FEATURE | BEARING DIAGNOSTICS

Methods for diagnostics of bearings in

non-stationary environments
Submitted 13.09.11
R Klein, E Rudyk and E Masad Accepted 29.02.12

Bearing failure is one of the major causes of breakdown in The failing bearing produces typical noise and vibration
rotating machinery. One of the key challenges in bearing patterns, known as bearing tones. In contrast, ‘healthy’ bearings
diagnostics and prognostics is to detect the defect as early as exhibit virtually no vibration. Degraded bearings exhibit a unique
possible, when the failure signs are weak. vibration pattern (typical to each failure mode) allowing for the
Vibration-based early detection of bearing failure requires an detection and determination of damage progression.
improvement of the signal-to-noise ratio by effective signal Diagnostics of rotating machinery during its regular operation
de-noising and extraction of the weak failure signs that involves, in many cases, analysis of non-stationary signals as the
can be obscured by other vibration sources and noise. The rotating speed, loads and environmental conditions vary (in
challenge is to enhance the weak signature in the early stages some cases rapidly) with time. Therefore, the algorithms should
of defect development. allow the analysis of non-stationary or quasi-stationary signals.
The task of enhancing the weak failure signs is complicated Monitoring bearing degradation by vibration analysis is an
by the fact that changes in operating conditions influence established technique and the various methods of analysis have
vibration sources and change the mixture recorded by been widely published(1,2,3,4,5). The prevalent methods for bearing
the sensors. As a result, the recorded signal becomes non- analysis are orders representation and enveloping. Most of the
stationary. The proposed technique suggests a solution to this condition indicators are based on the peak values or energies in
problem. The technique adapts the basic dephase method for the frequency or order spectra.
the analysis of non-stationary signals recorded from a system This study presents the capabilities of the different techniques
operating under changing operating conditions. for automatic diagnostics of bearings and, therefore, will
The adapted dephase method was applied to vibrations concentrate on the discrimination capabilities of the extracted
measured in rotating machinery, including systems with features (condition indicators) and potential false or missed
healthy bearings and damaged bearings. The same feature alarms.
extraction procedure was applied to signatures in the orders The methods compared are feature extraction of the order
domain before and after the application of the adapted spectrum, dephased spectrum and envelope spectrum. Only
dephase. The results show the effectiveness of the method for methods that can be implemented in an onboard autonomic
diagnosis, both in the orders representation and in the orders system were considered in this study.
of the envelope representation. The paper presents different analysis techniques, the types of
features extracted and, finally, a comparison of the capabilities of
the different methods.
1. Introduction
Monitoring vibrations can be used to detect machine faults, 2. Bearing monitoring
including unbalance, misalignment, oil film bearing instabilities,
roller bearing degradation, mechanical looseness, structural The pattern of a defective bearing is spread across a wide
resonance and cracked rotors. The technique works by comparing frequency band (several harmonics and amplitude modulations)
vibration levels to baseline values. and, therefore, can be easily masked by the excitations of the
Attention to the condition of bearings in rotating equipment other components (other rotating components and the structure
is of high importance, since many machine problems eventually are considered as generating ‘noise’). Signal de-noising and
cause bearing defects. Bearings can fail due to initial production separation of the bearing’s weak signature are crucial to bearing
quality, improper or contaminated lubrication, heavier loading diagnostics since the other rotating components and the structure
than anticipated (caused by other machinery problems such as commonly induce much higher vibration levels.
imbalance, misalignment, bent shaft, etc), improper installation, The signal processing methodology for the detection of
surface fatigue or other factors. damaged bearings includes signal synchronisation according to
rotating speed, orders representation of the original and dephased
signal and enveloping signals(6,1,2,7,3,4,8). These techniques are
usually applied to stationary signals for derivation of one-
Dr Renata Klein, Eduard Rudyk and Dr Eyal Masad are with dimensional vibration signatures. For non-stationary signals, the
R K Diagnostics, Misgav Industrial Park, PO Box 66, D N Misgav 20179, same algorithms are applied in running segments considered as
Israel. Tel: +972 4 9990750; Fax: +972 4 9990751; Email: renata.
quasi-stationary, thereby generating two-dimensional signatures,
[email protected]
ie spectrograms in the time-frequency or cycle-order domains.

