Journal of http://jvc.sagepub.

com/
Vibration and Control

Bearing Diagnostics Based on Pattern Recognition of Statistical Parameters

Fengfeng Xi, Qiao Sun and Govindappa Krishnappa
Journal of Vibration and Control 2000 6: 375
DOI: 10.1177/107754630000600303

The online version of this article can be found at:

http://jvc.sagepub.com/content/6/3/375

Published by:

http://www.sagepublications.com

Additional services and information for Journal of Vibration and Control can be found at:

Email Alerts: http://jvc.sagepub.com/cgi/alerts

Subscriptions: http://jvc.sagepub.com/subscriptions

Reprints: http://www.sagepub.com/journalsReprints.nav

Permissions: http://www.sagepub.com/journalsPermissions.nav

Citations: http://jvc.sagepub.com/content/6/3/375.refs.html

>> Version of Record - Jan 1, 2000

What is This?

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

Bearing Diagnostics Based on Pattern Recognition
of Statistical Parameters

FENGFENG XI
Integrated Manufacturing Technologies Institute, National Research Council Canada, 800 Collip
Circle, London. Ontario N6G 4X8, Canada

QIAO SUN
Department of Mechanical and Manufacturing Engineering, University of Calgary, 2500 Univer-
sity Drive, N. W., Calgary, Alberta T2N 1N4, Canada

GOVINDAPPA KRISHNAPPA
Integrated Manufacturing Technologies Institute, Western Laboratory, National Research Council
Canada, 3250 East Mall, Vancouver, B. C. V6T 1 W5, Canada
(Received 31 August 1998; accepted19 January 1999)

Abstract: In this paper, a new bearing defect diagnostic and classification method is proposed based on pattern
recognition of statistical parameters. Such a pattern recognition problem can be described as transformation
from the pattern space to the feature space and then to the classification space. Based on trend analysis of
six commonly used statistical parameters, four parameters, namely, RMS, Kurtosis, Crest Factor, and Impulse
Factor, are selected to form a pattern space. A 2-D feature space is formulated by a nonlinear transformation.
An intraclass transformation is used to cluster the data of different bearing defects into different regions in the
feature space. The classification space is constructed by piecewise linear discriminant functions. Training the
classification space is performed, in this paper, by using data of bearings with seeded defects. Diagnosis of
the defected bearings in the classification space then becomes straightforward. Numerical experiments show
that the proposed method is effective in indicating both the location and the severity of bearing defects.

Key Words: Diagnostics, bearing defects, pattern recognition, statistical parameters

1. INTRODUCTION

The subject of rolling bearing diagnostics has been studied over the past 25 years because they
are used in all rotating machinery from small, medium, to large size (Rao, 1996). Common
failure of rolling bearings includes spalling, corrosion, brinelling, and so on. These defects
induce repetitive vibrations when bearing elements encounter them (McFadden and Smith,
1984a, 1984b). The objective of bearing diagnostics is to identify the type of defects and the
severity by means of the technologies involving measuring and processing of these defect-
induced vibration signals.

375-

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

376

There are anumber of available vibration measurement methods, including proximity

probes, velocity pick-ups, accelerometers, and vibrometers. Velocity pick-ups are portable
and thus commonly used in field measurements for scheduled maintenance checkup. Perhaps
due to this, vibration standards for most machinery are specified in terms of velocity.
Proximity probes and accelerometers normally need to be mounted on the object to be tested.
Tendency is to use embedded sensors on the machine, either capacitor-based or piezoceramic-
based. Compared to the others, the laser-based vibrometer is more expensive, hence limited
to experimental studies. Since vibration induces sound, acoustic measurement methods can
be used to measure vibroacoustics. Basically, there are two methods, acoustic emission and
microphone, with the former in the high-frequency range from approximately 50 up to 2 MHz,
and the latter in the low-frequency range from 0 to 40 kHz. Generally, acoustic emission (AE)
method is very sensitive to initial metal cracking (Yoshioka, 1993), whereas vibration method
is effective for monitoring defect growth (Rao, 1996).
Signal processing is to extract the characteristics of measurement signals. It can be
performed in time domain, frequency domain, and time-frequency domain. Time domain
analysis calculates the statistical parameters including Peak, RMS, CrestFactor, Kurtosis,
ImpulseFactor, and ShapeFactor (Howard, 1994). Frequency domain analysis transforms
time-domain data into frequency domain, and then compares the results with the analytical
defect frequencies (Braun, 1986). These methods include spectrum analysis, envelope
analysis, and cepstrum analysis. Time-frequency domain analysis is used for processing
transient and nonstationary signals, such as the methods including short-time Fourier
transform, Wgner-Ville distribution, and wavelets.
In practice, amplitude demodulation (envelope analysis) followed by spectrum analysis
is the most common method, used together with trend analysis, for bearing diagnostics. This
method detects the frequency of the repetitive impulses generated by the bearing defects.
The method is effective for detecting defects that exhibit distinct defect frequencies, such
as single spall. It becomes less effective when defect frequencies are not distinct, such as

