Mechanical Systems and Signal Processing 144 (2020) 106908 Contents lists available at ScienceDirect Mechanical Systems and Signal Processing journal homepage: www.elsevier.com/locate/ymssp Review Signal based condition monitoring techniques for fault detection and diagnosis of induction motors: A state-of-the-art review Purushottam Gangsar a, Rajiv Tiwari b,⇑ a b Department of Mechanical Engineering, Shri G S Institute of Technology and Science, Indore, Madhya Pradesh 452003, India Department of Mechanical Engineering, Indian Institute of Technology Guwahati, Guwahati, Assam 781039, India a r t i c l e i n f o Article history: Received 20 December 2018 Received in revised form 11 August 2019 Accepted 13 April 2020 Available online 24 April 2020 Keywords: Induction motor (IM) Mechanical and electrical faults Vibration and current signal Multi-fault diagnostic Machine learning algorithms a b s t r a c t Uninterrupted and trouble-free operation of induction motors (IMs) is the compulsion of the modern industries. Firstly, the paper reviews the conventional time and spectrum signal analyses of two most effective type of signals, i.e. the vibration and the current for various IM faults. The vibration and the current signal analyses (time and spectral) is performed using the signals measured from different faulty IMs from a laboratory setup. Subsequently, the advantages and difficulties associated with these conventional procedures are discussed. Next, this paper presents and summarizes the existing research and development in the field of signal based automation of condition monitoring methodologies for the fault detection and diagnosis of various electrical and mechanical faults of IMs. Nowadays, artificial intelligent (AI) methods are being employed for the IM and other machine fault diagnosis. Advancements of the AI based fault diagnosis including the popular approaches are reviewed in details. These techniques are being integrated with traditional monitoring techniques. The AI based fault monitoring and detection techniques for IMs published up to 2000 are briefly described, however, more attention is paid to the techniques that are introduced in roughly past two decades, i.e. during 2000–2019. In overall, this paper includes review of system signals, conventional and advance signal processing techniques; however, it mainly covers, the selection of effective statistical features, AI methods, and associated training and testing strategies for fault diagnostics of IMs. Finally, dedicated discussions on the recent developments, research gaps and future scopes in the fault monitoring and diagnosis of IMs are added. Ó 2020 Elsevier Ltd. All rights reserved. Contents 1. 2. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Various types of faults in induction motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1. Mechanical faults in induction motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1. Bearing fault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.2. Rotor related fault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2. Electrical faults in induction motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ⇑ Corresponding author. E-mail address: [email protected] (R. Tiwari). https://doi.org/10.1016/j.ymssp.2020.106908 0888-3270/Ó 2020 Elsevier Ltd. All rights reserved. 4 4 5 5 5 7 2 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 Nomenclature C CF fs fr kðx; xi Þ K L Pm Rmsd Rpp T1 T2 T3 w x(t) xi yi Soft margin or penalty parameter for support vector machine Crest factor Supply frequency Rotational frequency of rotor Kernel function Number of classes Lagrangian Probability distribution of energy Mean to standard deviation ratio Peak to peak ratio No load Light load High load A normal vector to the hyperplane Time domain signal or data Training vector Fault’s class or label Greek ai c j lr l1 ni r rb /ðxÞ v Langrange multiplier Kernel parameter Kurtosis r-th Statistical moment Mean or first moment Slack variable for support vector machine Standard deviation Width of the RBF kernel Transformation function Skewness Abbreviations AC Alternate current AI Artificial intelligence ANN Artificial Neural Network ART adaptive resonance theory BBS Best basis selection BEF Ball element fault BF Bearing fault BP Back propagation BR Bowed rotor BRB Broken-rotor bar CART Classification and regression tree CBM Condition based maintenance CWT Continues wavelet transform CV Cross validation CVA Common vector approach DAG Direct acyclic graph DAQ Data acquisition system DC Direct current DWT Discrete wavelet transform EOP Emergency operating procedure ERM Empirical risk minimization EWN Evolving wavelet network FL Fuzzy logic FFT Fast Fourier transform FNN Fuzzy neural network GA Genetic algorithm GDA Generalized discrimination analysis HMM Hidden Markov model P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 HHT HOS HT ORF IACO ICA IR IRF LDA LDB LIBSVM MCSA MFS MLBS MLP MR MRA MUSIC ND OASYS ORF OVA OVO PCA PD PDF PSWT PUF PUF1 PUF2 PVM RBF RFE RMS RUWPT RWE SFAM SLBS SOM SRM STFT SVM SWF SWF1 SWF2 SWPT SVs TDA UMP UR VFD WPT WT WVD 3. 3 Hilbert-Huang transform Higher order statistics Hilbert transform Outer race fault Improved ant colony optimization Independent component analysis Infrared Inner race fault Linear discriminant analysis Local discriminant basis A library for support vector machine Motor current signature analysis Machine fault simulator Multi-level basis selection Multilayer perception Misaligned Rotor Multi resolution analysis Multiple signal classification No defect condition of induction motor On-line operator aid system Outer race fault One versus all One versus one Principal component analysis Partial discharge Probability distribution function Pitch synchronous wavelet transform Phase unbalance fault Phase unbalance fault level-1 Phase unbalance fault level-2 Park vector machine Radial basis function Recursive feature elimination Root mean square Recursive un-decimated wavelet packet transform Relative wavelet energy Simplified fuzzy ARTMAP Single-level basis selection Self-organizing map Structural risk minimization Short time Fourier transform Support vector machine Stator winding fault Stator winding fault level-1 Stator winding fault level-2 Stationary wavelet packet transform Support vectors Time domain averaging Unbalanced magnetic pull Unbalanced rotor Variable frequency drive Wavelet packet transform Wavelet transform Wigner-Ville distribution 2.2.1. Stator winding fault (SWF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2.2. Broken rotor bar faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2.3. Phase unbalance and single phasing fault. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Fault monitoring of induction motors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 3.1. Time domain analysis of IM faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 3.2. Frequency domain analysis of IM faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 4 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 4. 5. 6. 3.2.1. Healthy motor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2. Broken rotor bars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.3. Stator winding faults or armature faults. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.4. Bearing faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.5. Air-gap eccentricity related fault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3. Challenges in the time and frequency analyses of vibration and current signals for IM faults . . . . . . . . . . . . . . . . . . . . . . 