2
The International Journal of Condition Monitoring | Volume 2 | Issue 1 | March 2012
 BEARING DIAGNOSTICS | FEATURE

The goal is to emphasise the bearing vibrations by localising The spectrum of phase average represents mean amplitudes
the different excitations or by separating the natural frequencies of the raw signal in the harmonics of the corresponding rotating
and the bearing impact rates from the excitations of other rotating speed:
components. The ‘background noise’ generated by other rotating
components is reduced significantly by the adapted dephase
{ 1 1
}
ŷ ( kR1 ) = E ν̂ [ mR ,(m+k )R ] ( kR1 ) m ≤ M − k ................(2)

process, which is proposed in the following sections. The where ŷ = |FFT(y)| and ν̂ =|FFT(ν)|.
separation of excitations in the signal processing stage is crucial The phase average’s capability to filter out the asynchronous
since only the patterns distinguishable in the signatures can be elements is proportional to the number of cycles averaged (M).
revealed in the feature extraction process. In order to remove all the asynchronous elements it is preferable
In automatic diagnostics, the process of feature extraction to average over as many cycles as possible, ie over a long period
plays a very important role. The proposed process includes of time. On the other hand, when dealing with non-stationary
comparison of the analysed signatures to the ‘baseline’ population signals the phase average varies with time (as a result of changing
(a model of normality), ie determination of the Mahalanobis operating conditions) and averaging over a long period will
distance between the analysed signature and the baseline. The distort the result.
distance is evaluated around orders corresponding to the pattern The phase average y is replicated in order to get a vector of the
expected to be generated by one of the bearing’s failure modes. same size as ν (the resampled signal):
Yn = yn mod N n ∈Ν .................................(3)
2.1 Signal processing
Now removing the phase average, thereby generating the
2.1.1 Orders representation dephased signal, removes all the effects synchronous with the
The orders representation is generated by estimating the spectral rotating speed R1:
content of a signal synchronised according to the rotating speed. ν = ν − Y .........................................(4)
The synchronisation operation (resampling) transforms the In some cases, there are several asynchronous shafts and the
signal from the time domain to the cycle domain. It removes procedure needs to be repeated for each shaft rotating speed
the frequency non-stationarity due to changes in the rotating separately. Removing the phase average of R1 from the signal
speed. Resampling transforms harmonic signals (with frequency resampled by R2 is a little trickier, since the R2-resampled signal is
proportional to the rotating speed) into constant harmonic not periodic with any period associated with R1.
functions and, consequently, into sharp peaks (ideally delta Instead of taking the raw data ν, we resample ν̂ by R2. This
functions) in the corresponding spectrum that is designated as signal has already had the R1 effects removed from it. After
orders representation. resampling, the signal elements that are synchronous with R2 can
be removed. This process can be repeated with all shafts. Finally,
2.1.2 The principle of the dephase algorithm
the signal is resampled back to the desired cycle domain (Rk).
The dephase algorithm is designed to increase the signal-to-noise
ratio (SNR) of the bearing-induced vibrations (in this case the 2.1.3 The adapted dephase algorithm
noise represents the vibrations from other sources). The goal is to For synchronous signal elements with varying amplitudes, the
automatically suppress the synchronous elements of the vibration phase average will represent the mean amplitude; therefore, the
signal commonly associated with shaft and gear vibrations, phase average removal will not eliminate these elements and
leaving the asynchronous vibrations that in most cases represent locally can even amplify it.
the structure and the bearings. The basic dephase algorithm removes the phase average of
The proposed method(7,9) is based on the removal of the the entire signal and it is appropriate for stationary signals. Its
phase averaged signal. The phase average (or synchronous time performance depends on the accuracy of the measurement and
average) is a known method applied for cyclo-stationary signals. analysis of the rotating speed (Ri)1.
The phase average removes the asynchronous components by The adapted dephase removes the phase average in running
averaging the resampled signal over a cycle of rotation. All the frames (which may or may not overlap). It is able to remove the
signal elements that are not in phase with the rotating speed are phase average from non-stationary signals where the amplitudes
eliminated, leaving the periodic elements represented in one cycle. of the synchronous elements are changing (due to varying
These are the elements at the harmonics of the corresponding rotating speeds and loads).
rotating speed. The phase average can be calculated in the cycle The recommended frame size depends on the amount of
domain (signal after resampling) or by averaging FFTs calculated variation in the operating conditions, such as load and rotating speed.
over one cycle. Typically, for slow transients with rotating speeds of 100-300 rps, the
In the cycle domain, the phase average y corresponding configured frames were of 30-100 cycles representing ~0.5 s.
to rotation frequency R1 of the cycle history ν is calculated as The spectrum of a dephased signal is expected to reveal peaks
follows: related to the asynchronous tones of a deteriorated bearing. It
1 M −1 may also present peaks related to aero-dynamical phenomena
yn :=
M
∑ν n+mN n = 1,…, N .........................(1) (for example a rotating stall).
m=0

