Broken Rotor Bar Fault Diagnosis Techniques Based on Motor Current Signature Analysis for Induction Motor—A Review

Abstract

The most often used motor in commercial drives is the induction motor. While the induction motor is operating, electrical, thermal, mechanical, magnetic, and environmental stresses can result in defects. Therefore, many researchers who are involved in condition monitoring have been interested in the development of reliable and efficient fault diagnostic technologies. This paper’s goal is to provide an overview of available fault detection methods for the broken rotor bar problem, one of several defects associated to induction motors. Despite the fact that it is less common than bearing or insulator failure, this fault may cause electrical machines to fail catastrophically. It can be quite harmful, especially in large motors, and it can develop as a result of manufacturing faults, repeated starting of the machine, mechanical stress, and thermal stress. Hence, a review on rotor defect diagnosis was conducted. In order to confirm rotor bar fracture, this research provides probable defect signatures that can be extracted from the current signal. Each defect signature is reported according to (a) loading level, (b) the number of BRBs, (c) validation, and (d) methodologies.

1. Introduction

Induction motors consume roughly 60% of industrial electric energy [1]. Being the backbone of the industrial world, induction motors need to be guaranteed that the machine runs for its intended lifetime without any major failure. In studies by IEEE [2] and EPRI [3], it is reported that 10% of faults are rotor-related (Figure 1).

This damage disrupts the motor’s ability to function safely, puts regular production at risk, and, as a result, incurs severe financial penalties. An efficient incipient defect detection technique can lower maintenance costs by averting costly failures and unplanned downtime. As a result, industrial organizations are working very hard to use an equipment maintenance plan to discover early faults. In order to diagnose the existing failure at an early stage, that is, before it causes the IM to halt, the maintenance plans primarily rely on the IM condition monitoring [4]. The sensor-based approach is one way to investigate faults in electric systems. Due to the benefits of acquiring measurement data that may be gathered for in-depth analysis, modern vehicles, particularly electric vehicles [5], use a lot of sensors. The signals obtained by the sensors and auxiliary instrumentation techniques are used to assess the state of a machine. Signals are the graphical trends of the IM parameters that, when processed properly, can provide a failure signature. Several ways for monitoring the condition of IMs have already been established, which watch over a specific parameter of the IM and allow its health to be assessed. Fundamentally, a condition monitoring method’s effectiveness is determined by its price, accuracy, and of course its capacity to quantify the issue. However, to effectively apply condition monitoring techniques, the user must be knowledgeable and skilled enough to differentiate between a normal operating condition and a probable failure state. The most accurate knowledge of the mechanical and electrical properties of the machines in working order and malfunction depends on the accuracy of the fault detection procedures. Despite the fact that broken rotor bar faults are less frequent than other faults such as bearing faults or insulation damage, they can nonetheless cause enormous damage to electrical machines. Broken bar issues in induction motors are frequently caused by air bubbles that are formed inside the bars during casting, which eventually result in hot spots and small cage cracks with cast aluminum cages [6]. Larger industrial IMs with copper-fabricated cages typically have a different mechanism for this issue. Corrosion, vibrations, and expansion of the bars along the shaft direction due to heat are some of the primary reasons for a broken bar fault [7], and it can be particularly harmful in large motors. In general, the rotor is subjected to a variety of stresses from thermal, mechanical, dynamic, and magnetic sources, which together lead to rotor failures. The following causes can be found related to the breakage of rotor bars [8,9,10,11,12]:

• Thermal stresses can heat the rotor cage during the direct starting of the motor.
• The rotor may experience mechanical stresses as a result of loose lamination and bearing degradation, which could lead to rotor bar cracking.
• Due to centrifugal forces, oscillating shaft torque, and mechanical load oscillation, dynamic stresses can occur on the rotor.
• Vibration, electromagnetic forces, and uneven magnetic pulls may cause magnetic stresses.

When the motor is required to execute rigorous duty cycles, breakage of rotor bars might be a big issue. In this situation, certain dangerous consequences are developed, such as [13,14]:

• Sparking due to broken rotor bars is a major worry in hazardous environments.
• This fault leads to oscillations in the rotor’s speed and torque that hastens the deterioration of the bearings and other driving components.
• Broken rotor bars may lift out of the slot when the rotor rotates at a high radial speed and strike the stator winding, resulting in a disastrous failure of the motor.

