2. Health Monitoring of steady speed/load
machinery a well established practice
However, few techniques are available for
monitoring unsteadily operating equipment
Techniques required for advanced equipment
such as electromechanical shovel, variable
duty hoists, etc.
4. Machinery Maintenance Policy driven by:
◦ Availability of resources (spare parts, pers., capital)
◦ Importance of equipment
◦ Availability of technology and expertise
Modern Maintenance Policy evolved through:
◦ Run-to-Failure
◦ Periodic Maintenance
◦ Predictive Maintenance
Maintenance is delayed until some monitored
parameter of the equipment becomes erratic
Proactive
Balances resources
5. Benefits:
◦ Environment
◦ Safety
◦ Production
◦ Staff Shortages/Costs
◦ Scheduling
◦ Spare Parts (JIT)
◦ Insurance
◦ Life Extension
6. Faults in rotating
machinery have very
representative features in
the frequency domain
Consider bearing:
◦ Frequency Response a
function of Fault, Slippage,
Noise
Diagrams from: Randall, B. State
of the Art in Machinery Monitoring, JSV
7. One branch of artificial-intelligence domain
Usually involves representing a state or object
to be indentified with a vector of
commensurate numerical values
◦ E.g. In classifying fruit: weight, spectroscopic
values, etc.
Representative vector called a “pattern” or
“classification object”
Classification achieved by computing decision
surfaces around classes of objects
10. Explore a technique developed for monitoring
health of structures first established in
◦ K Worden, H Sohn, CR Farrar. Novelty detection in a
changing environment: Regression and inter
polation approaches, J.Sound Vibrat. 258 (2002).
Essential idea: clustering of patterns will vary
with modal parameters (speed/load/temp)
Technique improves the segmentation step;
rendering classification almost trivial
11. Variable speed machinery
◦ Elements of a machine‟s vibratory response are
assumed to have a strong relation to the speed of
the given machinery
Distribution for speeds:
◦ Means vary with speed *C30
◦ Variances vary with resonance response
*C20
y
* C10
x
12. Segment vibration signal
Group segments according to the machine‟s
speed
Calculate Gaussian parameters for small
segments of speed (sample statistics
assumed to be population statistics)
◦ Curse of Dimensionality
Interpolate or Regress each component of
statistical parameters
Decision boundary function of speed (and
other modal parameters)
13. Sub problem of pattern recognition
◦ Rather than train a classifier on all classes, we train
on the “normal” class and then signal an error when
behaviour deviates from it
◦ Employed where knowledge of all classes (of faults)
not practical to attain
Decision boundary encircles normal patterns
A wide variety of techniques available
Examine two:
◦ Boundaries containing a certain quantile of data (i.e.
a discordance test)
◦ Boundaries derived by Support Vectors
14. For uni-variate data – a simple task:
◦ Classify as normal if test pattern falls within nth
quantile of training data
◦ Think confidence level
P(| x | R) 0.95 | x| 1.96 x 1.96
1.96
15. For multi-variate data:
◦ Build multi-variate model from multiple uni-variate
ones – assuming independence
P( A B) P( A)* P( B)
16. Assuming independence
1 x
d d ( i i )2
1
p ( x ) p ( xi ) e 2 i
i 1 i 1 2 i
d
x
1
( i i )2 1
( x )t 1 ( x )
1 2 i 1 i 1
d
e e 2
(2 ) d i (2 ) | |
d
i 1
17. Example:
◦ Distr. #1 µ= 5 and σ=4
◦ Distr. #2 µ= 10 and σ=8
4 0
◦ Joint Distr. therefore has µ= [5,10] and
0 8
◦ The level curves of the distribution are determined
by the Mahalanobis squared distance given by
t 1
r (x ) (x )
2
19. Is independence a reasonable assumption in
the context of variable load/speed
machinery?
◦ Many spectral components of vib. Machinery
strongly related
Consider Bearings
Consider Gear Meshing
Etc.
◦ Gaussian fit depends on independence of
probabilities of individual parameters
◦ May prove poor in this context
20. This ellipsoidal boundary is very rigid and will
not work well if the data is not perfectly Gaussian
Rather than computing the quantile for a test
patterns given speed
◦ Center each speed bin‟s data about the origin and alter
its distribution from ellipsoidal to spherical with the
whitening transform
◦ Consequence: All modal data is centered at the origin
with faulted data orbiting the healthy data
◦ Now draw a decision boundary around the healthy data:
use Support Vectors
N.B. There is still some dependence on the
assumption of a Gaussian fit
21. *C30
Faulted Data
y
Healthy Data
*C20 for all Speeds
y x
* C10
x
22. Support Vector Technique: Tax‟s Support
Vector Data Description (for Novelty
Detection)
◦ Attempts to fit a sphere of minimal radius around
normal data
◦ But a in a higher dimensional space (using the
“kernel trick”)
Generates a very flexible decision boundary in the
input space
23.
