Tools for Advanced Sound & Vibration Analysis

Transcription

1

2 Tools for Advanced Sound & Vibration Ravichandran Raghavan Technical Marketing Engineer

3 Agenda NI Sound and Vibration Measurement Suite Advanced Signal Processing Algorithms Time- Quefrency and Cepstrum Wavelet AR Modeling Application Examples Bearing fault detection, dashboard motor testing, speaker testing, Prognostics 3

4 NI Sound and Vibration Measurement Suite Minimize development time with ready-to-run application examples Get started quickly with the Sound and Vibration Assistant (LabVIEW not required) Build custom data acquisition systems faster than ever with DAQ configuration XControl Avoid the expense of verification with NI ANSI- and IEC-compliant octave and sound-quality analysis Decrease test time with parallel processing 4

5 Sound and Vibration Signals Can indicate the condition or quality of machines and structures Cooling fans with faulty bearings produce louder noise You can analyze sound and vibration signals to Optimize a design Ensure production quality Monitor machine or structure conditions 5

6 Signal Characteristics Plane Frequen ncy Short time but wide band Long time but narrow band Short time & narrow band Time 6

7 Signal Processing Algorithms Overview Time Domain Domain Time- Domain Quefrency Domain (Cepstrum) Wavelet Model-Based 7

8 How to Select the Right Algorithms Order Order Order Order Time Time Time Time- Quefrency Quefrency Quefrency Quefrency Wavelet Wavelet Wavelet Wavelet Model Model Model Model Based Based Based Based t f 8

9 Select the Right Algorithms Order Order Order Order Time Time Time Time- Quefrency Quefrency Quefrency Quefrency Wavelet Wavelet Wavelet Wavelet Model Model Model Model Based Based Based Based t f 9

10 Limitations of the FFT No information about how frequencies evolve over time Not suitable for analyzing impulsive signals 10

11 Power Spectrum A power spectrum does not contain time information 11

12 Transients It is difficult to detect presence of transients in a signal by its power spectrum 12

13 Time- 13

14 Time- The short-time Fourier transform (STFT STFT) is the most popular time-frequency analysis algorithm STFT 14

15 Advantages of Time- Time-frequency representation shows how frequency components of a signal evolve over time Reversed in time domain 15

16 Application Example: Speaker Test Speakers play a log chirp for quality test 16

17 Select the Right Algorithms Order Order Order Order Time Time Time Time- Quefrency Quefrency Quefrency Quefrency Wavelet Wavelet Wavelet Wavelet Model Model Model Model Based Based Based Based t f 17

18 Quefrency 18

19 Cepstrum and Quefrency Cepstrum is the spectrum of a decibel spectrum Quefrency is the independent variable of cepstrum IFFT 19

20 Cepstrum Property The cepstrum reveals the periodicity of a spectrum A peak in the cepstrum corresponds to harmonics in power spectrum Rahmonics 10Hz harmonics A peak at 0.1s quefrency 20

21 Cepstrum Property Cont. 13Hz harmonics 10Hz harmonics 10Hz and 13Hz harmonics 1/13 = 0.078s 1/10 = 0.1s 21

22 Application Example: Bearing Fault Detection Use a cepstrum to detect a bearing fault Characteristic frequency for an outer ring fault of a bearing f f outer inner N = Bf D 1 B cos( α) 2 DC Characteristic frequency for an inner ring fault of a bearing N = Bf D 1+ B cos( α ) 2 DC D B N B : Number of balls : Ball diameter α : Ball contact angle f : Rotation frequencyd : Retainer diameter C 22

23 Bearing Fault Detection Example Geometry parameters of the bearings under test are: NB = 7 f = 30Hz D C = 70mm D B = 10mm α = 0 Characteristic frequencies of the bearings are: Outer ring fault Inner ring fault N f D B B fouter = 1 cos( ) = 3.0f = 90Hz 2 α D C N f D B B finner = 1 cos( ) = 4.0f = 120Hz 2 + α D C 23

24 Harmonics of Bearing Signals Use harmonics to detect bearing faults The outer ring fault signal has harmonics of 90Hz The inner ring fault signal has harmonics of 120Hz 24

25 Cepstrum of Bearing Signals A peak in the cepstrum means harmonics exist in the power spectrum 25

26 Select the Right Algorithms Order Order Order Order Time Time Time Time- Quefrency Quefrency Quefrency Quefrency Wavelet Wavelet Wavelet Wavelet Model Model Model Model Based Based Based Based t f 26