3.4. Advanced condition monitoring methods for IM faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Artificial intelligence based fault diagnosis of IMs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1. ANN based diagnosis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2. Fuzzy logic based diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3. SVM based fault diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4. Other hybrid AI method based diagnosis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5. New challenges in AI based fault diagnosis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Observations, research gaps and ideas for future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Concluding Remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 12 13 14 16 18 18 19 20 22 22 24 25 30 31 31 1. Introduction Induction motors (IMs) has become the backbone of the industrial world, since they play a vital role in the mechanical power generation, manufacturing and transportation industries. The most common applications of IMs are the pumps, compressors, blowers, fans, machine tools, cranes, conveyors, and electric vehicles. Over 50% of the total electrical energy generated globally is utilized by motors, mainly induction motors (IMs), which consumed around 60% of the industrial electricity [24]. Deployment of IMs in all types of industries is growing steadily owing to their low cost, ruggedness, high power-toweight ratio, and adaptability to a wide variety of operational conditions [231]. Reliability and availability of IMs are crucial to ensure a trouble free and continuous operation in the industry. However, IMs are exposed to many unavoidable stresses, such as the mechanical, electrical, thermal and environmental stresses, throughout the operation due to variation in the external loading, deviation in power supply, surplus heat, inadequate lubrications and incompetent sealing, dusty environment, manufacturing defect and natural aging. This creates some modes of unexpected faults in different components of the motor. These faults usually left unnoticed in its primary stage; and this end up with the catastrophic motor failure. Consequently, the industry has to bear a huge financial loss due to the process shutdown or sometimes severe human wounds. Consequently, to avoid the likelihoods of catastrophic motor failure, early fault detection and diagnosis of the IM components that begin to degrade is done in the industry [83]. The condition monitoring provides a continuous assessment of various components of IMs throughout its serviceable life. Early diagnosis of IM faults, offers adequate warning of imminent failures, condition based maintenance and minimum downtime [56,98,54,182,202,23,107]. 2. Various types of faults in induction motors The IM mainly comprises of three components which are the stator, the rotor and the bearings (as shown in Fig. 1). The IM fails due to damage in any of its components or subcomponents. The faults in IM are broadly categorized as either electrical or mechanical faults, as shown in Fig. 2 [139,231,83,6,66]. Stator core with field winding Cover plate 1 Rotor shaft Fan Housing Stator Cover plate 2 Rotor Fan d Foundation Bearing 1 Rotor bar End ring Fig. 1. The schematic exploded view of an induction motor. Bearing 2 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 5 Induction motor faults Electrical faults Stator faults Turn-to-turn fault Coil-to-coil fault Phase-to-phase fault Phase-to-ground fault Rotor faults Broken rotor bar Broken end ring Mechanical Faults Electrical supply faults Phase unbalance Single phasing Bearing faults Outer race fault Inner race fault Rolling element fault Cage/train fault Rotor faults Unbalanced rotor Bowed rotor Misaligned rotor Fig. 2. Types of common induction motor faults. The expected percentage of failure of IM components is described as [5,28] (i) Bearing damages: 41–42%, (ii) Stator winding damages: 28–36%, (iii) Rotor related damages: 8–9% and (iv) Other damages: 14–28%. The bearing fault and the stator winding fault together account for at least 69% of total motor faults. Although, the percentage failures of rotor and shaft are very low, the maximum bearing damages occur due the eccentric rotor, misaligned and unbalanced shaft, and occasionally winding failure that results due to rubbing of the faulty (misaligned and broken) rotor. The most of winding failures occur due to unbalanced power supply. If the fault coverage is extended to include the rotor related faults (such as the broken rotor bar, misaligned rotor, and unbalanced and bowed rotor) and other IM faults (such as, the phase unbalance and single phasing), more than 90% of all fault modes are covered. Now different mechanical and electrical faults of IMs are discussed in detail. 2.1. Mechanical faults in induction motors About 45–55% of IM failures occur due to mechanical faults. Now, various mechanical faults are described. 2.1.1. Bearing fault The most of IMs use either the ball or roller bearing, which consists of the rolling element, the outer race, the inner race, and the train (or cage) as shown in Fig. 3. The main causes of the bearing failures are: (i) high vibration of the rotor due to large output load torque that ultimately leads to high fatigue stress, (ii) improper installation of bearing, (iii) deterioration of lubrication due to shaft voltage directing to high current in bearing, (iv) due to heat conducted through the shaft, and (v) ultimately friction and contamination. The bearing faults make the rotor eccentric that produces unbalanced magnetic pull (UMP), it is described in next section, and gives additional load on the bearings. The bearing fault is one of the reasons of excessive vibration in the motor as the shaft dynamics are influenced by the distorted air–gap between the stator and the rotor, and also the due to change in the bearing stiffness. The ultimate effects of bearing faults are rotor bar failures, which produce a premature breakdown of IMs. For more details of rolling bearing faults and associated vibration characteristic frequencies readers may refer to Tiwari [202]. 2.1.2. Rotor related fault For no defect in IM, the air–gap between stator and rotor remains same as the axis of rotation of the rotor is similar as the geometrical axis of stator, and the rotor is centrally aligned with the stator. However, if the rotational axis is different from the geometrical axis of stator then the air–gap varies, and the condition is stated as the air–gap eccentricity. Indeed, it is a most frequent rotor fault of IMs. Eccentricities is defined as the static and the dynamic, which are shown in Fig. 4. In the static eccentricity, the position of the minimal radial air–gap length is fixed in space; while in the dynamic eccentricity, the position of minimal air–gap rotates with the rotor Antonino-Daviu et al. [11]. Large eccentricity produces appreciable 6 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 Outer race Rolling element Outer race fault Bearing cage or train Inner race Fig. 3. A solid model of the bearing fault in the IM. unbalance radial forces, especially at high speeds, which result in the rub between the stator and the rotor, and ultimately the stator and/or the rotor damage. The eccentricity may be caused by a relative misalignment of the stator and the rotor in the commissioning stage, the misalignment between the shaft axis of motor and the shaft axis of load, incorrect installation of bearings, rotor unbalance load, bearing wear, and mechanical resonance during passing through the critical speed [59]. 2.1.2.1. Rotor misalignment. The rotor misalignment or misaligned rotor (MR) is of two types, i.e. the parallel and angular misalignments (as shown in Fig. 5) or its combination may take place. In the parallel misalignment, the minimal air–gap between the stator and the rotor is fixed, and it is caused by inappropriate positioning of the rotor and the stator core at the commissioning time. In the case of angular misalignment, the centre of rotor does not coincide with the centre of rotation of the rotor, and the minimal air–gap rotates with the rotor; and it is due to several reasons, like the bent rotor shaft, unbalance, and bearing wear. Maximum 10% air–gap eccentricity is permitted in IMs. The MR creates the static air–gap eccentricity. When it crosses a permissible limit, the resulting unbalance force can cause rotor to the stator rubbing, core damage and ultimately the destruction of the IM. 2.1.2.2. Bowed rotor. The bowed rotor (BR) and bent rotor are actually same phenomenon, but only difference is that the bowed rotor deformation is measurable inside the machine housing (or on mounting the rotor on bearings) and it is due to static weight of rotor, while the bent rotor deformation can be observed outside the machine also (without mounting on bearing), which is due to permanent deformation of the rotor (as shown in Fig. 6). It creates the dynamic air–gap eccen- Stator Rotor Rotor geometric center Stator geometric center Fig. 4. The air–gap eccentricity in the IM during rotor motion (a) normal motor (b) motor with static eccentricity (c) motor with dynamic eccentricity. Bearing end (a) Stator axis Rotor axis (b) Fig. 5. The parallel and angular rotor misalignment in the IM. (a) parallel misalignment (b) angular misalignment. P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 7 tricity. The BR is caused by local rubbing (resulting in permanent metallurgical changes), local expansion and yielding (resulting in permanent bowing and cracking), weight of the rotor on sitting the rotor stationary for a long time (creep), and residual stresses. The ultimate effect of BR is the rotor misalignment, rotor-to-stator rubbing, and ultimately damage of the motors [159,158]. 