As a result, preventing this defect has also grown to become a major priority in the field of condition-based maintenance of electric machines [15,16]. There are numerous condition assessment strategies that can be divided into invasive and non-invasive approaches for rotor fault diagnosis [17,18,19]. Special sensors must be mounted within the motor for invasive techniques such as vibration and magnetic flux. Contrarily, non-invasive methods such as measuring the motor’s current, acoustic signal or instantaneous power have no impact on the device’s internal design. Non-invasive methodology approaches based on motor current signature analysis (MCSA) [14,20,21,22] have dominated for years and can eliminate the requirement for additional hardware complexity. Hence, reviewing these non-invasive diagnosis methods is necessary since they provide an inexpensive and highly sensitive way to monitor a variety of heavy industrial machinery online. The analysis of sideband frequency components (SFCs) connected to the fault is the foundation of the majority of MCSA approaches. The theoretical background regarding the appearance of SFCs has been discussed in [23].

Due to rotor bar breakage, the current sharing in two neighboring bars increases [24].

The irregular MMF can therefore be represented as

$$\begin{array}{c}\hfill F_s=F_{m}cos(n{\theta }^{{}^{\prime }}\pm s\omega t)\end{array}$$

(1)

here, $\omega$ is the supply frequency, the space angle with respect to the rotor is ${\theta }^{{}^{\prime }}$, s denotes the slip, and n is an integer (1, 2, 3 …). The induced voltage due to MMF components in the stator winding can be expressed as per (1), as follows:

$$\begin{array}{cc}\hfill v_s& =v_{m}cos(n\theta +(n\left(\frac{1-s}{p}\right)\pm s)\omega t)\hfill \end{array}$$

(2)

Here, p refers to the number of pole pairs, and n can have values of p, 5p, 7p, 11p, 13p, and so forth to match stator pole pairs. The space angle in reference to the stator can be determined as:

$$\begin{array}{cc}\hfill \theta & ={\theta }^{{}^{\prime }}+\left(\frac{1-s}{p}\right)\omega \hfill \end{array}$$

Therefore, in general, the components of the current spectrum can be represented as:

$$\begin{array}{cc}\hfill f_s& =\left(k\right(1-s)\pm s)f,k=1,5,7,\dots \hfill \end{array}$$

(3)

The fault frequency components $(1-2s)f$ are obtained by k = 1. On the other side, the slip frequency component interacts with the air gap field and causes oscillation at twice the slip frequency ($2sf$) in the electromagnetic torque as well as in speed. The motor current then has both the $(1-2s)f$ and $(1+2s)f$ components as a result of the phase modulation by $\pm 2sf$. This magnetic phenomenon results in the introduction of additional current harmonics as:

$$\begin{array}{cc}\hfill f_b& =(1\pm 2ns)f,n=1,2,3,\dots \hfill \end{array}$$

(4)

The frequency sideband associated with the speed ripple, $(1+2s)f$, is less dominant than the sideband associated with $(1-2s)f$, since it is the secondary effect.

A categorization based on the loading level, quantity of broken bars, validation method, and signal processing techniques is presented for each fault signature to discuss and analyze the fault diagnosis methodologies, as shown in Figure 2.

The rest of the article is divided into the following sections. The proposed classification criteria are presented in Section 2. Following that, Section 3 discusses the fault detection methods corresponding to the aforementioned classes. Then, in Section 4, the primary difficulties for rotor fault detection are covered. A few research gaps are listed in Section 5. Finally, Section 6 presents the conclusion.

2. Review Categories

Here, a review report on broken rotor bar fault analysis methods is presented. The number of BRBs, the loading level, the validation methodology, the signal processing techniques, and the existence of additional faults were all taken into account when classifying various research studies.

2.1. Loading Level

The motor loading level is crucial for examining rotor faults, particularly when sideband frequencies are observed in the steady-state current spectrum. These frequencies are located far apart from one another under full load conditions, making them simple to detect. On the other hand, these overlap with the fundamental frequency under no load or light load conditions. In general, three stages of loading—namely, no load (NL), medium load (ML), and full load—are taken into account while examining the rotor defect (FL). The NL term is regarded as a case with a light load as well. Despite the difficulties, a no-load study must be performed to investigate a BRB failure to avoid the influence of load fluctuations. Numerous studies have demonstrated that load fluctuations occasionally may have an impact similar to that of a broken rotor failure [25,26,27]. Hence, in industrial applications, the no-load motor analysis may be used to minimize the expense of the machinery health monitoring, while also achieving the benefit of keeping faults and load-induced current oscillation separate.

