Archives of Computational Methods in Engineering

https://doi.org/10.1007/s11831-020-09446-w

ORIGINAL PAPER

Review on Machine Learning Algorithm Based Fault Detection

in Induction Motors
Prashant Kumar1 · Ananda Shankar Hati1

Received: 20 September 2019 / Accepted: 30 May 2020

© CIMNE, Barcelona, Spain 2020

Abstract
Fault detection prior to their occurrence or complete shut-down in induction motor is essential for the industries. The fault
detection based on condition monitoring techniques and application of machine learning have tremendous potential. The
power of machine learning can be harnessed and optimally used for fault detection. The faults especially in induction motor
needs to be addressed at a proper time for avoiding losses. Machine learning algorithm applications in the domain of fault
detection provides a reliable and effective solution for preventive maintenance. This paper presents a review of the machine
learning algorithm applications in fault detection in induction motors. This paper also presents the future prospects and
challenges for an efficient machine learning based fault detection systems.

1 Introduction powerful tools for classification problems and classification

problems [99]. Due to this reason, Artificial Intelligence
Early fault detection in induction motor provides the lever- (AI) based on machine learning algorithm has attracted
age of minimum downtime and maximum production. The researchers in the domain of condition monitoring.
study of maintenance of induction motors and its failure The fault detection in IMs is divided into multiple stages
analysis is a vast area of research for the researchers for like data acquisition, data processing, feature extraction and
many years [4]. Traditionally, the fault detection for the IM implementation of ML algorithms for fault recognition [34,
was dependent on current and vibration [67, 70]. The simple 103]. Most AI-based fault detection system requires the fea-
methods which were earlier used were overcurrent, overvolt- tures for developing the input vector for the ML algorithms.
age and earth-fault. However, with the modernisation and The data is preprocessed by feature extraction algorithms
advancement in techniques, the combination of traditional [103] to develop an input feature vector for ML algorithm
and modern approach provides an efficient technique for for classification of faults. Sometimes, independent compo-
the fault detection in the induction motors. The predictive nent analysis (ICA) is used to convert the high dimensional
maintenance based on the condition-based monitoring of the vector to a low dimensional vector for easier analysis and
induction motor is the dire need of the modern industries recognition [10, 48]. The feature vector is used as an input
for minimal downtime [15, 27, 107]. The preamble of fault vector for developing AI enabled fault identification system.
detection has four components, namely: (a) Identifying fault The techniques like Genetic algorithm, Particle swarm opti-
location; (b) Determining faulty parts; (c) learning incipient mization, convex optimization and statistical learning have
failure and their causes; (d) predicting the pattern of faults. been used for feature selection and extraction [13, 53, 75,
The fault detection can be treated as classification problem 86]. For developing AI based system, ML algorithms like
and pattern recognition problem too. ML algorithms are k-Nearest Neighbor (k-NN) [11], Artificial Neural Network
(ANN) [26, 87, 89], Support Vector Machines (SVM) [56,
59], Decision trees [52], Bayesian Classifier [3], random
* Ananda Shankar Hati
anandashati@iitism.ac.in forest [68] and deep learning [33] techniques have been effi-
ciently used. In recent times, deep learning have also been
Prashant Kumar
prashant88344@gmail.com explored by researchers for fault detection [49].
This paper has been divided into different sections. Sec-
1
Department of Mining Machinery Engineering, Indian tion 2 enumerates the faults and their causes in the induction
Institute of Technology (Indian School of Mines), Dhanbad, motors. Section 3 enlists the machine learning algorithms
India

13
Vol.:(0123456789)
 P. Kumar, A. S. Hati

technique for the induction motors. Section 4 includes the

( )
frot bd cos𝛽
machine learning algorithm based condition monitoring and ftd = 1− (4)
2 pd
fault diagnosis. Section 5 presents the discussion, limitations
and future scopes of machine learning based fault detection where frot is the rotational frequency, Nb is the number of
algorithms. Finally, Sect. 6 concludes the work. balls, bd is the ball diameter, pd is the pitch diameter, for is
the vibration frequency for outer raceway fault, fir is the
vibration frequency for inner raceway fault, fdb is the vibra-
2 Faults and their Causes
tion frequency for the ball defect, ftd is the vibration fre-
quency for the train defect and 𝛽 is the contact angle. Schoen
The faults in the IM can be segregated as a mechanical faults
et. al. [81] have revealed that vibration frequency component
and electrical faults [1, 42, 43, 91]. Mechanical faults are cat-
is noticeable in the current spectrum. Current spectrum has
egorised as bearing fault, broken rotor bar fault (BRB) and
vibration frequency component and is given by:
eccentricity. Electrical faults are stator winding faults and rotor
winding faults. These fault in the IM have several symptoms fbng = |fr ± m ⋅ fv | (5)
which includes: (a) Unbalance magnetic pull; (b) Excessive
heating; (c) Increased losses; (d) Efficiency reduction; (e) where fbng is the vibration component frequencies in the
Decrease in average torque; (f) Enhanced torque pulsation. The current spectrum, m = 1, 2, 3, ….
causes of faults can vary from the operating conditions and
ambient temperature. Some of these causes have been enlisted,
namely: (a) Overloading; (b) Overspeed; (c) Manufacturing 2.2 Broken Rotor Bar
defect; (d) Dirt, debris and corrosion; (e) Components failure;
(f) Overheating in winding and bearing; (g) Excessive dielec- There has been a lot of changes in the design of the sta-
tric stress; (h) Shock loads. tor core, stator windings and construction. However, rotor
bar has been left out and no significant changes have been
2.1 Bearing Fault introduced in it. The rotor failure accounts for 5–10% of
total induction motor faults [7, 44, 45, 63]. The BRB related
Bearing is among the basic components of the IM which fault is pre-dominant in medium or high rating motor over
helps in separating the stationary part with the moving part. the small rating motors as the starting torque demand and
IM mostly consist of rolling or ball element bearings. The thermal stress is high. Due to a BRB, it produces an enriched
bearing has two rings, namely outer ring and inner ring. A set field in the vicinity of the faulty region in IM. It is because
of balls or rolling element rotates between the two rings. The of local demagnetising slip frequency and induced current
bearing fault can occur due to fatigue under balanced condi- in the damaged rotor slots. Flux density is more in ampli-
tion too. The fault can take place due to wear and tear and due tude in the vicinity of faulty area. The result shows that for
to ageing, shock loads, overheating, imbalances and improper an IM with 40 rotor bars, damage of one broken bar causes
lubrication. Bearing fault has a majority stake (40–50%) of degradation of 2–4% in the steady state torque performance
all IM failures [46]. The ball bearing fault is characterised as whereas, for three and five broken bars, it is between 10
outer raceway defect, inner raceway defect, ball defect and and 15% [18]. Few possible reasons for damaged broken bar
train defect [79]. Each defect has its own characteristic vibra- have been enlisted below [63]:
tion frequency depending on the construction and mechanical
dimension of the bearing and is given as follows [63, 93]: 1. Stresses like magnetic stress and thermal stress caused
( ) due to electromagnetic forces, unbalanced magnetic
Nb frot bd cos𝛽 pull (UMP), noises, thermal overload, excessive losses,
for = 1− (1)
2 pd sparking;
2. Dynamic stresses;
Nb frot
(
bd cos𝛽
) 3. Mechanical stresses;
fir = 1+ (2)
2 pd
If a BRB exists, flow of current in that bar will be obstructed.
( As a result, a field will be unavailable near the faulty bar
[ ]2 )
pd frot b cos𝛽 in the rotor. Due to this imbalance, UMP is created which
fdb = 1− d (3)
2bd pd rotates at one time the rotational speed. It modulates at a
frequency which is equal to multiple of slip frequency and
number of poles. This frequency is called as pole pass fre-
quency. This causes increased magnitude in the vibration

