A LABVIEW mini-expert to identify bearing
defects automatically
The rolling-element expert virtual instrument.
This is an example how advanced data processing of shocks produced by surface defects
leads to a robust mini-expert telling if a bearing surface defect is present on one or some of
the components like inner races, outer races of rollers (balls) even at very slow rotational
speeds. This limits the rate of false alarms characterizing simpler methods like SPM, HF,
etc.. The diagnosis is displayed by a mini-expert vi in a straightforward fashion, thus
bypassing the need of a vibration expert to analyze the bearing signatures, like the CSI
approach below which explains how to diagnose bearing failures.
Virtual instruments (abbreviated as vi's) consist of a PC with a data acquisition card. They
are developed under Labview. Their front panels include controls, displays, etc. on the PC
screen. Manipulating the controls (switches, rotary knobs, sliders, etc.) occurs with the
pointing device (mouse) instead of your fingers. The same PC can accommodate lots of
different instruments. The bearing expert is one of them. Some other dedicated instruments
follow in the next web pages. They concern monitoring rotor vibrations, gear failures,
flexible coupling reactions among other applications.
Virtual instrumentation opens the door to inexpensive transportable dedicated.
Problems with anti-friction bearings (from CSI seminars)
Background information on roller-element (anti-friction) bearings
Bearing life-cycle depends on two basic sets of variables: Application variables such as
load, speed, lubrication, mounting, temperature, and shaft quality, and configuration
variables such as design, materials of construction, method of construction, and quality
control.
Approximately 90% of all bearing failures are caused by improper applications, and
about 10% are caused by tone of the configuration variables. Our experience has shown
that one out of every 10 bearings taken off the shelf and installed will have a defect when
installed. This means that 10% of the time we have introduced a new problem into the
machine at overhaul.
Anti-friction bearings are constructed using several distinct parts--the inner race, outer
race, cages or retainers, and the balls or rollers. Three distinct parts aid us in detecting
bearing defects. Each component of an anti-friction bearing develops a unique
characteristic vibration signature. Typical defects detected in a vibration signature
include those found in the outer race, inner race, retainer or cage, and the balls or rollers.
By calculating the discrete frequencies associated with the four components, we can
identify which bearing component is defective. Other data contained in the vibration
signature includes information on internal clearances, looseness, adequacy of lubrication,
and the effect of thrust and radial loads. Most of the above bearing problems have
characteristics that are readily discernible from the fundamental running speed of the
shaft (1xRPM). By analysing the signature, we are able to determine whether bearing
failures are caused by an application variable or a configuration variable.
Anti-friction bearing frequencies.
The predominant frequencies generated by anti-friction bearings are classified and
computed as follows, based on a few characteristics of the bearing
Bearing parameters to compute bearing frequencies:
Formulas have been developed to calculate critical anti-friction bearing frequencies. Five
parameters must be known to accomplish these calculations:
Bd ball or roller diameter
˝
Pd
pitch diameter
Nb ˝
α number of balls or rollers
f ˝
contact angle
˝
RPM/60
BPFO Bearing Outer Race BPFO = f . (Nb/2) . [1-(Bd/Pd).cos α]
Frequency
BPFI BPFI = f . (Nb/2) . [1+(Bd/Pd).cos α
Bearing Inner Race Frequency
BSF ˝ BSF- =f /2. (Pd/Bd) . [1 - (Bd/Pd).cos α)2]
Ball Spin Frequency
FTF FTF = f/2 . [1-(Bd/Pd).cos α]
Fundamental Train Frequency
Analysing bearing signatures
When analysing bearings for surface defects, use the following guidelines:
˝
1. Bearing components normally fail in the following order: race defects, ball or roller
defects, cage defects (unless the bearing was defective when installed).
2. Inner race defects and failures occur at much lower amplitudes than outer race
defects.
3. BSF is usually generated when a ball or roller is defective. When multiple bars are
defective, multiples of BSF appear, i.e., if BSF is at 800 RPM and four balls have
defects, you should also see a peak at 3200 RPM or 4xBSF.
In all cases, a surface defect on an inner race, an outer race or on a roller (ball) generate
shocks at the bearing characteristic frequencies. In a frequency spectrum, these
correspond to train of frequency pulses extending from the 0-1000 Hz range normally
devoted to vibration analyses to higher frequency ranges in the domain of vibro-acoustics.
Such families of peaks merge with the peaks due to other causes. A real specialist must
then deal with the bearing analysis to sort out other causes present in the frequency
spectrum.