27 Wavelet 27

28 Wavelet vs Sine Wave Wavelet = Wave (Oscillatory ) + let (Compact) 28

29 Wavelet Transform: Look at the FFT First 29

30 Wavelet Transform 30

31 Application Example: Dashboard Motor Production Test A dashboard motor is a stepper motor that has an angle constraint Oil pressure, tachometers, and speedometers use dashboard motors Dead zone 31

32 Dashboard Motor Faults There are two kinds of faults Fault 1 Knock at turning angles Fault 2 Rub noise Good Motor Knocks Larger Knocks and Rub 32

33 Why do Wavelets Work? Knocks generate spikes and resonance Spikes and high frequency resonance result in larger wavelet coefficients 33

34 Select the Right Algorithms Order Order Order Order Time Time Time Time- Quefrency Quefrency Quefrency Quefrency Wavelet Wavelet Wavelet Wavelet Model Model Model Model Based Based Based Based t f 34

35 Model-Based 35

36 Auto-Regressive (AR) Modeling A sample in a time series can be considered as the linear combination of past samples plus error x( n) = M k= 1 a k x( n k) + e( n) Deterministic part (Model Coefficients) Stochastic part (Modeling error) 36

37 Power Spectrum Estimation The AR model spectrum has higher resolution than the FFT based spectrum 37

38 Application Example: Hard Disk Drive Production Test AR modeling errors indicate different types of HDD faults. Good Pitch Crack Zee 38

39 Application Example: Engine Knock Detection Optimized ignition timing results in a higher degree of engine efficiency Earlier ignition results in a lower engine temperature and reduced efficiency. Late ignition might result in auto-ignition and cause engine knocks, which are shock waves on the cylinder. Engine knocks are transient events and can be detected by the AR modeling error. 39

40 Engine Knock Detection - Sample 1 Constant Speed You cannot see knocks in the signal, even though you can hear them clearly Peaks indicate the existence of knocks 40

41 Engine Knock Detection - Sample 2 Run-up and Run-down 41

42 Highlights of AR Modeling Good mathematical description of stationary signal. The AR modeling error indicates transients in the signal 42

43 Select the Right Algorithms Order Order Order Order Time Time Time Time- Quefrency Quefrency Quefrency Quefrency Wavelet Wavelet Wavelet Wavelet Model Model Model Model Based Based Based Based t f 43

44 Watchdog Agent Prognostics Toolkit

45 Prognostics Confidence Value for performance degradation assessment (CV ~ 0-1) Health Radar Chart for multiple components degradation monitoring Health Map for potential issues and pattern classification Risk Radar Chart to prioritize maintenance decision 45

46 Watchdog Agent 46

47 Prognostics & Forecasting Methods Start of Performance Degradation Feature 2 Normal Behavior Current Situation Model of Failure Predicted Probability of Failure Predicted Confidence Value 0 Evolution of ARMA Prediction ARMA Prediction Feature 1 Prediction Uncertainty 47

48 Signal Processing & Feature Extraction Raw Vibrations Stationary Raw Vibrations Signal Processing & Feature Extraction Time Domain Domain Time- Wavelet/Wavelet Packet Principal Component (PCA) Non-Stationary Time- synchronous Average FFT + Envelope Wavelets scales a CWT of time synchronous signal for gearbox with broken tooth (File 105) Time (for 1 revolution) 48

49 Health Assessment Health Assessment Logistic Regression Statistical Pattern Recognition Feature Map Pattern Matching (Self-organizing Map) Neural Network Gaussian Mixture Model (GMM) Normal Behavior in Operating Condition 1 Recent Behavior in Operating Condition 1 Raw Vibration in each Operating Condition Normal Behavior Most Recent Behavior Feature Space Confidence Value (CV) 49

50 Health Diagnosis Health Diagnosis Support Vector Machine (SVM) Feature Map Pattern Matching (SOM) Bayesian Belief Network (BBN) Hidden Markov Model (HMM) Evidence-based Holo-Coefficients Normal Behavior in Operating Condition 1 Recent Behavior in Operating Condition 1 Features Gearbox Health Map Normal Gearbox C1 C1 C3 C3 C2 C2 Need to train 1 health map for each operating condition! C1 C1 C3 Recent Behavior Gear 1 Broken Tooth C4 C4 C4 C4 50

51 PHM Analytics for a Wind Farm WIND FARM INFO: CURRENT CONDITIONS: N Equipment Health Equipment Risk 51

52 WIND TURBINE INFO: CURRENT CONDITIONS: Component Health N 1 Wind Turbine Efficiency Wind Turbine Efficiency Time Wind Turbine CV History Component Risk Health CV History Time 52

53 NI LabVIEW Watchdog Agent Toolkit Signal Features Health Assessment Confidence Value Health Prediction Future Health Health Diagnosis 53