2.2. Number of Broken Bars

The different numbers of BRBs were taken into account when detecting IM BRBs. One, two, or even more broken bars have been considered by several researchers. However, the goal should be to identify the motor’s problem at a very early stage, thus it is crucial to conduct studies using a motor with one broken rotor bar. Advances signal processing based diagnostic tools are required because it is very difficult to detect a malfunction at this early stage, since the machine operates almost normally in this circumstance.

2.3. Validation Methodology

It is suggested that the validation for the BRB defects diagnostic be divided into three groups. While some researchers only use simulations of systems, others only utilize experimental systems. The most effective approach is to simulate the system and empirically validate the method. The MATLAB software package or FEM in ANSYS workbench is frequently used to confirm the simulation-based diagnosis procedure. For data acquisition, the LabView software is generally used for direct interference with the system when performing practical validation using a test rig [28,29].

2.4. Signal Processing Technique

A number of signal processing methods, including the fast Fourier transform (FFT), Hilbert–Huang transform (HHT), short-time Fourier transform (STFT), Wigner–Ville distribution (WVD), Discrete wavelet transform (DWT), continuous wavelet transform (CWT), etc., can be used to identify BRB. For improved diagnosis, the required signal processing techniques should be used in accordance with their purpose. The most common signal processing techniques are briefly explained in this section.

2.4.1. Fast Fourier Transform

The Fourier transform is the most popular technique for signal analysis in the frequency domain. It is a mathematical method that results in a function of frequency, X($\omega$), from a function of time, x(t). All of the frequency-related information can be easily investigated in this domain. An inverse Fourier transform allows for the reconstruction of the original signal. The discrete Fourier transform (DFT) was developed to study the frequencies in time-domain signal. However, it needs a large number of data points, so FFT was developed as a modification of DFT for faster computation [30,31]. The equation of DFT for N data samples can be given as:

$$\begin{array}{c}\hfill X\left(k\right)=\sum _{n=0}^{N-1}x\left(n\right)e^{-j2\pi k(n/N)}k=0,1,2\dots ,(N-1)\end{array}$$

(5)

After creating the signal’s FFT coefficients, the amplitude and phase versus frequency plot can be created using the equation below:

$$\begin{array}{cc}\hfill Amplitudespectrum& =\frac{\sqrt{Re{\left(X\left[k\right]\right)}^2+Im{\left(X\left[k\right]\right)}^2}}{N}\hfill \end{array}$$

(6)

$$\begin{array}{cc}\hfill Phasespectrum& =arctan(\frac{Re{\left(X\left[k\right]\right)}^2}{Im{\left(X\left[k\right]\right)}^2})\hfill \end{array}$$

(7)

The fault frequencies associated with the occurrence of rotor faults in the amplitude spectrum of stator current are depicted in Figure 3. These frequencies are distinct when the motor is fully loaded, but when there is no load, they overlap with the fundamental component and are therefore impossible to detect.

2.4.2. Hilbert–Huang Transform (HHT)

The Hilbert–Huang transform (HHT) is a combination of the Hilbert spectral analysis (HSA) and empirical mode decomposition (EMD). Using the EMD methodology, the signal is divided into intrinsic mode functions (IMFs), and the HSA method is then applied to the IMFs to obtain instantaneous frequency data. Any function can be referred to as an IMF if it meets the following criteria: First, the number of extrema and the number of zero-crossings in the overall data set must either be identical or deviate by no more than one. Second, the envelope obtained from the local maxima and the envelope obtained from the local minima have zero mean values at every point. IMFs can be extracted by shifting, which can be described as follows:

• Identification of all local extremas in the test data.
• Upper envelope detection by connecting all local maxima.
• Lower envelope detection by connecting all local minima.

