See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/339123818

Detecting Rolling Elements Bearings Faults

Chapter · February 2020

CITATIONS READS

0 909

1 author:

Jaafar Alsalaet
University of Basrah
21 PUBLICATIONS   27 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Developing mechanical ventilator View project

Smart machine health monitoring system View project

All content following this page was uploaded by Jaafar Alsalaet on 08 February 2020.

The user has requested enhancement of the downloaded file.

 Detecting Rolling Elements Bearings Faults
Dr. Jaafar Alsalaet
College of Engineering-University of Basrah

Abstract
Rolling elements bearings are the most important machine elements. Failure of these elements
may cause catastrophic breakdown and result in costly downtime. Bearing condition monitoring
thus plays an important role in the current machinery maintenance strategies. Due to the complex
structure of rolling elements bearings, the observed vibration signals are normally corrupted by
noise and random patterns especially in the early stages of bearing defects. It is important to early
discover the incipient bearing anomaly before problem development which may lead to
catastrophic failure. A number of signal analysis techniques, both in time and frequency domains,
have been introduced to extract useful information from the noisy vibration signal. In this
workshop, some traditional and advanced bearing vibration analysis techniques will be explored
with some practical examples.

Contents
1. Introduction .................................................................................................................................. 3
Bearing Vibration Patterns............................................................................................................... 4
Sources of Vibration in a Roller Bearing .......................................................................................... 5
Bearing Failure Stages ...................................................................................................................... 5
2. Condition Based Maintenance and Precision Maintenance ...................................................... 11
3. Time Domain Techniques ........................................................................................................... 14
RMS, Crest and Kurtosis Factors .................................................................................................... 14
Time Waveform ............................................................................................................................. 15
4. Frequency Domain Techniques .................................................................................................. 16
FFT Spectrum ................................................................................................................................. 16
FFT Waterfall .................................................................................................................................. 17
5. Power Cepstrum ......................................................................................................................... 18
6. Shock Pulse Method ................................................................................................................... 20
7. Spike Energy ............................................................................................................................... 22
8. High Frequency Detection .......................................................................................................... 24
9. Envelope Detection and Demodulation ..................................................................................... 24
Lowpass Filter Enveloping.............................................................................................................. 24
Hilbert Transform Enveloping ........................................................................................................ 26
PeakVue Technique Enveloping ..................................................................................................... 27
Case Study: Multistage RO Water Pump ....................................................................................... 28
Case Study: Catacarb Pump P202-B............................................................................................... 30
10. Signal Enhancement Techniques ............................................................................................ 33
Signal Decomposition .................................................................................................................... 33
Adaptive Noise Cancellation .......................................................................................................... 34
Adaptive Linear Enhancer .............................................................................................................. 35
11. SPM High Definition and Symptom Enhancing ....................................................................... 36
HDm and HDc ................................................................................................................................. 37
SPM HD Time Signal ....................................................................................................................... 37
SPM HD Spectrum .......................................................................................................................... 38
12. Wavelet Transform ................................................................................................................. 41
Case Study: P304 Screw Oil Pump ................................................................................................. 43
13. Automated Diagnosis and Advanced Techniques .................................................................. 46
Artificial Neural Networks.............................................................................................................. 46
Principal Component Analysis ....................................................................................................... 48
Support Vector Machine ................................................................................................................ 50
Closing Remarks-Gap between academic research and professional practice ............................. 54
1. Introduction
All rotating machines have bearings to support the rotating parts and isolate them from the
stationary parts. Most of these bearings are rolling-element bearings. A rolling element bearing
comprises of inner and outer races, a cage and rolling elements. There are five types of rolling
elements that are used in rolling-element bearings: balls, cylindrical rollers, spherical rollers,
tapered rollers, and needle rollers.

Figure 1.1 Deep Groove Ball Bearing

Figure 1.2 Types of Rolling Bearings

Bearing Vibration Patterns
Defects can occur in any of the parts of the bearing and will cause high-frequency vibrations. In
fact, the severity of the wear keeps changing the vibration pattern. In most cases, it is possible to
identify the component of the bearing that is defective due to the specific vibration frequencies
that are excited. Raceways and rolling element defects are easily detected. However, the same
cannot be said for the defects that crop up in bearing cages. Though there are many techniques
available to detect where defects are occurring, there are no established techniques to predict
when the bearing defect will turn into a functional failure.
In section 3 and next section, we will introduce modulation process and how vibration is excited by
bearing faults. We will see how bearing defects generate both the bearing defect frequency and
the ringing random vibrations that are the resonant frequencies of the bearing components.
Bearing defect frequencies are not integrally harmonics to running speed. However, the following
formulas are used to determine bearing defect frequencies. There is also bearings database
available in the form of commercial software that readily provides the values upon entering the
requisite bearing number.

N  Db 
BPFI = 1 + cos θ  × RPM

2  Dp 
N  D 
BPFO = 1 − b cos θ  × RPM
2  D p 

(1.1)
1 D 
FTF = 1 − b cos θ  × RPM
2  D p 

D p   Db  
2

BSF = 1 −  cos θ   × RPM

2Db   D p  
 


Where Db : ball or roller diameter

Dp : pitch circle diameter of the bearing
N : number of balls
θ : contact angle
BPFI : Ball Pass Frequency (Inner Race)
BPFO : Ball Pass Frequency (Outer Race)
 FTF : Fundamental Train Frequency (Cage)
BSF : Ball Spin Frequency (Rolling Element)
Rules
• The sum of BPFI and BPFO equals to the number of rolling elements
• BPFO = No. of rolling elements x FTF
• Simpler equation to estimate the frequencies (for N between 8 and 12 element):
BPFO ≈ 0.4 × N
BPFI ≈ 0.6 × N

Sources of Vibration in a Roller Bearing

1. Friction: when two surface are sliding one on another, there will be a friction regardless
whether there is a lubricant or not and the specifications of the lubricant. Vibration due to
friction is random and not periodic. It generates broadband noise, vibration at wide range
of frequencies.
2. Stress waves: if we have metal-to-metal contact, we will get very short duration pulses
which excite high frequency. The pulses may be periodic or not.
3. Resonance: when surfaces impact, wide range of frequencies is generated. This will excite
the natural frequencies of the machine components. The natural frequencies may be very
high for roller bearings (above 20 kHz). These very high frequencies can be detected using
special techniques and will be helpful in determining bearing condition as will be discussed
later.

Bearing Failure Stages

Detecting rolling element bearing faults is the highest priority for most vibration analysts.
Detecting the fault at the earliest stage is a priority for maintenance engineer. Before we get into
the specifics of the four stages of bearing failure, I would like to describe how the vibration
changes in general during operation.
If a bearing is poorly lubricated, we can detect an increase in the level of “noise” at very high
frequencies. It is not a specific, single frequency; instead it will depend on a number of factors to
do with the machine’s construction. Suffice to say that you cannot hear it; it is well above your
hearing range.
As the state of lubrication worsens, the level of the noise will increase, but the frequency of the
noise will slowly reduce – it will move from very high frequencies to high frequencies. That is not
to say that you can’t detect the condition at lower frequencies; it is stronger at the higher
frequencies.
As the film of lubricant between the bearing surfaces is reduced further, we will have more and
more metal-to-metal contact, causing “stress waves” to be generated. Stress waves (also referred
to as “shock pulses”) are like ripples in a pond; the moment the metal surfaces make contact, a
wave of energy races away from the point of contact at the speed of sound. It all happens very
quickly.
The subsurface defects will slowly develop due to the extreme forces experienced within the
bearing. The difference is that these defects are likely to be localized; at the bottom of the outer
race for example. The “noise” from the bearing due to poor lubrication is relatively constant (it is
random, therefore not periodic), whereas when a fault condition develops (e.g. a crack or spall), a
new source of periodic vibration will be introduced; that are bearing defect frequencies in eq.
(1.1). If the damage was on the outer race, each time a rolling element passes that location there
will be a spike in the vibration. When the point of damage is between rolling elements, there is no
vibration (well, less vibration). The good news is that we can calculate the frequency of this
vibration (we can determine how often the rolling element will pass that point). The bad news is
that the vibration is very, very low in amplitude.

