Rotating machine fault detection using principal component analysis of vibration signal

Rotating machine fault detection using principal
component analysis of vibration signal
Tristan Plante, Lucas Stanley, Ashkan Nejadpak, Cai Xia Yang
Department of Mechanical Engineering
University of North Dakota
Grand Forks, United States
ABSTRACT — Current vibration based maintenance methods
can be improved by using principle component analysis to
identify fault patterns in rotating machinery. The intent of
this paper is to study the effects of using principle component
analysis in a vibration based fault detection process and to
understand the capability of this method of maintenance.
Because vibration-based maintenance practices are capable
of identifying motor faults based on their respective vibration
patterns, principle component analysis observed in frequency
domain can be used to automate the fault detection process.
To test this theory, an experiment was set up to compare
health conditions of a motor and determine if their patterns
could be grouped using principle component analysis. The
result from this study demonstrated that the proposed
method successfully identified healthy, unbalance and
parallel misalignments of rotary rotor. Therefore, it is
capable of detecting faults in early stages and reducing
maintenance costs.
Keywords—Principal Component Analysis; FFT; Fault Detection;
Vibration Analysis; Severity
I. INTRODUCTION
Within the modern day industrial and energy sectors, basic
electric motors are the backbone in a large variety of applications.
Thus, the maintenance and upkeep of these machines are an
essential component to keeping up with production demands.
When plants experience an unexpected shutdown due to motor
failure, the costs associated with lost production time can be
astronomical. It would be ideal if shutdowns are kept to a
minimum if not avoided entirely. Because of this, it is common
practice to hire vibration analysts to perform both preventive and
predictive maintenance [1]. These forms of maintenance are
considered to be highly informative and beneficial when
performed correctly.
Typically, vibration analysts use Fast Fourier Transformation
(FFT) to process and interpret vibration data from motors and
other rotating systems. During this process, most of their time is
spent observing data and identifying patterns. The skill of an
analyst is usually dependent on their past experience and training.
While relatively effective, these costs are still considerably high
and greatly depend on the analyst’s training, experience and
equipment. Even when analysts have the capability and means to
promote effective maintenance practices, mistakes can still be
made that affect production. This highlights the limitations of
traditional vibration analysis and outlines the need for
improvements in predictive maintenance. One such improvement
is the application of principal component analysis (PCA) to FFT
data. This process is explored in [2].
Due to substantial improvements in computer processing, the
concept of autonomous pattern recognition has become more
effective. As mentioned earlier, PCA has an application to this
process. Applying PCA would reduce the training needed by
analyst and hopefully increase the accuracy of predictive
maintenance. Rather than analyze vibration data manually, PCA
can be used computationally to find the patterns associated with
mechanical faults. For example, a motor that is misaligned can be
grouped with other motors that share the same fault condition.
The same applies for mechanical faults of other types. This notion
points towards the automation of the vibration analysis process.
This would reduce a large portion of the costs associated with the
maintenance industry. For this reason, an experiment was
conducted to apply PCA towards fault detection and determine if
this method could be deemed both efficient and reliable enough
to merit further study and perhaps see an application within an
industrial setting.
The tests conducted involved three operating conditions of a
typical axle system connected to an induction motor. The axle
itself was under load supported by two bearing couples whose
respective vibration data was recorded using a set of
accelerometers mounted at four locations. Resultant FFT data was
processed using PCA which helped outline any patterns that
emerged from the experimental data. The validity of these
patterns were tested using several unknown data sets to determine
if they could be identified correctly.
The remainder of this paper is structured as follows. Section II
introduces the fundamentals of fault detection and principal
component analysis. Experiment setup, including data collection
process, experimental system faults, PCA experiment conditions,
and Matlab program process will be discussed in section III.
Experiment data displacement and analysis using PCA method
will be described in sections, IV and V, respectively. The paper