Considering $m_1$ to be the mean of the upper and lower envelope, the first component $m_1$ can be given as the difference between the data ($X\left(t\right)$) and $m_1$:

$$\begin{array}{cc}\hfill h_1& =X\left(t\right)-m_1\hfill \end{array}$$

(8)

$h_1$ can be treated as a proto-IMF if it meets the criteria for an IMF. Now, it can be used as data in the subsequent sifting procedure. $h_{11}$ can be generated in the next stage as:

$$\begin{array}{cc}\hfill h_{11}& =h_1-m_{11}\hfill \end{array}$$

(9)

After k times shifting, h${}_1$ becomes an IMF:

$$\begin{array}{cc}\hfill h_{1k}& =h_{1(k-1)}-m_{1k}\hfill \end{array}$$

(10)

Hence, the first IMF component is denoted as $c_1=h_{1k}$.

The number of sifting steps required to create an IMF is determined by the stopping condition. The four stopping criteria generally in use are as follows:

(a)

The sifting should stop when the sum of differences (SD) is lower than a pre-specified value, where SD can be given as:

$$\begin{array}{cc}\hfill S{D}_k& =\sum _{t=0}^T\frac{{|h_{k-1}\left(t\right)-h_k\left(t\right)|}^2}{h_{k-1}^2\left(t\right)}\hfill \end{array}$$

(11)

(b)

The second condition is based on the S-number, which states that the sifting process will only come to an end if zero-crossing events and the number of extrema remain the same, or only vary by one throughout the course of S consecutive siftings.

(c)

The third criterion is known as the threshold method, where shifting should stop when global fluctuations remain in between two predefined threshold values.

(d)

The final criterion was developed based on the tracking of energy differences. By using $X\left(t\right)-c_1=r_1$, the first IMF c1 may be distinguished from the remaining data. Here, $r_1$ is the residue. If $r_1$ still has a longer period fluctuations in it, it is categorized as fresh data and put through a similar sifting procedure. The process can be carried out for each subsequent $r_j$’s. If $r_n$ (the residue) turns into a monotonic function and extraction of IMF is not further possible, the sifting process comes to an end.

After obtaining the IMF components, the Hilbert transform may be used to calculate the instantaneous frequency. After applying the Hilbert transform to each IMF component, the original data can be obtained as the real part, as follows:

$$\begin{array}{cc}\hfill X\left(t\right)& =Real\sum _{j=1}^n{a}_j\left(t\right)e^{i\int {\omega }_j\left(t\right)}dt\hfill \end{array}$$

(12)

The EMD-based signal breakdown into the IMFs is shown in Figure 4 and Figure 5 for both the healthy and defective cases, respectively.

2.4.3. Short-Time Fourier Transform (STFT)

Gabor developed windowed Fourier atoms to analyze the frequency changes in sounds [32]. The expression for a real, symmetric window $g\left(t\right)$, which is time-shifted by u and modulated by frequency $\xi$, is

$$\begin{array}{cc}\hfill g_{u,\xi }\left(t\right)& =e^{i\xi t}g(t-u)\hfill \end{array}$$

(13)

This window must be normalized so that:

$$\begin{array}{c}\hfill \left|\left|g_{u,\xi }\right|\right|=1forany(u,\xi )\in \mathbb{R}\end{array}$$

(14)

It is possible to calculate a windowed Fourier transform of a signal $f\left(t\right)$ by first multiplying the original signal by the window and then finally taking the Fourier transform.

$$\begin{array}{cc}\hfill Sf(u,\xi )& ={\int }_{-\infty }^{+\infty }f\left(t\right)g(t-u)e^{-i\xi t}dt\hfill \end{array}$$

(15)

As the Fourier integral being concentrated is close to $t=u$ when the transform is multiplied by $g(t-u)$, this transform is sometimes referred to as the short-time Fourier transform (STFT). The following formulas can be used to determine the energy density ($P_s$), also referred to as a spectrogram:

$$P_{s}f(u,\xi )={\left|Sf\right(u,\xi \left)\right|}^2$$

(16)

$$\begin{array}{cc}\hfill & \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}={|{\int }_{-\infty }^{+\infty }f\left(t\right)g(t-u)e^{-i\xi t}dt|}^2\hfill \end{array}$$

(17)

The STFT-based current analysis is presented in Figure 6. The ‘V’ pattern can be used to demonstrate the existence of a BRB fault. It is visible that the fault characteristic is prominent when there is a rotor fault, whereas in the case of a healthy condition, it is absent.

