CHAPTER 1 Introduction and Literature Review
1.15 Machine Learning Techniques
1.15.5 Conclusions – Machine Learning
using discrete wavelet transform (DWT) to three levels using Daubechies wavelets.
Classification accuracy of 96.67% was obtained. Tabar et al.(2017) developed an algorithm based on SVMs to identify different intensities of cavitation in a CP. Features used for the classification were selected based on their sensitivity towards the fault using compensation distance evaluation approach. Kumar and Kumar (2017) used the GA optimized SVM algorithm for the fault detection in CPs. The faults considered were the clogged impeller, broken impeller, bearing inner race and outer race defects. The features that are sensitive to the faults are acquired from the raw vibration signals and scale marginal integration graph (from CWT). An accuracy of 96.66% was found in the identification of various CP faults.
Table 1.8: Comparison of researches on application of machine learning to fault diagnosis of CPs
References Faults considered Signal type Domain of signal
Machine learning algorithm used
Data sampling rate/
sample length
Features
considered Multiclass Accuracy obtained
Perovic et al., 2001
Discharge blockage faults, impeller defects and both discharge blockage and impeller defects
Motor current FFT Fuzzy logic n/a
Slip, current amplitude, amplitude of the shaft speed component, accumulative noise
Yes 83.7%
Zouari et al., 2004
Loosening of front or rate CP attachments,
misalignment, cavitation, partial flow, and air injection
Vibration Time- domain
NN, neuro-
fuzzy network 50kHz Fischer, PCA Yes n/a
Yuan and Chu, 2006
Gear damage, structure resonance, rotor radial touch friction, shaft crack, bearing damage, body join looseness, bearing looseness, rotor part looseness, pressure pulse, cavitation, vane rupture, rotor
eccentricity, shaft bend
Vibration FFT SVM n/a Relative spectral
amplitude, PCA Yes Precession ratio >95%
Azadeh et al., 2010
Seal damage,
misalignment, bent shaft, cavitation, bearing damage, plugged seal, dirty seal, rubbing, eccentric shaft
Flow rate, discharge pressure, NPSHR, BHP (Brake Horse Power), efficiency, vibration and temperature
n/a Fuzzy logic n/a n/a Yes n/a
Table 1.8: Comparison of researches on application of machine learning to fault diagnosis of CPs (contd.)
References Faults considered Signal type
Domain of signal
Machine learning algorithm used
Data sampling rate/
sample length
Features considered Multiclass Accuracy obtained
Sakthivel et al., 2010b
Bearing faults, seal defects, impeller defects, impeller and bearing faults, cavitation
Vibration Time- domain
decision tree fuzzy, rough set fuzzy
24kHz/
1024
Mean, standard error, median, standard deviation, sample variance,
kurtosis, skewness, range, minimum, maximum, and sum
Yes 99.33%
Saberi et al., 2011 and
n/a Vibration Time-
domain
SVM and ANN n/a n/a No 100%
Zhao et al., 2010
Impeller leading edge and trailing edge damage
Vibration FFT Probabilistic neural network, k nearest
neighbours, SVM
9kHz Noise based features Yes 99.56%
Sakthivel et al., 2010a
Bearing faults, seal defects, impeller defects, impeller and bearing faults, cavitation
Vibration Time domain
C4.5 decision tree
24kHz/
1024
Standard error, standard deviation, variance, kurtosis, skewness, range, minimum, maximum, sum
Yes 100%
Nasiri et al., 2011
Cavitation with varying severities
Vibration Discrete Fourier transform (DFT)
NN 10kHz Kurtosis, crest factor Yes n/a
Muralidharan and
Sugumaran, 2012
Bearing faults, Impeller faults, both bearing and impeller faults, cavitation
Vibration Wavelets (DWT)
Naïve Bayes, Bayes Net
24kHz/
1024
DWT coefficients Yes 100%
Table 1.8: Comparison of researches on application of machine learning to fault diagnosis of CPs (contd.)
References Faults considered Signal type
Domain of signal
Machine learning algorithm used
Data sampling rate/
sample length
Features considered Multiclass Accuracy obtained
Yunlong and Peng, 2012
Mass unbalance, rotor misalignment, foundation looseness
Vibration FFT LSSVM 800Hz/
4096
EMD, IMF Yes 92.6%
Farokhzad et al., 2012
Impeller defects, seal defects and cavitation
Vibration FFT ANN n/a Mean, standard deviation, variance, skewness, kurtosis, crest factor, slippage, root mean square
Yes 100%
Sakthivel et al., 2012b
Bearing faults, seal defects, impeller defects, impeller and bearing faults, cavitation
Vibration Time- domain
Roughset – fuzzy
24kHz/
1024
mean, standard error, median, standard deviation, sample variance,
kurtosis, skewness, range, minimum, maximum, and sum
Yes 97.5%
Sakthivel et al., 2012a
Bearing faults, seal defects, impeller defects, impeller and bearing faults, cavitation
Vibration Time domain, Wavelets
Gene expression programming, SVM, wavelets – gene expression programming
24kHz/
1024
mean, standard error, median, standard deviation, sample variance, kurtosis, skewness, range, minimum, maximum, and sum
Yes 99.93%
Azadeh et al., 2013
n/a Flow rate,
discharge pressure, suction pressure, vibration, velocity
n/a SVM-GA,
SVM-PSO, ANN
n/a n/a No 93.3%
Table 1.8: Comparison of researches on application of machine learning to fault diagnosis of CPs (contd.)