Figure 1.3 Crack and Spall Development [2]

 Figure 1.4 Defect in the Bottom of Outer Race [1]

Bearing failure stages are divided classically into four stages.

Stage One
In this stage the defect is minor. If the bearing is removed, there will be no signs of damage as the
damage is predominantly sub-surface. The metal-to-metal contact causes random vibration of high
frequencies (5 to 40 kHz). The high-frequency detection techniques such as Shock Pulse, Spike
Energy, ultrasound, and PeakVue and ultrasound measurement (acoustic emission) are effective in
this stage.
 Figure 1.5 Stage One of Bearing Failure [1]

Stage Two
As the fault develops, the sub-surface defects will grow, eventually breaking through to the
surface, causing spalls, cracks, flakes, etc. The forces of the impacting will be greater, and there
will definitely be periodicity to the vibration. Bearing resonance frequencies may be excited that
will show peaks in the frequency range 1000 to 5000 Hz. The high frequency techniques will
continue to be effective. Enveloping (demodulation) will also be effective, with peaks visible at the
bearing forcing frequencies (BPFO, BPFI, BSF and FT – depending on the nature of the fault) along
with harmonics. Harmonics of the bearing forcing frequencies may also be visible in the
acceleration spectrum.

Figure 1.6 Stage Two of Bearing Failure [1]

Stage Three
In this stage, wear is usually now visible on the bearing and may expand through to the edge of the
bearing raceway. The high frequency techniques will still indicate the presence of a fault. The
peaks in the envelope spectrum will continue to grow in amplitude. There will be peaks in the
velocity spectrum that correspond to the bearing forcing frequencies (BPFO, BPFI, BSF and FT) and
their harmonics depending on the fault condition. In some cases, sidebands around the bearing
defect frequency may appear due to amplitude modulation of fault.

Figure 1.7 Stage Three of Bearing Failure [1]

• Outer race defect: When the outer race become defective, such as a spall in the outer race,
there will be BPFO component and when the fault is developed the BPFO harmonics may
become significant in the spectrum. If the outer race in rotating, there will be sidebands at
outer race rotating frequency due to load variation on the defective portion.
• Inner race defect: If the inner race is defective, there will be BPFI and its harmonics when
the fault is developed. Since the inner race is rotating, the load area is also rotating and,
hence, vibration amplitude is varying according to the position of the defect or spall.
Sidebands will exist around the BPFI and its harmonics, spaced at the rotation frequency.
• Ball defect: If the ball or rolling element is defective, then BSF component is generated
with its harmonics. Vibration at 2xBSF may be high due to contact with the inner and outer
race. There will be also sidebands at FTF (cage) because the balls are rotating at FTF around
the outer race. If the outer race is fixed, the sideband spacing is below 0.5x, but if the outer
race is rotating spacing will be higher than 0.5x.
Stage Four
The bearing has considerable damage in this stage. The high frequency detection techniques may
still be used but the amplitude of the high-frequency vibration is now reduced due to the
smoothing of the sharp edges. There is now so much damage that the vibration loses its
periodicity. The peaks at the BDF will drop down, and seemingly random peaks will appear and
the noise floor will rise up. This is true for the velocity spectrum and the envelope spectrum.
As more metal is removed, the clearance between bearing parts is increased and the 1X
component and its harmonics with noticeable noise floor may be presented due to looseness. The
overall vibration velocity RMS is increased.

Figure 1.8 Stage Four of Bearing Failure [1]

2. Condition Based Maintenance and Precision Maintenance
If we do a survey of the maintenance philosophies employed by different process plants, we
would notice quite a bit of similarity despite the vast variations in the nature of their operations.
These maintenance philosophies can usually be divided into four different categories:
• Breakdown or run to failure maintenance
• Preventive or time-based maintenance
• Predictive or condition-based maintenance
• Proactive or prevention maintenance.

Run to failure maintenance

The idea behind run to break maintenance, as the name imply, is to run the machine until failure
or breakdown occurs and then replacing the parts that are damaged when the machine is
completely stopped. This maintenance procedure is inefficient due to a number of factors:
• Maintenance department perpetually operates in an unplanned ‘crisis management’ mode.
• It may also lead to sub-sequent failures and, hence, other parts may be damaged resulting in
higher maintenance cost.
• When unexpected production interruptions occur, the maintenance activities require a large
inventory of spare parts to react immediately.

Preventive maintenance
Also named "planned maintenance", "calendar-based maintenance" and "scheduled
maintenance". The philosophy behind preventive maintenance is to schedule maintenance
activities at predetermined time intervals, based on calendar days or runtime hours of machines
so that they will not fail. This philosophy has less impact on the criticality of the plant used for
equipment that does not run continuously, and where the personnel have enough skill, knowledge
and time to perform the preventive maintenance work. The main disadvantages are:
• Scheduled maintenance can result in performing maintenance tasks too early or too late.
• It is possible that, without any evidence of functional failure, components are replaced when
there is still some residual life left in them.
• It is possible that reduced production could occur due to unnecessary maintenance.
• In many cases, there is also a possibility of diminished performance due to incorrect repair
methods.
Condition based maintenance
Mechanical and operational conditions are periodically monitored, and when unhealthy trends are
detected, the troublesome parts in the machine are identified and scheduled for maintenance.
The machine would then be shut down at a time when it is most convenient, and the damaged
components would be replaced.

When the plant runs in predictive mode, the condition of the running machines is known.

The unexpected failure can be prohibited or at least reduced so that it is not frequently occur.

We can plan maintenance schedule and we can plan production schedule around the
knowledge of the plant condition. That means "reactive" action is no longer a problem since
most work is planned.

There is less need of staff overtime and stress is reduced

Proactive or prevention maintenance

This philosophy is based on some strategies such as "Precision Maintenance", "Reliability Based
Maintenance" and "Reliability Centered Maintenance". This philosophy emphasis on tracing all
failures to their root cause. Each failure is analyzed and proactive measures are taken to ensure
that they are not repeated. Proactive maintenance is the strategy used to anticipate the failure
before it occurs. It includes a wide variety of practice and strategies and utilizes various
technologies. It utilizes all of the predictive/preventive maintenance techniques discussed above
in conjunction with root cause failure analysis (RCFA).
RCFA detects and pinpoints the problems that cause defects. It ensures that appropriate
installation and repair techniques are adopted and implemented. It may also highlight the need for
redesign or modification of equipment to avoid recurrence of such problems. As in the predictive-
based program, it is possible to schedule maintenance repairs on equipment in an orderly fashion,
but additional efforts are required to provide improvements to reduce or eliminate potential
problems from occurring repeatedly. Again, the orderly scheduling of maintenance allows lead-
time to purchase parts for the necessary repairs. This reduces the need for a large spare parts
inventory, because maintenance work is only performed when it is required.
RCFA requires additional tools and skilled staff in order to be implemented. Signal analyzers such
as FFT, signal enveloping, high frequency detection, phase analysis and other techniques may be
required to accomplish the tasks.
Regarding bearings problems, precision maintenance implies that in addition to the bearing
condition monitoring, it is also important to consider:
• Early detection of bearing faults using HFD techniques for example.
• Fault localization, is it in the outer race, inner race, balls or cage. Is it in the loading area or
other area?
• Causes of the fault. Is it due to over-loading of the bearing or other reason such as
contamination, lack of lubrication or sparks. Additional tools such as ultrasound measurement
to detect sparks and oil analysis to assess oil condition are required to accomplish the task.
• Root cause elimination by proper design/re-design, spark elimination and proper lubrication
have to be done before replacing the bearing.