13
Review on Machine Learning Algorithm Based Fault Detection in Induction Motors

spectrum and occurs at rotational frequency and its side- eccentricity and mixed eccentricity. In static eccentricity,
bands are given by: there is a persistent offset between centreline of stator and
rotor. Dynamic eccentricity caused from mutable offset
fbrb = frot ± fp (6) between the centreline of stator and rotor. Eccentricity may
be caused due to bearing wear and tear, bent rotor shaft,
fp = (fsyn − fr ) ⋅ P (7) mechanical resonance etc. An air-gap eccentricity of up to
10% is permissible. Static and dynamic eccentricity gener-
where fp is the pole pass frequency, P is the number of poles ally co-occur.
and fsyn is synchronous speed [21]. These side-bands are
predominantly present in higher harmonics of rotational
speed ( 2frot , 3frot,...). The broken rotor bar detection using 3 Machine Learning Algorithms Theory
Eqs. 6 and 7 have been used in several literatures and has
high practical implementation [60]. The low value of slip or Machine learning algorithm have become renowned in fault
slight loading causes the rotational frequency frot to be close detection system due to their reliability, adaptability and
to the synchronous speed and side-bands is closer to the robustness. The application of these algorithms have helped
principle slot harmonics makes it difficult to identify in the in developing an efficient system. These systems do not need
vibration signal spectrum. Broken rotor bar (BRB) fault is prior knowledge for their operation [56]. The algorithms
among the most common fault occurring in the underground like Support Vector Machines (SVM), k-Nearest Neighbors
mines. Due to BRB fault, side-bands are visible around the (k-NN), Artificial Neural Network (ANN), Decision trees,
principle slot harmonics in the current spectrum and is given Bayesian Classifier, random forest and Convolutional Neural
by Gyftakis et al. [24]. Network (CNN) are commonly used in fault prognosis and
flsb = f (1 − 2ks)Hz (8) diagnostic.

3.1 Support Vector Machine

frsb = f (1 + 2ks)Hz (9)
where k = 1 for principle slot harmonic, f is the supply fre- SVM is a powerful tool for classification problems. Vladimir
quency, flsb is the left side-band frequency and frsb is the Vapnik introduced SVM in 1994. SVM algorithm constructs
right side-band frequency. a hyperplane or set of hyperplane in an infinite dimensional
space. The hyperplane for which the margin is maximum is
the optimal hyperplane. Thus, SVM tries to make a deci-
2.3 Stator Faults sion boundary in such a way that the separation between the
two classes is as wide as possible. For small database cases
These are primarily associated to winding faults and insula- and machine fault detection, SVM is an effective tool for
tion failure. These faults may be line-to-ground or line to classifications. SVM is popular in the area of machine fault
line fault. It accounts for 30–40% of all reported IM failures detection due to its attractive features and good empirical
[45]. Factors leading to stator insulation failures are: performance. Binary classification is a basic model of SVM
[98]. It is basically two class problem and hyperplane is
1. electrical discharges; drawn to separate two classes optimally in the space. This
2. leakage in the cooling system; hyperplane is used to separate new data once given to the
3. heating in winding leading to a higher temperature; algorithm. The basic idea of SVM is shown in Fig. 1.
4. loose bracing for end winding [63]. There are two classes represented by positive (+) sign
and negative (−) sign. It draws a boundary (or hyperplane)
2.4 Eccentricity to separate two classes. The boundary is done in such a way
that the distance is maximised between the boundary and
The non-uniform air-gap distribution between the stator and nearest data points of both classes. The margin is kept as
rotor leads to eccentricity [9, 28, 93]. The eccentricity at an bargain between the margin level and error. There can be
incipient stage doesn’t affect much but with large eccen- number of hyperplane drawn for separating two classes
tricities, it can lead to non-uniform flux distribution, result- (as shown in Fig. 1), however, the hyperplane having the
ing unbalanced magnetic pull that can cause stator to rotor maximum distance from the nearest data point on the either
rub. Furthermore, varying inductances due to eccentricity side of classes will be chosen for the best fit and accuracy.
cause an unbalanced magnetic flux in the air gap. It cre- Figure 2 shows the hyperplane which segregate the two
ates fault frequencies in the line current. Airgap eccentricity classes optimally i.e. positive (+) sign and negative sign
is basically of three types i.e. static eccentricity, dynamic (−). Finally, SVM classification problem is solved by an

13
 P. Kumar, A. S. Hati

dimension of classified vector does not influence the perfor-

mance of SVM and it differentiate it from conventional clas-
sifiers. Generally this method provides better generalization
capabilities with respect to methods based on empirical risk
minimization. This is advantageous in fault classification
problem as number of features need not be limited.
Generally, these classifiers have good classification prop-
erties as compared to conventional classifiers. The support
vector contains the information needed for defining the clas-
sifier and when support vectors are properly selected, rest
feature is not of much interest. It can be used for solving
linear and non-linear classification problems. For non-linear
classification, kernel functions like Gaussian, quadratic and
cubic are required.