References Faults considered Signal type
Domain of signal
Machine learning algorithm used
Data sampling rate/
sample length
Features considered Multiclass Accuracy obtained
Muralidharan and
Sugumaran, 2013c, 2013a
Bearing faults, Impeller faults, both bearing and impeller faults, cavitation
Vibration Wavelets (DWT)
Decision tree, J48
24kHz/
1024
DWT coefficients Yes 99.84%
Muralidharan and
Sugumaran, 2013b
Bearing faults, Impeller faults, both bearing and impeller faults, cavitation
Vibration Wavelets (DWT)
Roughset fuzzy
24kHz/
1024
DWT coefficients Yes 99.84%
Muralidharan et al., 2014
Same as above Vibration Wavelets (CWT)
SVM 24kHz/1024 CWT coefficients Yes 99.84%
Xue et al., 2014
Varying severities of cavitation, unbalance and misalignment
Vibration Frequency – domain
SVM 50kHz NSP Yes 0.9649/1
Sakthivel et al., 2014
Bearing faults, seal defects, impeller defects, impeller and bearing faults, cavitation
Vibration Time- domain
decision tree, Bayes Net, Naïve Bayes and kNN classifiers
PCA, LLE, LTSA, HLLE, Isomap, kPCA, Laplacian Eigenmap, Manifold chart, MVU
Yes 100%
Zhao et al., 2016
Inner race wear, outer race wear, roller wear, and impeller wear
Vibration Time- domain
Deep learning, softmax regression
10.24 kHz n/a Yes >99%
Table 1.8: Comparison of researches on application of machine learning to fault diagnosis of CPs (contd.)
References Faults considered Signal type
Domain of signal
Machine learning algorithm used
Data sampling rate/
sample length
Features considered Multiclass Accuracy obtained
Azizi et al., 2017
No cavitation, limited cavitation and developed cavitation
Vibration Time-domain GRNN 16 kHz/
n/a
RMS, mean, variance, standard deviation, skewness, kurtosis, crest factor, upper bound, lower bound, range, shape factor, margin factor, impulse factor, entropy, zero crossing rate
Yes 100%
Bordoloi and Tiwari, 2017
Varying severities of suction blockages and cavitation
Vibration Time-domain SVM-GA and SVM- ABCA
20kHz/
2000
Standard deviation, skewness, kurtosis
Yes 94.6%
Ebrahimi and Javidan, 2017
Impeller faults and seal defects
Vibration Wavelets (DWT)
SVM n/a Mean, standard deviation,
sample variance, RMS, crest factor, skewness, slippage, kurtosis, fifth and sixth central moments
Yes 96.67%
Shervani- Tabar et al., 2017
5 levels of cavitation Vibration Time-domain SVM and ANN
n/a Mean, peak, root mean square, standard deviation, kurtosis, skewness, kurtosis, crest factor, clearance factor, shape factor, impulse factor
Yes 97%
Kumar and Kumar, 2017
Broken impeller, clogged impeller, bearing inner race defects and bearing outer race defects
Vibration Time-domain and CWT
SVM 70kHz/
7000
Kurtosis, skewness, mean, RMS, variance, peak, impulse factor, shape factor, crest factor, scale corresponding to highest energy, ration of peak energy to total energy
Yes 96.66%
Despite so many advantages offered by machine learning approaches, there are a few drawbacks, and these are,
In the case of supervised learning strategies, the pre-requisite is the availability of labeled fault data. However, in the most of industries many times either they do not have any data available, or even if they have the data, they do not know which fault it belongs to. In such circumstances, visual inspection techniques serve to be better options than supervised machine learning. Otherwise, unsupervised machine learning techniques may be used, like, Zhao et al. (2016) presented an unsupervised self- learning fault diagnosis method using deep learning to classify various CP faults including bearing faults (i.e., the inner and outer race wear, roller wear) and impeller wear. To demonstrate the robustness of the feature recalculation and effective high- level feature extraction, the noise was introduced in the input data of every auto- encoder. Classification accuracy of over 99% was found.
The classifier designed for one system may not apply to another system, which is of different geometry.
When designing a highly sophisticated machine learning algorithm, it may be difficult to maintain and debug it.
The developed algorithm may be sensitive to the noise.
The algorithm may be operating speed dependent
The data sampling resolution seems to have a significant effect on the algorithms classification performance. From Table 1.8, it can be seen that there has not been a systematic choice of the sampling rate of the sensor signals.