2.1.2.3. Unbalanced rotor (UR). The unbalance in a rotor is defined as an uneven distribution of mass about centre of rotation of the rotor. This is also a primary cause of vibration in IMs. The UR consists of the static and dynamic unbalances. The static unbalance is due to unbalanced forces in a single transverse plane, while the dynamic unbalance is due to unbalance forces in different transverse planes, which as well generate the unbalance couple. The UR generates extreme centrifugal forces and vibrations during operation that degrade the life of rotor, bearings, coupling, seals, and gears. A small amount of UR may cause severe problems in high speed induction motors. In actual practice, a rotor has unbalances because of the manufacturing defects, uneven material density, and loss or gain of material while the rotor is in running condition. The unbalance rotor in IMs is caused by mainly due to the broken rotor bar, resulting in total damage of the motor. 2.2. Electrical faults in induction motors About 35–40% failures of IM take place due to electrical faults. Various electrical faults are described below: 2.2.1. Stator winding fault (SWF) The SWF can occur due to the turn-to-turn, coil-to-coil, phase-to-phase or phase-to-ground fault as shown in Fig. 7. Most of the winding faults are the consequence of the growth of undetected turn-to-turn defects. The main cause of turn-to-turn defect is long term thermal aging and ultimately insulation failure. The SWF can occur due to the excessive heating (thermal stress), the unbalance power supply (electrical stress), hitting by the broken, unbalanced, or misaligned rotor bar (mechanical stress), the vibration, the failure during installation, and the contamination by oil. The SWF may lead to the opening, shorting or grounding of one or more circuit of winding, excessive heating and total damage of machines. These faults produce non-uniform magnetic field in the air–gap of IMs that causes unbalanced air–gap voltage and line current, high mechanical vibration and increased torque pulsation. Bearing end Stator axis Centrally bent rotor axis Fig. 6. The bowed rotor in the IM. b Phase to phase c s Turn Coil to coil Phase to ground a Fig. 7. Various possible faults in stator winding of an IM. 8 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 BRB Rotor shaft Cage (bars) Cage (end ring) Fig. 8. A solid model of the broken rotor bar in an IM. 2.2.2. Broken rotor bar faults Though the rotor of a squirrel cage IM is extremely rugged, the broken rotor bar (BRB) occurs due to failure of rotor bars and end rings. IM rotor faults are mainly BRB (as shown in Fig. 8), and the main cause of these faults is the fluctuating load and direct online starting [174]. The BRB can occur due to frequent start at rated voltage, thermal unbalance and over-loaded during starting (thermal stress), unbalanced magnetic pull, electromagnetic force and noise (magnetic stress), and fabrication defect (mechanical stress)[61]. The main cause of BRB is electrical related, therefore it is characterized in the electrical faults. These faults produce localized heating or arcing in the rotor, vibration due to expansion of rotor and bowing, stray leakage flux, fluctuation of speed, supply current and torque pulsation [93]. 2.2.3. Phase unbalance and single phasing fault The phase unbalance produced when the voltages are not equal (unbalanced) between three phases. The small amount of voltage variation causes drastically increase of current in the motor winding, if allowed for some time, then the total damage of the motor may occur by overheating. The single phasing occurs when the one phase of the IM is suddenly open-circuited during operation. This is caused by loose connections, blown fuse and partial damage of switch gears. The single phasing effect is similar to the voltage unbalance and the worst possible case of voltage unbalance is the single phasing. It causes the overheating, shaft vibration noise, ultimately insulation damage and stator winding failures. 3. Fault monitoring of induction motors This section presents a literature survey to shed light on various techniques used and progress made in the fault detection and diagnosis of IMs. The catastrophic IM failure can occur due to failure of any of its components such as the stator, rotor and bearing. The fault in the different components of IMs produces one or more of the symptoms [28], Zhongming and Bin [236], including (i) unbalanced air–gap, (ii) flux, (iii) increased torque and speed variation, (iv) decreased average torque, (v) excessive heating, (vi) excessive mechanical vibrations, (vii) decreasing efficiency and increasing losses, and (viii) deviation and asymmetry of currents and voltages. In order to detect these fault symptoms of IMs, a number of condition monitoring techniques have been developed [139,136,6], (i) magnetic flux (or axial leakage flux), including (i) air gap torque monitoring, (ii) acoustic noise measurement, (iii) thermal monitoring, (iv) partial discharge measurement, (v) instantaneous angular speed, (vi) instantaneous power, (vii) surge testing, (viii) chemical analysis (lubricating oil debris; cooling gas), (ix) vibration monitoring, and (x) current monitoring. The magnetic flux monitoring can be used to identify the turn-to-turn SWF, BRB, and rotor eccentricity by analyzing axial magnetic flux in the shaft using a search coil looped concentrically around the motor shaft [84]. This method can be used to locate the turn-to-turn fault position by mounting at least of four coils axis-symmetrically to motor shaft [149]. However, this method cannot be employed widely in the industry due to installing search coils. The air–gap torque is used to monitor the BRB, SWF and eccentricity related faults [86]. It denotes the combined effect of entire flux linkages and the currents in both rotor and stator of the motor [25,116]. It is very sensitive to any unbalance created by the unbalanced voltages as well as by the defects [167,79]. The acoustic noise monitoring is performed by investigating the noise spectrum [4]. The noise is produced as a result of Maxwell’s stresses that act on the iron surfaces. The most commonly used noise measurement sensors are microphones, or specialized equipment to measure noise (sound level meter). This monitoring technique are applied as a supplementary technique in the existence of strong noise and interferences for detection of the air–gap eccentricity and stator structure related faults [126,50]. The thermal monitoring is implemented by measuring the local temperature of the IM, and the model based or parameter based approaches [186]. In order to measure the local temperature, the sensors are usually installed on the motor winding or inserted in the insulation, which is electronically isolated from its instrumentation. Thermocouples, embedded temperature P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 9 detectors, or resistance temperature detector are utilized for this purpose. These signals are generally utilized to monitor particular regions of the stator core and the bearing. In addition, thermal images are also used to identify a number of IM faults, such as the SWF, BF, misalignment and mass-unbalance [226,127]. The partial discharge (PD) method can be only used for the insulation related fault. The PD are explained as a small discharge in a gas filled void or a dielectric surface of a liquid or a solid insulation system. The PD occurs owing to the insulation imperfection. The major reasons of PD are the manufacturing defect, delamination within the ground wall insulations or overheating, consequently air pockets or voids in the insulation. A degraded winding produces