In the vibration frequency range (typically 10-1000 Hz), the patterns of frequency spectra
may indeed be complex, due to problems of rotor dynamics; pumps & ventilators (blade
passing frequencies, vanes, etc). Add to this problem contributions to the frequency
spectra that are specific to rolling-element bearings that have nothing to do with the
signatures associated to surface defects in the vibro-acoustic range (>1000 Hz):
• Internal looseness in a bearing has a component at l XRPM and several multiples of
shaft RPM. A bearing running on the shaft or in the housing shows 3xRPM and
possibly higher multiple's of RPM.
• Bearing misalignment can result in vibration at the number of balls or rollers times
shaft RPM.
Trending may be another way to monitor bearing and other failures
˝
• Vibration amplitude increases as bearing degrades and may disappear prior to failure.
• Baseline energy may increase across the entire spectrum and the band width broaden
as the bearing degrades.
This again requires quite a lot of expertise to interpret frequency spectra, as shown below
with some results from actual failures.
Frequency analyses of actual bearing failures in the frequency
spectra (from CSI seminar).
Suppose you want to pinpoint whether a bearing has some surface defects on either of its
components: inner race, outer race or roller. You perform a frequency analysis from an
accelerometer placed as near to the bearing as possible. Here is what one can get (from CSI
seminar).
Between 0 and 2000 Hz, one gets plenty of peaks. Are these peaks caused by bearing
defects? One then displays a list of all peaks vs. frequency (modern versions should
probably use a cursor). In this list one tries to spot contributions from bearing surface
defects after computing them from the rotational speed and the bearing characteristics or,
else, from such programs as ATLAS from SKF (see Appendix).
One spots in the list frequency rays with a dot for a defect on the outer race, while stars are
for a defect on the inner race. That is undoubtedly the job of a specialist and is no
automatic diagnosis to put in every hand.
The idea behind the bearing expert.
The above procedure some specialist to do the job. One should try to find an easier way
and streamline the diagnosis so that one only needs to connect the accelerometer to an
instrument, enter the bearing characteristics and let it run to obtain an automatic
diagnosis. That is what basically was done hereafter using a test rig with calibrated bearing
defects to validate the results.
Before describing the method in detail, let us outline what the method consists of.
1. One performs the analysis in the vibro-acoustic domain between 1(2) and 12 KHz. This
eliminates vibration contributions.
2. Shocks due to surface defects occur at the characteristic bearing frequencies. They do
not generate single frequencies. Instead one gets families of equally spaced frequency
rays. That is why this can cause the frequency spectra to be loaded with peaks. A better
way to reduce the number of peaks to examine is to use cepstra. Cepstra differ from
normal frequency spectra by the fact that they perform an analysis not in terms of
frequencies but in term of periods of shocks. Well-used, they reduce a signal to very few
peaks, paving the way to an automatic diagnosis of bearing defects.
3. Cesptra are somewhat insensitive to the attenuation of the sound waves produced by
the defects. In this way, the accelerometers need not be located very close to the bearing
to produce a reliable diagnosis.
4. Cepstral techniques deserve some care when implementing them. This is further
discussed in this document.
As a result, one presents a bearing expert to diagnose surface defects in rolling-element
bearings. It is extremely easy to use and does not require a vibration specialist to run.
The bearing expert was validated in the lab using a full-scale test rig.
Details of its inner workings are given below.
Test rig
The rig features variable speed from 60 to 1550 rpm, radial loads from 0 to 10 Tons,
completely separable roller or ball bearings on which calibrated defects are introduced on
races and/or rollers (balls). These defects have increased severities with widths ranging
from 0.1 to 0.4 mm. Bearing tested are SKF22NU15EC and FAG 1215TV
The test rigs with hydraulic jack, the test One of the bearings tested FAG 1215TV test
bearings and the variable speed drive. bearing: completely separable to introduce
Not shown: a hydraulic unit for lube and calibrated defects on races and balls. Inner
high-pressure jack (100bars). bores are 75 mm. Shocks were measured
with resonant accelerometers at first (freqs
at ca 5 KHz). Later normal off-shelf
general-purpose accelerometers were used
instead to use standard accelerometers from
conventional monitoring system. Their
responses were digitally high-passed to
eliminate low-frequency contributions of
vibrations and get true shock signatures.