concludes with future work and a summary recapping the main
advantages of the proposed method in Sections IV and V.
II. BACKGROUND INFORMATION
A. Fault detection
In short, fault detection is the common practice of identifying
machine faults before they become severe enough to cause harm
to itself or the surrounding area. Experiments performed in
previous work show that vibration analysis can be used to detect
faults in motors [3]. Like most practices, there are a number of
methods to do this. However, vibration analysis is the most
common. It is capable of detecting the widest range of mechanical
faults. It can identify faults from mechanical looseness, to bearing
faults, to misaligned rotors and shafts [4]. Vibration analysis can
be used to determine the type and severity of faulty systems. This
is possible because mechanical faults produce patterns in vibration
data that are considered to be specific to their identity. For
instance, unbalanced rotors resonate a large amplitude peak in FFT
at the same frequency of the running speed also known as the 1X
peak [1]. Additionally, mechanical looseness responds with
several peak multiples of the running speed 1X, 2X, 3X etc. Of
course, these are just a few examples of faults that can be detected
using vibration analysis.
B. Principal Component Analysis (PCA)
Principal component analysis is a statistical approach to
identifying and isolating large amounts of data with multiple
variables. In other words, PCA is a feature extraction technique
capable of distinguishing data values based on their respective
variance to the rest of the data set [5]. This is done by calculating
the amount of variance between several data sets and assigning
each variable its own dimension to determine which variables have
the largest impact on the variance within the data [6]. By doing
this, data of different types are separated into “groups” which can
be recognized by identifying one data set within that group. This
is commonly used to find patterns in statistical information as seen
in [7] which explains the application of PCA to organize a large
number of cells into groups based on the genes they possess.
Fault detection, in its nature, involves the identification of
vibration patterns. This paper explores how PCA could be applied
to determine these patterns computationally without the need of
manual data analysis. By comparing large numbers of data sets,
PCA can be used to group known FFT vibration patterns based on
their relevant trends. At which point, unknown data can be
introduced so it can be grouped into their respective fault
categories. If successful, this paper may provide a new method of
maintenance within industrial applications and could stand as the
foundation for future research regarding this application.
III. EXPERIMENT SETUP
A. Data Collection Process
Three health conditions of a motor were studied in a series of
experiments: healthy, unbalanced and parallel misalignment. Each
of these experiments were conducted using the Machinery Fault
Simulator (MFS) purchased from SpectraQuest Inc. Operating at
25Hz (1500 RPM), data was recorded using four PCB
Piezotronics’ IMI model 603C01 accelerometers. Fig. 1 shows an
image of this setup.
Data acquired from the accelerometers were processed using a
PCI-4498 data acquisition device (DAQ) purchased from
National Instruments. This device’s output was then analyzed
using the “Sound and Vibration Assistant” software produced by
National Instruments set up to read 2000 samples at a rate of
20,000 Hz. The raw time-based vibration data was then converted
into an FFT using the zoom power spectrum function, and a
Hanning window with a magnitude based setting. This was set up
to record data over a 15 second period. Throughout this period,
the FFT used a root mean squared average to acquire a reliable
estimate of the vibration condition. Because most vibration
analysts use the low frequency spectrum for fault detections, the
frequency range was limited from 0 Hz to 1000Hz and the number
of line set to 1000 to produce an adequate spectrum resolution of
1 Hz.
Fig.1: Machinery Fault Simulator setup
Two accelerometers were placed on each bearing housing in
the horizontal and vertical axis, perpendicular to the direction of
the shaft. Fig. 2 shows the placement of the two accelerometers on
the rear bearing housing. This placement is identical to the
placement of the accelerometers on the bearing housing closest to
the motor used to operate the system. In fig. 2, the shaft is going
into the page. For the experiments conducted in this paper, no
accelerometers were placed in the axial direction because data
taken in that direction produced results that were less informative
than their radial counterparts.
Fig. 2: Placement of accelerometers on the rear bearing housing

