Robustification of fault detection algorithm in a three‐phase induction motor using MCSA for various single and multiple faults

K. C. Deekshit Kompella, EEE Department, SNIST, Hyderabad, India. Email: kkcd10@gmail.com Abstract Fault diagnosis in induction machines, particularly at a premature stage, has become necessary to avoid unexpected interruption of the industrial process. Moreover, a reliable and cost‐effective fault detection process is highly essential to avoid inappropriate shut down of the machine. Thus, a critical comparison is done between two popular pre‐fault component cancellation techniques and a reliability test is performed to develop a robust algorithm for fault diagnosis in three‐phase induction motor. To achieve this, various single and multiple faults experienced by induction motor are created and tested to examine the effectiveness of developed algorithm. Evaluation of the proposed topology based on motor current signature analysis is done by repeating the process for several times and tested for consistency. As a part of this, various feature extraction parameters are computed and compared to identify the best estimate of fault and its seriousness. The proposed technique is examined in MATLAB and LabVIEW environment and experiments on a three‐phase, 1.5‐kW induction motor.


| INTRODUCTION
Induction motors are major contributors to industrial applications because of their high power to weight ratio, less maintenance, economical and rugged structure [1]. However, these motors are susceptible to unexpected failure due to long usage [2]. If these failures are not recognized in the incipient stage, they may become catastrophic to the industrial process [2]. Therefore, induction motor fault detection has become a challenging task for researchers. According to an IEEE survey on large motors of above 200 HP, bearing failure accounts to 41%, stator faults to 37%, rotor faults to 10%, and other faults to 12% out of total faults experienced [3][4][5][6]. Therefore, continuous condition assessment of these faults is of high interest to prevent the unexpected shut down of industrial process [7]. According to the location and the nature of the fault, the bearing faults can be classified into (1) cyclic faults (single-point defect) and (2) non-cyclic faults (generalized roughness [GR]), rotor faults into (1) broken rotor fault and (2) end ring faults, stator winding faults into (1) turn-turn (inter-turn) faults, (2) phase to phase faults and (3) winding to winding faults and other faults into (1) eccentricity fault and (2) gear faults etc. [8].
Chemical monitoring, noise monitoring, temperature monitoring, torque monitoring, acoustic emission monitoring, flux monitoring, vibration monitoring, and current monitoring are the most popular techniques to access the state of the machine. However, all the methods except vibration and current monitoring listed above have its own merits and drawbacks which are suitable for very few applications of motor. For instance, noise monitoring is not suitable for mining and petrochemical industries as they are working under huge noise. Contrary to this, few monitoring techniques, namely vibration and current monitoring, are very popular for all categories of applications and have few challenges as following.
Many articles [9][10][11] have discussed the detection of bearing faults using a vibration analysis. In many cases, like noisy environment with shocks, harsh environments, and external vibrations, the vibration signals cannot be obtained directly [12]. In addition to this, accelerometers and costly sensors are required to process the signals for fault detection and their installation is justified to large rating machines [13]. Furthermore, vibration monitoring requires huge manpower to capture the signal by placing the sensors at various positions of the machine, namely horizontal and vertical, and it becomes hectic in the above-mentioned environments. To circumvent these problems in fault detection, a cost-effective, non-invasive, non-intrusive, easy access, requirement of less sensors and less manpower monitoring technique is highly desirable. Current monitoring will meet all these requirements and attracts the concentration of recent researchers.