Left-hand side: The test rigs with hydraulic jack, the test bearings and the variable speed
drive. Not shown: a hydraulic unit for lube and high-pressure jack (100bars). One of the
bearings tested FAG 1215TV test bearing. Inner bores are 75 mm. Shocks were measured
with resonant accelerometers at first (frequency at ca 5 KHz). Later normal off-the-shelf
general-purpose accelerometers were used instead to use standard accelerometers from
conventional monitoring system. Their responses were digitally high-passed to eliminate
low-frequency contributions of vibrations and get true shock signatures.
Raw defect signatures: see Table below along with miniexpert diagnoses
Raw signatures :Sound bearing: typically gaussian
noise. Surface roughnesses in bearings exhibit very
little spatial correlation. As a result, sound bearings
tend to produce acoustic responses that are almost
gaussian noise. Such a noise is characterized by a
Kurtosis (random variable) close to 3.
Outer race defect: Succession of shocks with
about the same amplitude whose spacing can be
derived from the
B(all)P(assing)F(requency)O(uter) race
characteristic frequency of the bearing,
depending on the pitch diameter, the ball (roller)
diameter, the number of the rollers (balls), the
angle of contact (if any) and the shaft rpm.
Rotational speed is 1500 rpm in this example
obtained with a calibrated defect on the outer
race of a SKF22NU15EC As all other responses
they are digitally high-filtered from 2 KHz on.
Inner race defect: Succession of shocks
modulated over a period corresponding to a shaft
rev. Shock spacing corresponds to a
B(all)P(assing)F(requency)I(inner) race that can
be computed from standard formulae, knowing
the pitch diameter, the ball diameter, the speed
and the contact angle. Shaft rpm is again 1500
rpm like in all following signatures.
Roller (roller)surface defect: Same remark as
before except that shock spacing corresponds to
twice the characteristic B(all)S(pin)F(requency)
and is modulated with a period corresponding to
one cage rev.
Expert system as a LABVIEW vi: Basic idea behind data processing
Suppose that a surface defect appears at the surface of a ball (roller) as shown below. This defect
will cause successive shocks as it encounters either the inner or the outer race. If the load carried
by the bearing is vertical downward, such shocks tend to increase with the pressure between
roller and races. This pressure reaches a maximum when the contact between the defect and the
races are aligned with the load zone. That is the case when the defect is below. As the positions
of the contacts depart from the vertical, the shocks they produce decrease in amplitude.
Shocks altogether disappear when
the ball is located in the upper
part of the outer race since there
is no longer any pressure between
the ball and either race. In this
reasoning, the outer race is
supposed fixed and the bearing is
not pre-constrained (look at suffix
C which is C3 in the test
bearings).
That is what one conveys in the
left-hand-side representation of
the bearing for one cage
revolution. Blue circles indicate
the positions of the ball defect
with the inner race, whereas red
circles have the same meaning
with the outer race.
Circle radii are proportional to shock intensities. They depend on some distribution of the load
zone. In order to account for the attenuation of shock waves through the roller as sensed by an
outboard accelerometer, one could further reduce the intensities of the blue circles, although this
was not really observed with the experimental results from the test rig. The right-hand side
bearing shows the position of contacts of the defects with the races over several cage revs.
Now the main red circle can be viewed as the source of the acoustic wave generated by the
defect and the subsequent blue and red circles like echoes of this main shock waves. These are
not real echoes as one encounters in the propagation of acoustic shock waves. The latter would
be much closer spaced. The former are pseudo echoes caused by the dynamic and kinematics of
bearings and their surface defects.
Attenuated echo-like shock waves lend themselves to cepstral techniques when one wishes to
separate them from the main incident shock wave and reach a high data compression ratio. To
this aim, one must exert some care to obtain sound cepstra according to L.F. Pau [Failure
Diagnosis and Performance Testing Series Control and Systems Theory vol. 11, Marcel
Dekker]:
Centering the buffer of data within a load cycle and windowing it for
proper cesptral analyses.
Steps:
1. It is important to start the data acquisition when the defect enters its load zone and
terminates it when leaving the zone. Therefore one samples the acoustic response over a
time span extending over several load zones for ball defects (the longest of all
corresponding to successive cage revolutions). One then narrows the time window to
correspond to a maximum energy in its center and to extend over half a cage revolution.
As shown in the picture below, one samples the shocks over a few cage revs and then
extract a shorter part of the data acquisition window to center it over half a cage rev.
2. one then applies a double exponential window w(t) simulating an attenuation of shock
waves when departing from the window center. Like other defects, outer race defects
lacking specific load zones now generate attenuated echoes fit for cepstral analyses.