3.2 k‑Nearest Neighbors

k-Nearest Neighbors is among the simple machine learn-

ing methods used for the classification problems. k-NN is a
non parametric and instance based learning in which simi-
Fig. 1 Positive and negative class
lar items are placed together. It belongs to the supervised
learning category and is significantly used in pattern rec-
ognition problem and data mining. In this algorithm, the
specimen or samples within the class will be near to the
samples with similar features [14]. In this algorithm, all
the samples belongs to points in n-dimensional space ℜn .
Euclidean distance is used for determining the samples with
similarities. Let a random sample x given by the feature vec-
tor a1 (x), a2 (x), … an (x) shows the distribution of the sample
x. The distance between two samples xi and xj is defined to
be d(xi , xj ),where
√∑
d(xi , xj ) = r = 1n(ar (xi ) − ar (xj ))2 (10)

The target function in the k-NN can be real valued or dis-

crete valued. For each training example (x, f(x)), it is added
to the list of training examples. For a query instance xq to
be classified. Let x1 , x2 , … xk denotes the k instances from
training samples that are nearest to xq , then
k
̂f (xq ) ← arg maxv∈V
∑
𝛿(v, f (xi ) (11)
i=1
Fig. 2 Optimal separating hyperplane
where 𝛿(a, b) = 1 if a = b and where 𝛿(a, b) = 0 otherwise.

optimization technique so that margin is kept maximum 3.3 Artificial Neural Network (ANN)
and error is minimum. Lastly, SVM models is developed
which can predict the class of new data. The traditional ANN is among the handiest tool for non-linear analysis.
approach in machine learning is to minimize the error on ANN is an information paradigm that is motivated by the
the training data set and it is called empirical risk minimiza- way the biological nervous system works and processes
tion. However, support vector machines (SVM) is based on information with the help of the brain. The most common
structural risk minimization [40, 92]. SVM can handle large form of ANN has three layers: input layer, hidden layer
data sets efficiently. The training of SVM is done so that the and output layer. ANN can be represented by a set of links

13
Review on Machine Learning Algorithm Based Fault Detection in Induction Motors

between input layer and hidden layer connected with the individual tree in the random forest tells about a class pre-
multiple of weight and associated bias. ANN has connected diction. The class with maximum votes are selected as our
units called artificial neurons which emulates the neurons of prediction model. The low correlation between the model is
the biological brain. The connected units transmit a signals the key. The random forest provides an optimum classifica-
like the synapses in biological neurons. A simple structure tion performance because the tree protect each other from
of ANN in shown in Fig. 3. their individual errors (as long as they don’t constantly all
ANN takes an input x1 , x2 , x3 and output is given as: err in the same direction). Decision trees are very sensitive
( ) to changes to the data they are trained on. Even the small
T
∑ changes to the training set results in different tree structure.
y = A(W x) = A Wi xi + b (12)
i=1
So, the performance also varies [29].

where A is an activation function (can be sigmoid, logsig,

purelin etc.), W is a weight matrix, b is a bias (scalor). Ini- 3.5 Bayesian Classifier
tially, weights (W) are initialised randomly and then opti-
mized by an iterative training procedure, based on relation- Bayesian classifier is among the most practical learning
ship between input-output patterns. methods used in activity recognition and pattern recognition.
It belongs to the category of probabilistic classifiers. These
3.4 Decision Trees and Random Decision forest classifiers are based on the Baye’s theorem and their work-
ing lies in basic assumption that features are independent of
A decision tree is a type of machine learning algorithm each other. Consider a training set S = (I1 , O1 ), (I2 , O2 ), … , (
which has a pyramid like structure or we can say an inverse IN , ON ) with label o, oh = m1 , m2 , … , mG , h = 1, 2, … N,,
tree like structure. It is used to solve classification and assuming Zt possible values for it , t = 1, 2, … , n ; and G
regression problems. It emulates tree like structure. Each possible vales of O. These classifiers learns the probabil-
non-leaf node represents an input feature. The arcs branch- ity distribution between input and output with the help of
ing from internal node (non-leaf nodes) are marked with an conditional probability distribution based on assumption of
input feature. These arcs are mapped with possible values conditional independency:
of the target or output feature. It might also lead to a sub-
P(I = i|O = od ) = P(I (1) = i(1) , … , I (n) = i(n) |O = md ) (13)
ordinate decision node on a different input feature. The top
node of a tree is called as root node. Each leaf is marked n
with a class or a probability distribution over the class. It ∏
P(I = i|O = od ) = P(I (t = i(t) ) d = 1, 2, … , G (14)
signifies that dataset has been classified into an unique class t=1
or a probability distribution.
Random forest is a combination of multiple decision trees Based on the learning model, the output value o with the
for the decision making. It has multitude of decision trees highest probability for given input i can be calculated by
at a training time. It outputs the the mode of the class of the Baye’s Theorem:
individual tree. It also resolve the problem of over-fitting P(I = i�O = md )P(O = md
of decision trees. It is an ensemble learning method. Each P(O = md �I = i) = ∑ (15)
dP(I = i�O = md )P(O = md )

and
∏
O = arg maxcj P(O = md ) P(I (t) = i(t) |O = md ) (16)
t

3.6 Deep Learning

Deep learning (DL) belongs to the family of machine learn-

ing and is based on ANN. DL has deep architectures for
solving the complex problems. There are many deep learn-
ing architectures namely, deep neural networks (DNN), con-
volutional neural networks (CNN), deep belief networks and
recurrent neural networks (RNN). DL also falls under the
Fig. 3 Structure of simple ANN subclass of machine learning. DL uses deep architectures

13
 P. Kumar, A. S. Hati

and multiple layers to extract features from the raw input have used poly-coherent composition spectrum (p-CCS)
signals [5]. along with data fusion technique for fault diagnosis [105].
Deep learning refers to the deep and multiple layers for
handling the complex structures. DL has substantial credit 4.2 Support Vector Machine Based Fault Detection
assignment path (CAP) which is basically about finding
weight that makes the network to exhibit desired behaviour. SVM is among the powerful tool for the classification of
CAP is the sequence of transformation from input to out- faults in IMs. The selection of kernel functions directly
put. In a feed-forward NN, depth of CAP’s is equal to the affects the performance of SVM classifiers. For developing
network and hidden layers plus one for output as it is para- an efficient fault detection system, several structures of SVM
metrised. In the network where signals might propagate mul- has been developed in recent times. The particle swarm opti-
tiple times through a layer, potentially CAP depth is incal- mization algorithm and least square-SVM has been proposed
culable. Deep learning models have CAP > 2 for efficient to diagnose the faults of the motor bearing [16]. Authors
extraction of features and better performance than shallow have used higher order statistics technique along with the
structures. The multiple layers in DL based networks helps SVM for condition monitoring and fault diagnosis in the
in learning features eloquently [80]. rotating machines [73]. Acceleration signal based full spec-
trum are being used as a feature input vector for SVM for
analysing rotor faults [20]. A hybrid intelligent technique
based on wavelet transform, principal component analysis
4 Machine Learning Algorithm Based and twin SVM has been presented for classifying multiple
Condition Monitoring of IMs faults in rotating machines [58]. An empirical mode decom-
position (EMD) along with the weighted least square SVM
4.1 Pre‑processing of Data has been used for improving the performance of fault detec-
tion system [57]. High-frequency intermittent components
The application of machine learning algorithm for condition and non-Gaussian noises have been suppressed with the help
monitoring of IMs require signal processing of raw data. of weighted least square SVM. The c-SVC and mu-SVC
The implementation of AI technique in real world requires model of SVM is presented with four kernel functions for
the feature extraction and feature vector as an input vector faults classification using statistical features calculated from
[55]. Analysis like frequency-domain analysis, time-domain vibration signal under healthy and faulty condition [74]. A
analysis, time frequency analysis has been done for develop- methodology based on pseudo Wigner–Ville Distribution
ing the pattern for fault recognition. Statistical features like (PWVD) and SVM in combination has been demonstrated
mean, standard deviation, kurtosis and skewness is estimated for time-domain feature extraction and fault identification
for generating the feature vector. Frequency domain analysis with acoustic emission signals [51].
like Fast Fourier Transform (FFT) and bi-spectrum analysis The practical application and need of industries motivated
are used for analysis of the signals. Time-frequency analysis the researchers for improved and more efficient SVM for
like wavelet analysis [101], short-time fourier transform [2], fault prognosis and diagnosis. Authors have used multiclass
wavelet packet transformation [41], Hilert–Huang tranform SVM and features were extracted for SVM based on binary
[66] and sparse decomposition [102] are used for feature particle swarm optimization algorithm from the vibration
extraction. The availability of highly efficient computational data [110]. An empirical mode decomposition and multi-
machines have contributed a lot in applying machine learn- class transductive SVM are used in combination to diag-
ing algorithms in fault detection. The concepts like inter- nose the faults in gear reducer [85]. Feature extraction based
net of things, remote diagnostics, wireless monitoring have on statistical parameters from a wavelet packet transform
changed the way in which fault detection is done. Due to this (WPT) has been used along with distance evaluation tech-
development, there is a large amount of data accumulation nique (DET) and a support vector regression (SVR)-based
for developing more accurate fault detection system [49]. generic multi-class solver for fault detection [84]. SVM
The availability of humongous amount of data is also pos- based on modified shuffled frog-leaping algorithm (MSFLA)
ing a challenge to data pre-processing task. The data fusion has been presented for an efficient fault classification [104].
and dimensionality reduction has become an integral part Features for modified SVM were extracted from wavelet
of fault detection system. In [8], authors have used princi- energy time spectrum and power spectrum of maximum
pal component analysis for dimensionality reduction of data wavelet energy level. A comparison is done based on dif-
obtained from fast fourier transform. In [82], preprocessing ferent machine learning algorithms to SVM for developing
of vibration signal has been done using principal compo- efficient system [25]. A comparative study based on ANN
nent analysis for reducing the dimension of feature vector and SVM for detecting bearing faults based on statistical
without compromising the performance of system. Authors features [37].