Conventional fast Fourier transform (FFT) is popularly employed for spectral analysis of the stator current, which has many disadvantages namely resolution problem due to limited data, spectral percolation due to single window, not suitable for non-stationary signals, lacking the ability to provide time frequency relation, and side band issue due to dominant components in the spectrum. In order to obtain high resolution in FFTbased methods, zoom FFT (ZFFT) is proposed in the literature [14,15]. However, high resolution by ZFFT is affected by acquisition time. Furthermore, many high-resolution techniques like Multi signal Classification (MUSIC) algorithm have been presented in [16][17][18]. Despite its advantages like high resolution and reduction of noise influence compared to FFT, it is computationally too complex to implement in real time. In continuation to this, resolution and time-frequency relation problem of FFT are addressed using short-time Fourier transform (STFT) [19,20] and Wigner-Ville distribution (WVD) [21,22], but these have few drawbacks. For example, STFT uses fixed size window for all range of frequencies which leads to spectral leakage whereas WVD is complex due to cross terms. This situation leads to the development of variable sized window-based spectral transform and became very popular as the wavelet transform (WT) addressed in [23][24][25]. However, it is suitable only for few categories of faults like cyclic faults related to bearing due to a greater number of coefficients especially at high-level decomposition.
On the other hand, the extreme problem of FFT in detecting fault frequencies is the domination of regular components like fundamental, harmonics, and noise under the normal condition of the machine. Due to this, the fault frequencies in the current spectrum are invisible and may lead to misinterpretation about fault. This is a serious issue especially for broken rotor faults and non-cyclic fault in bearing damage. Therefore, many recent authors [21,[26][27][28][29][30][31][32][33][34][35][36][37][38][39][40][41] have proposed the removal of these regular dominant components named as noise cancellation. In the present work, it is termed as pre-fault component cancellation. In continuation to this, a Notch filter is proposed in [21,26,27] to remove these pre-fault components, but it requires detailed knowledge about the spectral components and has main lobe width issues leading to the removal of fault frequencies from the spectrum. In [28], finite impulse response filter banks are used to remove these components which require detailed information regarding fundamental and harmonic components for correct removal and, it is difficult especially at variable frequency condition. In order to overcome this issue, Teager-Kaiser energy operator is proposed in [29,30] and is a non-linear filter which assumes the stator current fundamental frequency as a constant and all harmonic frequencies are displayed on the right side of the constant frequency. However, it is impossible to identify noise which is a severe issue in the fault detection. In continuation to this, Park's vector transformation is presented in [31,32], which converts the three-phase stator current into d-q axis components and elevates the fault frequencies, but it requires a greater number of current sensors and mathematical calculations are complex. To overcome this issue, a Kalman filter is employed in [33,34] and is used to estimate different states which are not measurable directly and assumes all states as linear which is not practically existent. In contrast to this, pre-fault component cancellation using frequency spectral subtraction (FSS) with an adaptive filter (Wiener filter) [35][36][37][38] and without filter [39,41] is presented. In [35,36], the pre-fault component cancellation is carried out by predicting these components under healthy condition and canceled in a real-time fashion. After cancellation in [35], the residual current is employed for computation of root mean square (RMS) value as the feature extraction parameter to estimate the bearing fault. The remaining current after cancellation in [36] is used for spectral analysis using FFT and is severely affected by noise especially at early stage of fault, and in [37,38] it is used for wavelet decomposition and computation of standard deviation (SD) as fault indexing parameter (FIP) to estimate and categorize the fault. On the other hand, FSS without filter in [39] is carried out by modelling a healthy current depending on machine specification and the frequency components using FFT are subtracted from faulty component and resultant signal is used for the feature extraction parameter of energy as FIP. In [40,41], the FSS is done using WT and SD is calculated as FIP. In [42,43], a detailed comparison is carried out between FSS with and without a filter and SD is calculated as FIP. In these types of pre-fault component cancellation techniques, the fault existence is estimated based on feature extraction parameters and is strongly affected by noise especially after cancellation. This leads to development of disturbances in FIP signal and may cause reduced fault indexing even for increment in fault magnitude. This shows a lot of misinterpretation about fault and leads to unreliable diagnosis of machine which causes unexpected maintenance.