Description of the Labview bearing mini-expert front panel
The control window to A «bearing characteristic window» to enter the the «file saving window» to chose between on-line
select the number of geometric data of the bearing in order to compute or delayed analysis.
averages to be used for the characteristic frequencies from the rotation Several switches are available:
identification speed Run a diagnosis from either a file or real
These data can be obtained with standard programs measurements with a MIO card from National.
like SKF ATLAS for most standard bearings in the Save the result from real measurements or not
nomenclature. Auto file name generates automatic filenames with
encoded dates and times. Otherwise the user must
enter his choice.
This display is more important when one runs the
expert with a number of averages bigger than one.
The speed can vary between successive acquisitions.
In case the vi does not measure speed, For automatic speed measurements, the The indicator mentions whether the speed
one can enter the speeds of the inner race instant speed is monitored. With the variations are OK for the diagnosis.
and the outer race manually. One assumes switch one can reject too fast speed
these speeds to be correct and steady. variation for the diagnosis.
The raw signal from the accelerometer (here unscaled) in time Cepstrum for the range of quefrencies associated to bearing
doamin (sec) defects. Green for BPFO, cyan for BPFI, yellow for ball or
Gain cursor for the MIO card to get the full dynalics: 0.5 fro +- roller, red highest value outside these quefrencies. An algorithm
10V, 1 for +-5V, etc. cleans the cesptrum from noise.
This panel provides a summary of the diagnosis. Some other parameters describing the
Four bar graphs show values that are derived from ruggedness of various displays.
weighed contributions to the cepstrum caused by the Pk-pk values of 1KHz-12KHz filtered signatures
standard bearing defects. Basically it sums up the of the accelerometers provide some idea about
amplitudes of the cepstrum around the first few the severity of the shocks.
quefrencies for shocks caused by the typical defects The vi does not provide an absolute severity
BPFO BPFI and BSF (2X) according to some appropriate scale, but a good identification of a defect.
weighing. The last bar graph corresponds to the same
weighing applied to the highest remaining peak outside The experience shows that absolute severities are
those for standard defects. hard to achieve, despite what some vendors
pretend. With brand-new healthy bearings, their
Sound bearing emit signatures with Kurtosis around 3 due scale has quite often led to diagnose ghost faults.
to normal spatially uncorrelated surface roughness.
Shocks raise Kurtosis. When it exceeds the "Kurtosis
limit" entry, an automatic diagnosis follows if the switch
next to it is low. Otherwise the expert declares "no shock
no diagnosis"
If the switch is up an automatic diagnosis occurs. It may
still consistently detect the same type of fault repeatedly,
because cepstral techniques are very sensitive. Beware
nevertheless that they are also sensitive to noise that
could cause random diagnoses. If this occurs, discard the
diagnosis.
Examples of diagnoses
These are examples from actual rig measurements of almost fool-proof diagnoses. The last one corresponds
to a sound bearing with a close to 3 Kurtosis and the "Expert always" switch down. In case shocks are caused
by another component than the bearing and yet at a frequency close to those of a characteristic bearing
failure, another diagnosis is "Default not on bearing". These diagnoses are only what matters to the end
user.
Gallery of case studies with full Labview bearing expert panels
Ball defect at low speed
The bearing was damaged with a 0.2mm groove on one roller. SPM says it is slightly or not damaged above
400 rpm. At lower speeds, SPM severity scale indicates the bearing is healthy.
Inner race defect
The bearing was damaged with is a calibrated 0.2 mm groove on the inner race. At 1500 rpm SPM would
specify that the bearing is slightly damaged.
Outer race Defect
With 0.4 mm groove at the surface of the outer race.
Sound bearing
The "Expert Always" is up. This overrides a too low Kurtosis and carries out the diagnosis al the same. It tells
that one detects an inner race defect. This may correspond to a very minute defect and should be believed
with some caution. Next comes the diagnosis for the same response and the "no shock no diag(nosis)" message
Conclusions
Identification of defects is superior and top easy and validated on a
full-scale test rig.
One need not be a vibration and/or acoustics specialist to run the expert. The expert
focuses on surface defects and nothing else. That is what makes it efficient and easy to use.
Using signatures obtained with the test rig with bearings fitted with various types of
surface defects and increasing defect severities, the mini-expert consistently identified the
defects correctly.