24. Dr. Timusk‟s PhD data
Spectraquest gear dynamics simulator
◦ Variable frequency drive
◦ Gearbox (two stage parallel reduction)
Subject to variable loads (particle brake)
Data acquisition system: NI PXI
◦ Ceramic Shear ICP Accelerometers (0.5 to 6500 Hz)
◦ Sampling 4kHz/channel
Faults:
◦ motor with bearing faults, broken rotor bars, rotor unbalance
◦ gear faults: missing tooth, chipped pinion, outer race bearing
25. Feature Post-
Sensing Segmentation Classification
Extraction Processing
Segment vibration data into segments of
„steady‟ speed and load
◦ Segments defined by n-shaft rotations
Accounts for varying speed
Ensures coherent signal
Windowed (Gaussian Window – 70% overlap)
26. Steady speed/load not guaranteed
◦ But can generate segments with reasonable steadiness
and variance can be computed
Group vibration segments into bins of a selected
size
◦ Size effects how many classification objects in each bin
curse of dimensionality balanced against need for very fine
modal resolution
27. Feature Post-
Sensing Segmentation Classification
Extraction Processing
A number of parameters could be employed to
represent a vibration segment
◦ Crest factor, average power, kurtosis, impulse
factor, etc.
◦ Autoregressive Models (AR)
AR models
◦ Think Root Locus Method from Control Systems: You
determine the placement of poles to shape the
frequency response of the CS
AR models control placement of poles to shape model‟s
frequency response to be representative of a signal‟s
frequency response in the least squares sense
◦ User selects the number of poles
The more poles, the more representative the signal is
Balanced against the curse of dimensionality
28.
29. Feature Post-
Sensing Segmentation Classification
Extraction Processing
Segmentation step makes data almost
perfectly separable
30.
31. Fit each component of each statistical parameter
(mean and covariance matrix) to model
Components of mean vector could be fit with
polynomial
Components of covariance matrix not traceable
32. Covariance matrix components vary wildly
Additional concern:
◦ Covariance matrix derived from regression may not
be positive semi-definite
x x 0
t
Method available to deal with issue (added complexity)
Classification results are poor
33. Instead, we must store each bin‟s statistical
parameters
◦ Any bins which are ill-conditioned or under
sampled could then simply be interpolated over
◦ Positive semi definitenessguaranteed
◦ Good classification results
34. High acceptance rate of healthy data generates poor
rejection rate of faulted data (ellipsoidal boundaries)
35. Interpolating over missing/ill-conditioned
bins
◦ One missing bin: interpolated statistics almost the
same as those of measured values
◦ Three bins missing:
36. SVDD has one
parameter – sigma
◦ Integer value [1,inf)
◦ Low values – Tight
bound
Choice of sigma
has very little effect
No frustrating
trade off between
classification error
on normal and
faulted data
Superior
classification
37. Too good to be true?
◦ Tax explored variable load/speed machinery
without our segmentation steps
Training SVDD over all speeds, he achieved an average
error of 8%
Our average error of 2% is very plausible!
Segmentation step removes overlap between
faulted data of one speed bin and healthy
data of others
38. The errors shown on
the right are based on
data from one
accelerometer
Faults are not all
located near this
accelerometer
Segmentation has
made classifications
sensitive enough so
that accelerometers
can measure spatially
disparate faults
Plausible: Underwater
warfare analogy
39. For a fixed amount of data, increasing the
dimensionality of the space increase
classification error
Statistical Parameterization is doubly cursed
40. Statistical parameterization
◦ Approach extends well to variable speed machinery
Gaussian/independence assumption not theoretically
correct but the data cluster well anyway
◦ Prefer interpolation over regression
Memory requirements not a concern
(but might try piece-wise linear regression in the
future)
◦ Interpolation possible over missing/ill-conditioned
bins
41. ◦ Whitened data with Support Vectors
Statistics for each bin still required
Produces a less rigid decision boundary
Better classification results
Still somewhat dependent on assumption of
Gaussianaity
◦ Segmentation essentially renders classification
stage trivial
◦ Segmentation makes it possible for sensors to
detect faults on physically distant machinery
components
◦ Suffers doubly from the curse of dimensionality
42. Verification of methodology on real world
machinery (diamond drill head with dyno)
Develop classifier variants for multi-modal
processes which are less susceptible to the curse
of dimensionality
Develop ONLINE prognostics techniques
◦ When will failure occur?
◦ What is the probability a machine will fail at time x?
Develop economic means of measuring torsional
load for this application
Develop complete software architecture (software
engineering principles) and prototype