In view of this, the work proposes detection of various faults, namely bearing damage with different faulty bearings like the outer race fault, cage damage, ball defect and GR faults, broken rotor fault with different magnitudes, and combined multiple faults with the use of current signature analysis. Moreover, a detailed comparison between two popular pre-fault component cancellation techniques, namely FSS and Wiener filter cancellation (WFC), is presented and the residual current is employed for spectral analysis using Matrix Pencil method (MPM) to effectively elevate and estimate the incipient fault component. The fault indexing is calculated using multiple feature extraction parameters and is compared for better performance. Even though this work consists several existing methods, the innovativeness is involved in selecting method, improvement, and combination in the following way.
Many recent authors in the literature discussed only particular type of faults like bearing related or rotor related independently. The proposed work presents both the faults along with their combination simultaneously.
A systematic method of current signal processing by comparing popular pre-fault component cancellation techniques with and without filter to get reliable fault indication is presented which improves the performance.
The reliability and accuracy of the proposed fault detection topology is crucial in testing the Wiener filter performance, since the noise is added to develop a synthetic data which varies randomly for every iteration. A new Wiener filter is designed to overcome this problem by training it with white noise of large bandwidth that covers all fault frequencies. It is further improved by multiplying the filter coefficients with the Kaiser window function. This unique design improved the results considerably.
In FSS, the pre-fault component cancellation is done with generated synthetic data which represents the healthy current which should match the actual current of the machine. The actual current spectrum of the machine depends on physical design aspects, surroundings, and working environment. Hence, the generated current may not be same for all the machines. Therefore, the modelling of the healthy current is carried out by using white Gaussian noise (WGN) with suitable signal-to-noise ratio and normalized to per unit value which is suitable for different load conditions. This model has the flexibility to be used for any kind of machine environment by suitably modifying the current spectrum.
The proposed algorithm uses wavelet de-noising, where the threshold function is defined based on SD and decomposition level which is not same as general spectral analysis. Furthermore, the mother wavelet Daubechies of the order 44 (db44) is selected based on its good correlation coefficient with the current signal. This improves the performance of both the pre-fault component cancellation techniques and leads to good indication for fault.
In general, either MPM or pre-fault component cancellation techniques with conventional spectrum analysis are adopted to improve the resolution of the current spectrum but a combination of these two is not preferred so far. In the proposed method, the MPM is employed after the above steps which leads to good resolution. The FIP graphs are plotted to examine the accuracy and reliability of the proposed method.
To examine the robustness of proposed method towards noise, each faulty signal is acquired for 10 to 11 times and the above-mentioned steps are repeated. The FIPs are calculated for every iteration and are plotted to examine the accuracy and reliability of the proposed method.
The effectiveness of proposed fault detection topology is evaluated by a quantitative analysis with the existing methods and is discussed in the results section.
The concept, implementation, and performance of proposed technique are presented in Section 2. The experimental setup and procedures are discussed in Section 3. The experimental results and the quantitative analysis are discussed in Section 4, and conclusions are described in Section 5. Figure 1 shows the stepladder of the proposed fault detection methodology using the current spectral analysis. The stator current is acquired from the three-phase induction motor using data acquisition system and processed for wavelet denoising to remove the higher order noise components. Now the de-noised stator current has both healthy and faulty components and the dominant components (pre-fault) are removed using FSS and WFC. In the next stage, spectral analysis using MPM is done to elevate the fault frequencies.

