Condition Monitoring of Variable State Machinery

1

LOGO

Condition Monitoring of Machinery Subject to Variable States

Condition Monitoring of Machinery Subject to Variable States

Jordan McBain, P.Eng.

Monitoring of Mobile Underground Mining Equipment

2 - 68

Acknowledgments

Vale IncoCEMIDr. TimuskCommittee, External/Internal Reviewers

3 - 68

Problem Overview

Maintenance options have advanced considerably from reactive policies

Modern sensors, computers and algorithms have set the stage

Health monitoring of steady machinery widely available Few techniques are available for monitoring unsteadily

operating equipment Techniques required for advanced equipment such as

electromechanical shovel, variable duty hoists, etc. Subject to variable loads, speed, temperatures, etc.

4 - 68

Problem Definition

Enable condition monitoring (CM) of mobile underground mining machinery Multiple Artificial Intelligence (AI) techniques

• As validated on laboratory test bench – gearbox and bearing faults

• Extensible to real-world applications?

Administration issues of automated computerized monitoring systems• Sporadic network availability• Bandwidth limited environments• Enterprise level integration

Extensible Software Engineering Architecture

5 - 68

Outline

Background Condition Monitoring (CM) Artificial Intelligence (AI) for CM Monitoring of Variable-State Machinery

Methodology and Limitations Statistical Parameterization Augmented Novelty Detection System Identification Cross-Correlation Software Architecture Conclusion

6

Background

7 - 68

Maintenance Management

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

• Only 15% of failures follow MTBF model (Lihovd, 1998)– Naval/air study

Predictive Maintenance• Maintenance is delayed until some monitored parameter of the

equipment becomes erratic• Proactive• Balances resources

8 - 68

Condition Monitoring

Thrust State of equipment determined by variations in

monitored parameters

Benefits Environment Safety Production Staff Shortages/Costs Scheduling Spare Parts (JIT) Insurance Life Extension

9 - 68

AI for CM

Savvy technicians employ(ed) a screw driver set atop a vibrating machine Resultant vibration of screw driver used by

technician to classify health

AUTOMATE THIS! More sensitive Earlier detection of faults Consistent, reliable measurements

• Consistent, reliable classification

10 - 68

Pattern Recognition

One branch of AI domainPatterns used to compute decision ruleGeneralization(Double) Curse of Dimensionality

11 - 68

Pattern Recognition

Sensing• Accelerometers, tachometers, acoustic emission

sensors, thermocouples, etc.

Segments• Choose time intervals for division of data• Synchronous intervals (fixed # of samples)• Asynchronous intervals (fixed # of shaft rotations)

Feature Extraction• N Parameters combined to form “patterns” or “feature vectors”• Statistics, Auto-regressive (AR) models, MUSIC Spectrum, etc.

Classify

• Generate decision rules from training data• Apply decision rules to test data• Fault detection: Novelty Detection (support vectors, neural

networks, etc.)

Post-Processing

• Diagnostics, prognostics• Health reporting• Sensor failure analysis• Etc.

12 - 68

Monitored Parameters

Vibration, Thermography, Oil Analysis, NDT, et

Vibration Heavily used in literature Non-destructive, online,

sensitive Faults in rotating machinery

have strongly representative features in the frequency domain

Diagram: (Randall, 2004)

13 - 68

Novelty Detection

Motivation: addresses imbalance of data from one class in relation to that of others Data from faulted states are difficult to collect (economics,

operations)

Sub problem of pattern recognition train on the “normal” class and then signal error when

behaviour deviates from the decision boundary

A wide variety of techniques available Examine two:

Boundaries containing a certain quantile of data (i.e. a statistical discordance test)

Boundaries derived by Support Vectors

14 - 68

Support Vectors

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

15 - 68

Variable-State Machinery

Primary aggravators: load and speed Referred to as nuisance variables in the literature (Gelman,

2005)

In vibration monitoring Power of vibration a product of the effects of load and

speed• Relation between power and speed non-linear• Resonances!• Vibration a function of health and mechanical state (speed,

load, etc.)

16 - 68

• When machine is healthy, deviations in consequent vibrations are small

• When health is poor, deviations due to speed become significant

• Stack: Damping in undamaged machinery is largely insensitive to speed/load changes – damaged machine

Diagram: (Stack, 2003)

17

Methodology and Limitations

18 - 68

Test Bench

19 - 68

Limitations

Test bench realism Mass of shaft

• Inertia of rotor system– Signal-to-Noise Ratio (SNR) of fault signals

Gear type• Helical gears

– Increased mesh strength» SNR of fault signals

Lack of complexity• Variable Frequency Drive (VFD) and/or particle break

control on load/speed– Torsional vibrations (typical in diesel engines) not

evaluated

20 - 68

Limitations

Challenging control problem Closed loop on speed

VFD Open loop on load VFD Torque profile fed forward

to speed VFD Torque control

superimposed on speed control

Noisy torque signal Inconsistent effect on

algorithms

21 - 68

Limitations

Applicability to Underground Environment Harsh conditions not present in laboratory Temperatures