2.4.4. Wigner–Ville Distribution

When analyzing non-stationary signals, the Wigner–Ville distribution, which offers information on the time–frequency plane, is extremely helpful. It has the lowest time–frequency resolution of any other time–frequency distribution. However, the WVD experiences the so-called cross-terms when the signals contain several frequency components [33]. For a given time series $s\left(t\right)$, the non-stationary autocorrelation function can be expressed as:

$$\begin{array}{cc}\hfill R_x(t_1,t_2)& =\langle (x\left[t_1\right]-\mu \left[t_1\right]){(x\left[t_2\right]-\mu \left[t_2\right])}^{*}\rangle \hfill \end{array}$$

(18)

where the $\langle ..\rangle$ operator takes the average of all possible realizations of the process, and $\mu (.)$ denotes the mean. Now, the Wigner function $W_s(t,f)$ can be defined by the autocorrelation function as a function of the average time $t=(t_1+t_2)/2$ and the time lag $\tau =(t_1-t_2)$, finally taking the fourier transform as:

$$\begin{array}{cc}\hfill W_s(t,f)& ={\int }_{-\infty }^{\infty }{R}_x(t+\tau /2,t-\tau /2)e^{(-2\pi j\tau f)d\tau }\hfill \end{array}$$

(19)

Cross-terms are produced by this distribution because it is not a linear transform, which is a significant drawback in the Wigner–Ville distribution. Several strategies have been put forward in the literature [34,35,36] to lessen the cross-term difficulty, some of which have resulted in new transforms suh as the Smoothed pseudo-Wigner–Ville distribution (SPWVD), Choi Williams distribution, Cohen’s class distribution, etc. The BRB characteristic generated by the Wigner–Ville distribution (WVD) is shown in Figure 7. Despite displaying the ‘V’ pattern associated with the BRB fault in the T–F plane, cross-terms are also evident, which is the fundamental disadvantage of WVD-based analysis.

2.4.5. Discrete Wavelets Transform (DWT)

DWT [37] for a discrete signal $X\left[k\right]$, $k=0,1,2,\dots \dots N-1$ having a frequency bandwidth [0,$\frac{F_s}{2}$] (here $F_s$ is sampling frequency and N is the number of samples) decomposes into approximation and detail coefficients.

At the lth level, the detail and the approximation coefficients will be $D_l$ and $A_l$ with frequency bandwidths of [$\frac{F_s}{2^{l+1}}$, $\frac{F_s}{2^l}$ ] and [0,$\frac{F_s}{2^{l+1}}$ ], respectively. Figure 8 illustrates how rotor faults impact various decomposition levels in DWT-based current analysis when three-level decomposition of motor current is implemented.

2.4.6. Continuous Wavelet Transform (CWT)

The convolution-based algorithm and the FFT-based algorithm are the two widely used techniques for computing wavelet coefficients [38]. At a time location $\tau$, for a mother wavelet $\psi \left(t\right)$, CWT coefficients obtained from the first method can be represented as follows:

$$\begin{array}{cc}\hfill T_x(\tau ,s)& =(x\ast {\overline{\psi }}_s)\left(\tau \right)\hfill \end{array}$$

(20)

here, ${\overline{\psi }}_s$ = $\frac{1}{\sqrt{s}}{\psi }^{*}\left(\frac{-t}{s}\right)$. The computation begins at $\tau$ = 0 and $\tau$ must be increased up until the signal’s termination. For each time step, the CWT coefficients must be calculated.

On the contrary, the FFT-based approach is more computationally efficient than the first, since it does not call for the inner-loop calculation associated with the translation parameter $\tau$. In this approach, the wavelet transform can be expressed as:

$$T_x(\tau ,s)={\int }_{-\infty }^{\infty }x\left(t\right){\psi }_{\tau ,s}^{*}\left(t\right)dt$$

(21)

$$\begin{array}{cc}\hfill & \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}={\int }_{-\infty }^{\infty }x\left(t\right)\frac{1}{\sqrt{s}}{\psi }^{*}(\frac{t-\tau }{s})dt\hfill \end{array}$$

(22)

where, respectively, $\tau$ and s stand for the translation and dilation parameters (scale). The wavelet transform coefficients can be used to recover the signal using the following equation.