approximately thirty times higher PD activity than a healthy winding [189,196]. The instantaneous angular speed waveform is used in time and frequency domain for monitoring of specific IM faults such as the BRB [174]. Fault diagnostics of IMs, mainly the rotor cage related, can be done based on instantaneous power signature analysis [53,3]. Other methods, like the surge testing, which uses impedance matching under an applied surge, have also been used for monitoring of winding and eccentricity related faults [205,90]. The cooling gas method are also applied for the fault identification of insulation and stator winding. The degradation of insulation produces carbon mono-oxide gas, which goes into the cooling air circuit and are detected by an infrared (IR) absorption method [195]. The lubricating oil debris method can be used for the bearing damage by analyzing the wear debris particles and impurities in the lubricating oil [81]. The vibration is generally used for the diagnosis of mechanical faults in IMs, for example the BF, and the unbalanced, bowed and misaligned rotors [96]. Besides, this can also be used to detect the electrical faults of IMs; for example, the BRB, SWF and phase unbalance [39,134]. In order to acquire vibration for the vibration monitoring, the sensor like the accelerometer can be mounted at the appropriate location on machines [208,225]. When a small fault occurs, it varies the dynamics of the systems, consequently large deviations appears in the vibration patterns. These deviations by faulty components can be detected using a suitable data algorithm, and the motor status can be assessed. The current monitoring or machine current signature analysis (MCSA) is generally used to diagnose electrical faults of IMs, such as the broken rotor bar, stator winding fault, etc.[46,47]. In addition to electrical faults, the MCSA can be applied to detect mechanical faults [30,34]. Different severity levels of IM faults can also be detected by a suitable stator current spectrum analysis [198]. In the case of MCSA, current signals are measured using sensors, like current probes, by attaching to motor current supply cables. The general principle of using the current based monitoring involves the high variation in current signals caused by faulty components of IMs [26]. Other types of techniques such as extended Park’s vector approach can also be used for analysis of the SWF using the stator current [140,2,78]. In other work, Corne et al. [45] presented the relation between three types of evolving bearing faults with their corresponding reflection in the current of IM, analyzed by the extended Park vector approach. Finally, it can be mentioned that similar to the vibration monitoring, the current variation can be detected by using a suitable data analysis method. Various monitoring techniques have been applied to diagnose different IM faults. In general, the efficient diagnosis of IM faults depends on the choice of suitable monitoring techniques. However, the most of methods are expensive, complex, invasive and/or not capable of providing rich information about motor conditions. Besides, the most of these methods can easily identify a particular IM fault that means other faults cannot be detected using the same methods. Nowadays, vibration and current monitoring are most preferable techniques for IMs [200,6]. In a research work, Thomson and Orpin [197] showed that an integrated method of vibration and current monitoring can effectively diagnose the mechanical and electrical faults in IMs. Kral et al. [115] advised the vibration monitoring for diagnosing the bearing as well as other mechanical faults. Later, several researchers have recommended the vibration monitoring for the detection of mechanical as well as electrical faults [70,69]. In order to acquire current signals for the MCSA, sensors like current probes can be attached to the motor cables. Initially, the current monitoring was only preferred for the detection of SWF, BRB and rotor eccentricity. Li and Mechefske [126] and Timusk et al. [201] related the MCSA with vibration monitoring for the identifying various mechanical and electrical IM faults, and finally presented that the vibration monitoring is effective for bearing fault detection, however, the MCSA is effective for the BRB detection. From the literature, it is found that the vibration and current based monitoring are preferred for the IM fault diagnosis [199,155,70,69] because of following reasons, (i) these techniques are non-intrusive, reliable and inexpensive, (ii) their effectiveness and high accuracy in signal analysis which signifies present condition of the machine, (iii) they are easily measurable for further signal processing, (iv) their ability of detecting and distinguishing the most mechanical and electrical faults, and (v) motor current and vibration can be acquired online, thus, it is possible to perform fault detection online. By using suitable signal processing techniques, it is possible to identify deviations in the current and vibration signals produced by different IM fault. To analyze the signals obtained from the condition monitoring of IMs, signal processing techniques has been employed in three domains namely the time, frequency and time–frequency domain [52]. The choice of these domains depends on the information required for the motor fault diagnosis [136]. 3.1. Time domain analysis of IM faults When faults occur in different components of IMs, the vibration and current signals change in time-domain. The distribution and amplitude of signals vary with different fault conditions [66] and the difference between the vibration signals of normal and faulty motors cannot be segregated easily. Moreover, there are many impulses on the waveform of vibration for the bearing fault. For current signals, the variation for different faults is not clearly detectable as the line frequency is the main components; and the fault signal is modulated on the sine wave of the line frequency. However, in the case of phase 10 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 unbalance and stator winding faults, a high variation of the current magnitude among three phases can be clearly seen. Changes or variations in both vibration and current signals for different faults cannot be predicted or segregated directly looking into these signals, due to the irrelevant information or high noise and machine unbalances, especially at the early stage of fault. In other word, the variation in the raw time domain current as well as vibration signals are too small to be detected, therefore the comparison of raw time domain signals of faulty and healthy IMs is not effective in order to determine whether the component is behaving normally or exhibiting signs of a failure. In order to reveal the significant information from time signals, a signal processing method is needed, which converts raw signals into appropriate condense form. The time domain analysis of IMs have been improved using different signal processing method. In a study, Kral et al. [114] presented a time domain based technique for mechanical unbalance detection in IMs. In this technique, first the unbalance specific oscillation of the electric power is extracted by a band-pass filter. Then the averaged pattern of this component is determined by means of an angular data clustering technique. In this way, the oscillation of the electric power in time domain is mapped into a discrete waveform in the angular domain. The amplitude of the fundamental harmonic of these discrete data served as the unbalance indicator of the proposed scheme. The time domain signal analysis of IM have been further improved by extracting useful statistical features including RMS value, standard deviation, skewness, kurtosis, higher statistical moments etc. [226,222,141,155]. However, various factors affect the effectiveness of the feature; thus it is a big challenge to estimate, which feature(s) is/are more sensitive to the particular machine faults. Therefore, a number of statistical features are extracted from time domain signal in order to select the most efficient features that can effectively characterize the signals of healthy and defective IMs [62]. These features can input to other Artificial Intelligence (AI) based system to improve and automate the IM fault diagnosis [92,72,74–75]. 