B. PCA Experiment Conditions
For any health condition of a motor, a change in spectral
patterns can be observed in FFT when a system experiences
varying degrees of fault severity. To account for the variance in
vibration data in the PCA method, fault severity was altered for
every test. Since a healthy motor cannot experience different
degrees of healthiness, additional load was added to the motor
under healthy conditions in order to change the response of the
spectral pattern, thus simulating changing circumstances.
Although load does not directly create a larger fault severity, for
the purpose of this experiment it was assumed that an increased
severity or load would intensify the response in the vibration data.
For the series of healthy system tests, washers were added
uniformly around the rotor to increase the amount of load applied
to the system. This was done in multiples of four, starting zero
and going until there were a total of 20. Unbalanced tests were
conducted by adding washers in sets of two, starting at zero and
going to 20. Fig. 3 shows the rotor used to add load and simulate
unbalance. Misalignment was simulated via the misalignment dial
in place on the MFS (see fig. 4). These tests increased by five
milli-inches (mils) starting at 10 mils and ending at 35 mils. The
unknown healthy test was taken with nine washers arranged
uniformly the rotor. The unknown unbalanced test was taken with
seven washers. The unknown misalignment test was conducted at
27 mils.
Fig. 3: Rotor used to add load and simulate unbalance
Fig. 4: Dial used for introducing parallel misalignment
By changing the severity of the faults and the load of the
healthy condition, the robustness of the PCA method was tested.
This was done to determine if PCA could correctly group faults
of the same type together that had different operating conditions.
This concept was represented in test set seven which was recorded
in a similar fashion. The only difference was that the severity/load
of the unknown data set lied between the severities of the known
faulty test sets. This allowed for the possibility of trend mapping
the data set and approximating the severity of an unknown fault.
C. Experimental System Faults
As mentioned earlier, the health conditions of the motor that
were tested were healthy, parallel misalignment, and rotor
unbalance. The healthy data set was taken to act as a control group
for the other test sets. The fault conditions studied were tested to
determine if the PCA method could correctly group health
conditions by type and identify fault severity. With that said, the
severity of one fault cannot be accurately compared with that of
another. Instead, the intention of this experiment was to evaluate
the trend of each fault group individually. For example, the most
severe instance of misalignment cannot be compared with the most
severe instance of unbalanced because the systems were operating
under different conditions. Rather, it was intended that for each
fault, a trend in fault severity would emerge. This would allow for
an unknown test set to have the severity of its fault estimated.
D. Matlab Program Process
Once data was collected from each experiment, text files were
created for each data set displaying the running speed and
amplitude values at each frequency. For the majority of the
experiment, data was left unedited for the purpose of
distinguishing the application of PCA towards vibration based
fault identification. However, for comparison purposes, this paper
also explains the benefits/limitations of adjusting data sets to
improve fault groupings. This process is explained further in the
“Experimental Data” and “Analysis” sections of this report.
After being formatted, the data associated with their respective
accelerometers was assigned to a cell containing similar data from
other data sets. Once each data set was organized, the program
processed them using Matlab’s “pca” function to produce a PCA
using the correlation matrices. This process was repeated for each
accelerometer.
Because each recorded test was controlled, the location of
unknown tests could be compared to that of each known test by
creating a region in which the fault was known. By doing this,
unknown tests could be identified by determining if the test lied
within the proximity of the region specific to that fault condition.

IV. EXPERIMENT DATA
A. Three Fault Spectrum Results
Figs. 5-7 displays three examples of the FFT data recorded
during the experiment. Each graph shows the FFT of a healthy
rotor, misaligned rotor, and an unbalanced rotor. Traditionally,
these are the types of graphs vibration analysts used to identify
faults and their corresponding severities.
Fig. 5: Healthy Rotor Data in FFT
Fig. 6: Parallel Misaligned Rotor in FFT
Fig. 7: Unbalanced Rotor Data in FFT
While it is easy to identify the differences between each graph
from a broad perspective, it is quite difficult to identify the specific
peaks that represent the fault as a whole. With the exception of
unbalance, which has a 1X peak that dominates the graph by a
significant margin, it becomes increasingly difficult to identify the
more subtle differences between the healthy and misaligned FFT
data set. These subtle changes are what make vibration analysts
training and experience so important. PCA presents an alternative
approach to this method, which may be advantageous to analyzing
such subtle differences.
B. PCA of Fault data PC 1 and PC 2 (2D)
The processed data sets for each accelerometer are represented
in figs. 8-11. Each data set displays both the scores (red) and
loadings plot of the PCA. In short, the square plot is represented
by a large number of points each of which represents a single
frequency value in the graph. On the other hand, the points
representing the loadings portion of the graph are each a single test
(unbalance, healthy, etc.). Additionally, each point in the loadings
portion of the graph are marked with either an “H” (healthy), a “U”
(unbalanced), or an “M” (misalignment) as well as the test number
for that fault type in order of occurrence.
Fig. 8: Accelerometer 1 PCA
0 100 200 300 400 500 600 700 800 900 1000
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
FFT of Motor Data
Frequency (Hz)
Magnitude(grms)
0 100 200 300 400 500 600 700 800 900 1000
0
0.05
0.1
0.15
0.2
0.25
FFT of Motor Data
Frequency (Hz)
Magnitude(grms)
0 100 200 300 400 500 600 700 800 900 1000
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
FFT of Motor Data
Frequency (Hz)
Magnitude(grms)

