DIGITAL SIGNAL PROCESSING IN

FAULT DIAGNOSIS:

PROFESSOR:
DR. ABDOLLAHI
BY:
BENYAMIN RAMEZANPOUR

1
CONTENTS:

DSP in fault
Introduction Conclusion
diagnosis

2
INTRODUCTION:
Corrective

Maintenance

Predictive Preventive

Introduction DSP in fault diagnosis conclusion

3
INTRODUCTION:

Introduction DSP in fault diagnosis conclusion

4
INTRODUCTION:

5
Introduction DSP in fault diagnosis conclusion
ROTATING MACHINERIES:

6
Introduction DSP in fault diagnosis conclusion
TECHNIQUES:

7
Introduction DSP in fault diagnosis conclusion
DSP IN FAULT DIAGNOSIS:

Data Feature Fault

Denoising
acquisition extraction diagnosis

8
Introduction DSP in fault diagnosis Conclusion
DENOISING:
CNN
Neural
networks
denoising Autoencoders
Digital signal
processing
9
Introduction DSP in fault diagnosis Conclusion
DSP FOR DENOISING:
DSP Techniques

Decomposition
Wavelet Filters
signal
10
Introduction DSP in fault diagnosis Conclusion
TRADITIONAL FEATURE EXTRACTION:

Traditional feature
extraction
• sound • SVM
• vibration • Time domain • KNN
• Frequency • MLP &…
domain
Data aquation Classification

11
Introduction DSP in fault diagnosis Conclusion
DISADVANTAGES :
➢ Weak generalization
➢ Time consuming
➢ High computation cost
➢ High signal processing knowledge needed
➢ Low Accuracy when different fault occur
simultaneously
➢ With the smallest change in the system, the whole
method should be revised
12
Introduction DSP in fault diagnosis Conclusion
INTELLIGENT FEATURE EXTRACTION:

Deep neural nets

CNNs RNNs

1D-CNN 2D-CNN LSTM

13
Introduction DSP in fault diagnosis Conclusion
2D-CNN:

14
Introduction DSP in fault diagnosis Conclusion
TIME-FREQUENCY DOMAIN:
 Gabor Transform and the Spectrogram(STFT):

15
Introduction DSP in fault diagnosis Conclusion
STFT:
Example: Quadratic Chirp

16
Introduction DSP in fault diagnosis Conclusion
UNCERTAINTY PRINCIPLES:
 In time–frequency analysis, there is a fundamental uncertainty principle that
limits the ability to simultaneously attain high resolution in both the time and
frequency domains. In the extreme limit, a time series is perfectly resolved in
time, but provides no information about frequency content, and the Fourier
transform perfectly resolves frequency content, but provides no information
about when in time these frequencies occur. The spectrogram resolves both time
and frequency information, but with lower resolution in each domain.

 This has implications for the Heisenberg uncertainty principle , as the position
and momentum wave functions are Fourier transform pairs. 17
Introduction DSP in fault diagnosis Conclusion
UNCERTAINTY PRINCIPLES:

18
Introduction DSP in fault diagnosis Conclusion
WAVELETS:
 Mother wavelet:

 Continues wavelet transform (CWT):

 inverse continuous wavelet transform (iCWT):

19
Introduction DSP in fault diagnosis Conclusion
MOTHER WAVELETS:

20
Introduction DSP in fault diagnosis Conclusion
DWT IN FAULT DIAGNOSIS:

21
Introduction DSP in fault diagnosis Conclusion
WIGNER_VILLE DISTRIBUTION SPECTRAL :

22
Introduction DSP in fault diagnosis Conclusion
WIGNER_VILLE DISTRIBUTION SPECTRAL :

23
Introduction DSP in fault diagnosis Conclusion
WIGNER_VILLE DISTRIBUTION SPECTRAL :

24
Introduction DSP in fault diagnosis Conclusion
OTHER METHODS:

Hilbert-haung Mel- Cochlea gram

transform spectrogram
25
Introduction DSP in fault diagnosis Conclusion
SIGNAL DECOMPOSITION METHODS:

VMD

EEMD

EMD

26
Introduction DSP in fault diagnosis Conclusion
WHAT DOES VMD DO?
➢ VMD decompose a signal into a number of
modes.
➢ Each mode is band limited about its center
frequency.
➢ It decompose a real valued signal into finite
number of modes.
➢ Principle components of these modes are
selected to reconstruct the original signal
27
Introduction DSP in fault diagnosis Conclusion
HOW DOES VMD WORK ?
Wiener Filter properties
• It is a low-narrow band filter whose frequency(w) is close to zero
• Narrowband means the variation across the central frequency will be very less
• The bandwidth around the central frequency (i.e the frequency which we are interested in) is kept minimum.

Forming analytical signal

• The analytical signal is the addition of the real-valued signal and its Hilbert transform
• to remove this negative frequency, we need to compromise with the complex domain(because Hilbert transform is in
the complex domain).

Frequency mixing
• 2Cos(w1*t)Cos(w2*t)=Cos((w1+w2)*t)+Cos((w1-w2)*t)

28
Introduction DSP in fault diagnosis Conclusion
CONSTRAINT OPTIMIZATION:
A constrained optimization problem contains two major aspects:
➢The objective function
➢Constraints
Note: The original optimization happens in the frequency domain. But the
energy remains the same in both domains.

29
Introduction DSP in fault diagnosis Conclusion
 CONVERT THIS CONSTRAINED PROBLEM TO AN UNCONSTRAINED ONE BY
CREATING THE LAGRANGIAN FUNCTION:

This Lagrangian function is solved using a method known as the

Alternating Direction Method of Multipliers(ADMM). It initializes
the variables with random values and searches for the optimal
values.

30
Introduction DSP in fault diagnosis Conclusion
VMD LIMITATIONS:
1.The number of modes is a parameter to fix.
2.When two frequencies are too close, VMD might
recognize both as one which is still an issue to solve.
3.The whole process of VMD is very time-consuming and
totally depends upon the initialization.

31
Introduction DSP in fault diagnosis Conclusion
VMD METHODS FOR FAULT DIAGNOSIS:

32
Introduction DSP in fault diagnosis Conclusion
GEARBOX FAULT DIAGNOSIS:

 a) Microphones b) Vibration sensors

33
Introduction DSP in fault diagnosis Conclusion
GEARBOX FAULT DIAGNOSIS:
Performing VMD on raw signal

Mode selection

Splitting the signal and performing

wavelet transform

classification
34
Introduction DSP in fault diagnosis Conclusion
VIBRATION RESULT:
 Accelerometers 3:

35
Introduction DSP in fault diagnosis Conclusion
VIBRATION RESULT:

 Accelerometers 3:

 Acc. =64%

36
Introduction DSP in fault diagnosis Conclusion
VIBRATION RESULT:
 Accelerometers 3 after VMD:

37
Introduction DSP in fault diagnosis Conclusion
VIBRATION RESULT:
 Accelerometers 3 after
VMD:

 Acc. =95.5%

38
Introduction DSP in fault diagnosis Conclusion
39