| PROPOSED TOPOLOGY
The residual components of stator current are used for computation of multiple feature extraction parameters as FIP to estimate the fault and its seriousness. Total procedure is repeated for 10 to 11 times to test the robustness of proposed algorithm. Each stage of proposed fault detection scheme is explained in the following.

| Stator current acquisition
The machine stator current is acquired by means of the Hall-effect transducer. Motor supply lines are wounded on the current sensor, and output terminals of the sensor are connected to data acquisition card (DAC). The DAC will provide interface between stator current and MATLAB. The acquired current signal through DAC is processed in MAT-LAB by choosing a sufficient sampling frequency to avoid the missing data. Similarly, the stator current is processed using the proposed technique for all conditions of the motor.

| Discrete wavelet transform-based wavelet de-noising
Stator current de-noising is used to eliminate higher order frequencies to diminish the noise floor problems. Discrete wavelet transform (DWT) is used to decompose the current in view of eradicating the higher order frequencies. De-

F I G U R E 1 Stepladder of proposed fault detection in induction motor
noising of a signal involves three different stages: first, the stator current is decomposed into required level depending on sampling frequency, fundamental frequency, and slip speed which is not same as normal spectral analysis using WTs. After removing the unwanted components from the wavelet coefficients using threshold value, a new signal is reconstructed using modified coefficients. Proper threshold value will play a crucial role in de-noising the signal without missing original data [44]. The threshold value is selected using below expression.
where σ j , N j are standard deviation and length of the coefficients at the jth level, respectively. The threshold value selecting from (1) is changed from level to level. This way of selecting threshold will reduce the probability of noise presented in reconstructed signal. The latest coefficients are used to reconstruct the signal which is used to remove pre-fault components.
The DWT decomposition is done using a mother wavelet of Daubechies of order 44 (db44), which is selected based on its the good correlation coefficient with current signal. This results in a high-frequency resolution at low time and high time resolution in low frequencies due to sampling of scaling and shifting parameters [44]. The decomposition is carried out using following equations [45].
where G 1 is the low pass filter coefficient, G 2 is the high pass filter coefficient, and i(n) is the acquired stator current.

| Pre-fault component cancellation
Third stage of signal processing is pre-fault component cancellation and is carried out using two ways, namely FSS and WFC. The regular components in the stator current spectrum under normal situation are fundamental, its harmonics, and noise by other sources. These components remain in the stator current spectrum even after fault exists and will dominate the fault component especially at early stage of the fault. Therefore, these components are removed using two proposed techniques and are presented in the following subsections.

| Frequency spectral subtraction
In this method of cancellation, stator current under healthy condition is modeled in MATLAB using machine specifications and the spectral analysis is computed using FFT. Furthermore, the FFTs of acquired current under various faulty conditions of the motor are computed and subtracted from modelled components. Residual components after subtraction are used to reconstruct the signal, which is used for spectral analysis using the MPM. The signal obtained after MPM is processed for feature calculation and decision algorithm. The detailed process is shown in Figure 2 and the mathematical analysis is explained in the following. The stator current under healthy condition is modelled as where Therefore, where I 0 is magnitude of fundamental, ω s = 2πf s , f s is fundamental frequency, WGN is the white Gaussian noise, and 5ω s , 7ω s , 11ω s are 5th, 7th, and 11th harmonics, respectively. After wavelet de-noising, the i h (n) becomes i hd (n) and is processed for computation of FFT as shown in Figure 2. Furthermore, the acquired current under various faulty conditions of the machine is i(n) , which becomes i d (n) after denoising.
FFTs of both the signals are given by where k = 0, 1, …N − 1, and I hd (k), I d (k) are spectral components of modeled and acquired signals, respectively. The resultant components after FSS are obtained using Reconstruction of stator current after spectral subtraction can be done using inverse FFT of ζ(k).
where i FSS is the stator current after FSS, which is used for spectral analysis using MPM in the next section.