3.2. Frequency domain analysis of IM faults In order to illustrate frequency domain analysis of IMs, the frequency domain data was acquired from the laboratory experiments as shown in Fig. 9. Experiments were carried out on a test-rig that involved a machine fault simulator (MFS). The test-rig consists of test IMs with different seeded mechanical and electrical faults, measurement sensors (AC current probe and tri-axial accelerometer), a constant DC power source, a data acquisition system (DAQ) and a monitor. The test-rig used for the experimentation is as shown in Fig. 9. The MFS consists of an IM (three phase, 0.373 kW, 50 Hz, 2-pole, pre-wired self-aligning mounting system), a sliding shaft, a flexible coupling, a split bracket bearing housing, rotors with split collar ends, a multiple belt tensioning, pulleys, a variable frequency drive (VFD) with multi-featured front panel programmable controller, a magnetic brake, and a photovoltaic sensor. In the basic setup, an IM (test machines in the present illustration) was attached to the shaft using a flexible coupling. The shaft was mounted on the bearings with the split bracket bearing housing. One pulley was attached at the end of the shaft and the other with the gearbox shaft. They were connected using a multiple belt tensioning mechanism. The gearbox mechanism connected the shaft with a magnetic brake. The magnetic clutch was used to apply mechanical load to the IM externally. A VFD was employed to control the motor speed. A tachometer was attached close to the motor shaft end for measuring rotational speed of the shaft. A constant DC power supply source was used to operate the tachometer. The baseplate of the simulator was attached with the vibration isolators and the base stiffener. These all mechanisms were installed in such a way so that they could easily shift, removed and replaced for different experiments [70,69]. A number of IMs with different simulated faults were available with the test rig (as shown in Fig. 10). Five fault conditions (i.e., the broken rotor bar (BRB), bearing fault (BF), unbalance rotor (UR), bowed rotor (BR) and misalignment rotor (MR)) were simulated in five different motors; however, four fault conditions (i.e. stator winding fault: level 1 (SWF1) and level 2 (SWF2), and phase unbalance: level 1 (PUF1) and level 2 (PUF2)) were simulated in two motors only. In the present anal- Test-motor Accelerometer Constant DC Source Current probes Monitor MFS DAQ Fig. 9. Test-rig used for experimenation. 11 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 Motor shaft Rheostat-1 Rheostat-2 end BR BF UR MR BRB SWF PUF (a) (b) Fig. 10. (a) An IM with no defect (b) IMs with various seeded faults. b Ibs Short Circiut Rf Fuse if1 Ias a c s as2 Ics as1 Fig. 11. Turn fault on a single phase of the IM. ysis, overall, ten different faulty conditions of IMs have been considered. The procedure of generation of the considered faults in IMs is discussed now. The BRB defects have been manually produced in the workshop by drilling a number of bars in the rotor cage. The SWF may be due to the turn-to-turn, coil-to-coil, and phase-to-phase or phase-to-ground fault as shown in Fig. 7. In the most of cases the SWF is the consequence of the growth of undetected turn-to-turn faults. The catastrophic failure of IMs can be avoided by early diagnosis the turn-to-turn fault of a coil. Therefore, in this study, the turn-to-turn SWF is considered with two severity levels (i.e., SWF1 and SWF2), which was simulated by inserting an additional load via an external control box (or rheostats) to the winding as shown in Fig. 11. In this figure, ‘‘a”, ‘‘b” and ‘‘c” denote the three phases of IM, Ias, Ibs, and Ics denote the current in three phases, if1 denotes the current flowing through the short circuit. as1 denotes the normal turns and as2 denotes the shorted turns in phase ‘‘a”, and Rf denotes the external resistance. A variable resistor was used to introduce a varying amount of resistance in the turn-to-turn short between the windings. The control box consists of 0–2 X variable resistor. By regulating the value of control box, the severity levels of turn-to-turn fault can be established, which are reflected through loop current variations. High resistance simulates a low severe faults and vice versa. The control box also avoids permanent motor winding damage by restricting the circulating current in the shorted portion of winding. However, when the control box is disconnected, the motor becomes a normal. Moreover, other SWF can be simulated using the proposed method. In order to simulate the PUF and their severity levels (i.e., PUF1 and PUF2), similar procedure as the generation of SWF was adopted in a separate IM. The BF may be due to the outer race fault (ORF), inner race fault (IRF) and ball element fault (BEF). In this illustration, the ORF is considered, which was artificially developed in the bearing near the motor shaft end by spalling out or drilling out some material of the outer raceway. The UR was achieved by taking a balanced rotor from the manufacturer and intentionally removing balance weight and/or adding weight. Here, the balanced weight was attached to small aluminum pins extending from both ends of rotor. The BR was achieved by carefully bending the rotor at the center, which when produced creates the dynamic air–gap eccentricity. The MR is classified as the parallel and angular misalignments. The parallel misalignment can be obtained by dislocating the both motor bearing with same amount and it can be achieved by adjusting four jack bolts, which are attached to the motor. However, angular misalignment can be created by dislocating one end more 12 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 than the other end of the motor. In this work, the rotor misalignment is considered as an angular misalignment because the probability of generation of angular misalignment are higher than parallel one. For the acquisition of current and vibration signals from the test IM, three AC current probes were fastened with the power leads of the IM and a tri-axial accelerometer was installed on the top of the test motor near the motor shaft end, respectively. The vibration and current probes were connected to the DAQ. A signal monitor with a virtual instrument was utilized for data collection. The data was collected for further signal analysis. When faults occur in IMs, the frequency spectrums of vibration and current signals, and its distribution changes, where it indicates new frequency components which appears in the spectrum. For analyzing current and vibration spectrum, these signals were acquired from different fault conditions of IMs. The frequency domain data collected from the laboratory test is used for the illustration. 3.2.1. Healthy motor An ideal healthy IM or motor with no defect (ND) has physical constructional symmetry, such as, an equally spaced and constant air–gap length, equal rotor resistances in the rotor and stator windings, and a balanced rotor. However, there are inherent construction asymmetries and imperfections in an actual healthy IM, for example, the air–gap length is not perfectly spaced and as the rotor rotates the air–gap length varies, and the rotor and stator winding resistances for each phase are not the same. These minor physical asymmetries generate unequal magnetic flux and as a result, magnetic force induced vibrations. Hence, a healthy IM is expected to generate some low magnitude vibrations. In addition, due to above problems, some variation can occur in the current signal of healthy motors also. Fig. 12 shows the power spectrum of stator current of a healthy motor, when motor was rotating at 40 Hz under the full load condition. In the current spectrum of motor with ND, the maximum amplitude comes at 40 Hz; however, the line current frequency is 50 Hz, because the supply frequency applied to the motor is 40 Hz using a VFD. In current signals, sidebands of lower amplitudes around the supply frequency exist even when the machine is healthy, as can be seen in Fig. 12. This could be due to uneven rotor bar resistance because of the die-casting process, rotor asymmetry, etc. The amplitude and number of sidebands tend to increase if any faults occur and this is an indication of faults. 