$$\begin{array}{c}\hfill x\left(t\right)=\frac{1}{C_{\psi }}{\int }_{-\infty }^{\infty }{\int }_0^{\infty }{T}_x(\tau ,s){\psi }_{\tau ,s}\left(t\right)\frac{1}{s^2}dsd\tau \end{array}$$

(23)

where $C_{\psi }$ is defined as ${\int }_0^{\infty }\frac{{\psi }_{\omega }^{*}{\psi }_{\omega }}{\omega }d\omega$. The wavelet should meet the zero average criterion, i.e., ${\int }_{-\infty }^{\infty }\psi \left(t\right)dt=0$, and the admissibility criterion, i.e., $C_{\psi }<\infty$, for both these analysis and synthesis purposes. Along the time dimension, Fourier transform of the CWT can be expressed as

$$\begin{array}{c}\hfill ϝ\left[T_x(\tau ,s)\right]=\sqrt{s}X\left(\omega \right){\psi }^{*}\left(s\omega \right)\end{array}$$

(24)

Finally, the CWT coefficients can be obtained by taking the inverse Fourier transform as:

$$\begin{array}{cc}\hfill T_x(\tau ,s)& =\frac{1}{2\pi }{\int }_{-\infty }^{\infty }\sqrt{s}X\left(\omega \right){\psi }^{*}\left(s\omega \right)e^{j\omega \tau }d\tau \hfill \end{array}$$

(25)

Two loops are required to carry out the computation in the first (convolution-based) method. The scales (wavelet’s dilation parameter) must be tracked in the first loop, and another loop is needed to track the time position $\tau$. The computation must be performed for each scale starting at $\tau$ = 0 and continuing until the signal reaches its end location. However the benefit of using the second method (FFT-based) is the exclusion of the second loop in contrast to the convolution-based technique. CWT is in this case is evaluated for all values of t in a single run for a certain scale. As a result, the FFT-based method speeds up the total CWT calculation while reducing computing complexity. CWT offers a time-scale representation of a signal. A specific frequency range will be present in the reconstructed signal using coefficients for a given scale. Figure 9 displays an example of this kind of analysis, where the expected fault feature can be observed, which becomes more intense with an increase in the number of broken rotor bars.

The benefits and limitations of each of these signal processing tools are detailed in

Table 1. Comparison of signal processing methods related to BRB detection [8].

Signal Processing Tool	Computational Complexity	Representation	Advantages	Drawbacks
Fast Fourier transform	$NlogN$	PSD	Easy to implement.	Poor resolution.
			Fast and computationally efficient.	Not applicable for non-stationary signal.
Hilbert–Huang transform	$NlogN$	Hilbert Spectrum	Valid for non-stationary signal.	EMD is sensitive to stopping criteria.
			Avoid dynamic frequency decomposition.
			The uncertainty principle does not impose limitations on time–frequency resolution.
Short-time Fourier transform	$NlogN$	Spectrogram	Valid for non-stationary signal.	Basis function is fixed.
			Easy to implement compared with other time–frequency analysis.	Problem of energy smearing in T-F plane.
Wigner–Ville distribution	$N^{2}logN$	Spectrogram	Provides best time–frequency resolution.	Presence of cross terms if the signal is not a mono component.
			Basis function is adaptive in nature as it is derived from the signal itself.	Sensitive to noise.
Discrete wavelets transform	$N^{2}logN$	Scalogram	Provide perfect reconstruction of the signal upon inversion	DWT requires the data sample size to be an integer multiple of two for full decomposition of signal
			DWT is more computationally efficient than other transformations because of its excellent localization properties.	DWT coefficients are sensitive to the time shifting of signal.
			The DWT provides a sparse representation for the signals.
Continuous wavelet transform	$N^{2}logN$	Scalogram	Valid for non-stationary signal.	Selection of optimal mother wavelet is a challenge.
			Higher time–frequency resolution compared with STFT	There is overlap between frequency bands.

. Hence, a particular tool has to be chosen based on the applications.

3. Fault Analysis

The BRB fault can be found using a variety of fault signatures. The motor current, which is used in non-invasive methods, is the most typical defect signature. An overview of the detection procedure employing the stator current fault feature is provided in two subsections. Diagnoses for the steady-state and transient start-up cases are presented in that order.