13
Review on Machine Learning Algorithm Based Fault Detection in Induction Motors

4.3 k‑Nearest Neighbors based fault detection method for online rotor bar detection using feature repre-
sentation as an input vector to ANN network [71]. A fea-
k-NN is an instance based learning and has worked well in ture extraction based on WPD and EMD has been used as
fault classification problems. Authors have presented a bear- an input to backpropagation neural network for early fault
ing health monitoring based on feature extraction from spec- detection in rotating machinery [6]. Authors have analysed
tral kurtosis and cross relation along with principal com- the performance of ANN and SVM in faulty rotor bearing
ponent analysis and k-NN [90]. An advanced k-NN based systems [38]. It also showed that ANN performance was
fault severity detection model has been developed using better than SVM classifier. A RBF-NN based electrical and
redundant statistical features estimated from wavelet packet mechanical fault detection has been proposed for induction
transform [94]. k-NN algorithm based on feature extraction machine fault diagnosis [100]. Feature extraction was done
from vibration signal has been presented for unbalanced from power spectra of vibration signal. A high-dimensional
fault detection [61]. Authors investigated multiscale energy hybrid model for machine fault diagnosis based on multi-
analysis of discrete wavelet transformation to generate low ple feature selection and RBF neural network is presented
dimensional feature vector for k-NN based classifiers for [106]. Authors have presented an intelligent fault diagnosis
machine fault diagnosis [36]. A condition monitoring system method for bearing faults based on probabilistic theory and
based on k-NN classifiers has been developed using load and a fuzzy neural networks using frequency features as an input
acceleration indices as feature from vibration signal along vector [95]. The online fault detection in induction motors
with principal component analysis for dimensionality reduc- is developed by combining the discrete wavelet transform
tion [72]. A bearing fault diagnosis system has been demon- (DWT), feature extraction, genetic algorithm (GA), and
strated using low dimensional feature extracted by Hilbert neural network (ANN) techniques [26]. Samanta et. al.
Huang Transform (HHT) as an input vector to the k-NN clas- [76] have used feature extraction from vibration signal with
sifier [65]. A multi-fault detection system based on feature selection of optimal feature by genetic algorithm as an input
extraction from acoustic signal and using it as an input to to ANN, MLP and RBF network for bearing fault detection.
Nearest Neighbour [23]. Lein and Zuo presented a weighted The performance of three networks has been compared too.
k-NN for identifying the severity of gear crack based on A recurrent neural network has been used for bearing fault
feature extraction by two-stage feature selection and weight- diagnosis in the form of an autoencoder [54]. A fault detec-
ing technique (TFSWT) via Euclidean distance evaluation tion for rotating machinery has been developed with the help
technique (EDET) [50]. A k-NN based fault detection is of multi-sensor data fusion and bottleneck layer optimized
presented by developing isolation index by decomposing CNN [96].
the k-NN distance [109]. The selection of k is an important
parameter in an efficient k-NN algorithm implementation 4.5 Decision Trees and Random Decision Forest
and it should be wisely selected. Based Fault Detection

4.4 Artificial Neural Network (ANN) Based Fault Sun et. al. [88] proposed the fault diagnosis system based on
Detection decision trees using principal component analysis as way to
reduce features after feature extraction. It has been shown
ANN is a powerful tool for pattern recognition in several that decision tree and PCA based system has good accuracy
applications. It is among the most commonly used classi- and reduced training time than backpropagation neural net-
fiers in developing an intelligent fault detection systems. A work systems. Decision tree algorithm is used for faults clas-
fault detection system based on ANN and radial basis func- sification and feature extraction has been done from wavelet
tion (RBF) classifier is presented by feature extraction from transformations [78]. In most of the cases, decision trees are
multiresolution analysis of vibration signals for automatic used along with other classifiers for an efficient fault detec-
detection of cracks in rotors [12]. Multi-layer perceptron is a tion system. Standalone use of decision trees are less in fault
kind of ANN consists of only forward connections to units in detection systems.
subsequent layers [31]. A feature vector from vibration sig-
nal is developed by calculating standard deviation of wavelet 4.6 Naive Bayes Classifier Based Fault Detection
packet coefficients which is used an input vector for ANN
classifiers [69]. An artificial neural networks based robust This method is suitable only for independent feature vector
fault detection system is developed using outer bounding i.e. mostly the statistical features. Authors have proposed a
ellipsoid algorithm for estimating model uncertainty [62]. model to extract the frequency based features from wavelet
A multilayer perceptron neural network has been developed packet transform [19]. This feature is used as an input to
using as statistical features derived from wavelet transform Naive Bayes classifier for bearing fault detection in induc-
as input to the network [77]. Authors have presented a tion motors. A Naive Bayes classifier is used along with