100%

75%

50%

25%

Operating Life
High misalignment Tolerable misalignment No misalignment
Hi unbalance Acceptable unbalance Accurately balanced
High contamination Low contamination No contamination
Extreme loading Acceptable loading Proper loading
High vibration Acceptable vibration Low vibration

Figure 2.1 Expected machine life vs. operating conditions

3. Time Domain Techniques

RMS, Crest and Kurtosis Factors

Root Mean Square
RMS value is an indication of the power contained in a signal and it is widely used in vibration
severity measurement.

N −1

∑x
1 2
RMS = i (3.1)
N i =0

The overall RMS value of vibration in velocity units is not suitable to detect bearings problems
when common 10 to 1000 Hz filter is applied due to the effect of common vibration problems. The
RMS of the filtered vibration signal in acceleration can be used to assess bearing condition. For
example, the high frequency band of 1 to 20 kHz is used for acceleration severity measurement.
The frequency band of 20 to 50 kHz is used in ultrasonic measurement in dB [3]. Spectral emitted
energy in the band of 250 to 350 kHz is used by some instruments to measure acoustic emission in
dB [4].

Crest Factor
Crest factor is the ratio of the peak value (positive or negative) in a signal to its RMS value. It
ranges from 1 for square waves to 1.414 for pure sinusoidal signals and higher values for signals
with pulses or short duration activities. Trending of this factor for roller bearings shows an
increase of Crest factor (rise of short duration pulses with respect to RMS value) then eventual
drop as the fault develops and gets broader. However, there is no vibration analysis software
package currently trends Crest factor.

Kurtosis Factor
The kurtosis factor or forth statistical moment is given by the following equation:
 N −1
1
N ∑( x − x )i
4

K= i =0
2
(3.2)
1 N −1
2

N

∑
i =0
( i ) 
x − x


High kurtosis factor indicates the presence of repeated impulses. It lends itself to the detection of
whether the spectrum contains small peaks distributed along a broad frequency range or several
peaks positioned at certain locations.

Time Waveform
The time waveform is a plot of the time signal as it is measured. It is rather a complex topic in
vibration analysis and used to confirm the results of FFT in some situations such as in bearings and
gears fault diagnosis.
Any defect in one (or more) of the bearing parts will excite short duration pulses when the rollers
pass over the defect. These impulses will, in turn, excite the bearing natural frequencies as a
consequence of impacting. The repetition rate (frequency) of these impulses depends on the
rotating speed, bearing geometry and the location of the defect. The duration of these impulses is
very short as compared with the interval between them (period); therefore, the energy is
distributed at a very low level over the period. As a result, the impulses can easily be obscured by
noise or other frequency components.

Figure 3.1 Typical time waveform due to defective outer race

Figure 3-1 shows a typical time waveform for a bearing with outer race defect. The repeated
pulses are corresponding to the impacts generated when balls passes over the defect (spall or
pitfall). The duration between each two successive pulses equals to the reciprocal of BPFO
frequency in Hz. However, in many cases bearing vibration pattern is obscured by other patterns
and noise and the resulting signal will be random and noisy such that it is difficult to detect
bearing patterns unless using other techniques such as enveloping or Wavelet transform. More
about time waveform will be coming in the signal envelope detection section 9.

4. Frequency Domain Techniques

FFT Spectrum
FFT spectrum is useful to separate the different frequency signals or components. The spectrum is
the plot of the signal components amplitudes vs. frequency. It is obtained from the time waveform
by a process called Fourier Transform. Each component in the FFT spectrum has its own frequency
and amplitude. The order of the component is simply its frequency divided by the shaft rotational
frequency.
FFT spectrum is used to explore the bearing fault stages earlier in sec. 1. It is powerful tool to
analyze machinery vibration and diagnose most of problems including bearing problems. However,
one must remember the following facts about using FFT spectrum in bearing diagnosis:
1. In the earlier stages of bearing fault, the very high frequency range of 5 to 50 kHz may not
be noticeable unless using filtration due to obscuring of the signal by other frequencies.
Also, the acceleration units must be used.
2. As the damage develops, bearing defect frequencies begin to appear in the spectrum
according to the damage location.
3. Since the signal of vibration due to impacting is not sinusoidal, there will be harmonics of
the fault frequency in the spectrum in most cases.
4. If the inner race is rotating, the BPFI component will be amplitude modulated with 1X
frequency according to the position of the defect with regard to loading area. This will be
reflected as sidebands on both sides of BPFI in the spectrum.
5. If the outer race is rotating, the BPFO component will be amplitude modulated.
6. If one or more ball is defected, the BSF will be amplitude modulated with 1X cage
frequency since the balls pass through loading area at a rate of cage frequency.
Figure 4-1 shows the FFT of a bearing with defected rotating inner race. The harmonics of BPFI are
more visible when using acceleration units.

Figure 4.1 FFT of bearing vibration with defected inner race

FFT Waterfall
FFT waterfall is a cascaded view of the successive FFT spectra taken at different intervals. In other
words, it is frequency-time representation of a signal. The x-axis represents frequency, y-axis is the
amplitude and z-axis is time.
Trending waterfall is the plot that can be used to show the progress of vibration along extended
periods of time by comparing FFT spectra along the entire monitoring period, for example along
one year or more.
Another waterfall plot is the plot used to represent the signal in time-frequency such that the
signal is decomposed to a number of blocks with a constant time interval. The FFTs of these blocks
are plotted in cascade fashion. This representation is useful to track changes of the signal along a
short period such as during machine run-up or coast-down. We can track the frequency and
amplitude of each component and obtain useful information such as resonance conditions. Figure
4-2 shows the FFT waterfall of a faulty bearing where vibration is collected along 8 seconds during
run-up and the FFT of signal blocks is obtained each 0.1 sec interval successively and plotted. The
variable frequency and amplitude modulation are clearly seen on the plot.
 Figure 4.2 FFT waterfall of a faulty bearing during run-up

5. Power Cepstrum
Cepstrum is simply the inverse Fourier transform of the logarithmic power spectrum. The words
cepstrum, quefrency, rahmonics and gamnitude in cepstrum analysis correspond to spectrum,
frequency, harmonics and magnitude respectively in FFT analysis [5]. The quefrency is measured in
seconds and it is actually a “delay time” or periodic time rather than absolute time. According to
the above definition, cepstrum can be expressed by:

C A A (τ ) = F −1
{log X 2 (f )} (5.1)

Where CAA(τ) are the cepstral components while X(f) are the spectral components. The
importance of cepstrum analysis comes from the fact that it can detect repeated patterns (family
of harmonics) present in the vibration signal. This feature is very important in the analysis of roller
bearing vibration. Harmonics in the frequency domain are represented by a single quefrency
component in the cepstrum domain. Hence, early detection of harmonics is possible even when
the fundamental frequency component is not detectable. Furthermore, the harmonic spacing can
easily be indicated from the reciprocal of the quefrency value. Also, the value of the main cepstral
peak is good trend parameter because it represents the average over a large number of individual
harmonics.
Some researchers showed that detection of incipient faults using Cepstrum is possible even when
the fundamental defect frequency is absent in the spectrum. The fundamental frequency may be
absent due to two reasons; an average and shift effect which causes a slow migration of the
fundamental impact frequency from its computed value; and an inter-modulation effect which
translates defect related information to frequency locations unrelated to the fundamental impact
frequency [6].

Figure 4.3 (a) FFT Spectrum, (b) Cepstrum plot

Figure 4.3 shows the FFT and cepstrum analysis for a ball bearing with defected balls in the earlier
stage. Despite that BSF, corresponding to 2.6X = 129 Hz, and its harmonics are not detectable on
the spectrum, the repeated patterns of BSF are successfully detected on the cepstrum plot by the
peak quefrency component of 7.8 msec = 1/129.

6. Shock Pulse Method

The original “Shock Pulse Method”, patented by Eivind Sohoel (SPM Instruments) in 1969. In 1970
SPM Instruments began to exploit products that are based on this method. The SPM approach
measures the mechanical shock speed, measuring the compression wave produced when rolling-
element and race interact and, with damage or failing lubrication, eventually, collide. When
contact occurs, a surface reaction manifests as a compression wave and travels at the indicated
speed of sound within the given material.