• Degradation of lubricants• Thermal expansion of components

– Alters vibratory signature– Time-varying parameter not considered

Heavy shock/vibration• Noise for vibration-based CM

– Inclusion in training » Overly broad decision boundary

– Exclusion» Additionally processing required

22 - 68

Understanding Classification Results

23

Statistical Parameterization

24 - 68

Statistical Parameterization

Established technique from the literature (Worden, 2001)

Motivation: Distribution of vibration parameters will change

according to time-varying parameters

Experiments with variable speed only

25 - 68

Statistical Parameterization

Established Thrust: Develop a decision boundary that changes according to

speed Double Curse of Dimensionality Restrictive Gaussian assumption

x

y

* C10

*C20

*C30

26 - 68

Statistical Parameterization

27 - 68

Statistical Parameterization: Improvement

Contribution: Develop a rule to first center and whiten data

• Eigenvalue problem Center/whiten all training data

• Train SVDD Center/whiten test data according to rule

• Apply SVDD decision boundary to determine faults

x

y

* C10

*C20

*C30

x

y

Healthy Data for all Speeds

Faulted Data

28 - 68

Statistical Parameterization

Choice of AR Model Order with Standard Statistical Parameterization(Interpolation)

29 - 68

Statistical Parameterization

Statistical Paramterization with Whitening

30 - 68

Statistical Parameterization

Interpolation over (4 consecutive) missing bins

Smaller number of missing bins Minimal impact

31 - 68

Statistical Parameterication

Curse of Dimensionality Measured by increasing feature vector dimension

32 - 68

Statistical Parameterization

Established approach Double curse of dimensionality Gaussian Assumption Excellent classification results

Statistical Parameterization with Whitening Mitigates double curse Provides more flexible boundary

• Reducing effect of Gaussian Assumption

Classification results at least as good

33

Augmented Novelty Detection

34 - 68

Augmented Novelty Detection

Previous limitations Varying degrees of curse of dimensionality Gaussian Assumption

Motivation Intuition gained from Statistical Parameterization

• Include time-varying parameter in feature vector– Trivial but not established in the literature

Problem reduced to standard novelty detection

Experiments with variable speed only

35 - 68

Background: Order Tracking

Ordinarily: Vibration sampled at constant intervals Order tracking: vibration sampled at constant shaft

rotational intervals Use pulse train from tachometer to indicate sampling

interval Irregular resampling

Question: How many samples per shaft rotation are

appropriate to gain good classification results?

36 - 68

Sensitivity Analysis: Order Tracking

37 - 68

Order Tracking

Statistics

AR10

With OT Without OT

38 - 68

Interesting Feature Vector: Acoustic Emissions

Statistical ParameterizationMulti-Modal Novelty Detection

39 - 68

Baseline: Statistical Parameterization

AR10 Feature VectorStatistical Features

40 - 68

Results

Statistical Parameterization Multi-Modal Novelty Detection

41 - 68

Curse of Dimensionality

Multi-Modal ND Statistical Paramterization

42 - 68

Validation: Experimental Procedure

Procedure:- Train with on one healthy gear- Validate on a different healthy gear and faulted components

43

System Identification

44 - 68

System Identification

Shifts problem to the feature vector rather than adapting decision boundary

Feature vector composed of elements of a gear’s transfer functions

Analysis with both varying speed and load

45 - 68

System Identification

Assume a gear can be modeled as a torsional spring

Use system identification to model the transfer function with MIMO

System:

Gearbox Speed

LoadVibration

( )mx cx kx f t

x x u

y x u

B

C D

A 1 2

1 2

( ) ( )( ) ( ) ( )

( ) ( )

B z B zV z S z T z

A z A z

46 - 68

Omitting Time-Varying Parameters

No Adaptation for Speed or Load No Adaptation for Load

Employing Multi-Modal Novelty Detection

47 - 68

Sensitivity Analysis: Model Order

Changing Number of ZeroesChanging Number of Poles

48 - 68

Curse of Dimensionality

49 - 68

Generalization

50

Cross-Correlation Analysis

51 - 68

Cross-Correlation Analysis

SysID Failings Must measure all time-varying parameters Must develop transfer functions for each

• Susceptibility to the double curse of dimensionality?