3.1. Steady-State Analysis

Although typical FFT-based approaches are the most straightforward way to find fault frequencies, the issue is that if these frequencies are located very closely, conventional methods for spectrum analysis lack the necessary resolution to distinguish these, so tools for high-resolution frequency-domain analysis have been employed. Multiresolution Taylor–Kalman Approach [39], Prony analysis [22] Modified Prony Method [40], ESPRIT [41], and Root-MUSIC analysis [42] have been demonstrated to be effective in resolving spectral resolution issues. For better diagnosis, several advanced signal processing technologies have been deployed. Using DWT to identify broken rotor bars yields an accurate result. However, the primary challenge is choosing the best wavelet. Different wavelet functions were compared [43] for both NL and VL circumstances. Time–frequency approaches, such as STFT, CWT, Wigner–Ville representation, and HHT, were utilized to detect BRB in variable-speed turbine generators [33] while outlining the benefits and drawbacks of each representation.

Table 2. Diagnostic classification based on MCSAs using steady-state current.

Number of BRB	One and Two									Multiple
Validation	Simulation			Experimental			Simulation and Experimental			Simulation			Experimental			Simulation and Experimental
Loading Level	NL	ML	FL	NL	ML	FL	NL	ML	FL	NL	ML	FL	NL	ML	FL	NL	ML	FL
FFT				[44,45,46]	[44,45,46]	[45,46,47]	[48]	[48]	[48]					[49]
HHT				[50]	[50]	[50]			[51]				[50]	[50]	[50]		[52]
STFT									[53]
WVD	[33]								[51]
DWT	[43]	[43]	[43]		[54]						[55]			[56]	[56]	[57]	[57]	[57]
CWT	[58]	[58]	[58]

summarizes the diagnostic classification based on MCS utilizinging steady-state current.

When multiple defect detection is necessary, the use of artificial intelligence and machine-learning-based techniques has also been shown to be quite useful. Fuzzy logic [59], artificial ant clustering [60], the hybrid FMM-CART model [61], and convolution neural networks (CNN) [62] have all been employed as useful diagnostic tools in this context to find faults such as broken rotor bars, eccentricity issues, imbalanced voltage, bearing damage, etc. Despite the high efficiency of neural-network-based diagnosis techniques, especially for online defect detection, they have a number of limitations, including the following:

• Due to the enormous number of parameters that need to be tuned, network parameter tuning is challenging.
• The lengthy learning process as a result of the significant computing load.
• It occasionally becomes trapped on local optima, which reduces effectiveness.

Finding optimum membership functions necessitates strong domain knowledge, such as in the case of fuzzy-logic-based techniques.

Table 3. Recent MCSAs based on artificial intelligence.

Reference	Method	Detected Fault	Analyzed Signal	Accuracy Rate
[63]	Hilbert Transform and Fuzzy decision tree	1 BRB	Steady state current signal	98.75%
		2 BRB
[64]	PCA and multivariate relevance vector machine with multiple Gaussian kernels	1 BRB	Steady-state current signal	80–95%
		2 BRB
		3 BRB
[65]	Fuzzy-logi- based approach	1 BRB	Steady-state current signal	98.30%
		2 BRB
[66]	Hilbert transform and statistical analysis	0.5 BRB	Start-up current signal	99%
		1 BRB
		1.5 BRB
[58]	Spectral entropy and tuned SVM	1 BRB	Steady-state current signal	91–100%
		2 BRB

represents some of the recent MCSAs developed for fault classification using artificial intelligence.

3.2. Transient Analysis

In the energy industry, it has been demonstrated that studying transients in terms of frequencies is an effective way to investigate disturbances, which in turn aids in system monitoring [67]. In the field of machinery, condition monitoring also utilizes methods to study start-up current, which has a high signal-to-noise ratio [68] and high slip, suggesting that spectral components related to faults may be easily separated [69]. Advanced Transient Current Signature Analysis (ATCSA) [70,71] has become popular recently in this field. Even in scenarios where MCSA may generate erroneous alarms [26] during machine diagnosis, its value has been proven. Generally, it involves looking at the motor current during the startup transient [72]. The Wigner–Ville distributions (WVD), Discrete wavelet transform (DWT), and Atom-Based Transforms are the most used transform types for determining the time–frequency trajectories. WVDs are not appropriate because, when the evolutions of the frequency components are too close together, the cross-terms produced by this method cannot be removed [73] by a kernel.