3.2.2. Broken rotor bars A rotor bar can be failed due to various stresses such as the thermal, magnetic, mechanical, dynamic, residual, and environmental stresses [35,179]. When the rotor asymmetry appears due to the broken rotor bar, it creates in an addition of a forward rotating field a backward rotating field that turns with the speed of ðsf s Þ (where s is the operating slip, f s is the supply frequency). The result of this is an additional frequency component f b ¼ ð1 2ksÞf s in the stator current. This cyclic variation in the current implies a speed oscillation and a torque pulsation at the twice slip frequency ð2sf s Þ. This speed oscillation induces the upper sideband components at f b ¼ ð1 þ 2ksÞf s (where, f b is the broken rotor bar frequency, and k = 1, 2, 3, . . .). So broken rotor bars tend to induce additional components in the stator current at frequencies given by: f b ¼ ð1 2ksÞf s ð1Þ These sidebands are function of the slip and the supply frequency; therefore, these sidebands are dynamic in nature and vary with the operating condition of IMs [232]. 20 Stator current (dB) 0 -20 -40 -60 -80 -100 10 20 30 40 Frequency (Hz) 50 60 Fig. 12. Stator current power spectra of a healthy IM at T3 when VFD is set to 40 Hz. 70 13 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 The BRB also excites the electromagnetic field disturbance and thus intensifies the torque modulations and vibrations of the motor. If a bar is broken in an IM, then there will be no flow of current in that bar. Consequently, the field in the rotor around that particular bar will not exist. Thus, the force applied to that side of the rotor would be different from that on the other side of the rotor. It creates an unbalanced magnetic force that rotates at one times rotational speed and modulates at a frequency equal to slip frequency times the number of poles, which is known as the pole pass frequency. That means in the vibration spectrum, increased amplitudes will occur at the rotation frequency f r and its sidebands [101] f brb ¼ f r f p ð2Þ where f p is pole pass frequency defined as: f p ¼ ðf s f r Þp ð3Þ where p is number of poles and f s is the supply frequency. In addition, the BRB also produces additional sidebands around 1X and higher harmonics ð2f r ; 3f r ; ::::Þ in the vibration spectrum. In order to study the BRB fault, the stator current power spectra for the BRB when the motor is rotating at 40 Hz under full load is added in Fig. 13. The sidebands arise around 40 Hz rather than the line frequency, i.e. 50 Hz, because the VFD was set to 40 Hz. The slip at this rotating speed and full load is near about 4.29%, which is found by s ¼ ðN s N r Þ=N s , where Ns is the synchronous speed of magnetic field or electrical speed, and Nr is the actual rotor speed or mechanical speed. Fault frequencies or sidebands (as calculated from Eq. (1)) are 37 Hz and 43 Hz when k = 1, 34 Hz and 46 Hz when k = 2, and 31 Hz and 49 Hz when k = 3. These sidebands are easily visualized in the current spectrum as shown in Fig. 15. The rise in sequence of such sidebands as compared to the healthy motor spectrum is actually due to broken bars. From Fig. 13, it can be observed that sidebands for the BRB are visible when the motor is heavily loaded. From Eq. (1), it is evident that, these sidebands are closely related to motor slip. If there is no load on the motor, the slip will be almost zero. In that case, the sidebands will be overlapped by the supply frequency and thus make fault diagnosis a difficult task. Thus, it is necessary to heavily load the motor in order to separate the sidebands [52]. Overloading a motor is undesirable since it reduces the operating life of motor and is not generally under control of the operator. Accurate diagnosis of the BRB, therefore is difficult at light or no load because of the low slip operation. In addition, the actual rotor speed is required for calculating slip of the BRB motors, which is not always available. In order to study the BRB fault, the vibration spectra for the BRB, when the motor is rotating at 40 Hz under full load, is added in Fig. 14. Fault frequencies or sidebands (as calculated from Eq. (2)) are 41.72 Hz and 34.84 Hz. In addition, the sidebands occur in higher harmonics of rotational speed are 76.56 Hz and 119.84 Hz. These sidebands can be visualized in the vibration spectrum; however, they are of small amplitudes. Similar to the current spectrum analysis, the low slip due to low load is undesirable in the vibration spectrum analysis also. The BRB is more difficult to detect in the vibration spectrum than the current spectrum because of small amplitudes. 3.2.3. Stator winding faults or armature faults Asymmetrical short circuit in the stator winding are usually related to the insulation damage. These faults occur due to hot spots in the stator core, causing the high temperature, electrical discharges, loosening of structural components, moisture, and oil contamination [76]. The additional frequency components appear due to this fault in the current spectrum are given by: f st ¼ fk nð1 sÞ=pgf s ð4Þ 0 Stator current (dB) 34 Hz -20 31 Hz 37 Hz 43 Hz 46 Hz 49 Hz -40 -60 -80 -100 10 20 30 40 Frequency (Hz) 50 60 Fig. 13. Stator current power spectra of IM with the BRB at T3 when VFD is set to 40 Hz. 70 14 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 0 119.84 Hz 76.56 Hz 41.72 Hz -20 Acceleration (dB) 34.84 Hz -40 -60 -80 -100 0 20 40 60 80 Frequency (Hz) 100 120 140 160 Fig. 14. Vibration (acceleration) power spectra of IM with the BRB at T3 when VFD is set to 40 Hz. where f st short turns frequency, f s is the supply frequency, p is the number of pole pairs, k = 1, 3, 5. . .; n = 1, 2, and so on. The stator asymmetry results in the negative sequence components that appear in the input current. These faults produce asymmetry in the motor impedance causing the motor to draw an unbalance phase current. This results in induction of negative sequence voltage in the rotor. Since the rotor is short-circuited, this will result in abnormal current flow in the rotor and results in the damage of the rotor. In addition, the SWF also produces fault frequencies as ð pf r ; 2pf r ; 4pf r ; :::Þ in the vibration spectrum. In order to study the SWF, the stator current power spectra of an IM with the SWF, when the motor is rotating at 40 Hz under full load, is added in Fig. 15. The slip at this rotating speed and load is near about 4.29%. The fault frequency (as calculated from Eq. (4)) occurs at 78.5 Hz and 1.5 Hz (when k = 1 and n = 1); 158.5 Hz and 81.5 Hz (when k = 3 and n = 1), 238.5 Hz and 161.5 Hz (when k = 5 and n = 1), 117 Hz (when k = 1 and n = 2), and 155.5 Hz (when k = 1 and n = 3). It is observed from Fig. 15 that fault frequencies are noticeably visible, which indicates the stator winding fault in the IM. However, the supply voltage (or phase) unbalance and the load unbalance could also produce the unbalance current and current harmonics at frequencies described by Eq. (4). This is one of the major problems in the SWF detection based on the current spectrum analysis. In order to study the SWF, the vibration spectra of an IM with the SWF when the motor is rotating at 40 Hz under full load, is added in Fig. 16. The fault frequencies, i.e. 76.56 Hz, 153.12 Hz and 306.24 Hz (as calculated using ð pf r ; 2pf r ; 4pf r ; :::::Þ, can be observed from the vibration spectrum. However, these frequencies are not clearly identifiable as there are many other frequencies of higher magnitudes. The SWF cannot be identified by vibration signal alone the current spectra are also necessary. 