Figure 6.1 Elastic wave front as a result of metal to metal collision [7]

The SPM transducer is an accelerometer having 32kHz resonance frequency used to capture the
stress waves. A fault-free bearing, that is well designed and properly selected for this application,
is used to generate the “Carpet Value” or baseline relative to RPM and shaft diameter. Empiric
testing established a table of such normal noise level expectations, or a quantifiable “carpet value”
measured in dB. A brand new bearing entering service already has an expected Shock Pulse value
measured in dB. Also, it would generate a regular and unvarying noise to our ear.
Even at that point in time, minor noises from impulses will rise above the average carpet noise.
The maximum observed excursion reflects the number of shocks from all sources. A phenomenon
such as the slight “catching” of an out-of-round imperfection would generate a rise of the
maximum (peak value) rather than in the overall noise. Both the carpet level and maximum value
develop over time. This yields not two, but rather, three values: the carpet, the peak, and the
difference between them.
The amplitude of a shock pulse is proportional to the speed of the colliding. In the case of a rolling
element bearing, the speed of the rolling elements is proportional to the diameter and the
rotational speed of the shaft. The dependence of the shock pulse amplitude on the relative speed
of the objects at the moment of impact creates problems when establishing alarm acceptance
levels. To solve this problem, a normalization factor (dBi or HDi) is introduced that is effectively
normalizing the shock pulse amplitude regardless of the rolling element speed. The result is that
the shock pulse reading is presented on a normalized scale. By defining the diameter of the
bearing and by measuring RPM (or by manually entering RPM) the normalization factor called dBi
for different diameters and RPMs (or HDi in the SPM HD method) can be calculated.

Figure 6.2 Description of SPM measuring parameters [7]

Based on the above, the SPM measuring parameters are (see Figure 6.2):
dBsv or HDsv: this is the unit of non-normalized (raw) shock pulses.
dBi or HDi: normalization factor, depends on RPM and bearing dimensions and it can be
calculated from empirical relations.
dBn or HDn: this is the unit after normalization.
dBm or HDm: the is the highest shock pulse found during measurement time in the normalized
scale (referenced to dBi).
dBc or HDc: this is the threshold level where there exits 200 shock per second expressed in the
normalized scale (referenced to dBi).
In Figure 6.2, the peak shock pulse before normalization is 40 dB, while dBi is calculated to be 10
dB. Hence the normalized dBm = 40 – 10 = 30 dB indicating some problem according to the
severity level. Likewise, the non-normalized carpet level is 22 dB, hence the normalized dBc = 22 –
10 = 12 dB.
Recent SPM meters involve not only level detection but also shock pulse spectrum for bearing
defect frequency confirmation.

7. Spike Energy
Spike energy (SE) measurement was proposed by IRD Mechanalysis (currently Entek IRD). This
method employs the overall vibration in the frequency range 5 – 50kHz to assess the condition of
rolling elements bearings. The IRD accelerometer type 970, which has a natural frequency of
27kHz, is utilized to collect vibration signal.
Currently, the SE measurement is based on a bandpass filtered signal with lower cut-off (highpass
filter) set at 100Hz, 200Hz, 500Hz, 1kHz, 2kHz or 5kHz. While the upper frequency (lowpass filter)
is permanently set at 65kHz [7]. The highpass filter selection depends on the machine speed and it
tends to remove the low-frequency components resulting from common problem such as
unbalance and misalignment which may obscure the pulses resulting from a bearing defect.
Usually acceleration peak-peak value in g’s (1 g = 10 m/s2) is employed as monitoring parameter
and, hence, denoted as gSE. The following table shows the recommended filter setting for gSE
measurement as advised by Entek IRD:
HP Filter setting Speed range (RPM)
500 Hz 0 to 100
1000 Hz 100 to 1000
2000 Hz 1000 to 1500
5000 Hz 1500 and above

Figure 7.1 shows the flow chart of Spike Energy signal processing. Spike Energy is a demodulation
process that tends to detect the pulses in the filtered signal by a peak-peak detector followed by a
decay circuit to hold the peak and slowly decay it. The decay time constant is manually selected by
the user or automatically selected by the processing software according to measurement Fmax.
The peak-to-peak detector in SE circuit is very sensitive to the defect frequency as compared to
other envelope detection or demodulation methods. Unlike other demodulation techniques
which may cause distortion to the detected peaks due to the use of low-pass filter in envelope
detection, Spike Energy detection circuit preserves the severity of defects by holding the peak-to-
peak amplitude of the impulses and enhances the fundamental defect frequency and its
harmonics by applying a proper decay time constant.
Many users have observed the fluctuation of Spike Energy amplitude, when observed over time
(whether using analog or modern digital meters). Unlike the quickly reacting classical parameters,
gSE levels show a laborious rise time and slow decay. This feature is inherent to gSE processing.
What it means: cumulative or multiple event impulse energy is required to push the gSE level up. A
period of quiescence is needed to settle back down. The delay period is required to prevent the
amplitude modulation in the signal from imposing fluctuations in the measured gSE value.
Recently, the decay time becomes proportional to the frequency manifest within the signal.
Despite the fact that gSE is a good indicator for bearing fault, it can easily be compromised. Also,
measurement must be trended to identify the acceptable limits.

Figure 7.1 Spike Energy processing [7]

8. High Frequency Detection
This technique was introduced by SKF as a competitive to Spike Energy and SPM. It is based on
measuring the peak of the signal in the frequency range 5 to 60kHz. Unlike gSE, there is no further
processing to the signal other than bandpass filtering. This technique provides reasonable
performance as a condition detection system.

9. Envelope Detection and Demodulation

Envelope may be defined as the outer shape of the signal. When a high-frequency signal is
amplitude modulated with a low-frequency signal, the envelope would be waveform of the latter
signal. Envelope detection is the process of demodulation in order to extract the low-frequency
content. For bearing vibration, envelope detection is useful in identifying the intensity of the
pulses and also in finding the repetition rate of these pulses. Repetition rate is related to the
bearing characteristics frequencies (BPFI, BPFO, BSP and FTF) and can be found by spectrum
analysis of the demodulated signal.

Lowpass Filter Enveloping

Enveloping has special importance in identifying bearing faults in the earlier stages. The process of
extracting the envelope is shown in Fig. 9.1. The signal is bandpass filtered to remove the low-
frequency content (related to common machinery problems) and very high frequency noise. Then,
the signal is rectified and low-pass filtered or integrated to obtain the envelope [9]. It is important
to select the appropriate filter setting for envelope spectrum processing such that no details are
discarded.
As compared to Spike Energy method which detects peak-peak value, envelope processing tends
to average the rectified peaks during low-pass filtration resulting in smoothed out peaks with
altered peak values. Hence, the envelope processing is more suitable to obtain the demodulated
time waveform or demodulated spectrum. Figure 9.2 shows a typical envelope spectrum for a
faulty ball bearing. The bearing characteristic frequency will be shown as a peak in the spectrum
(50 Hz in this example). There will be multiples of this frequency due to the impulses in the
envelope time waveform.
 Figure 9.1 Envelope processing by low-pass filtering

Figure 9.2 Envelope spectrum for a faulty roller bearing

The maximum frequency of the envelope spectrum (Fmax) is normally set at 1:5 of the highpass
filter setting or sometimes at different ratio (1:2 for example). This is because that the bearing
characteristic frequencies are far less than resonance frequencies which are smoothed out during
envelope detection. It is important to select the appropriate filter setting for envelope spectrum
processing such that no details are discarded. The following settings, which are provided by SKF,
are used in some vibration analyzers and were adopted in HiDAC-8 portable machinery fault
diagnostic platform as predefined settings.
Table 9.1 Enveloping Setting suggested by SKF

Hilbert Transform Enveloping

As an alternative to the rectification and low-pass filtration used in demodulation process, the
Hilbert transform can be used to obtain envelope which will preserve the actual pulses
amplitudes. The steps to obtain the demodulated signals are as follows:
1. The signal is bandpass filtered as usual
2. The Hilbert transform of the filtered signal is obtained as follows;
• Take the FFT of the signal
• Cancel out the negative spectrum (the frequency components beyond 0.5 x Number of
points) and double the positive spectrum.
• Take the inverse FFT to obtain the Hilbert transform.