13
 P. Kumar, A. S. Hati

frequency spectrum estimation for optimising the shaft volt- been developed by using combination of Stochastic gradient
age condition monitoring [17]. Authors have presented a descent, restricted Boltzmann machines and particle swarm
broken rotor bar detection based analysis and results were optimization [83].
compared from Naive Bayes classifier [68]. Authors have Convolutional neural network (CNN) is a class of deep
proposed a combinational method for fault detection in neural networks which has been used for fault detection
transformers based on Bayesian network and AdaboostMI in recent times. A 1-D convolutional neural network hav-
algorithm [108]. Authors have compared fault detection sys- ing self feature extraction and classification properties has
tem based on feedforward ANN and Naive Bayes for induc- been designed [32]. CNN provides a functionality of fea-
tion motors. The features for classifiers were extracted by the ture extraction form raw data and enable fault detection in
dual tree complex wavelet transform [82]. motor bearings [97]. CNN has been used for bearing fault
The comparison has been made in several papers in detection without feature extraction based on conversion of
between Naive Bayes classifier and other classifiers. Authors 1-dimensional vibration signal to 2-dimensional vibration
[64] have discussed a pattern classification methods for fault image [30]. Deep learning is comparatively new and fault
detection based on Naive Bayes classifiers, k-NN, SVM and detection system based on it has been developed recently.
ANN using time domain features as input vector. A fault Deep learning has great potential owing to its deep architec-
diagnosis based on automated segmentation method of ther- tures and its use will help in developing a better and reliable
mographic images has been developed using Naive Bayes fault diagnostic systems.
classifiers and C-45 decision trees [39].

4.7 Deep Learning Based Fault Detection 5 Discussion, Limitations and Future Scopes

Deep learning methods learns the features automatically at The machine learning algorithm based fault detection system
multiple levels of abstraction in the network. The abstraction has more reliability and adaptability. It is being used widely
at multiple level allows the automatic learning of the com- in fault detection. Moreover, a decision of choosing the right
plex function directly from the raw data. ANN is often used algorithm depending on the data is foremost important. The
to develop a deep neural networks. Deep Neural Networks pros and cons of algorithms are mentioned in Table 1.
(DNNs) have overcome the disadvantage of ANN by using SVM works really well for small datasets with an excel-
deep architectures instead of shallow one as in case of ANN. lent generalization properties. SVM can also classify non-
DNN allows the processing of raw data without much pro- linear datas with the help of kernel functions. SVM provides
cessing and can handle complex non-linear functions [35]. high accuracy in fault detection and diagnosis. k-NN is an
A deep belief network (DBN) based diagnostic network instance based learning and is often called lazy learner. In
has proved to be more reliable for fault classification than k-NN algorithm, k denotes the data points of training set
SVM and back propagation neuron networks (BPNN) and lying in the vicinity of the data points of test data sets. It is
it also overcomes the problem of disturbances and noises faster during training process, however it is slow during clas-
[22]. A deep belief network for bearing fault detection has sification process. ANN is among the efficient classifiers and

Table 1 Pros and cons of ML Algorithms Pros Cons

algorithms
SVM 1. High Accuracy 1. Sluggish for large data
2. Handles outliers better 2. No physical meaning
k-NN 1. Easy implementation 1. Slow in real time
2. Robust to noise 2. Sensitivity to Outlier
ANN 1. Parallel processing capability 1. Hardware
2. High accuracy 2. Overfitting
Decision trees and random 1. Easy to understand and interpret 1. Overfitting
forest
2. Longer Training Period
Naive Bayes 1. Robust to noise 1. Assumption of inde-
pendent features
2. Low computation cost
Deep learning 1. Robustness to variations in data 1. Large data requirement
2. Self feature extraction 2. Large time for training

13
Review on Machine Learning Algorithm Based Fault Detection in Induction Motors

has been used significantly in fault detection systems. The Compliance with Ethical Standards
structure of ANN can be adjusted for achieving better classi-
fication properties. It can be used for an efficient fault detec- Conflict of interest The authors declare that they have no conflict of
tion systems. A decision tree represents a tree-like structure. interest.
Decision trees are often used for fault classification proper-
ties. A higher version of decision trees, i.e. random forests
can be used as it is immune to external noises and easier to
interpret. Naive Bayes classifiers depends on the probability References
of two independent sets. It rather tells about the probability
1. Acosta G, Verucchi C, Gelso E (2006) A current monitoring sys-
of an instance belonging to a particular class rather then the tem for diagnosing electrical failures in induction motors. Mech
classification. Deep learning is relatively a new technique in Syst Signal Process 20(4):953–965
the field of fault detection systems. It overcomes some of the 2. Banerjee TP, Das S, Roychoudhury J, Abraham A (2010) Imple-
disadvantages of other algorithms. Generally, SVM, ANN mentation of a new hybrid methodology for fault signal clas-
sification using short-time fourier transform and support vector
and deep learning performs well in multi-dimensional and machines. In: Soft computing models in industrial and environ-
continuous data and k-NN, decision trees and Naive Bayes mental applications, 5th international workshop (SOCO 2010),
works well for discrete data [47]. Springer, pp 219–225
The future of AI is immense and needs more attention, 3. Baraldi P, Podofillini L, Mkrtchyan L, Zio E, Dang VN (2015)
Comparing the treatment of uncertainty in Bayesian networks
especially in the field of deep learning. Some of the future and fuzzy expert systems used for a human reliability analysis
scope and possibilities are enlisted below: application. Reliab Eng Syst Saf 138:176–193
4. Benbouzid MEH (2000) A review of induction motors signature
analysis as a medium for faults detection. IEEE Trans Ind Elec-
1. With the availability of high computational machines,
tron 47(5):984–993
a hybrid deep learning based fault detection system can 5. Bengio Y et al (2009) Learning deep architectures for AI. Found
be developed which will be fail-safe and can improve Trends® Mach Learn 2(1):1–127
diagnostic performance. 6. Bin G, Gao J, Li X, Dhillon B (2012) Early fault diagnosis of
rotating machinery based on wavelet packets-empirical mode
2. Current system are dependent on feature selection, fea-
decomposition feature extraction and neural network. Mech Syst
ture extraction, data collection and so many processes. Signal Process 27:696–711
Deep learning has the potential to develop a complete 7. Bonnett AH, Soukup GC (1986) Rotor failures in squirrel cage
diagnostic system and needs more attention. induction motors. IEEE Trans Ind Appl 6:1165–1173
8. Cai B, Zhao Y, Liu H, Xie M (2016) A data-driven fault diagno-
3. Smart-hybrid fault detection system can be developed
sis methodology in three-phase inverters for PMSM drive sys-
for a real time applications based on deep learning. tems. IEEE Trans Power Electron 32(7):5590–5600
9. Cameron J, Thomson W, Dow A (1986) Vibration and current
monitoring for detecting airgap eccentricity in large induction
motors. In: IEE Proceedings B (electric power applications), IET,
6 Conclusion vol 133, pp 155–163
10. Cao L, Chua KS, Chong W, Lee H, Gu Q (2003) A comparison
This paper has presented a detailed review of machine of PCA, KPCA and ICA for dimensionality reduction in support
learning-based algorithm for fault detection. This review is vector machine. Neurocomputing 55(1–2):321–336
11. Casimir R, Boutleux E, Clerc G, Yahoui A (2006) The use of
fruitful for developing an effective fault monitoring system
features selection and nearest neighbors rule for faults diagnostic
for induction motors. The fault detection based on machine in induction motors. Eng Appl Artif Intell 19(2):169–177
learning has improved the reliability and efficiency. The 12. Castejón C, García-Prada J, Gómez M, Meneses J (2015)
machine learning algorithms like ANN, k-NN, SVM, deci- Automatic detection of cracked rotors combining multireso-
lution analysis and artificial neural networks. J Vib Control
sion trees and deep learning have been discussed thoroughly.
21(15):3047–3060
The machine learning algorithm based fault detection has 13. Chandrashekar G, Sahin F (2014) A survey on feature selection
been given due attention in this paper. Also, latest deep methods. Comput Electr Eng 40(1):16–28
learning algorithms like convolutional neural network and 14. Cover TM, Hart P et al (1967) Nearest neighbor pattern clas-
sification. IEEE Trans Inf Theory 13(1):21–27
deep belief network have been reviewed for fault detection.
15. Cusido J, Romeral L, Espinosa AG, Ortega JA, Riba Ruiz
Application of deep learning gives an advantage of self-fea- JR (2011) On-line fault detection method for induction
ture extraction and selection which consequently, reducing machines based on signal convolution. Eur Trans Electr Power
the risk of human error in feature extraction and their selec- 21(1):475–488
16. Deng W, Yao R, Zhao H, Yang X, Li G (2019) A novel intelligent
tion. Along with it, future scopes have been proposed too for
diagnosis method using optimal LS-SVM with improved PSO
developing an effective system. The future of AI is immense algorithm. Soft Comput 23(7):2445–2462
and needs more attention to developing a complete and self- 17. Doorsamy W, Cronje W (2014) Optimisation of shaft voltage
learning package for unified fault prognosis and diagnosis. based condition monitoring in generators using a Bayesian
approach