Computational expensive

Cross-correlation based feature vector Sensors on disparate machinery components will

behave in a time-correlated manner Use statistical correlation signal

• Generate feature vectors from it

Eliminates failings of SysID*

( )[ ] [ ] [ ]m

f g n m g n mf

*( )( ) ( ) ( )f g t g t df

52 - 68

Results

53 - 68

Curse of Dimensionality

54 - 68

Generalization

55

Software Engineering Architecture

56 - 68

Challenge

No silver bullet for condition monitoring (pattern recognition) Multitude of techniques for multitude of problems Wide variety of (transient) machinery Similar CM problems: prognostics, sensor failure analysis Extensible beyond rotating machinery

Pattern recognition problem generates multiple possible combinations of Sensing Segmentation Feature vector generation Classification techniques Post-processing requirements

57 - 68

Software Design

Design for change! Recognize

broader-scoped problem• Intelligent

Signal Processing and Analysis

Smart Signal Processing

Structural Monitoring

Aircraft Monitoring

Stationary Equipment Monitoring

Seismicity Monitoring in Mines

Wind Turbine Monitoring

Ship Propulsion and Auxiliary System

Biomedical Monitoring

Automotive Part and Test Bench Monitoring

Vehicle Monitoring

Process Monitoring

Measure while Drilling

58 - 68

Scope of Present Work

Design Object-Oriented (OO) Data Processing Layer Online, flexible and dynamic routing of signals Augmentable with user/programmer defined

techniques Design for intelligent signal processing

• Implement for CM

Create MATLAB prototypeReview and make recommendations for

integration with mining enterprise systems International Rock Excavation Data Exchange

Standard (IREDES)

59 - 68

Use Cases

Hand-held, portable monitoring system Cheaper, economies of scale Intermittent monitoring

Dedicated online monitoring system Costly Equivalent problem to intermittent monitoring

• Intermittent functionality/benefits achievable by “wheeling” this system around

Capabilities to monitor more than one (physically proximate) machine at a time

Data Connectivity Limited bandwidth Intermittent network connectivity

60 - 68

Design

Dynamic online signal routing Supports online selection of algorithms Subscription based

Multiple data sources From

• disc • DAQ • networked sensors

Varied sensor types Support n-dimensional signals

+MUserSamplesQueue()+register() : RegistrationToken+hasBeenCleared() : bool+addToQueue()+clearData()+getData()+unregister()

-Data-Time-AbsoluteTime

MUserSamplesQueue

+DataSource()+MonitorChannel() : RegistrationToken+getData()+clearData()+updateQueue()

-ChannelQueues : MUserSamplesQueue-SampleRatesArray-ChannelNames-updateQueues

DataSource

1 *

NetworkedSource AsynchronousDataSource StoredSensorDataDAQSystem

61 - 68

Design

Signal conditioning strategy Typical signal processing techniques “Signal” representing time intervals for segmentation

Signal conditioner Does the actual work of

• getting sensor data • passing it through selected algorithms

Feature generator Requests conditioned signals from conditioner Segments signals according to segmentation strategy Combines multiple feature vectors into one

62 - 68

Design

+DataSource()+MonitorChannel() : RegistrationToken+getData()+clearData()+updateQueue()

-ChannelQueues : MUserSamplesQueue-SampleRatesArray-ChannelNames-updateQueues

DataSource

-SegStrategy-SigConditioner-Name

FeatureGenerator

+SegmentationStrategy()+getSegmentTiming()

SegmentationStrategy

+SignalCondionter()+getConditionedData()+register()+unregister()+clearData()-recurseForConditionedData()-recurseRegister()-recurseClearData()

-SigConditoningChain : SignalConditioner

SignalConditioner

«uses»

«uses»

1

*

keyPhasor constantNumRotations constantTimeInterval«uses»

63 - 68

Design

Intelligent Analyzer Strategy Requests feature

vectors from feature generator

Does the classification work• Depending on

“state” of classification problem

+DataSource()+MonitorChannel() : <unspecified>+getData()+clearData()+updateQueue()

-ChannelQueues-SampleRatesArray-ChannelNames-updateQueues

DataSource

-SegStrategy-SigConditioner-Name

FeatureGenerator

IntelligentAnalyzerStrategy

ExpertSystem FaultTree PatternRecognition

NoveltyDetection

SVDD

1

*

CombinationOfClassifiers

StatisticalParameterization

1 *

64 - 68

Integration with Enterprise Layer

IREDES in need of augmentation for CM CM standards already exist

Don’t reinvent the wheel

Two options of differing granularity Open Systems Architecture for Enterprise Application

Integration (OSA-EAI) Open Systems Architecture for Condition-Base

Maintenance (OSA-CBM)

Wide industrial support US Navy Caterpillar Rockwell Automation Systems

65

Conclusion

66 - 68

Conclusions

No silver bullet for CM Wide variety of techniques for a wide variety of

applications Advances in CM for variable-state machinery

• Must consider time-varying parameters to optimize operations

Techniques Limitations:

• Normal Distribution• Double Curse of Dimensionality• Sensors to measure time-varying parameters

Extensible to other mining and non-mining applications

67 - 68

Conclusions

Software architecture Recognize broader problem of Intelligent Signal Processing

• Subsumes CM, prognostics, sensor failure, etc. Design for change

• Greater breadth of marketability• Extensibility/Maintainability of Software Design

Integration at the Enterprise level• Rich standard exists to augment IREDES

Future work Take the solutions to the underground environment Validate in harsh environment

68

LOGO