3. Calculate the envelope A( t ) = xR2 (t ) + xI2 (t ) , where xR (t ) is the real part of the Hilbert
transform and xI (t ) is the imaginary part of it. (Not that xR (t ) is exactly the same as the input
real time signal).
PeakVue Technique Enveloping
Another envelope detection technique developed by Emerson is the so called PeakVue technique
[15. The technique is based on digitally sampled vibration signals at high sampling rate to capture
the stress waves. The steps include bandpass filtering as in previous methods, rectification and
then application of PeakVue envelope detection on the sampled data. Instead of using low-pass
filtration of signal which could severely alter the amplitudes of peaks, the rectified signal is divided
into time slots of ∆t equal the reciprocal of 2.56 times the maximum demodulated frequency.

1
∆t = (9.1)
2.56 Fde mod
The peak at each time slot represent the sample at the given time as shown in Fig. 9.3.

Figure 9.3 PeakVue Enveloping

The above technique is simple, yet powerful and proved to be able to preserve the actual peaks
even when using low values for the demodulated frequency as shiwn in Fig. 9.4 [15]. Moreover,
autocorrelation may be used to enhance the SNR and clearly show the repetitive patterns in time
waveform. The autocorrelation coefficients Cj are ranging from -1 to +1 and are calculated from
equation (9.2) for sampled data signal.
Rj
Cj =
R0
N /2 −1
(9.2)
Where R j = ∑ xx
i=0
i i+ j
In the autocorrelation process, the random noise will be averaged out due to the fact that
negative and positive values are subtracted while repetitive patterns are summed up. To reduce
the calculations of autocorrelation function, FFT-based autocorrelation technique can be used.

Figure 9.4 PeakVue vs. Lowpass Filtering Techniques [15]

Case Study: Multistage RO Water Pump

This is high pressure multistage pump belongs to Reverse Osmosis unit in SCPI company in Basra.
Motor power is 22 kW, running at 1470 RPM and directly coupled to a multistage pump. The pump
is installed in early of 2016. After few months of operation, the spike energy readings of motor
drive end bearing begin to increase. In March 2017, the spike energy reading increased to 0.53 gSE
as shown in Figure 9.5. Few days later, it reaches 0.94 gSE with noticeable high frequency sound.
The envelope time waveform and FFT are collected to provide more information about the case.
The four characteristic frequencies of the installed bearing (type FAG-6310) are as follows; Ball-
Pass Frequency Inner race (BPFI) = 4.95x, Ball-Pass Frequency Outer race (BPFO) = 3.048x,
Fundamental Train Frequency (FTF) = 0.381x, and Ball Spin Frequency (BSF) = 1.98 x.
Figure 9.6 shows the envelope time waveform and FFT spectrum of bearing vibration using
PeakVue technique. The time waveform of the demodulated signal clearly shows high amplitude
peaks. The FFT spectrum of demodulated signal shows large peak at exactly BPFI 4.95X and its
multiples with multiples of 1X sidebands.

Figure 9.5 Trend of motor bearing spike energy vibration for RO pump
 Figure 9.6 Envelope time waveform and FFT spectrum for RO pump bearing vibration
When the bearing is replaced and a complete maintenance procedure is performed on the pump,
vibration level decreased to 0.23 gSE.

Case Study: Catacarb Pump P202-B

This is a heavy duty pump used to pump hot catacarb solution in the Southern State Company of
Fertilizers (SCF) in Khor Elzubayr in Basrah. The pump consists of 570kW induction motor running
at 2930 RPM and a three-stage centrifugal pump directly coupled to the motor. The pump is
shown in Figure 9.7. The pump main rotor is periodically replaced each several months due to
severe vibration problems that cause its failure. There was high vibration in the pump in both
drive-end and free-end with vibration velocity RMS of 14.2 mm/s in the horizontal direction of the
pump non-drive-end.
Vibration signatures from the horizontal direction on the non-drive-end are collected using
velocity, acceleration and PeakVue envelope methods. Inspecting the acceleration spectrum, there
are harmonics of 1X component up to 7X with high peak at 5X as shown in Figure 9.8. The 5X
component coincides with Blade Passing Frequency (BPF = No. of blades x shaft speed) of the
impeller. This frequency is a characteristic of a pump and it is the result of pumping pulsation
when the blades inject fluid through the outlet. However, high amplitude BPF indicates some
pump related problems. The suspected problem is that the impeller is axially eccentric inside the
pump casing.

Figure 9.7 P202-B Catacarb pump

Figure 9.8 Acceleration spectrum of non-drive-end of P202-B

Another problem is detected in the non-drive-end of the pump. The PeakVue peak-peak value is
9.38g which is above alarm limit and indicate bearing problem. Double angular contact bearings
type FAG 7311B are used in the non-drive end. The bearing has BPFI = 7.086X, BPFO = 4.915X, BSP
= 2.046X and FTF = 0.409. When envelope demodulated spectrum is used to study the problem,
the bearing tone of BPFO 4.915X and its multiples are clearly shown in the demodulated spectrum
as shown in Figure 9.9. This tone was completely obscured by the nearby 5X component of the
blade pass frequency in the normal acceleration spectrum.

Figure 9.9 PeakVue demodulated spectrum of non-drive-end in P202-B

The lack of lubrication due to insufficient oil level in the oil sump was the problem. Bearings are
lubricated by oil ring as shown in Figure 9.10 where the ring sprays the oil on the bearings. The
maintenance staff indicated that the oiler bulb was replaced with incorrect type which was the
reason of oil level being insufficient in the sump.

Figure 9.10 Lubrication system of non-drive end bearings in P202-B

10. Signal Enhancement Techniques
When bearing vibration signal is contaminated with noise or other vibration sources, fault
detection process needs more advanced techniques. When the fault develop, it becomes easier to
detect the fault by common techniques such as FFT. When the SNR (signal to noise ration) is poor,
normal time waveform and FFT become ineffective in detecting faults. To improve the
performance of signal analysis techniques, the SNR has to be enhanced. Some signal enhancement
techniques will be explored here.

Signal Decomposition
The time domain signal acquired from a machine consists of periodic signal and additive noise.
Some techniques are used to cancel out noise and extract the useful data. Among these
techniques are:
Leakage-free sampling
In this case, sampling period is locked to the rotational speed such that the analysis length is
integer multiples of shaft rotations. Encoders are used to achieve this. However, encoders cannot
be installed to every machine and need special hardware provision that make this method
impractical for everyday vibration data collection.