| Wiener filter subtraction
In this, the pre-fault components in the stator current are removed using the Wiener filter in a real-time manner as shown in Figure 3. In this type of cancellation, instead of modeling healthy stator current, it should be acquired from the machine. Therefore, these components are taken under healthy condition and are cancelled using Wiener filter to extract the fault component. The stator current is assumed to be a wide sense stationary and the filter coefficients are calculated using minimum mean square sense to minimize the prediction error [46].
The mean square error ξ from Figure 4 is Equation (13) can be modified as where E{.} is the expectation and the error in (14) is minimized by differentiating ξ with respect to w M (k) and equating to zero. Then After some mathematical simplifications, the Wiener filter coefficients can be obtained as  : : υ id ðpÞ υ id ð1Þ υ id ð0Þ : : υ id ðp − 1Þ : : : : : : : The simplified version of Equation (16) is where ϒ id is a Toeplitz matrix of autocorrelation, W is a vector of the filter coefficients, and υ did cross-correlation matrix between the desired signal and the observed signal [37,44]. The filter coefficients can be obtained as The Wiener filter coefficients obtained from (18) have removed the faulty components which are nearer to dominant components and may show poor performance. Therefore, the filter coefficients were modified using the following function.
where K(n) is the Kaiser window function and is defined as ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi where I 0 (x) is zeroth-order Bessel function and β is the adjustable parameter. Now the performance of the system is tested by evaluating the prediction error from (14). The prediction error from the equation is modified as follows and evaluated for both healthy and faulty bearing conditions.
Under normal condition of the machine, the autocorrelation and cross-correlation of the signal will be same and the prediction error will be where ξ min is prediction error, which is minimum under healthy condition of the motor. As the motor fault develops, the prediction error starts increasing and becomes high at the advanced stage of the fault. Whenever fault develops, the prediction error is modified as where I l is the fault magnitude and w l is fault frequency with l = 1, 2, 3… Under the healthy condition of motor, fault magnitude I l is zero and (23) becomes (22) as explained before. Whenever bearing fault develops, the second and third parts of (23) start increasing. The severity of the fault can be estimated by evaluating the prediction error in (23). For cyclic type bearing fault, the fault frequencies are predictable and magnitudes are also detectable due to severe impact. Therefore, the prediction error in (23) significantly changes and indicates the fault clearly. For non-cyclic and broken rotor fault, the fault frequencies are unpredictable and nearer to fundamental component with very small magnitude. Correspondingly, the prediction error is very less and difficult to identify the fault component.
In view of this, the advanced spectral tool MPM is adopted to estimate the fault with clear index especially at the incipient stage. The mathematical analysis of stator current after prefault component cancellation has been discussed in the following subsection.

| Matrix pencil method
The stator currents after pre-fault component cancellation using FSS and WF are i FSS (n) and i ζ (n) respectively and are treated in section as i f (n) for mathematical simplification. The mathematical model for MPM is taken from [47] and is modified as follows. The signal after the pre-fault component cancellation in the time domain is expressed as where T s is the sampling period, q is the harmonic number of fault frequency, f m is the fault frequency, and I m is the magnitude of fault component. Equation (24) may also expressed in terms of the exponential component as 598 - Now the resultant signal after decomposing into vector components using MPM is employed for feature extraction parameter cancellation and is discussed in the following subsection.

| Fault indexing parameter
This study presents computation of various features of stator current after spectral analysis to avoid misinterpretation about fault and to make fault detection algorithm as robust. Stator current after spectral analysis is treated as IMPM(k). The list of various feature parameters, such as RMS value (RMS), mean value (M), standard deviation (SD), variance (V), kurtosis (K), skewness value (SK), crest value (C), and entropy (E), is defined in Table 1 [48].
The fault severity is estimated by taking the ratio of the feature value of faulty signal to the healthy signal. Under the healthy condition of machine, FIP should be 1. However, due to the presence of noise, it varies around 1. As the fault becomes severe, the indication of this parameter increases correspondingly.

| EXPERIMENTAL SETUP AND PROCEDURE
The investigation arrangement is mentioned in Figure 5 and consists of 1.5 kW (2 HP), 440 V, 50 Hz, three-phase induction motor with speed range of 1415-1500 rpm, NI MyDAQ, LA55P current transducer made by Life Energy Motion (LEM), measured inmeters, and a PC with LabVIEW software. The motor is connected via three-phase autotransformer to apply the voltage of 440 V, and the supply lines are passing through a current sensor to sense the stator current. The sensed stator current is acquired by a data acquisition card made by National Instruments by a sampling rate of 10 kHz through the LabVIEW software. In the present work, SKF 6205ZZ single row deep grove is used as test bearing. The test bearing is mounted on driving end of the shaft and the specifications are as follows: pitch diameter (P) = 0.0389 m, ball diameter (B) = 0.00792 m, number of balls (N b ) = 9 and contact angle (θ) = 0 0 .
In the proposed work of fault diagnosis, the following bearing faults with both categories are tested.