where N = 1/R1 and M is the number of cycles in the signal. Note 1

A more detailed sensitivity analysis of the dephase algorithm is
that y is a vector in RN representing a single cycle. presented in(7).

3
The International Journal of Condition Monitoring | Volume 2 | Issue 1 | March 2012
FEATURE | BEARING DIAGNOSTICS

2.1.4 Envelope analysis 2.2.2 Searching for bearing patterns

Envelope analysis is used to reveal the repetition rate of the The expected rate of shocks generated by a degraded bearing is
bearing shocks(1,2,7,3,6,8). calculated based on the bearing’s geometry, assuming perfect
In many applications the envelope analysis is applied on the kinematic motion without slippage. However, in practice the
band-pass filtered synchronised vibration signal. The filter band motion is not perfect and the axial-to-radial loads ratio varies with
should be around the dominant resonance frequency carrying the rotating speed and operating conditions. Therefore, the bearing
the bearing shocks. tones can only be approximated. To compensate for the inaccuracy
The envelope e(t) of the signal x(t) is the absolute value of of the bearing tones, tolerances are defined, for example a few
the analytic signal and it is usually derived using the Hilbert percent of their nominal values. The location of the peaks must
transform (H): therefore be searched for in the signature (order domain relevant
e = x + iH ( x ) ....................................(5) for the diagnosed bearing) within the tolerance range.
Each failure mode of the bearing has a well-defined pattern
The spectrum of the squared envelope is estimated in order to
of expected peaks. This pattern consists of a carrier (BPFO,
reveal the repetition rate of the shocks.
BPFI and BSF) with several harmonics and sidebands of the
Envelope analysis enhances the bearing vibrations but
shaft or the cage rotating speeds. The algorithm searches for
its capability is limited by the SNR of the bearing’s induced
the entire pattern, taking into account the provided tolerances,
vibrations compared to the other sources. Using a band-pass filter
and selecting the location k, with the highest corresponding
may increase the SNR. However, there are several problems with
z-score (Mahalanobis distance from the baseline). This selection
this method:
represents the most probable location of the peaks.
n The filter frequency band cannot be determined when the
The score P is calculated by:
bearing is in good condition
n The frequency band would differ from one bearing to ⎧⎪ ⎛ p − µ k ⎞ ⎫⎪
P = max ⎨ median ⎜ k ⎬ ..........................(6)
another k
⎪⎩ ⎝ σ k ⎟⎠ ⎪⎭
n For automatic health monitoring purposes, the filter frequency where k denotes the index of the pattern locations, pk are the peak
band must be predetermined2. levels, μk is the mean level in the baseline and σk is the standard
deviation of the baseline.
We are proposing to replace the band-pass filtering by dephase When the score P is low it means that no such pattern was
prior to envelope calculation. identified and the features are meaningless. If the score is high
it means that the searched failure mode was found and its most
2.2 Feature extraction probable location is the point where the maximum P is attained.
Since our goal is to compare the discrimination capabilities Once the most probable peak location was found, the energy
of the different processing methods, the feature extraction z-scores corresponding to each peak in the pattern can be calculated.
was performed using the same algorithms for all the orders Since all the signatures considered for detection of bearing deterioration
representations considered. The parameters were similar for the are spectra in the order domain, the energy z-scores ZEk,i are calculated
order of the original signal and the order of the dephased signal, according to Equation (7) (see also the illustration in Figure 1):
and a separate configuration was applied for the order of the
⎪⎧ n ⎪⎫ ⎪⎧ n ⎫⎪
envelope signal. (
Ek,i = 10 log10 ⎨ ∑ PSDi+ j ⎬ − 10 log10 ⎨ ∑ Trend PSDi+ j ⎬ )
The feature extraction process for bearings includes several ⎩⎪ j=− n ⎭⎪ ⎩⎪ j=− n ⎭⎪ ....(7)
steps: characterisation of the normal operating population Ek,i − µ k,i
ZEk,i =
(the baseline), application of an algorithm that searches for σ k,i
the peaks related to the bearing and calculation of the energies where i and j are spectrum indices, n is half the size of the energy
corresponding to the specific failure mode pattern. window, Trend(PSDi) is the background level (see Figure 1) of
2.2.1 Baselines the power spectral density (PSD), and μk,i and σk,i are the mean
and the standard deviation of the baseline calculated similarly to
Baselines are signatures in the considered domains derived Ek,i (see Equation 7).
from healthy recordings. Each set of such signatures represents
the statistics of healthy signatures in a certain predefined
operating mode. The baselines are usually the spectra (orders)
representing the average, variance, median and estimator of
the standard deviation (normalised inter-quartile range (IQR)
σ̃ = IQR ∙ 0.7413) of the signatures that compose it.