Time domain averaging

Time domain averaging is the technique based on doing averaging process in time domain such
that periodic signals are summed up while noise is cancelled out. Suppose that vibration signal x(t)
is composed of periodic signals f(t) and additive nose s(t). When summing several periods of the
signal, the repetitive signals add coherently while the noise incoherently [10].

xˆ (ti ) = Nf (ti ) + N s (ti ) (10.1)

It is clear that SNR will be enhanced by a factor of N . The averaged samples are synchronized
with trigger signal obtained from a tacho sensor such as Hall effect or photo sensor. Hence, this
technique is called Time Synchronous Averaging (TSA). When using digital sampling and to cope
with speed variation during rotation, the data may be re-sampled in order to obtain fixed number
of points for each revolution. The sample may span one trigger period or more. Figure 10.1 shows
TSA where averaging period is one revolution (one trigger period).
 Figure 10.1 Illustration of time synchronous averaging

Adaptive Noise Cancellation

The general concept of Adaptive Noise Cancellation (ANC) is depicted in Figure 10.2. The
monitored vibration signal v0 is contaminated with noise n0 picked by the primary vibration
sensor. A reference sensor is used to pick a reference noise n1 correlated to n0 in some way. The
reference noise n1 is filtered by adaptive filter in order to match n0 as close as possible. Adaptive
filter coefficients are estimated by Least Mean Square estimator which minimizes the output
power of the noise canceller. The output of the adaptive filter is subtracted from the combination
v + n0 in order to obtain best estimate of the required signal. ANC has been successfully applied to
monitor bearing vibration [10].
 Figure 10.2 Adaptive Noise Cancellation

Adaptive Linear Enhancer

The problem with previous scheme (ANC) is that one requires a reference signal which is strongly
correlated with the noise signal. The performance of an ANC scheme is usually limited by the
availability and quality of the reference signal.
An alternative mode of operation for an adaptive filter, based on ANC, is provided by an adaptive
line enhancer (ALE), as shown in Figure 10.3. This is a simple variation on the ANC requiring only a
single input signal. In this case the reference signal is obtained by delaying the input signal by a
fixed number of samples, ∆. The adaptive filter then endeavors to predict the signal xk from the
delayed samples. The result is that any input components which are predictable over the delay
appear at the filter output, yk , whilst the error signal contains those components which are
unpredictable over the delay [11].
 Figure 10.3 Adaptive line enhancer [11]

11. SPM High Definition and Symptom Enhancing

SPM High Definition (HD) is based on the same principles of traditional SPM method. Elastic wave
propagation in bulk material is detected by a tuned shock pulse transducer. Same type of
transducer is used in SPM HD method. However, the underlying interfacing circuitry and
processing algorithm is completely different from traditional SPM method.
The block diagram (black box) of SPM HD algorithm is shown in Figure 11.1. As SPM Instruments
developed commercial products that use this technology, minimal information is provided about
the underlying processing techniques.

Figure 11.1 SPM HD block diagram [7]

The inputs to the SPM HD device are the shock pulse vibration, RPM and bearing data. The RPM is
needed to calculate the normalization factor HDi as well as to synchronize data sampling as this
method uses sophisticated digital techniques. Bearing data is used to calculate normalization
factor and also to be used to extract SPM HD Time and HD Spectrum [7].
If the RPM is fixed, manual entry for its value is sufficient. But when the speed is varying or
fluctuating, it is essential to use accurate speed sensor to read the RPM. Single pulse per
revolution or multiple pulses can be used. When multiple pulses are used, such as when using
coupling holes or bolts as triggering means, the angular spacing must be accurate in order to avoid
smearing/distortion in the processed data.

HDm and HDc

In section 6, SPM is presented and HDm and HDc are described (corresponding to dBm and dBc).
These scalar values are described in dB scale and specify the operating conditions of the bearing.
SPM HD uses high efficiency A/D converter with digital filters to improve the SNR. Sufficient
measuring time must be used to ensure proper results. SPM Instruments recommend using
measuring time that spans 50 revolutions (10 rev. minimum) in order to obtain stable HDm
measurement.
Measuring time = 50 x (60/RPM)

HDm and HDc are not affected by RPM measurement accuracy since RPM is not used in the
calculation of these scalar values directly except when calculating HDi. To measure HDc, the total
measuring time is divided into 5 ms time slots. In each time slot, the strongest shock pulse is
identified and temporarily stored. When the measuring time is up, the weakest among these
stored values is selected as the HDc value (there are 200 5-ms slots in one second). This gives an
approximation of the level where more than 200 shocks per second are found.

SPM HD Time Signal

The time signal is a good tool to pinpoint bearing defects and it could be more useful than FFT
spectrum in many cases [7]. The HD technology aims to produce clear and crisp time signal.
Advanced digital processing techniques are used in this technology. Linear scale is used for time
signal.
SPM HD is analogous to time synchronous averaging as the SNR is enhanced by a factor equal to
square root of Symptom Enhancing Factor SEF. However, the algorithm is different as TSA cannot
be used with bearing faults because they are normally not synchronous with RPM (fractional
multiples of RPM). Also, the FFT averaging is used to obtain averaged spectrum and cannot be
used to average time signals. SPM HD looks for and enhances repetitive pulses even when they are
buried into noise. The algorithm involves digital sampling using efficient A/D converter,
rectification and envelope detection by low-pass filtering. The symptom enhancement algorithm is
applied to emphasis the repetitive pulses in the time signal. The SEF has considerable effect on the
clarity of the results and SNR but the measuring time will be, consequently, larger. Figure 11.2
shown the effect of using Symptom Enhancement on the time signal [2]. SPM provided no
information about the maths of Symptom Enhancement technique, but one can suspect that
autocorrelation is used just like the PeakVue technique of Emerson.

Figure 11.2 Effect of Symptom Enhancement on the SPM HD time signal [2]

SPM HD Spectrum
If a Fast Fourier Transform (FFT) is applied to the SPM Time Signal HD, an SPM Spectrum HD is
created. This is useful to interpret and locate bearing faults. The SPM Spectrum HD is also
presented on a linear scale.
Figure 11.3 shows the HD Spectrum for a bearing with cracked inner race. The severe amplitude
modulation of BPFI with 1XRPM is clearly shown on the HD Spectrum since there are multiple
sidebands (spaced at 1XRPM) around BPFI and its harmonics. It is also important to mention that
since HD Spectrum is obtained from the SPM Time HD, the rectification of the signal during
envelope detection resulted in a non-zero DC level. This DC level is reflected into a spectral
component at 0 Hz (or 0 Order) with its sidebands [7].

Figure 11.3 SPM HD Spectrum for a bearing with cracked inner race [7]

Hallsta Paper Mill [2]

This case describes using SPM HD method to study low-speed bearings problems in Hallsta paper
mill in Sweden. Application of this technique in condition monitoring and analysis of vibration
signals had led to important cost saving due to early detection of faulty bearings and changing the
philosophy of maintenance program. The paper mill is shown in Figure 11.4.
 Figure 11.4 Hallsta Paper Mill during maintenance [2]

For the lower side S-roller bearing, type SKF 22320 spherical roller bearing, vibration monitoring
showed high HDm readings (> 60 dBm) as shown in Figure 11.5.

Figure 11.5 S-roller bearing HDm values from Oct. 2009 onward [2]

The roller rotates at 15 RPM. For that low speed bearing, the SPM HD method is applied to obtain
time waveform of the demodulated signal with Signal Enhancement Factor of 10. The measuring
time for the setup extends from 5 to 8 hours due to large SEF and low frequency. The HD
demodulated time waveform before bearing replacement in December 2009 is shown in Figure
11.6. The time waveform clearly shows inner race pattern modulated with 1xRPM.
 Figure 11.6 Time waveform of S-roller bearing before replacement [2]

When the bearing is replaced, the HDm readings decreased to noticeable level (20 dBm) as shown
in Figure 11.5 above. The defective bearing is shown in Figure 11.7.