Although DWT is an appropriate remedy [74] for this, choosing the right wavelet and determining the ideal number of decomposition levels are its two main difficulties. If the correct wavelet is not selected for the analysis, the anticipated V pattern associated with the rotor failure will not be observed. Wavelet transforms [75,76] have also been employed in numerous research to diagnose induction motors. However, the disadvantage is that they only offer good resolution at higher frequencies. Diagnostic classification based on MCSAs using start-up current transients is summarized in

Table 4. Diagnostic classification based on MCSAs using start-up current.

Number of BRB	One and Two									Multiple
Validation	Simulation			Experimental			Simulation and Experimental			Simulation			Experimental			Simulation and Experimental
Loading Level	NL	ML	FL	NL	ML	FL	NL	ML	FL	NL	ML	FL	NL	ML	FL	NL	ML	FL
HHT				[50,77,78,79,80,81]	[50]	[50]							[82,83]
STFT		[84]			[23,85,86]		[23]	[23]	[23]							[85]
WVD				[87]				[88]
DWT				[74,77,78]	[86]			[89]	[89]	[90]	[90]	[90]	[74]
CWT				[28]			[91]	[91]	[91]

.

4. Major Challenges

In traditional MCSA, the fast Fourier transform (FFT) is used to analyse the current required by the machine during steady-state operation and determine sideband frequency components (SFCs) related to the fault. Even though MCSA is the most popular method for evaluating rotor conditions and is utilized in the majority of commercially available induction motors, the methods that make use of steady-state motor stator currents have some restrictions [92]. These are:

• The machine must run at a constant and known speed.
• Fundamental supply frequency to the stator must remain constant.
• The connected load should be very high to separate the broken bar frequencies from the fundamental one.

Another crucial drawback related to this conventional MCSA is the wrong diagnosis of the machine [93]. The incorrect diagnosis might happen in two different ways. First, there are the false-positive scenarios where load torque oscillations may add frequency components that are comparable to fault-related ones [25]. The presence of axial cooling ducts may also result in the introduction of frequency components that are comparable to those associated with rotor faults when the number of ducts and the number of motor poles are equal [26,27]. On the other hand, false-negative testing, in which the rotor problem goes unnoticed, is the alternative scenario. When machines are running at low slip, the sideband frequencies in the MCSA spectrum may almost overlap with the supply frequency. Due to this, it may be challenging to identify the fault frequencies [94].

5. Research Gaps and Future Research

Over the past few decades, research on the rotor fault diagnosis of IM has been ongoing. However, there are some research possibilities that might be taken into account in the future. These are listed below.

• In most studies, breakage of a single rotor bar is regarded as the early stage of a BRB fault. However, this defect begins with the beginning of a crack on the bar, and the diagnosis tool should be able to detect the fault at this incipient stage. The prognostics of such circumstances provide difficulties for researchers.
• The potential BRB detection methods based on acoustic emission can be investigated, as they are currently the least studied ones.
• In the literature review, there are fewer publications describing rotor fault identification in the presence of additional problems such as eccentricity faults, bearing faults, etc. Future research should focus on developing a diagnosis method that can identify BRB faults even when the machine has additional faults.
• Future research can also investigate the failure of a multimotor system with numerous coupled motors. In this situation, identification of the actual faulty machine is really challenging.
• The Internet of things (IoT) is becoming more applicable in the various engineering sectors in this era of Industry 4.0. Therefore, it would be beneficial for the IoT to be used effectively in the field of rotor health monitoring.

6. Conclusions

This study reviews the techniques for rotor fault detection which are non-invasive, i.e., that do not involve adding sensors to the machines. Fundamentally, the key findings of a successful rotor fault diagnosis approach are built on two concepts:

• Necessary spectral analysis to detect the fault signature with the appropriate sensitivity.
• Establishing a link between the fault signature and fault severity.

Though the analysis of the observed signal typically provides direct information regarding the presence of defects in the machine, a general quantification of their severity is still a challenge. The final section of this study also provides a summary of the different prospective areas of improvement that may be actively addressed and on which research could be concentrated in the future. Though there has been a significant improvement in this field, the commercialization of the practical fault detection technique remains a major problem in machine monitoring approaches. Online monitoring is frequently seen as the best option in this situation, but creating and implementing a system for online monitoring involves several difficult steps.