2
The frequency band can be determined using the kurtogram of the Figure 1. An illustration of the feature extraction: SI denotes
record with a damaged bearing. In an automatic system, the feature the corresponding rotating speed, the yellow surface
extraction process will require a large database of baselines covering the represents the baseline (μ+3σ) and the blue curve represents
frequency bands considered. the order spectrum

4
The International Journal of Condition Monitoring | Volume 2 | Issue 1 | March 2012
 BEARING DIAGNOSTICS | FEATURE

3. Dephase algorithm performance The time-frequency spectrogram of the simulated signal and
the asynchronous rotating speeds are shown in Figure 2.
In this section, the performance of the two dephase algorithms is
compared using a simulated signal. 3.2 Simulation results
3.1 Simulated signal The simulated signal was processed and the following
representations were derived: time frequency of the raw signals
The vibration spectra of healthy rotating machines consist of (see Figure 2), time-R2 orders (see Figure 3(a)) and both
two parts: in the lower frequency range the shaft vibrations are dephase algorithms (basic and adapted) were applied and their
manifested in several harmonics of the shaft speed; in the higher time-order representations were calculated (see Figures 3 and 4).
frequency range tooth mesh frequencies, blade pass frequencies, The dephase algorithms were applied in cascade on both rotating
etc show up at a few known frequencies. speeds, ie the first dephase of R1 was applied and subsequently the
The following model is used to simulate the vibration signal. resulting signal was dephased according to R2.
Let R1,…,RK be the rotating speeds of the simulated shafts and All the spectrograms are plotted using an identical logarithmic
denote by Φ1,…,ΦK the respective phase functions. For each colour scale (dB).
shaft k let pk,1,…,pk,Mk be the relative frequencies (orders) of the Since the signal did not contain asynchronous elements
parts which are mounted on shaft k. Denote by θk,1,…,θk,Mk the (besides the white noise), the dephase algorithm was expected
phase shifts of these rotating parts. The phase of each rotating to remove all the peaks.
part can be described by αk,i(t) = pk,iΦk(t)+θk,i. Denote by Ak,i
the amplitude function of the respective part, then the signal is
described by:
K Mk
x ( t ) = ∑ ∑ Ak,i ( t ) cos (α k,i ( t )) + n ( t ) ...................(8)
k=1 i=1
where n is white noise.
In a real machine, the rotating parts excitations X(ω) give
rise to responses at the sensor through the machinery’s structure
transfer function H(ω), which amplifies each frequency range
differently, Y(ω) = X(ω) · H(ω). The transfer function H(ω)
varies as a function of the rotating speed and varying loads. This
generates the respective variations in the amplitudes Ak,i(t).
The vibration signal was composed using two sets of rotating
speeds (R1 and R2) corresponding to asynchronous linear
accelerations (see Figure 2). The signal included components
synchronous with rotating speed R1 with different amplitudes
at harmonics 1, 2, 3, 17, 34, 51, 32 and 64. It also included
components synchronous with rotating speed R2 with different
amplitudes at harmonics 1, 2, 3, 4, 5, 24, 48, 72, 42 and 67.
In addition, a component at order 61 of R2 with frequency
modulation (FM) by the rotating speed R2 was added to simulate
a gear.
All the amplitudes of the synchronous elements (Ak,i(t) for
k = 1,2 corresponding to R1 and R2) were parabolic in order to
simulate the variations in amplitude as a function of the rotating
speed (see Figure 2). Figure 3. Orders by R2 spectrograms of: (a) original simulated
The added white noise SNR was 10 dB. No other asynchronous signal; (b) dephased signal using the basic dephase; and (c)
elements have been added to the signal. dephased signal using the adapted dephase