C. PCA of Fault data PC 1, PC 2, and PC 3 (3D)
Figs. 12 through 15 represent the same data presented in the
previous section of this report. The only difference is that a third
dimensions was added by including the third principal component
reference to the graph. Besides this, the data is presented in the
same manner with both the loadings and score plot displayed in
each graph.
D. PCA of Adjusted Data
Fig. 16 displays the results of vibration data that has been
scaled and ordered before being processed in PCA. This adjusted
data set was created by subtracting a spectrum produced by
averaging each of the healthy data sets. Amplitudes that were left
with negative values were set to zero to standardize the data set.
After this process, each data set was put in terms of order by
assigning the amplitude at a certain frequency to that of the
nearest order multiple of the running speed. This was done so
that data sets operating at different speeds could be compared
more readily with controlled tests. Next, data was scaled and

standardized using a square-root operation and computing the z-
score of the resultant amplitudes. Both these processes were done
to make larger peaks more distinct.
Fig. 16: Accelerometer 1 Adjusted PCA
Unknown data sets are represented with a square box
surrounding an “X” and are labeled in order of intended fault
grouping. That is “1” relating to healthy, “2” for parallel
misalignment, and “3” for an unbalanced rotor.
V. ANALYSIS
A. Fault Grouping in Two Dimensions
Based on the data presented in figs. 8 through 11, it can be stated
that PCA is capable of detecting mechanical faults based on
processed FFT vibration data. Within the loadings plot of each
graph, each of the fault conditions can be seen grouped together
with other points of the same fault type. However, there are some
exceptions to this. As seen in figs. 12 -15, there is some overlap
between the healthy and misaligned loading points. Because of
this, it would be difficult to correctly identify an unknown point
that lied within this region. This highlights the advantage of using
more than one accelerometer as well as the addition of another PC
dimension. Should one accelerometer produce a PCA graph that
does not clearly identify an unknown test set, other accelerometers
can help clarify the results by providing alternative views on its
fault type.
As for the score plot, there is less clarity about the information
shown compared to the loading plot. This is because each point
represents a single amplitude value with respect to each data set.
Many of these frequencies experience little change and are thus
destined to lie within the same area. The points located near the x-
axis of each graph could depict some type of trend. These points
may represent the increase in fault severity for unbalance or
misalignment. However, due to the setup of the experiment and
Matlab code, it is not possible to verify such a trend. Even if the
points were identified as a certain fault group, this is, in fact, the
only trend readily identified by the graph meaning the other two
conditions could not be identified based on score plot alone.
Needless to say, the loadings plot offers much more clear and
precise information.
B. Fault Grouping in Three Dimensions
Within figs. 12-15, the addition of a third PC dimension shows
a distinct improvement in the grouping of each fault condition.
Compared to the previous graphs involving only two PC
dimensions, the misaligned and healthy groups in figs. 12-15 do
not lie within the same region. Instead, each fault group becomes
more distinct with the addition of a third PC dimension and would
be easier to recognize computationally. This concept is explained
further by the incorporation of unknown data sets in each fault
grouping.
C. Unknown Data Representation
Looking once again at fig. 8 through 11, the unknown data sets,
represented by an (X), can be seen near or within each fault group.
Despite the fact that not every unknown data point lies within the
immediate proximity of each fault group, a simple observation
suggests that a point lying near a specific fault group can be
considered that same fault.
Whether or not user intuition can identify the unknown data set
is irrelevant. The purpose of this program is to determine the type
of fault computationally. To do this, the program was written to
incorporate an algorithm that judged the proximity of an unknown
data point with each of the fault groups. Each fault group would
be assigned a score based on the proximity of the unknown to the
fault group. The fault group with the highest score would be
assigned to the unknown data point.
Severity Trend
In most of the graphs displayed previously, the points within
the healthy region typically form a tighter cluster than that of the
other faults. Both the unbalance and misaligned data sets appear
to form more narrow regions that resemble a linear function. This
can be explained by looking at one of the fault groups in closer
detail as shown in fig. 17.
Fig. 17: Example of fault severity trend showing
Accelerometer 1’s Misaligned Fault Group
-0.1 0 0.1 0.2 0.3 0.4
-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
1
2
3
4
5
6
PCA of Motor Data
PC 1
PC2
1
2
3
45
6
1
23
45
6
Unknown 1
Unknown 2
Unknown 3
Healthy
Misaligned
Unbalanced