Figure 11.7 SKF 22320 bearing photo show cracked inner race [2]

12. Wavelet Transform

Although Fourier transform (FT) is very powerful tool in signal analysis, there is a major
disadvantage that limits its use in analyzing transient or non-stationary signals. Fourier expansion
has only frequency resolution, i.e. there is no time information in this type of analysis. Despite the
fact that FT can identify the frequencies contained in a signal, the time at which these frequencies
occur is unknown. In fact, FT detects the averages of the individual components over the
measurement time. On the contrary, Wavelet transform (WT) is capable to examine the signal in
both time and frequency simultaneously which permits its use to analyze signals that are
described to be non-stationary, aperiodic, noisy and transient [12].
Wavelet transform is a method of converting a function (signal) from one form into another for
the purpose of making the features of the signal more amenable to study. It has the ability of
representing the signals in both time and frequency. Given a function x(t) which is square
integrable, the Continuous Wavelet Transform (CWT) coefficient W(a,b) is the inner product of x(t)
and a normalized, dilated and translated wavelet function ψa,b(t) [12];

∞
W (a,b ) = x (t ),ψ a ,b (t ) =
∫
−∞
x (t )ψ a*,b (t ) dt

1  t −b 
Where the shifted and dilated wavelet function is given by: ψ a ,b (t ) = ψ 
a  a 
Where ψ(t) is called mother wavelet function. The selection of the wavelet function depends on
the type of the signal to be analyzed as well as the physical phenomenon or process being
interrogated. Examples of mother wavelets are the Haar, Morlet and Mexican hat.
In the CWT, the wavelet is dilated and shifted either continuously or at very fine resolution. This
will result in highly redundant coefficients. This redundancy can be removed by proper
discretization of the scaling and shifting factors given that the scaled wavelet functions satisfy the
condition of orthogonality. Dyadic grid scaling is normally used in Discrete Wavelet Transform
(DWT) with dyadic wavelet functions (father wavelet).
Application of wavelet transform in bearing signature analysis offers many advantages over the
traditional analysis techniques. The non-stationary and impulsive signals can be better analyzed by
using the time-frequency distribution of the wavelet transform. Moreover, the filtering and
denoising capabilities of the DWT can be found effective in improving the signal to noise ratio.
Figure 12.1 shows the 3D wavelet transform representation of vibration signal (scalogram) where
both time and frequency (scale) representations of the signal can be viewed.
 Modulus of Ca,b for a = 1 2 3 4 5 ...

1.5
W(s,m)

0.5

40
30
20 1000
800
10 600
400
200
0
scales a

Figure 12.1 Wavelet transform representation (scalogram)

Case Study: P304 Screw Oil Pump

This is a screw-type oil pump, shown in Figure 12.2. It belongs to the Southern State Company of
Fertilizers in Khor Elzubayr. It is used to pump the oil at high pressure for the sealing system in a
CO2 compressor. It has a 160kW electrical motor running at 2970 RPM and directly coupled to the
pump. This pump was installed in 2010 as a replacement for an old pump. After two years of
operation, the staff of SCF reported increased vibration acceleration level on the drive-end
support of the motor. The four characteristic frequencies of the bearing (type FAG-6219) are as
follows; Ball-Pass Frequency Inner race (BPFI) = 5.899x, Ball-Pass Frequency Outer race (BPFO) =
4.1x, Fundamental Train Frequency (FTF) = 0.41x, and Ball Spin Frequency (BSF) = 2.692x.
Inspecting the FFT spectrum in Figure 12.3, it is clear that multiples of 2XBSF are presented with
large amplitude at exactly 6XBSF. The presence of BSF is indication of defected balls as mentioned
above. However, 2XBSF is usually reported due to the fact that the defected ball rolls over the
inner and outer race each one revolution. Also, it has been observed that BSF is presented due to a
number of conditions such as ball inaccuracies, lack of lubrication and high preloading. These
harmonics of BSF are commonly accompanied by sidebands at FTF.
 Figure 12.2 P-304 screw oil pump

Figure 12.3 P304 Motor drive end bearing vibration spectrum

For further investigation, vibration time signal is viewed. Both direct and filtered signals are
displayed. Filtration is done by using CWT to obtain the filtered signal at frequency of 803 Hz,
corresponding to the largest harmonic of BSF. The original time signal and the cross-section plot of
CWT are shown in Figure 12.4. The original time signal is very noisy such that the amplitude
modulation can hardly be identified. On the contrary, the amplitude modulation of 803 Hz
component with the FTF is clearly visible on the cross-section plot. Since there are five peaks
within the time of 0.25 second, the modulation frequency is simply the reciprocal of the peaks
period or numerically 5/0.25 = 20 Hz.

Figure 12.4 Time waveform of bearing vibration for P304 motor; (a) direct, (b) filtered at 803Hz

Figure 12.5 shows the bearing that was taken out from P-304 motor. The metal oxidization and
rust clearly shown on the cage due to escape of water and other corrosive materials into the
motor. The pump undergoes breakdown periods from time to time which could be the reason for
contamination.
 Figure 12.5 P-304 motor bearing

13. Automated Diagnosis and Advanced Techniques

Artificial Neural Networks

All the techniques presented in the above sections for fault diagnosis need human interpretation.
Some techniques are easier to interpret while other techniques need deep knowledge and
experience. Regardless of the processing technique used, automatic interpretation can improve
the efficiency and reliability of fault diagnosis. Artificial neural networks (ANN) can identify and
classify real data such as vibration signals. ANN consists of a number of richly interconnected
artificial processing neurons called nodes, collected together in layers forming a network.
An Artificial Neural Network (ANN) is an information processing paradigm that is inspired by the
way biological nervous systems. ANNs, like people, learn by example. An ANN is configured for a
specific application, such as pattern recognition or data classification, through a learning process.
Learning in biological systems involves adjustments to the synaptic connections that exist between
the neurons. This is true of ANNs as well. Figure 13.1 shows a multi-layered ANN.

Figure 13.1 Multi-layered ANN

In artificial neurons, inputs are multiplied by weights and then calculated by an activation function.
Another function estimates the output of the artificial neurons. The number of nodes within the
input and output layers depends on the number of independent variables required to define the
problem. The number of hidden layers and their nodes are selected by trial and error methods.
The general procedure for building an ANN is:
1. Define the problem input variables. These are the features or patterns that are measured,
observed or estimated. For vibration diagnosis problems, these inputs can be the features
of vibration signal such as RMS value, peak, crest, amplitudes of certain orders or any other
processed value of the techniques discussed above or other techniques.
2. Define the problem output. These are the faults or conditions need to be detected.
3. Set the number of hidden layers and nodes.
4. Train the ANN by using training data of known inputs-outputs to find the appropriate
connection weights. Back propagation is the most applied technique to find the weights.
 5. Check the output of the ANN with training data and with testing data of known output to
confirm the accuracy of the network. If the performance is not satisfactory, then the
number of hidden layers and nodes can be adjusted and the whole process is repeated.

Drawbacks of ANN:
1. Need of precise training data that is not always available.
2. Human intervention still needed in the selection of training data and network topology.
3. Valid for already known and programmed faults. New faults or conditions cannot be
interpreted without using additional techniques such as expert systems.
4. Computational and programming complexity especially for large networks with multiple
inputs and outputs.

Principal Component Analysis

Principal Component Analysis (PCA) is a tool used to remove redundancy in possibly correlated
data set. PCA is one of the dimensionality reduction tools and it is the backbone of many methods
in features extraction and data retrieval. One of the most important perceptions which are
provided by PCA is to recognize which variables in the system are more important, which are
redundant, and which are noises. This is basically accomplished by projecting original correlated
variables into a new space of uncorrelated variables, which are called, Principal Components (PCs).
The few first Principal Components are selected because they contain most important data while
the other PCs are neglected because they contain an un-useful and a noisy data. To understand
the math of PCA, suppose we have a matrix of data X with M vector, where each vector represents
a measured (observed) variable with N time samples. The matrix is MxN sized.

 x11 x12 ... x1N 

x 
X =  21 
(13.1)
 M 
 
 xM 1 xMN 
The first step in PCA is to normalize the data. The zero mean and unity variance are examples of
scaling techniques used in this regard. The next step is multiplying the normalized data by its
transpose to produce a covariance matrix C:

C = X XT (13.2)
Matrix C, which is MxM sized, captures the correlations between all possible pairs of
measurements for which the diagonal and off -diagonal terms are called the variance and
covariance, respectively. More importantly, the correlation values reflect the noise and
redundancy in the measured data. Large values in the diagonal terms correspond to interesting
proper data while small diagonal terms indicate low SNR. On the other hand, large off-diagonal
values correspond to high redundancy (highly correlated data) while small off-diagonal terms
indicate low correlation.
The matrix C is de-factorized by using Singular Value Decomposition (SVD) to find the Eigen values
of the problem. Some of the highest Eigen values are considered and used to construct a loading
matrix P with the Q top Eigen values. So P is MxQ matrix. The aim of P matrix is to minimize the
dimensionality of the original data by projection of the data on the P matrix to construct scores
matrix T:

T = XTP (13.3)

The scores matrix is NxQ sized. The noise-free data can be constructed from the scores matrix as
follows:

Xˆ = (T PT )T (13.4)

The squared prediction error (SPE) between the original and estimated data, given in the below
equation, is good measure for fault development.