3.2.4. Bearing faults Bearings are the most affected component of the IM [124]. The bearing faults (BF) are categorized as the ball fault, outer race fault, inner race fault, and train fault. Defective bearings generate high vibrations at the rotational speed of each 20 0 Stator current (dB) 155.5 Hz -20 78.5 Hz 81.5 Hz 117 Hz 158.5 Hz 161.5 Hz -40 -60 -80 -100 0 20 40 60 80 100 120 140 160 180 200 Frequency (Hz) Fig. 15. Stator current power spectra of IM with the SWF at T3 when VFD is set to 40 Hz. 220 15 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 0 76.56 Hz 153.12 Hz Acceleration (dB) -20 306.24 Hz -40 -60 -80 -100 0 50 100 150 200 250 300 Frequency (Hz) 350 400 450 500 Fig. 16. Vibration (acceleration) power spectra of IM with the SWF at T3 when VFD is set to 40 Hz. component. These characteristic frequency components, which are related to the rolling element and raceways, are determined using the rotational speed and the bearing dimensions. Vibration analysis is commonly used to detect these frequency components. Vibration frequencies associated with different BF are given by [95,202,135] For inner race defect: d f i ¼ Zf r =2 1 þ cosa D ð5Þ For outer race defect: d f o ¼ Zf r =2 1 cosa D ð6Þ For ball defect: 2 f b ¼ Zf r =d 1 d D2 ! cos a 2 ð7Þ For train defect: d 2 f t ¼ f r =2 1 cosa D ð8Þ where Z represents number of rollers/balls, d represents rolling element diameter, D represents pitch circle diameter of bearing, a denotes the contact angle in radians, and f r is the rotational frequency. For the most of bearings that have 6–12 balls, the characteristic components are approximated byf o ¼ 0:4Zf r and f i ¼ 0:6Zf r . As the bearing supports the rotor, any BF will generate a radial motion between the rotor and the stator. The bearing damages produce mechanical displacement causing variation in the air–gap, which is defined by combination of rotating eccentricities varying in clockwise as well as anticlockwise direction. These variations of the air–gap produce current components at predictable frequencies, f bg related to the vibrational and electrical supply frequencies. Harmonic components produced by BF in the current spectrum are given by [94] f bg ¼ jf s kf v j ð9Þ where, k = 1, 2, . . ., f v represents one of the vibration characteristic frequencies (for example, f i is the inner race frequency and f o is the outer race frequency), and f s represents the supply frequency. In order to detect the characteristic frequency in the vibration spectrum, a bearing with the outer race fault is considered. The bearing (NSK 6203) with 29 mm pitch diameter and 6.75 mm rolling element diameter, and 8 rolling elements is considered. The characteristic bearing outer race frequency for the bearing, as calculated from Eq. (6) at 2341 rpm (39.02 Hz) of rotor speed, isf o ¼ 119:60 Hz. Fig. 17 shows the vibration spectrum for the considered bearing fault conditions. The bearing frequency component may be possible at the outer race frequency 119.60 Hz and multiples of it (i.e., 2: 239.2 Hz, 3: 358.05 Hz, and 4: 478.43 Hz). These components can be noticeably visualized in the vibration spectrum and hence a BF can be detected by the vibration analysis, however, other peaks can be seen throughout the spectrum. From the vibration analysis of the bearing fault, it can be concluded that for effective detection of fault frequencies, the measurement of rota- 16 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 20 478.43 Hz 239.2 Hz 0 358.05 Hz Acceleration (dB) 119.6 Hz -20 -40 -60 -80 0 40 80 120 160 200 240 280 320 360 400 440 500 Frequency in Hz Fig. 17. Vibration (acceleration) power spectra of IM with the BF at T3 when VFD is set to 40 Hz. tional frequency and accurate geometry of machine elements, for example, the pitch diameter, contact angle, and ball/roller diameter are required. These can cause problems for diagnosing the BF, if these parameters are not known [210]. In order to detect the BF using the current signal, a current spectrum is shown in Fig. 18. The characteristic fault frequency (as calculated by Eq. (9)) is 76.60 Hz, which can be seen in Fig. 20. Other frequencies of interest in the current spectrum are 159.6 Hz (when k = 1), 199.20 Hz and 279.20 Hz (when k = 2), 318.20 Hz and 398.80 Hz (when k = 3). Fig. 18 shows that the current signature analysis is also capable of detecting BF. However, similar to the vibration spectrum analysis, various factors affect the current harmonics that make the current spectrum analysis difficult for the BF. 3.2.5. Air-gap eccentricity related fault Uneven air gap exists between the rotor and the stator, and it is known as the air gap eccentricity. This causes an unbalance force on the rotor that pull the rotor even farther from the stator core center. The static eccentricity creates a steady pull in one direction, while the dynamic eccentricity creates an unbalanced magnetic pull, which acts on the rotor and rotates at the same angular velocity as the rotor. The air–gap eccentricity produces additional frequencies in both current and vibration spectra. The air gap eccentricity generates the characteristic frequency in the current spectrum, which is given by [109] f ect ¼ f1 mð1 sÞ=pgf s ð10Þ where f s represents the electrical supply frequency, p represents the number of pole pairs and s represents the slip. If both the static and dynamic eccentricities occur simultaneously, a low frequency component near the fundamental frequency is given by F ¼ jf s kf r j, wheref r represents the rotational frequency. In addition, in the vibration spectrum, the unbalanced rotor creates a large amplitude at 1 of rotating speed in radial directions, the bowed rotor creates 1 amplitude (and often 2 amplitude) in the axial direction, the angular misalignment (for the MR) creates 1 amplitude in the axial direction, small 2 and 3 amplitudes in the axial direction and small 1 amplitude in the radial direction. 20 Stator current (dB) 0 159.6 Hz 199.2 Hz 79.6 Hz -20 -40 279.2 Hz 398.8 Hz 318.8 Hz -60 -80 -100 -120 0 40 80 120 160 200 240 280 320 360 400 440 Frequency in Hz Fig. 18. Stator current power spectra of IM with the BF at T3 when VFD is set to 40 Hz. 480 500 17 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 20 Stator current (dB) 0 79.01 Hz -20 157.03 Hz 196.04 Hz -40 -60 -80 -100 -120 -140 0 40 80 120 Frequency in Hz 160 200 240 Fig. 19. Stator current power spectrum of IM with the BR at T3 when VFD is set to 40 Hz. In order to detect the air–gap eccentricity in the IM, a bowed rotor (BR) is considered. Fig. 19 shows the plot of the current spectrum for the BR. Characteristic fault frequencies (as calculated from Eq. (10)) are 79.01 Hz (when k = 1), 118.02 Hz (when k = 2), 157.03 Hz (when k = 3) and 196.04 Hz (when k = 4). It is observed that these frequency components can be easily visualized. Hence, the current signature analysis can be used in detecting the eccentricity related faults in IMs. However, it is noted that the SWF also produces the same harmonics in the current spectrum at frequencies obtained by Eq. (10). This is one of the great challenge in the eccentricity related fault detection. To improve the current and vibration monitoring of IMs, other frequency domain based methods, such as higher order spectrum, Hilbert transform, etc., have been applied [156,1]. Different faults generate a specific vibration and current spectra, which provide the fault harmonic components related to the particular fault. Chow and Fei [44] used the bi-spectrum analysis of vibration signals for the identification of machine faults, such as the asymmetrical fault, mechanical bearing movement, and stator winding fault. They concluded that the bi-spectrum is able to deliver sufficient and important spectral information for the condition monitoring and fault diagnosis, and is therefore advantageous than the power spectrum analysis. Speed; Load; Machine fault simulator with test IM (Healthy/Faulty) Data Acquisition Feature extraction Feature selection Total data sets (Extracted features) Training data set Testing data set AI Model Prediction accuracy Fig. 20. AI based fault diagnosis. 18 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 Thomson and Orpin [197] used the current and vibration spectra for the identification of various faults in IMs, such as the BRB, SWF, BF and eccentricity related faults. They concluded that the root cause of the problem can be established using a combination of the current and vibration analyses in comparison to only analyzing one signal. However, for complex systems, which involve various components, it is a challenging task to estimate characteristic frequencies of faults, since the vibration and current spectra can be influenced by several factors such as the electric supply, variable loading, noise and faults [27]. Even though characteristic frequencies are available, measured signals from motors are extremely nonstationary and the frequency spectrum analysis may not be suitable for such cases. Didier et al. [51] used the current spectrum to detect the BRB by analyzing the phase and concluded that the Fourier transform phase analysis allows to detect one BRB when motor runs under low load, but the robustness of the DFT method decreases for a half broken rotor bar, The Hilbert transform (HT) allows to detect a partially BRB, when the motor operates with the load torque equal or greater than 25% of rated load. In addition, some researchers tried different methods that can eliminate the load effect in the current spectrum. For example, Puche-Panadero et al. [152] used the HT for the identification of the BRB based on the MCSA, and concluded that this technique can accurately identify BRB frequencies at a very low slip condition by using only one phase current. 