This figure displays how the data sets are in the same order as
when they were read by the program. While only the fault trend
for misalignment is shown, each of the faults in the other graphs
display a similar pattern. For the purposes of this paper, fig. 17 is
analyzed to shed further light on this subject.
The severity levels of fig. 17’s specific fault cluster are shown
in points 1 through 6 starting at 10 mils and proceeding to 35 mils
in increments of 5 mils respectively. This arrangement is
significant because as each data set was collected, the severity of
that specific fault condition was increased. Meaning that data with
higher severity levels were located further from the healthy cluster
as seen in fig. 5. This concept holds true for each of the PCA
graphs and outlines the potential for PCA to determine the type of
fault present, and its approximate severity.
D. Adjusted PCA Interpretation
As mentioned earlier, fig. 16 corresponds to the results of a
loading graph whose respective FFT data has been scaled, ordered,
and standardized in an attempt to improve the fault identification
process. Based on the figure, each of the fault groupings appear to
be more independent of each other. That is, each group is more
distinct than the PCA graph produced without adjusting the data
set. With that said, it may beneficial to adjust the data set before
processing it in PCA and may provide a better result; however, the
downside to this method is that, based on this figure, no trend line
can be readily plotted and no estimate of the severity can be
determined.
VI. FUTURE WORK
In addition to actual identification of an unknown data set,
future work needs to be done on the relationship between data sets
from other motors. Specifically, if data from a variety of motors
can be processed in PCA to produce a program capable of
identifying the type of fault present in any type of motor running
at a certain speed. In principle, so long as the data was scaled and
filtered correctly, this concept would hold true for most motors
because of the fault based vibration patterns described in [1] and
proven in [3]. Engineers could then construct a large database in
which to refer to for machine maintenance for, potentially all
industrial applications involving rotating equipment. Therefore,
this concept should be studied in future experiments to reveal more
about the limitations and benefits of this process.
VII. CONCLUSION
In summary, PCA can be used to distinguish between motor
faults of varying types. This, of course, is done in conjunction with
several accelerometers as the use of supplementary data can
provide a more in-depth look at the data. By comparing data from
multiple accelerometers, steps can be taken towards automating
the fault identification process. This is further supported by the
similarities between the unknown data set and the other fault tests.
Additionally, when dealing with a system running at a constant
speed, the data has shown a correlation between the faulty data
set’s severity and its respective position in the PCA graph.
Meaning that the fault severity of an unknown data set could be
potentially estimated. With that said, PCA provides a simple and
inexpensive alternative to current maintenance techniques. Not
only can industries become more efficient by reducing production
costs, but they may also see an increase in productivity by avoiding
unplanned plant shutdowns thus improving the industry as a
whole.
ACKNOWLEDGEMENT
The work reported in this paper was funded by ND EPSCoR New
Faculty Start-up Award UND0019805 and ND EPSCoR
Advanced Undergraduate Research Awards UND0020387 and
UND0021402.
REFERENCES
[1] Pickens, A. S. Vibration Trouble-shooting Field Guide, PDM
Engineering.
[2] A. D. Nembhard, J. K. Sinha, A. J. Pinkerton and K. Elbhbah,
“Combined vibration and thermal analysis for the condition
monitoring of rotating machinery,” Structural Healthy
Monitoring, vol. 13, 2014.
[3] T. Plante, A. Nejadpak and C. Yang, “Faults detection and
failures prediction using vibration analysis,” IEEE
AUTOTESTCON 2015 Proceedings, 227-231.
[4] SpectraQuest, Inc., Applied vibration analysis training
manual & laboratory exercises, Richmond: SpectraQuest,
Inc..
[5] Z. Xia, S. Xia, L. Wan and S. Cai, “Spectral regression based
fault feature extraction for bearing accelerometer sensor
signals,” Sensors, vol. 12, no. 10, 2012.
[6] V. Powell and L. Lehe, “Principal component analysis
explained visually,” DISQUS, Available:
http://setosa.io/ev/principal-component-analysis/. [Accessed
11 July 2016].
[7] J. Starmer, StatQuest: Principle Component Analysis (PCA)
clearly explained, 2015.
[8] E. L. Russell, L. H. Chiang and R. D. Braatz, “Fault detection
in industrial processes using canonical variate analysis and
dynamic principal component analysis,” Elsevier, 2000.

Rotating machine fault detection using principal component analysis of vibration signal

More Related Content

What's hot (20)

Viewers also liked (8)

Similar to Rotating machine fault detection using principal component analysis of vibration signal (20)

Rotating machine fault detection using principal component analysis of vibration signal