multiple defects. Time domain analysis of the envelope signal may overcome the weakness
of spectrum analysis, as statistical parameters can provide information such as the shape of
the amplitude probability distribution and the energy level of the vibration signals. Much
research work has been done in using these parameters individually to detect the bearing
defects, and the results have shown that each parameter is only effective for certain defects.
For example, spikiness of the vibration signals indicated by CrestFactor and Kurtosis implies
incipient defects, while the high energy level given by the value of RMS and Peak indicates
severe defects.

An intuitive idea is to combine these statistical parameters into a weighted index to

take different defects into consideration. Based on this idea, we propose in this paper
a bearing diagnostic method based on pattern recognition of statistical parameters. This

pattern recognition problem is described as transformations of four spaces, namely, from

the measurement space to the pattern space, then to the feature space and finally to the
classification space. Based on the trend analysis of six commonly used statistical parameters,
four parameters, RMS, Kurtosis, CrestFactor, and ImpulseFactor, are selected to form a
pattern space. A feature space is formulated by a nonlinear transformation, and its dimension
is defined as two, for the purpose of displaying results on a two-dimensional plane. An
intraclass transformation is used to create clustering effect, which better facilitates the
classification of bearing defects. A supervised learning process is applied to formulate

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

 377

Figure 1. Mathematical pattern recognition.

discriminant functions that are further used to generate piecewise linear boundaries in the
classification space. The proposed method is tested with data provided by the Association of
American Railroads. Results show that it is an effective method for bearing diagnostics.

2. RUDIMENT OF PATTERN RECOGNITION

The subject of pattern recognition may be briefly stated as identification of classifiable

patterns from measurement data. This subject has been studied for many years in different
applications. literature, pattern recognition can be categorized as heuristic, linguistic, and
In
mathematical (Andrews, 1972). The technique adopted in this paper is the last one. In this
section, we briefly review the rudiment of mathematical pattern recognition. The reader may
refer to the textbook by Andrews (1972) for details.
As illustrated in Figure 1, we describe the mathematical pattern recognition as to
identify classifiable patterns through mathematical transformations of four spaces, namely,
the measurement space ~t, to the pattern space ~, then to the feature space ,~, and finally to
the classification space C, that is,

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

378

Note that this description is different from that given by Andrews (1972), in that we include
the measurement space. As will be discussed in the following, this allows us to construct the
pattern space by means of certain analysis methods, such as statistics, rather than simply by
the measurement data.

2.1. Measurement Space

As the title may suggest, measurement space is constructed directly by the measurement data.
If a finite dimension N is considered, the following column data vector

represents a series of measurements with N digitized data samples. Hence, the measurement

space is formed by N axes, and vector s represents a point in the space. To establish the
classifiable patterns, learning techniques should be applied. We adopt a supervised learning
procedure that uses prototype data of each class to be identified to establish the classifiable
patterns. For each class, we may need a set of prototype data samples. Therefore, a vector in
the measurement space representing the mth prototype of class k is expressed as

where m =
1, ... , Mk. M~ denotes the number of prototype data vectors for the kth class.
For given classes, the total number of data sets, Nd, is given by
K

2.2. Pattern Space

According to Andrews (1972), the pattern space was simply represented by the measurement
data. We extend it here, so that it could also be defmed by means of certain analysis methods
such as statistics. This means that after this process, the data vector s becomes the following
new vector

where subscript p indicates the pattern space. Dimension of the pattern space is denoted as

R, where R < N. A prototype in the pattern space is given by the following vector

The problem of pattern recognition is to correctly classify the known prototypes and
afford some degree of confidence based on certain criteria. The solution is based on the
assumption that the pattern space forms a metric space that satisfies the following conditions
(Andrews, 1972):

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

 379

where d denotes a distance function and a, b and c represent three vectors in the pattern space.
In terms of the metric space, the similarity of a point xp to the k-th class can be measured by

This similarity measure is an average of the squared distance between the pointx and the set
of prototypes (k) .
To ensure that we compare vectors with the same unit, normalization of
measurement data is required, which is referred to as squaring up the pattern space. A simple
way is to divide measurement data by their variance, known as variance normalization.