13
 P. Kumar, A. S. Hati

18. Elkasabgy NM, Eastham AR, Dawson GE (1992) Detection of 38. Kankar PK, Sharma SC, Harsha SP (2012) Vibration-based
broken bars in the cage rotor on an induction machine. IEEE fault diagnosis of a rotor bearing system using artificial neu-
Trans Ind Appl 28(1):165–171 ral network and support vector machine. Int J Modell Identif
19. Farajzadeh-Zanjani M, Razavi-Far R, Saif M, Rueda L (2016) Control 15(3):185–198
Efficient feature extraction of vibration signals for diagnosing 39. Karvelis P, Georgoulas G, Stylios CD, Tsoumas IP, Antonino-
bearing defects in induction motors. In: 2016 international joint Daviu JA, Ródenas MJP, Climente-Alarcón V (2014) An
conference on neural networks (IJCNN), IEEE, pp 4504–4511 automated thermographic image segmentation method for
20. Fengqi W, Meng G (2006) Compound rub malfunctions feature induction motor fault diagnosis. In: IECON 2014-40th annual
extraction based on full-spectrum cascade analysis and SVM. conference of the IEEE industrial electronics society, IEEE, pp
Mech Syst Signal Process 20(8):2007–2021 3396–3402
21. Finley WR, Hodowanec MM, Holter WG (1999) An analytical 40. Kecman V (2001) Learning and soft computing: support vector
approach to solving motor vibration problems. In: Industry appli- machines, neural networks, and fuzzy logic models. MIT press,
cations society 46th annual petroleum and chemical technical Cambridge
conference (Cat. No. 99CH37000), IEEE, pp 217–232 41. Keskes H, Braham A (2015) Recursive undecimated wavelet
22. Gan M, Wang C et al (2016) Construction of hierarchical diag- packet transform and dag SVM for induction motor diagnosis.
nosis network based on deep learning and its application in the IEEE Trans Ind Inf 11(5):1059–1066
fault pattern recognition of rolling element bearings. Mech Syst 42. Kia SH, Mabwe AM, Henao H, Capolino GA (2006) Wavelet
Signal Process 72:92–104 based instantaneous power analysis for induction machine fault
23. Glowacz A (2019) Fault diagnosis of single-phase induction diagnosis. In: IECON 2006-32nd annual conference on IEEE
motor based on acoustic signals. Mech Syst Signal Process industrial electronics, IEEE, pp 1229–1234
117:65–80 43. Kim K, Parlos AG (2002) Induction motor fault diagnosis based
24. Gyftakis KN, Antonino-Daviu JA, Garcia-Hernandez R, McCull- on neuropredictors and wavelet signal processing. IEEE/ASME
och MD, Howey DA, Cardoso AJM (2015) Comparative experi- Trans Mech 7(2):201–219
mental investigation of broken bar fault detectability in induction 44. Kliman G, Koegl R, Stein J, Endicott R, Madden M (1988) Non-
motors. IEEE Trans Ind Appl 52(2):1452–1459 invasive detection of broken rotor bars in operating induction
25. Han T, Jiang D, Zhao Q, Wang L, Yin K (2018) Comparison motors. IEEE Trans Energy Convers 3(4):873–879
of random forest, artificial neural networks and support vector 45. Kliman G, Premerlani W, Koegl R, Hoeweler D (1996) A new
machine for intelligent diagnosis of rotating machinery. Trans approach to on-line turn fault detection in AC motors. In: IAS’96.
Inst Meas Control 40(8):2681–2693 conference record of the 1996 IEEE industry applications confer-
26. Han T, Yang BS, Choi WH, Kim JS (2006) Fault diagnosis sys- ence thirty-first IAS annual meeting, IEEE, vol 1, pp. 687–693
tem of induction motors based on neural network and genetic 46. Kliman G, Stein J (1990) Induction motor fault detection via
algorithm using stator current signals. Int J Rotat Mach. https:// passive current monitoring-a brief survey. In: Proceedings of
doi.org/10.1155/IJRM/2006/61690 the 44th meeting of the mechanical failures prevention group,
27. Han Y, Song Y (2003) Condition monitoring techniques for elec- pp 49–65
trical equipment-a literature survey. IEEE Trans Power Deliv 47. Kotsiantis SB, Zaharakis ID, Pintelas PE (2006) Machine learn-
18(1):4–13 ing: a review of classification and combining techniques. Artif
28. Heller B, Hamata V (1977) Harmonic field effects in induction Intell Rev 26(3):159–190
machines. Elsevier, Amsterdam 48. LeCun Y, Bottou L, Bengio Y, Haffner P et al (1998) Gradient-
29. Ho TK (1995) Random decision forests. In: Proceedings of 3rd based learning applied to document recognition. Proc IEEE
international conference on document analysis and recognition, 86(11):2278–2324
IEEE, vol 1, pp 278–282 49. Lei Y, Jia F, Lin J, Xing S, Ding SX (2016) An intelligent fault
30. Hoang DT, Kang HJ (2017) Convolutional neural network based diagnosis method using unsupervised feature learning towards
bearing fault diagnosis. In: International Conference on Intel- mechanical big data. IEEE Trans Ind Electron 63(5):3137–3147
ligent Computing, Springer, pp 105–111 50. Lei Y, Zuo MJ (2009) Gear crack level identification based on
31. Hwang JN, Hu YH (2001) Handbook of neural network signal weighted k nearest neighbor classification algorithm. Mech Syst
processing. CRC Press, Boca Raton Signal Process 23(5):1535–1547
32. Ince T, Kiranyaz S, Eren L, Askar M, Gabbouj M (2016) Real- 51. Li X, Wang K, Jiang L (2011) The application of AE signal in
time motor fault detection by 1-d convolutional neural networks. early cracked rotor fault diagnosis with PWVD and SVM. JSW
IEEE Trans Ind Electron 63(11):7067–7075 6(10):1969–1976
33. Janssens O, Slavkovikj V, Vervisch B, Stockman K, Loccufier 52. Li Y, Yang Y, Wang X, Liu B, Liang X (2018) Early fault diag-
M, Verstockt S, Van de Walle R, Van Hoecke S (2016) Convolu- nosis of rolling bearings based on hierarchical symbol dynamic
tional neural network based fault detection for rotating machin- entropy and binary tree support vector machine. J Sound Vib
ery. J Sound Vib 377:331–345 428:72–86
34. Jardine AK, Lin D, Banjevic D (2006) A review on machinery 53. Lin SW, Ying KC, Chen SC, Lee ZJ (2008) Particle swarm opti-
diagnostics and prognostics implementing condition-based main- mization for parameter determination and feature selection of
tenance. Mech Syst Signal Process 20(7):1483–1510 support vector machines. Expert Syst Appl 35(4):1817–1824
35. Jia F, Lei Y, Lin J, Zhou X, Lu N (2016) Deep neural networks: 54. Liu H, Zhou J, Zheng Y, Jiang W, Zhang Y (2018) Fault diag-
a promising tool for fault characteristic mining and intelligent nosis of rolling bearings with recurrent neural network-based
diagnosis of rotating machinery with massive data. Mech Syst autoencoders. ISA Trans 77:167–178
Signal Process 72:303–315 55. Liu R, Meng G, Yang B, Sun C, Chen X (2016) Dislocated
36. Jung U, Koh BH (2015) Wavelet energy-based visualization and time series convolutional neural architecture: an intelligent fault
classification of high-dimensional signal for bearing fault detec- diagnosis approach for electric machine. IEEE Trans Ind Inf
tion. Knowl Inf Syst 44(1):197–215 13(3):1310–1320
37. Kankar PK, Sharma SC, Harsha SP (2011) Fault diagnosis of 56. Liu R, Yang B, Zio E, Chen X (2018) Artificial intelligence for
ball bearings using machine learning methods. Expert Syst Appl fault diagnosis of rotating machinery: a review. Mech Syst Signal
38(3):1876–1886 Process 108:33–47

