Wheel Health Monitoring Using Onboard Sensors

Transcription

1 Wheel Health Monitoring Using Onboard Sensors Brad M. Hopkins, Ph.D. Project Engineer Condition Monitoring Amsted Rail Company, Inc. 1

2 Agenda 1. Motivation 2. Overview of Methodology 3. Application: Wheel Flat Detection 4. Application: Wheel Wear Estimation 2

3 Agenda 1. Motivation 2. Overview of Methodology 3. Application: Wheel Flat Detection 4. Application: Wheel Wear Estimation 3

4 Motivation Wheel flat spots damage track Worn wheels affect truck performance and truck wear ~150 Wheel Impact Load Detectors in North America (irregular sampling intervals) Onboard monitoring provides a solution for monitoring every day, or several times per day 4

5 Agenda 1. Motivation 2. Overview of Methodology 3. Application: Wheel Flat Detection 4. Application: Wheel Wear Estimation 5

6 Methodology Raw Processed Collect Data Process Data 6

7 Onboard Data Collection Sensor Selection Low cost Low power Specs: range, bandwidth, sampling rate, resolution, etc. Sensor Placement High sensitivity to vibrations of interest Low sensitivity to vibrations that can corrupt signal 7

8 Onboard Data Collection Acceleration Time (s) 8

9 Onboard Data Collection Range used to characterize event Frequency (Hz) Range that can corrupt signal 9

10 Data Processing Preprocessing Feature Extraction Classification Machine Learning 10

11 What is Machine Learning? 1. Find features of interest that contain information related to the class or state 2. Classify the features using a set of rules and/or optimization routines 11

12 The Wide Field of Machine Learning 12

13 Example: Image Recognition Salmon or tuna? 13

14 Example: Image Recognition Features: 1. Length (# pixels) 2. Color (pixel intensity) 14

15 Example: Image Recognition Train Classifier: Salmon Tuna Color Length 15

16 Example: Image Recognition Deploy Classifier: Salmon Tuna Color Length 16

17 What the data might actually look like in the real world. Salmon Tuna Color Length 17

18 Overfitting the data is a problem! Salmon Tuna Color Length 18

19 We could linearly separate, and report confidence/ error band: Salmon Tuna Color Length 19

20 Or we could transform the features to a new space: Color (math) Abstract Feature #2 Length Abstract Feature #1 20

21 Or we could find new features that work better: Overall Height Salmon Tuna Fin Height to Length Ratio 21

22 Agenda 1. Motivation 2. Overview of Methodology 3. Application: Wheel Flat Detection 4. Application: Wheel Wear Estimation 22

23 Application: Wheel Flat Detection Accelerometer mounted to bearing adapter: 23

24 Application: Wheel Flat Detection What we hope the data will look like: Acceleration 36 inch wheel would generate a pulse 6.5 times per second traveling at 42 mph 24

25 Application: Wheel Flat Detection What the data really looks like: Acceleration 25

26 Application: Wheel Flat Detection Here is what it looks like in the frequency domain: 26

27 Application: Wheel Flat Detection Here is what it looks like in the frequency domain: Rotating frequency gets lost in noise!

28 Application: Wheel Flat Detection Original time domain signal: Acceleration 28

29 Application: Wheel Flat Detection Filtered around carrier frequency Acceleration 29

30 Application: Wheel Flat Detection Envelope signal Acceleration 30

31 Application: Wheel Flat Detection Frequency domain of envelope signal Frequency (Hz) 31

32 Application: Wheel Flat Detection Frequency domain of envelope signal Feature identified: rotating frequency Frequency (Hz) 32

33 Agenda 1. Motivation 2. Overview of Methodology 3. Application: Wheel Flat Detection 4. Application: Wheel Wear Estimation 33

34 Application: Wheel Wear Estimation Goal: determine level of wheel wear using onboard accelerometers 34

35 Application: Wheel Wear Estimation Rigid Body Modes Truck Performance Oscillations Measure Return to Service Maintenance Classify Extract features 35

36 Application: Wheel Wear Estimation The relevant information is encoded in various time domain, frequency domain, and statistical features Feature # Feature Description 1 Band Power (1) 0 5 Hz 2 Band Power (2) 7 12 Hz 3 Band Power (3) Hz 4 Magnitude at Fund. Frequency 5 Fundamental Frequency 6 Mean 7 Variance 8 Standard Deviation 9 Peak to Peak 36

37 Application: Wheel Wear Estimation The features are fed into a support vector machine (SVM) to classify An SVM seeks to maximize the distance between two classes 37

38 Application: Wheel Wear Estimation Results at 50 mph: Sample Simulation Predicted Class No Wear Med. Wear Fully worn Average Classification Accuracy of 79% Actual Class No Wear Med. Wear Full Wear

39 Application: Wheel Wear Estimation Results at 65 mph: Sample Simulation Predicted Class Average Classification Accuracy of 92% Classification accuracy using peak to peak and standard deviation: 55% Actual Class No Wear Med. Wear Full Wear No Wear Med. Wear Fully worn

40 Conclusions Sensor selection, placement, and data acquisition are not trivial Finding the features of interest requires expertise Machine learning provides a powerful way to classify features for health monitoring 40

41 Questions? 41