2.3. Feature Space

Feature space is the intermediate domain between the pattern space and the classification
space, with its dimension denoted by L. The data vector x f and the prototype vector
the feature space are defined as -
in y

and

where subscript f indicates the feature space. -

This space serves two purposes. The first is to reduce the dimensions of the pattern
space so that a classification algorithm can be efficiently computed and the results can be
readily presented. The second purpose is to combine the dimensions of the pattern space so
that classification characteristics can be extracted. The process of selecting a feature space
is to find a transformation either linear or nonlinear to reduce the dimension of the pattern
space, yet maintaining discriminatory characteristics for the classification purpose. There
are a number of transformation methods. The one adopted in this paper is the intraclass

transformation, which will be discussed later.

2.4. Classification Space _

The classification space is referred to as the separating surfaces in the feature space. For
given K classes, there will be K separating surfaces. The separating surfaces are hyperplanes
in the L - 1 dimension. A classification algorithm is to define the space partition in the L
dimensional feature space into disjoint regions, with each region associated with one class.
The separation is a point on a line for L 1, a line on a plane for L
=
2, and a plane in a =

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

380

volume for L =3. Since the data are usually not linearly separable, space partition generally
results in nonlinear surfaces. For example, the boundary of the classification space forL 2 =

may be formed by curves.

3. PATTERN RECOGNITION OF STATISTICAL PARAMETERS

The idea of pattern recognition of statistical parameters is first to construct the pattern space
based on statistics analysis and then perform pattern recognition analysis using statistical
parameters. In this section, we describe how to select the statistical parameters to form the
pattern space for bearing diagnostics. In the following two sections, we will describe how to
construct the feature space and classification space.

3.1. Selection of Pattern Space

Six commonly used statistical parameters for bearing diagnosis (Howard, 1994) are Peak,
RMS, CrestFactor, Kurtosis, ImpulseFactor, and ShapeFactor, which are considered here to
form the pattern space. In terms of the sampling data vectors, these parameters can be defmed
as follows:

where s denotes the mean value of the time domain signal.

Note that Peak and RMS have a unit, while the other four parameters are dimensionless.
For normalization, we use the RMS value of the undamaged bearing data, denoted by RMSo,

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

 381

as a normalizing parameter. Based on our experiment, we have found that RMSo can improve
the robustness of the energy level parameters, Peak and RMS, against variation in the bearing
operating conditions. RMSo is determined by considering all the possible loads and the
rotating speeds for the bearing under study In terms of RMSo, we can rewrite equations
(3) through (8) and define the following normalized vector z

It has been shown in the literature (Howard, 1994) that Peak and RMS values directly
reflect the energy level of the vibration signals. Since the localized bearing defects result in
structural vibrations, these two parameters are generally used to indicate the presence and
severity of the bearing defects. CrestFactor and Kurtosis are seen less dependent on the
vibration level but sensitive to the spikiness of the vibration signals. As such, they can provide
early indication of significant changes in the vibration signals. As the damage increases, the
vibration signals become more random, and the values of CrestFactor and Kurtosis could
decrease to the undamaged level. ImpulseFactor and ShapeFactors have similar effects like
CrestFactor and Kurtosis.

3.2. Trend Analysis

Since the statistical parameters of bearing vibration signals are affected by the bearing
operating conditions, for example, the rotating speed and load, trend analysis is conducted
to investigate the effect of the bearing operating conditions on the statistical parameters. We
intend to use two energy-level parameters ZI and z2, while examining the similarity among
the other four parameters Z3 to zs, which pertain to spikiness.
Figure 2(a) shows the variation of the four parameters Z3 to Z6 with the rotating speed,
and Figure 2(b) shows that with the load. In the figures, the lines with &dquo;*&dquo; represent the
undamaged bearings, while the lines with &dquo;+&dquo; represent bearing with a single spall on the
outer race. It can be observed from Figure 2(a) that CrestFactor and Kurtosis have a similar
trend, whereas ImpulseFactor and ShapeFactors share another similar trend. Furthermore,
these parameters are calculated for the same bearings under two different loads, and Figure
2(b) shows that all the four parameters have a similar trend corresponding to the different
loading conditions under the same speed. A similar trend was found in our study for other
types of bearing defects. Table 1 summarizes the trend analysis.
Based on the trend analysis, we choose Kurtosis against CrestFactor, and ImpulseFactor
against ShapeFactor. Note that Peak4lMSa (zl ) is similar to CrestFactor, and based on our
numerical experiment, we use the latter instead of the former. Hence, we define the following
vector xp to form the pattern space:

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

382

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

 383

Table 1. Summary of the trend analysis.