13
Review on Machine Learning Algorithm Based Fault Detection in Induction Motors

57. Liu X, Bo L, Luo H (2015) Bearing faults diagnostics based on 77. Sanz J, Perera R, Huerta C (2012) Gear dynamics monitoring
hybrid LS-SVM and EMD method. Measurement 59:145–166 using discrete wavelet transformation and multi-layer perceptron
58. Liu Z, Guo W, Hu J, Ma W (2017) A hybrid intelligent multi-fault neural networks. Appl Soft Comput 12(9):2867–2878
detection method for rotating machinery based on RSGWPT, 78. Saravanan N, Ramachandran K (2009) Fault diagnosis of spur
KPCA and twin SVM. ISA Trans 66:249–261 bevel gear box using discrete wavelet features and decision tree
59. Martínez-Morales JD, Palacios-Hernández ER, Campos-Delgado classification. Expert Syst Appl 36(5):9564–9573
D (2018) Multiple-fault diagnosis in induction motors through 79. Saruhan H, Saridemir S, Qicek A, Uygur I (2014) Vibration
support vector machine classification at variable operating condi- analysis of rolling element bearings defects. J Appl Res Technol
tions. Electr Eng 100(1):59–73 12(3):384–395
60. Mehrjou MR, Mariun N, Marhaban MH, Misron N (2011) Rotor 80. Schmidhuber J (2015) Deep learning in neural networks: an over-
fault condition monitoring techniques for squirrel-cage induction view. Neural Netw 61:85–117
machine-a review. Mech Syst Signal Process 25(8):2827–2848 81. Schoen RR, Habetler TG, Kamran F, Bartfield R (1995) Motor
61. Moosavian A, Ahmadi H, Sakhaei B, Labbafi R (2014) Support bearing damage detection using stator current monitoring. IEEE
vector machine and k-nearest neighbour for unbalanced fault Trans Ind Appl 31(6):1274–1279
detection. J Quality Maint Eng 20(1):65–75 82. Seshadrinath J, Singh B, Panigrahi BK (2013) Vibration analy-
62. Mrugalski M, Witczak M, Korbicz J (2008) Confidence esti- sis based interturn fault diagnosis in induction machines. IEEE
mation of the multi-layer perceptron and its application in fault Trans Ind Inf 10(1):340–350
detection systems. Eng Appl Artif Intell 21(6):895–906 83. Shao H, Jiang H, Zhang X, Niu M (2015) Rolling bearing fault
63. Nandi S, Toliyat HA, Li X (2005) Condition monitoring and diagnosis using an optimization deep belief network. Meas Sci
fault diagnosis of electrical motors-a review. IEEE Trans Energy Technol 26(11):115002
Convers 20(4):719–729 84. Shen C, Wang D, Kong F, Peter WT (2013) Fault diagnosis of
64. Palácios RHC, da Silva IN, Goedtel A, Godoy WF (2015) A rotating machinery based on the statistical parameters of wavelet
comprehensive evaluation of intelligent classifiers for fault iden- packet paving and a generic support vector regressive classifier.
tification in three-phase induction motors. Electric Power Syst Measurement 46(4):1551–1564
Res 127:249–258 85. Shen Z, Chen X, Zhang X, He Z (2012) A novel intelligent gear
65. Pandya D, Upadhyay S, Harsha SP (2013) Fault diagnosis fault diagnosis model based on EMD and multi-class TSVM.
of rolling element bearing with intrinsic mode function of Measurement 45(1):30–40
acoustic emission data using APF-KNN. Expert Syst Appl 86. Siedlecki W, Sklansky J (1993) A note on genetic algorithms for
40(10):4137–4145 large-scale feature selection. In: Handbook of pattern recognition
66. Peng Z, Peter WT, Chu F (2005) A comparison study of improved and computer vision, World Scientific, pp 88–107
hilbert-huang transform and wavelet transform: application to 87. Su H, Chong KT (2007) Induction machine condition monitor-
fault diagnosis for rolling bearing. Mech Syst Signal Process ing using neural network modeling. IEEE Trans Ind Electron
19(5):974–988 54(1):241–249
67. Pires VF, Kadivonga M, Martins J, Pires A (2013) Motor square 88. Sun W, Chen J, Li J (2007) Decision tree and PCA-based fault
current signature analysis for induction motor rotor diagnosis. diagnosis of rotating machinery. Mech Syst Signal Process
Measurement 46(2):942–948 21(3):1300–1317
68. Quiroz JC, Mariun N, Mehrjou MR, Izadi M, Misron N, Radzi 89. Sun W, Shao S, Zhao R, Yan R, Zhang X, Chen X (2016) A
MAM (2018) Fault detection of broken rotor bar in ls-pmsm sparse auto-encoder-based deep neural network approach for
using random forests. Measurement 116:273–280 induction motor faults classification. Measurement 89:171–178
69. Rafiee J, Arvani F, Harifi A, Sadeghi M (2007) Intelligent con- 90. Tian J, Morillo C, Azarian MH, Pecht M (2015) Motor bearing
dition monitoring of a gearbox using artificial neural network. fault detection using spectral kurtosis-based feature extraction
Mech Syst Signal Process 21(4):1746–1754 coupled with k-nearest neighbor distance analysis. IEEE Trans