N
SPE (i ) = ∑ ( xij − xˆij ) 2 (13.5)
j =1

When there is no fault, the SPE is normally small (below certain threshold). When fault is
developed, the squared error is increased due to presence of strong features or principal
components indicating some problem [13]. Also, the principal components can be used as input to
an ANN driven automated diagnosis system. Figure 13.2 shows the Eigen values and SPE for
normal and faulty bearing with different kinds of faults. High Eigen value, and consequently SPE, in
the faulty bearings are easily detectable. All faulty bearings are characterized with SPE beyond the
threshold limit set for a normal bearing [14].
 Figure 13.2 Eigen values and squared prediction error for normal and faulty bearings [14]

Support Vector Machine

SVM is computational learning method that based on the statistical foundations of learning
theory. It is one of the classification methods to classify linearly separable (initially) data. Suppose
some given data points each belong to one of two classes as shown in Figure 13.3, and the goal is
to decide which class a new data point will be in. The classification in this case is based on two
properties (features), so that each input data is 2-dimensional vector. The output is either Class1
(filled dots) or Class2 (empty dots). There are many hyperplanes that can classify the data, but in
case of SVM, the optimum hyperplane is chosen such that the distance from it to the nearest data
point on each side is maximized. More generally, a support vector machine constructs a
hyperplane or set of hyperplanes in a high- or infinite-dimensional space, which can be used for
classification, regression, or other tasks like outliers detection.
 Class-1 v
Positive f (x) = 0
Property2 v
f ( x ) ≥ +1

Class-2
Negative
v
f ( x ) ≤ −1

Property1

Figure 13.3 SVM classification of two classes

v
Suppose a set of input vectors xi that can be classified into two classes; Class1 and Class2. Let yi
be the class selector such that y = +1 for Class1 and y = −1 for Class2. The SVM can be trained to
find the function of hyperplane. The hyperplane is the plane whose data satisfy the following
condition:
v v v
f ( xi ) = ω ⋅ xi + b = 0 ∀ xi on the boundaries of hyperplane (13.6)

Where The function f() must satisfies the following (classification) conditions:

v v v  ≥ +1 for yi = 1
f ( xi ) = ω ⋅ xi + b =  (13.7)
≤ −1 for yi = −1
v
Where ω is the support vector whose dimension is the same as input vectors and b is a scalar.
The values of support vector and the scalar are estimated by exploiting optimization techniques so
that the classification error is minimum. The objective of optimization process is to obtain the
maximum separation or the widest hyperplane.
SVMs can also be used in non-linear classification tasks by exploiting kernel tricks. The data to be
classified is mapped onto a higher-dimensional feature space, where the linear classification is
possible. The n-dimensional input vectors are mapped into k-dimensional feature space.
Multi-class classification is also possible with SVM. The above discussion deals with binary
classification where the class labels can take only two values: 1 and −1. In the real-world problem,
however, we find more than two classes for examples: in fault diagnosis of rotating machineries
there are several fault classes such as bearing faults. The earliest used implementation for SVM
multi-class classification is one-against-all methods. It constructs q SVM models where q is the
number of classes. The ith SVM is trained with all of examples in the ith class with positive labels,
and all the other examples with negative labels. Another major method is called one-against-one
method. This method constructs q(q− 1)/2 classifiers where each one is trained on data from two
classes.
Figure 13.4 shows the flowchart of fault diagnosis using continuous Wavelet transform, PCA and
SVM techniques [14]. The CWT is used as filtration technique to obtain filtered time data at
specific scales that are corresponding to the bearing faults frequencies BPFI, BPFO, BSF and FTF.
PCA is utilized to reduce dimensionality of the data and extract features. The output of PCA is used
as input to the SVM classifier. The classifier must be firstly trained using sufficient known data.
Then the algorithm can be used to classify the fault from an unknown case.
Figure 13.4 Flowchart of fault diagnosis based on SVM
Closing Remarks-Gap between academic research and professional practice
Gap between academic research and ongoing practice is seen more or less in many engineering
and technology fields. While academic researches and scientific papers focus on developing
advanced techniques or new methods, business organization, in general, rely on well-established,
mature and traditional approaches. As related to the author profession, vibration analysis, and
from both the academic and professional practice, as university professor and consultant in many
industrial organizations, some of the reasons for that gap are:
1. The published researches are irrelevant to the practitioners.
2. Research papers are published in journals that are not accessible to practitioners (pure
scientific or academic journals) and reviewed by academic reviewers.
3. Business organizations and system developers have their own developed techniques and
do not want to invest in other (new) techniques. SPM and PeakVue are examples.
4. Some techniques are based on special and sophisticated software that are already covered
by license or cannot be incorporated in the existing systems. Examples are Matlab and
ANSYS based techniques.
5. Every new technique need to be well tested and take its way to the practical application.
6. Lack of practice. Some researches deal with specific cases in the laboratory and do not
cover other practical situations where conditions differ and cases overlap.
While bridging the gap is out of scope of this work, a suggestion is to make the researches more
practice-focused by applying Action Research, Collaborative Research and Decision Support
Systems.
References
[1] Vibration Analysis Reference, Mobius Institute, http://www.mobiusinstitute.com
[2] Accurate Condition Monitoring of Ultralow and Low Speed Machines using SPMHD and HDENV
Technologies with Case Studies, Tim Sundstorm, SPM Instrument, IMVAC2017, Dubai, 2017.
[3] Rolling Element Bearing Analysis, Brian Graney and Ken Starry, Material Education, Vol. 70, No.
1, pp 78-85.
[4] Enveloping and Spectral Emitted Energy, SKF Application Note CM3014.
[5] Frequency Analysis, Randall, R. B., Bruel & Kjaer, Denmark, 1987

[6] Effect of Inter-Modulation and Quasi-Periodic Instability in the Diagnosis of Rolling Element
Incipient Defect, C. C. Osuagwu and D. W. Thomas, ASME J. Mech. Des 104(2), 296-302 (Apr 01,
1982).

[7] An Introduction to the SPM HD Method, Tim Sundstorm, SPM Instruments AB, Oct. 2010.

[8] Spike Energy Measurement and Case Histories, Ming Xu, ENTEK IRD International Corporation.

[9] Nathan Weller, Acceleration Enveloping-Higher Sensitivity Earlier Detection, GE Energy,

http://www.ge-mcs.com/download/orbit-archives/2001-2005/2nd_quarter_2004/
2q04accelenvel.pdf

[10] Rolling element bearing fault diagnostics using the blind deconvolution technique, Mahdi
Karimi, PhD thesis, Queensland University of Technology, Sept. 2006.

[11] The enhancement of impulsive noise and vibration signals for fault detection in rotating and
reciprocating machinery, S. K. Lee and P. R. White , J. Sound and Vibration, 1998, 217(3), pp.
485-505.

[12] Design and Construction of a Multi Channel Wireless-Based Vibration Monitoring and Analysis
System, Jaafar K. Alsalaet, PhD thesis, University of Basrah, 2012.

[13] Multi Fault Diagnosis of the Centrifugal Pump Using the Wavelet Transform and Principal
Component Analysis, K. Berli, K. McKee, R. Entwistle, I. Mazhar, and I. Howard, Proceedings of
the 9th IFToMM International Conference on Rotor Dynamics, pp. 555-566. Springer
International Publishing, 2015.

[14] Advanced Vibration Analysis in Bearing and Gearbox Faults Detection, Sajjad N. Abdulkareem,
MSc. Thesis, College of Engineering, University of Basrah, 2015.
 [15] Description of PeakVue and Illustration of its Wide Array of Applications in Fault Detection
and Problem Severity Assessment, James C. Robinson and James E. Berry, Emerson Process
Management Reliability Conference 2001.

View publication stats