RMS value (RMS)
ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi All these faulty bearings are inserted in the machine one after other and the respective stator current is acquired with the above-mentioned equipment for 10 to 11 iterations. In the same way, rotor fault is tested with broken rotor fault. The rotor cylinder is drilled with holes in the following way with 5mm diameter using lathe machine in the workshop.
� Rotor with one hole � Rotor with two holes � Rotor with three holes First, the rotor is drilled with one hole and the respective current is measured for testing in the same way. Then, two and three holes are created one after other, and the corresponding currents are measured. Apart from this, two combined rotor and bearing faults are tested to examine the robustness of proposed methodology in the following way.
� Rotor with two holes and bearing outer race fault � Rotor with three holes and bearing outer race fault The construction of all these faulty equipment is shown in Figure 6. The acquired stator current under these faulty conditions of the machine is processed using prewritten algorithm in MATLAB. The total scenario of experimental procedure is shown in Figure 7. To test the reliability of fault detection topology, both the above-mentioned pre-fault component The plots of both methods are presented and compared in the following section. In view of length of the paper and size of the data, the proposed work is restricted to no load condition of the motor.

| EXPERIMENTAL RESULTS AND DISCUSSION
In the process of evaluating the proposed algorithm of fault detection in a three-phase induction motor, stator current under various conditions of the machine is acquired and plotted in Figure 8. It is observed that the stator current of various states of the machine is looking identical and is difficult to discriminate the faulty one. Therefore, the spectral analysis of these currents is computed and plotted in Figure 9, and it is observed that the fault components are difficult to elevate due to domination of regular components. In view of this, removal of pre-fault components using FSS and WFC is carried out and discussed in the following section as mentioned in the literature.

| Frequency spectral subtraction
In the FSS-based pre-fault component cancellation, the healthy current is modeled and is subtracted from acquired current spectrum as mentioned in Section 2. The spectral components of modelled, acquired, and reconstructed signal after FSS are shown in Figure 10, and is identified that the subtraction of regular components is done.
However, due to the mismatch between magnitudes of few components and noise in modelled and acquired currents, there exists fundamental and its harmonics with very less magnitude. This leads to the suppression of fault component especially at an early stage. To overcome this issue, advanced spectral analysis is carried out using MPM after FSS and the residual components are used to evaluate feature extraction parameters as mentioned in Section 2. To avoid the impact of noise, the algorithm is repeated for 10 times and the FIP indicator for healthy current is shown in Figure 11.
From Figure 11, it is understood that, under the healthy condition of the machine, the fault severity should be equal to 1. It is clearly observed in kurtosis, crest value (peak to RMS), and entropy. On the other hand, RMS, mean, SD, and variance F I G U R E 7 Experimental procedure KOMPELLA ET AL. parameters are slightly affected by noise. However, skewness parameter is very sensitive about noise and shows drastic change in graph. Therefore, skewness is not suitable for this case and may lead to wrong diagnosis of fault. FIP indicator of the outer race fault is shown in Figure 12, and it is clear that RMS, mean, SD, and variance have shown good indication for fault and represent the variation of fault magnitude for several repetitions of algorithm clearly. The remaining parameters such as kurtosis, skewness, crest, and entropy values have not shown any indication about fault and are greatly affected by minor changes in current due to several repetitions of the algorithm. The complete data of FIP for various faults after FSS is mentioned in Figure 13, where the parameters RMS value to entropy are represented with P1 to P8, respectively. From this, the parameters P1 to P3 have shown clear and sufficient indication about the existence of the fault and have very less oscillations. On the other hand, parameter P4 shows very high magnitude for low fault and shows very dangerous value for severe faults. This may result in easy identification of fault, but it is difficult to estimate its severity. Parameters P5 to P7 show no indication for low and medium faults but have indication for very severe faults, which leads to failure of machine due to unpredictable incipient faults. Finally, parameter P8 does not have any indication for all categories of faults.
In addition, the severity of the fault is identified using P1 as shown in Figure 14. In the case of rotor fault with different magnitudes, the fault severity is indicated in an incremental way and shows drastic change in a severe case. In the similar way, the fault severity indicated a high value for combined fault 2 than combined fault 1. However, for severe fault condition, the oscillations are somewhat high. Even though the oscillations are more, it is confident about existence of fault along with its severity.
On the other hand, pre-fault components are canceled using Wiener filter along with Kaiser window and is discussed in the following section.