Figure 2. Simulated signals: (a) time-frequency spectrogram

of the simulated vibration signal; (b) shaft rotating speed time Figure 4. Dephase algorithms comparison (cross-sections of
histories spectrograms)

5
The International Journal of Condition Monitoring | Volume 2 | Issue 1 | March 2012
FEATURE | BEARING DIAGNOSTICS

In Figures 3 and 4 it can be observed that the adapted

dephase removed the synchronous elements almost completely
in the entire spectrum (see Figure 3(c)). The basic dephase did
attenuate the levels of the synchronous elements but was less
effective compared to the adapted dephase (see Figure 3(b)).
The dephase algorithm’s ability to attenuate the high-order
synchronous elements is affected by leakage at high frequencies
in the phase average process. Nevertheless, at very high orders,
corresponding to frequencies above 15 kHz, the peaks were
attenuated by more than 60 dB by the adaptive dephase, leaving
practically only the noise.

4. Feature comparison
The goal of the feature comparison was to evaluate the capability
of the different types of signatures to reveal the deteriorated Figure 5. Feature extraction results of seeded fault test 1
bearings while avoiding false alarms in the case of healthy
bearings.
Three types of signature were considered: the orders of the
signal, the orders of the envelope (after adapted dephase) and
the orders of the adapted dephased signals. One of the goals
of the feature comparison was to estimate the rate of potential
false alarms in the signatures of complex machinery. False alarms
are usually generated by excitations from different components
that happen to overlap the bearing tone range. Since a reliable
simulation of such an environment is very complex, real data
from seeded tests on complex machines have been used. The
data included recordings with faulty bearings and recordings with
healthy bearings. Results from two different seeded fault tests
(on different machines) are presented: one with a local fault on
the outer race of bearing A and the second with a roller bearing
(bearing B) with a localised fault on one of the rollers. Figure 6. Feature extraction results of seeded fault test 2
The results of feature extraction for different recordings at
different operating modes of the machinery are shown in Figures the expected pattern, one may conclude that the orders and the
5 and 6. dephase orders have similar capabilities of detection, while the
In both graphs the results of three types of signature are envelope orders have an inferior detection capability. It should be
displayed and separated by white vertical lines (from left to right: noted that this inferiority can be the result of a narrow bandwidth,
the regular orders, the envelope orders and the dephase orders). the nature of the deterioration and limitations of the dephase
Several columns are plotted per signature type, representing algorithm etc, which affect the envelope. In the case of bearing
different fault types (for example patterns for faults such as outer B, the low scores in the envelope orders are a result of the narrow
race, inner race, rotating elements, etc). bandwidth of the measurement.
The colours are proportional to the scores (P in Equation
(6)) corresponding to each cell in the graph (see the colour bar). 4.2 False alarms
The rows represent the different segments (operating modes) In general, in the recordings with the healthy bearings most of the
and recordings that have been analysed. It should be mentioned scores were low. For the healthy records in the envelope orders
that in the different operating modes the bearings are loaded and in the dephase orders only low scores can be observed.
differently and therefore the fault manifestation is expected only However, in the regular orders domain there are some healthy
in part of the segments. The segments with a faulty bearing are records with relatively high scores, some of them with the same
separated from the rest by horizontal white lines (in the bottom magnitude as the faulty cases.
rows). It seems that from the point of view of potential false alarms
the regular orders domain is inferior to the envelope and dephase
4.1 Detection orders domain. The inferiority of the regular orders can be
It can be observed that the results of both regular orders and explained by the confusion of the algorithm that searches for the
dephase orders have high scores in the records with a faulty bearing peaks in the tolerance range. In the regular orders, the
bearing. The scores of the envelope orders are high only in the harmonics of the shaft are not removed and some of them may
records with a faulty bearing A. have high levels compared to the baseline and may cause this
Since high scores are obtained when the algorithm detects confusion.