,%,F- -. - 1, 1 - 0---A

*
+, #, and indicate similarity

The dimension of our pattern space is four, that is, R = 4.

4. SELECTION OF FEATURE SPACE

4.1. Featiire Extraction

Probably the most important aspect of pattern recognition is selection of feature space. Proper
and efficient feature extraction allows large dimension reduction, yet retaining as much as
possible the useful information. Constructing a feature space by a planar image is especially
advantageous owing to the fact that a planar image of data can be readily perceived and
analyzed by a human observer. For bearing diagnostics, we start by recording the vibration
signals using acoustic sensors or accelerometers. The raw data are a series of numbers
representing the amplitude of signals at discrete times. Although by using the aforementioned
statistical parameters and trend analysis, dimension is reduced to four, it is desirable that two
indices should be extracted, which gives a combination of the other parameters.
We consider two key factors that statistical parameters might give us about the bearing
vibration signals. One is the spikiness, and the other includes the shape of the amplitude
distribution density and the energy level of the vibration. The first one is straightforward
and can be provided by CrestFactor, Kurtosis, and ImpulseFactor. We choose Kurtosis in
view of its robustness in variation with the operating conditions, as indicated in Figures 2(a)
and (b). Based on the studies reported in the literature (Howard, 1994), the shape of the
amplitude density distribution can be reflected by the statistical parameters pertinent to the
impulsiveness of the signal, that is, CrestFactor, Kurtosis, and the ImpulseFactors. Therefore,
the second feature is selected as

By substituting equation (5) for CrestFactor into equation (11) and using a logarithmatic scale
for the second feature, the vector defming the feature space is given as

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

384

Figure 3. Feature space without intraclass transformation. ,

&dquo;I

The dimension of our feature space is two, that is, L = 2. Accordingly, the prototype vector
is defined as

Note that the actual statistical parameters used in the computation of our approach are Peak,
RMSo, Kurtosis, and ImpulseFactor.
Figure 3 shows the feature space constructed using equation (12) considering a range
of operating speed and load conditions for six different types of bearing defects. It can
be seen that samples representing different groups are scattered and overlapped. This will
impose difficulties in classification. For this reason, the intraclass transformation is applied
to introduce the clustering effects on the samples within the same class.

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

 385

4.2. Intraclass Transformation

The intraclass transformation is designed to increase the clustering of prototypes in the same
class. This is realized through minimization of a metric between the points defining the
class. For the kth class, there are a total of Mk prototypes yf> , each being a point in the
two-dimensional feature space. The intraclass transform in the feature space is defined as

where apostrophe indicates the intraclass transformation. Denoting o and or as the

variances of the two variables in the feature space, W~~~ is defined as the following diagonal
matrix:

Variances Q and or are determined, respectively, by the first and second row of the following
data matrix y(k) formed by M~ prototypes

Through the intraclass transformation, the mean square intraset distance of the kth class
is minimized (Andrews, 1972). It can be seen from equation (15) that the two coordinate
dimensions are inversely proportional to the variance of their own dimension. This can
be interpreted as stating that a small weight is to be given to those coordinates with large
variances because these particular coordinates have little in common over the prototypes of
the kth class. For those dimensions with near constant values, the variance will be small,
which will imply large weighting. Figure 4 shows the effects of the intraclass transformation
on the same set of data used to obtain Figure 3. Clearly, the samples belonging to the same

class are clustered.

For bearing diagnostics, the vector x f after the intraclass transformation becomes

where W is defined in the same way as equation (15); except in this case, the variance will
be determined based on x f spanning a number of samples.

5. FORMATION OF CLASSIFICATION SPACE

Recall that the problem of forming the classification space is to partition the feature space.
For given K classes, Sl, ... , Sr ... , SK , mathematically, this problem is to find a function
that can measure each point in the feature space in terms of its degree of membership to a
given class. This function is called the discriminant function in pattern recognition and is

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

386

Figure 4. Feature space with intraclass transformation.

defined such that for all points xf in the feature space within the region describing S~ , there
exists a function gk (x) such that

In other words, within the region, the kth discriminant function will have the largest value.
The piecewise linear discriminant function is used, which can approximate the nonlinear
boundaries separating the different class regions. This function is defmed by the minimum
distance between a point x f and the prototype points in class Sk

Point x f belongs to class Sk if the distance is minimum. The classification then comes to
determine the smallest distance between all of the prototypes of Sk and the unknown xf .
Mathematically, this can be written as

Through mathematical manipulation (Andrews, 1972), the piecewise linear discriminant

function canbe given as

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

 387

Figure 5. Classification space.

