance of equipment. The problem of static multicri- upgrading of cement manufacturing plants. These simula-
teria optimization is studied in Volin and Ostrovskii (2007), tions assist analytically based recommendations on the type,
where the choice of regime parameters during normal oper- amount and volume of averaging tanks, type and character-
ation is assumed at the stage of chemical processes design. istics of measurement tools, structure and parameters for
(Sharifzadeh, 2013) has already been mentioned for its review control system.
on integrated design practice. However, no general concurrent We have to emphasize that we need reliable mathematical
“TC + CS” design method is known so far. Therefore, engineers model of the processes at the stage of process design for the
have to rely on heuristics, where it means the blend of science, abovementioned simulations. It seriously complicates the
creativity and experience. design since the model identification requires some active
48 chemical engineering research and design 1 0 3 ( 2 0 1 5 ) 40–49
Srinivasan, R., Kraslawski, A., 2006. Application of the TRIZ Yakovis, L.M., 2010. Multi-level production control: modern state,
creativity enhancement approach to design of inherently problems, prospects. Autom. Remote Control 71 (10),
safer chemical processes. Chem. Eng. Process. 45 (6), 507–514. 2247–2258.
Volin, Yu.M., Ostrovskii, G.M., 2007. Multicriteria optimization of Yakovis, L.M., Sporyagin, K.V., 2011. Controller tuning for inertia
technological processes under uncertainty conditions. Autom. objects with time delay. Autom. Remote Control 72 (1),
Remote Control 68 (3), 523–538. 208–217.
Vyshnegradsky, I.A., 1877. On Controllers of Direct Action. Izv. Yakovis, L.M., Sporyagin, K.V., 2013. Typical controllers
SPB Tekhnolog. Inst. adjustment for multivariable objects of industrial automation.
Yakovis, L.M., 2008. Approximate solution for problem of In: Proceedings of Seventh IFAC Conference on Manufacturing
dynamical optimization of process control. In: Proceedings of Modelling, Management, and Control, 2013, June 19–21, 2013,
Sixth EUROMECH Nonlinear Dynamics Conference, June Saint Petersburg, Russia, pp. 1792–1797, vol. 7 part
30–July 4, 2008, Saint Petersburg, Russia (electronic no. 1.
publishing).