| Wiener filter cancellation
In the Wiener filter-based cancellation, the regular components are canceled, and the residual components are used for spectral  Figure 15. The periodogram of stator current under healthy condition before and after WFC is shown in Figure 16, and it is observed that the fundamental component is reduced drastically and does not have any impact on the spectrum. On the other hand, harmonics are completely removed due to well-trained filter coefficients. However, a countable noise will exist even after de-noising before WFC. This may lead to oscillations in fault index parameters and may cause misinterpretation about fault existence especially at nascent stage of fault. Therefore, MPM is employed and the FIPs are calculated similar to FSS.
The FIP indicator for stator current under healthy condition after WFC is shown in Figure 17, and it is observed that the parameters RMS, mean, SD, and variance are very sensitive about noise and show lot of oscillations at 4th, 6th, and 10th operations and may lead to confusion about existence of fault. On the other hand, remaining parameters, i.e. skewness, kurtosis, crest, and entropy, have shown good sign for healthy condition. Furthermore, the FIP for outer race fault is shown in Figure 18, and it is observed that fault is indicated by the first four parameters but have oscillations, whereas remaining four parameters does not have any indication as like FSS.
On the other hand, fault magnitude is clearly mentioned by parameters P1 to P3, which is shown for rotor and combined faults in Figure 19, and it is observed that the fault severity is sufficient enough in only few iterations; for example, in rotor fault the fault severity is exact in 1st, 4th, 5th, 7th to 10th. Similarly, the combined fault severity is clearly estimated in majority of the iterations but insufficient in 4th, 8th to 10th. The complete data of FIP are shown in Figure 20, and it is observed that parameters P1 to P3 have same indication about fault and shows good indication but have few oscillations. On the other hand, P4 shows good indication along with high index but has more oscillations, which lead to confusion in estimating the severity of the fault. However, parameters P5 to P8 do not have any indication about fault and are not useful as deciding factors.
Finally, a comparative analysis between FSS and WFC is carried out and presented as follows. Basically, two types of faults are compared, namely outer race fault and GR fault which are from two different categories of bearing damage and   are shown in Figure 21. As shown in Figure 21, oscillations are very less for FSS and almost consistent in all cases, whereas WFC has many oscillations and has a maximum peak to peak value. However, in estimating the severity of fault, FSS has poor performance as compared to WFC. The complete quantitative analysis between proposed WFC and FSS is presented in Tables 2 and 3.
The quantitative analysis of proposed methodology of fault detection is carried out by comparing the results with conventional Wiener filtering and FSS analysis [7,35,36,41], which are presented in Figures 22 and 23. The experimental results have proved the effectiveness of the proposed method.

| CONCLUSION
In this work, a robust methodology for fault detection in three-phase induction motor using current signature analysis has been proposed that makes use of MPM. A five-stage procedure of fault detection using advanced signal processing -607 tools is presented, which uses acquisition, de-noising, prefault component cancellation, MPM, and feature extraction. This work mainly focuses on tjhe elevation of the incipient fault component by removing pre-fault components using FSS and WFC, and a detailed comparison is carried out. Performance of those techniques is evaluated based on the oscillations and magnitudes of various FIPs. The FIP plots for each technique are developed by repeating the fault detection topology about 10 to 11 times for each fault. To test the robustness of the proposed fault detection topology, various faults related to bearing, rotor, and combination of these two at different magnitudes are examined. Finally, it has been observed that FSS has shown sufficient and consistent indication for fault in all the cases, whereas WFC is good enough for fault magnitude but has more oscillations. This leads to the confusion about fault existence. On the other hand, FSS requires a detailed knowledge about motor current behaviour to model a healthy current which is difficult for motors used in various situations. Therefore, it can be concluded that FSS is suitable for incipient faults identification, whereas WFC is preferable for both identification and estimation of fault. In future, the Wiener filter coefficients may be designed using better error minimization techniques for reduced oscillations and better identification fault. Furthermore, an artificial intelligence technique may be adapted for decision algorithm.

Rotor fault with two holes
Rotor fault with three holes     -613