The boundaries separating the different class regions are determined by the following

equation:

Figure 5 shows the boundaries determined by equation (22), providing a very distinct partition
of the different class regions for the different bearing defects.

6. CASE STUDY

6.1. Experiment Setup

The developed method was used to diagnose the defects of the tapered roller bearings used
in railway freight cars, as shown in Figure 7. Since the damaged bearings could cause train
derailment, railway bearing diagnostics is very important. A bearing test rig, as shown in
Figure 8, was set up in the Transportation Technology Center of The Association of American
Railroads. Two types of railway bearings, F and E, were tested with seeded defects under two

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

388

Figure 6. Results of bearing diagnostics.

Figure 7. Schematic of a railway bearing (Courtesy of the Association of American Railroads).

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

 389

Figure 8. Bearing test rig (Courtesy of the Association of American Railroads).

Table 2. Bearing condition classes.

different loads, 8K and 33K pounds for the first type and 8K and 27.5K pounds for the second
type. The wheel speed tested was from 25 to 80 mph. The data used were the vibration signals
measured from the bearing housing and sampled at 523 kHz. Table 2 summarizes the bearing
condition classes considered in our study, and Figure 9 shows the pictures of the defects.

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

390

Figure 9(a). Pictures of bearing defects (Courtesy of the Association of American Railroads): Broken
roller.

Figure 9(b). Pictures of bearing defects (Courtesy of the Association of American Railroads): Single
outer race spall.

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

 391

Figure 9(c). Pictures of bearing defects (Courtesy of the Association of American Railroads): Multiple
innerrace spall.

Figure 9(d). Pictures of bearing defects (Courtesy of the Association of American Railroads): Multiple
outer race spall.

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014

392

6.2. Bearing Diagnostics

Bearing diagnostics was carried out by using our test program written in MATLAB, according
to the computation procedure summarized in Figure 1. First, the statistical parameters were
calculated using equations (3) through (8). Each sample was then located in the feature space
according to equation (12) and shown in Figure 3. Through the intraclass transformation
(equation (14)), the prototypes were clustered in the feature space as shown in Figure 4. We
then constructed the six linear discriminant functions by equation (21) and determined the
classification space, as shown in Figure 5, through learning from the prototype data of the
six classes. It is interesting to note, from Figure 5, that severity of the bearing damage can
be shown by the migrating trend of class 2 to class 3, and of class 4 to class 5. Referring to
Table 2, class 2 and class 3 belong to the same type of defects, except that the latter represents
more severely damaged bearings. This is also true with class 4 and class 5. As such, severity

of the bearing damage is indicated by associating testing samples into different regions in the
classification space.
To test the effectiveness of this method, we used a set of data with an unknown type of
defect. We used our test program and located the testing data in the region of broken rollers,
as shown in Figure 6 (the &dquo;+&dquo; symbol). The result was verified by checking the bearing,
which indeed had a broken roller defect.

7. CONCLUSIONS .

A bearing diagnostics method has been developed based on pattern recognition of statistical
parameters. This method is unique in that it combines the commonly used statistical
parameters to provide an effective tool for bearing diagnostics, and it is simple both
conceptually and computationally. The method has been tested and the results show that
it is an effective method for bearing diagnostics, particularly in providing a simple way
for a human observer to visualize the diagnostics results on a two-dimensional plane of the
computer.

Acknowledgments. The msearrhreported here was possible under the financial support fium the Transport
Canada and the experiment support from the Association of American Railroads.

REFERENCES

Andrews, H. C., 1972, Mathematical Techniques in Pattern Recognition, Wiley-Interscience, New York.
Braun, S., 1986, Mechanical Signature Analysis, Academic Press, London.
Howard, 1., 1994, A Review of Rolling Element Bearing Vibration Detection, Diagnosis and Prognosis, Defense
Science and Technology Organization, Australia.
McFadden, Smith, J. D., 1984a, "Model for the vibration produced by a single point defect in a rolling element
P. D. and
bearing," Journal of Sound and Vibration 966
(1), 69-82.
McFadden, P. D. and Smith, J. D., 1984b, "Model for the vibration produced by multiple point defects in a rolling
element bearing," Journal of Sound and Vibration 966 (2), 263-273.
Rao, B.K.N., 1996, Handbook of Condition Monitoring, Elsevier Advanced Technology, Oxford, UK.
Yoshioka, T, 1993, "Detection of rolling contact sub-surface fatigue cracks using acoustic emission technique,Journal
of the Society of Tribologists and Lubrication Engineers 4 (4), 303-308.

Downloaded from jvc.sagepub.com at MEMORIAL UNIV OF NEWFOUNDLAND on August 2, 2014