Bearings were either of the roller type (SKF22NU15EC) or of the ball type (double row
spherical FAG1215TV). These bearings share the interesting property that one can easily
implement calibrated defects on the surface of all elements (races and rollers or balls)
because one can separate their components easily "lego"-style without breaking any of
their components.
Identification runs down to very low speed and is reasonably
insensitive to sensor location.
Remember that Kurtosis and cepstra (through the log transform) are a-dimensional
criteria. They perform extremely well to identify bearing defects despite somewhat "poor"
acoustic paths between the source of the shocks and the outboard accelerometer and/or
despite very low rotational speeds (down to 60 rpm and tiny 0.2mm wide defects, it never
failed). Of course, do not locate the accelerometer very far from the bearing to investigate.
Severity scales
Kurtosis and cepstra become somewhat useless to assess fault severities. The latter must
rely on the amplitudes of the shock waves as measured by the outboard sensor, eventually
with notions like carpet, peak levels as in SPM or other methods.
One could not establish reliable severity charts based on shock amplitudes. Some well-
established scales like from SPM also failed with sound bearings that were mistakenly
declared faulty, even at high speeds lie 1500 rpm.
Absolute severity scales require a careful placement of the accelerometers and these must
exhibit precise characteristics. Accelerometers used for vibration monitoring may exhibit
varying frequency responses in the vibro-acoustic domain and thus may prohibit the use of
severity charts.
Questioning the feasibility of fool-proof absolute severity charts
Severity charts like from SPM systems with SPM accelerometers (resonance at 32 KHz)
performed well to follow defects with increasing widths from 0.2 to 0.4 mm. That is fine for
trending. However, some sound bearings were declared as faulty by SPM right from the
start, thus causing a false. With SPM, faulty bearings systematically went unnoticed when
rotating below 400 rpm.
It would be nice to merge reliable severity charts, if any exist, in the above LABVIEW vi to
reach an unsurpassed bearing expert that, for sure, tells that the bearing is faulty (done
without false alarms with the present vi) and by how much (not done).
Using ATLAS to enter the bearing characteristics
The Appendix shows how to get bearing characteristics required by the bearing expert.
In order to eliminate a small inconvenience, one could link the above bearing mini-expert
to data banks like SKF ATLAS to save the user the transcript of the bearing dimensions.
If ATLAS does not hold the specific bearing type you read on the outer ring but rather a
generic type, you should verify if it comes up with the same number of rolling elements as
you actually have. They may be variations on the same theme, depending on the cage
execution.
Never blindly enter the engraved reference you read on the outer race of a bearing into
ATLAS. For example, the SKF bearing of the tests read 22NU15EC. In ATLAS you must
enter NU2215E. More o this I the Appendix.
Measuring speed
The mini-expert expects to know the rotational speed rather precisely. This was done from
one-top/rev TTL pulses fed to the 24-bit counters of National Instruments MIO cards. If
one cannot afford measuring the speed directly, one can enter its value read from other
devices in a plant, thus by-passing the module measuring speed.
Using standard accelerometers
Accelerometers were standard products as are available in common vibration monitoring
systems. That bypasses the need to install special types of accelerometers resonating at
higher frequencies like in other commercial solutions.
Accelerometers must be mounted properly to respond to frequencies up to ca 10 KHz. One
should not use highly resonant accelerometers to run the expert. Standard accelerometers
are just fine.
Hardware required
A PC with a National MIO-10 card or upward. Note that using cards like PCI-MIO-1
opens the door to all other virtual instruments developed in this research project: gear
defect analysis, rotor vibration monitoring with full spectra and some others like fast
digital data recorders.
Any question? Would you like to run the expert with a library of cases from the test rig? This
could be done upon request.
.mailto:[email protected]
Responsible Editor:, Gérard D'Ans, Research engineer at Laborelec and industrial collaborator of
ULB, tel 32 2 650 25 15 (ULB) or 32 2 382 0 568 (Laborelec)
Appendix: How to use ATLAS.exe from SKF to retrieve bearing
dimensions for SKF 22NU15EC
The number of rolling elements in the SKF22NU15EC of the test rig is 18 instead of 17 in
ATLAS and 22NU15EC is not a ball but a roller bearing. So be careful when entering the
type of bearing you want to investigate ad do not rely on what is written on the outer race.
Correct entry in ATLAS
Typing the prefix NU first despite what is engraved on the outer race (22NU15EC) yields
the correct number of rollers.
Another point: the diameters of the rollers and the pitch were approximated in the expert.
Despite this, its diagnosis was correct all the time. The expert is robust.