6
The International Journal of Condition Monitoring | Volume 2 | Issue 1 | March 2012
 BEARING DIAGNOSTICS | FEATURE

Conclusions technique – a review’, Tribology International, Vol 17,

pp 3-10, February 1984.
A new dephase algorithm derivative, which is adapted for 3. P D McFadden, ‘Condition monitoring of rolling element
separation of bearing tones in non-stationary signals, was bearing by vibration analysis’, Proceedings of the Inst Mech
presented and its performance evaluated. Eng Seminar on Machine Condition Monitoring, pp 49-53,
It was shown that the adapted dephase is superior to the basic 1990.
dephase and is much more efficient in removing the elements 4. N Sawalhi and R B Randall, ‘Semi-automated bearing
synchronous to the rotating speeds. The application of this diagnostics – three case studies’, School of Mechanical and
algorithm does not require any a-priori information on the nature Manufacturing Engineering, The University of New South
of the deteriorated bearing’s excitations and therefore it is suitable Wales, Sydney, Australia, 2008.
for automatic diagnostics. 5. Y F Wang and P J Kootsookos, ‘Modelling of low shaft speed
Three domains of signal analysis were proposed for automatic bearing faults for condition monitoring’, Mechanical Systems
bearing diagnostics: the orders of the signal, the orders of the and Signal Processing, 12 (3), pp 415-426, 1998.
envelope of the dephased signal and the orders of the dephase 6. J Antoni and R B Randall, ‘Differential diagnosis of gear and
signal. bearing faults’, Journal of Vibration and Acoustics, Vol 124,
A comparison of the analysis of the three domains was pp165-171, April 2002.
presented in terms of detection abilities and false-alarm rates. 7. R Klein, E Rudyk, E Masad and M Issacharoff, ‘Emphasising
Based on the examples shown, it seems that the orders of the bearings’ tones for prognostics’, The Sixth International
dephase signal present the best compromise, having the best Conference on Condition Monitoring and Machinery
discrimination capabilities. Failure Prevention Technologies, pp 578-587, 2009.
8. W X Yang and X M Ren, ‘Detecting impulses in mechanical
signals by wavelets’, EURASIP Journal on Applied Signal
References Processing, pp 1156-1162, 2004.
1. A A Barkov, Rolling Element and Fluid Film Bearing 9. N Sawalhi and R B Randall, ‘Localised fault diagnosis in
Diagnostics Using Enveloping Methods, Vibro Acoustical rolling element bearings in gearboxes’, Proceedings of The
Systems and Technologies Inc, 1994. Fifth International Conference on Condition Monitoring
2. P D McFadden and J D Smith, ‘Vibration monitoring of and Machinery Failure Prevention Technologies – CM/
rolling element bearing by the high-frequency resonance MFPT 2008.

7
The International Journal of Condition Monitoring | Volume 2 | Issue 1 | March 2012