70. Rodriguez PJ, Belahcen A, Arkkio A (2006) Signatures of electri- Ind Electron 63(3):1793–1803
cal faults in the force distribution and vibration pattern of induc- 91. Trutt FC, Sottile J, Kohler JL (2002) Online condition monitoring
tion motors. IEE Proc Electr Power Appl 153(4):523–529 of induction motors. IEEE Trans Ind Appl 38(6):1627–1632
71. Sadeghian A, Ye Z, Wu B (2009) Online detection of broken 92. Vapnik V (2013) The nature of statistical learning theory.
rotor bars in induction motors by wavelet packet decomposi- Springer, New York
tion and artificial neural networks. IEEE Trans Instrum Meas 93. Vas P (1993) Parameter estimation, condition monitoring, and
58(7):2253–2263 diagnosis of electrical machines, vol 27. Oxford University Press,
72. Safizadeh M, Latifi S (2014) Using multi-sensor data fusion for Oxford
vibration fault diagnosis of rolling element bearings by acceler- 94. Wang D (2016) K-nearest neighbors based methods for identifica-
ometer and load cell. Inf Fus 18:1–8 tion of different gear crack levels under different motor speeds
73. Saidi L, Ali JB, Fnaiech F (2015) Application of higher order and loads: revisited. Mech Syst Signal Process 70:201–208
spectral features and support vector machines for bearing faults 95. Wang H, Chen P (2011) Intelligent diagnosis method for roll-
classification. ISA Trans 54:193–206 ing element bearing faults using possibility theory and neural
74. Saimurugan M, Ramachandran K, Sugumaran V, Sakthivel N network. Comput Ind Eng 60(4):511–518
(2011) Multi component fault diagnosis of rotational mechani- 96. Wang H, Li S, Song L, Cui L (2019) A novel convolutional neu-
cal system based on decision tree and support vector machine. ral network based fault recognition method via image fusion of
Expert Syst Appl 38(4):3819–3826 multi-vibration-signals. Comput Ind 105:182–190
75. Samanta B (2004) Gear fault detection using artificial neural 97. Wen L, Li X, Gao L, Zhang Y (2017) A new convolutional neural
networks and support vector machines with genetic algorithms. network-based data-driven fault diagnosis method. IEEE Trans
Mech Syst Signal Process 18(3):625–644 Ind Electron 65(7):5990–5998
76. Samanta B, Al-Balushi KR, Al-Araimi SA (2006) Artificial neu- 98. Widodo A, Yang BS (2007) Support vector machine in machine
ral networks and genetic algorithm for bearing fault detection. condition monitoring and fault diagnosis. Mech Syst Signal Pro-
Soft Comput 10(3):264–271 cess 21(6):2560–2574

13
 P. Kumar, A. S. Hati

99. Witten IH, Frank E, Hall MA, Pal CJ (2016) Data mining: practi- models and radial basis function networks. Neurocomputing
cal machine learning tools and techniques. Morgan Kaufmann, 74(17):2941–2952
Burlington 107. Zhang P, Du Y, Habetler TG, Lu B (2010) A survey of condition
100. Wu S, Chow TW (2004) Induction machine fault detection monitoring and protection methods for medium-voltage induc-
using som-based RBF neural networks. IEEE Trans Ind Electron tion motors. IEEE Trans Ind Appl 47(1):34–46
51(1):183–194 108. Zhao W, Zhang Y, Zhu Y (2009) Diagnosis for transformer faults
101. Yan R, Gao RX, Chen X (2014) Wavelets for fault diagnosis based on combinatorial Bayes network. In: 2009 2nd interna-
of rotary machines: a review with applications. Signal Process tional congress on image and signal processing, IEEE, pp. 1–3
96:1–15 109. Zhou Z, Wen C, Yang C (2016) Fault isolation based on k-nearest
102. Yang B, Liu R, Chen X (2017) Fault diagnosis for a wind turbine neighbor rule for industrial processes. IEEE Trans Ind Electron
generator bearing via sparse representation and shift-invariant 63(4):2578–2586
K-SVD. IEEE Trans Ind Inf 13(3):1321–1331 110. Ziani R, Felkaoui A, Zegadi R (2017) Bearing fault diagnosis
103. Yang H, Mathew J, Ma L (2005) Fault diagnosis of rolling ele- using multiclass support vector machines with binary particle
ment bearings using basis pursuit. Mech Syst Signal Process swarm optimization and regularized fisher’s criterion. J Intell
19(2):341–356 Manuf 28(2):405–417
104. You L, Fan W, Li Z, Liang Y, Fang M, Wang J (2019) A fault
diagnosis model for rotating machinery using VWC and MSFLA- Publisher’s Note Springer Nature remains neutral with regard to
SVM based on vibration signal analysis. Shock Vib 2019:1–16 jurisdictional claims in published maps and institutional affiliations.
105. Yunusa-Kaltungo A, Sinha JK, Elbhbah K (2014) An improved
data fusion technique for faults diagnosis in rotating machines.
Measurement 58:27–32
106. Zhang K, Li Y, Scarf P, Ball A (2011) Feature selection for
high-dimensional machinery fault diagnosis data using multiple

13