3.3. Challenges in the time and frequency analyses of vibration and current signals for IM faults An effective fault diagnosis in IMs is not possible by analyzing only the time domain vibration and current signals. From vibration signals in time domain, the difference between the signal of normal and faulty motors can be visualized; however, these cannot be segregated. From time domain current signal, the difference is not visible among different faults. The reason is that the main component is the supply frequency and the fault signal is modulated or riding on the sine signal of the supply frequency. The vibration and current spectrum analyses can be implemented for IM fault diagnosis when the speed oscillation of motor is small, provided a higher accuracy in the diagnosis is not needed; however, various factors affect the spectrum analysis. Fault signatures, sometimes, have extremely low signal-to-noise ratio. So in order to capture sufficient harmonic information of individual fault, a suitable high sampling rate is needed and yet at the same time a sufficient frequency resolution is required. When the high frequency resolution is necessary, this poses a constraint that limits the low sampling rate to be used, in order to have a better frequency resolution while using the FFT analysis. In addition, different faults generate a specific vibration and current spectra, which provide a fault harmonic component related to the particular fault. However, for a complex system, which involve various components, it is a challenging task to accurately estimate the harmonic component of the fault and locate them in the spectrum [129]. In the case of current spectrum analysis, various difficulties are associated, such as the existence of characteristic harmonics due to the supply voltage distortion, load unbalance, and air–gap variation. In addition, harmonics of eccentricity due to both the design and construction of the motor, harmonics caused by variations of the load and the supply frequency, and background noise also make the current spectrum analysis difficult for the fault diagnosis. Moreover, for electrical faults the current spectrum is very sensitive to the accuracy of the motor slip and supply frequency measurements, because the fault frequencies or sidebands are function of the motor slip and the line frequency. However, it is not an easy task to estimate these parameters accurately. Even if characteristic frequencies of different faults are available, the measured vibration and current signals from motors are extremely non-stationary and it makes the spectrum diagnosis even more difficult for the fault detection. Additionally, other harmonics from the power electronics equipment, such as the VFD, also make the spectrum analysis of current more difficult. It is noted that the stator excitation frequency dynamically changes the position of the current harmonics appearing in the stator-current spectrum due to electrical faults, and is highly dependent on the mechanical motor load and the excitation frequency, which affects the slip frequency. As a consequence, the conventional current spectrum analysis must be amended to accommodate the new scenarios. The current and vibration spectra based methods may be effective for the fault detection and diagnosis of IMs only when the motor is almost fully loaded and running at a constant speed, however, there are many applications where these operating conditions are not achieved. 3.4. Advanced condition monitoring methods for IM faults However, some researchers have tried to overcome the aforementioned problems, for example in a study, Elbouchikhi et al. [55] developed a multiple signal classification (MUSIC) algorithm for the BF detection with the help of characteristic frequencies of the fault. In this work, the amplitude corresponding to fault characteristic frequencies are estimated and then fault indicators are derived for the fault severity measurement. Ameid et al. [9] used extended-Kalman filters with the FFT for improving the BRB fault detection. The BRB is detected through the estimation of rotor resistance using the extendedKalman filter (EKF) observer and the FFT for analyzing several mechanical and electrical parameters (i.e., the rotor speed, quadratic current components of control and stator phase current). The spectral analysis is effectuated only if the variation in estimated rotor resistance is significant. The EKF observer proved to be an excellent mathematical tool for the fault indicator in the variable speed drive. Moreover, the spectral analysis using FFT transforms can be a useful solution to ensure that the resistance variation indicates the BRB. In other work, Samanta et al. [170] proposed a fast and accurate spectral estimator based on the theory of Rayleigh quotient for detecting the spectral signature of the BRB. The proposed spectral estimator could precisely determine the relative amplitude of fault sidebands and had low complexity compared to available high- P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 19 resolution subspace-based spectral estimators. Detection of low-amplitude fault components was improved by removing the high-amplitude of the fundamental frequency using an extended-Kalman based the signal conditioner. The time–frequency analysis was preferred in the fault detection and diagnosis of IMs [234,235,148,8]. Singh and Ahmed [185] performed the wavelet transform (WT) analysis of vibration signals for the detection of electrical faults in IMs. Finally, they showed that the WT analysis of vibration signals in different frequency regions in time domain enhances the extraction of the information that can improve the ability of a diagnostic system. Awadallah and Morcos [13] used WT to extract diagnostic indices from the current waveform of the motor. Finally, they showed that, the advantage of WT for handling nonstationary signals make it more efficient for diagnosing and locating the fault. In a study, Gao and Yan [70] presented a comparison of different signal processing methods, namely the short time Fourier transform (STFT), discrete wavelet transform (DWT), wavelet packet transform (WPT), and Hilbert-Huang transform (HHT) for monitoring the bearing fault to handle nonstationary signals. The WPT could detect all transient components embedded in signals and perform better than all other signal processing methods. Zarei and Poshtan [227] applied the WPT energy of the stator current as a fault index for the monitoring of incipient bearing fault in IMs. Zolfaghari et al. [241] utilized the root-mean-square (RMS) of wavelet packet coefficients as a fault feature index for the detection of severity levels of the broken rotor bar in IMs. Wu et al. [217] proposed a diagnosis method of the three-phase voltage source inverter (VSI) fed open-circuit fault of IMs. The three-phase currents are sampled by current sensors and extracted by the DWT. In this work, they calculated Euclidean distance of energy vectors obtained from the approximate coefficients of the DWT to measure the similarity of the currents and accumulation values in order to locate the position of fault. Different methods are available that can identify certain faults in machines, such as the power spectrum graph, phase spectrum graph, cepstrum graph, auto regressive (AR) spectrum graph, spectrogram, wavelet scalogram, wavelet phase graph, etc. However, these methods often present several problems in terms of complexity and cost. The conventional signal based methods are not always reliable as they depend on operating conditions of motors [121]. In addition, other fault diagnosis methods are also available based on the mathematical modelling of machines [116]. Based on the explicit model, residual generation methods such as the Kalman filter, parameter estimation (or system identification) and parity relations are used to obtain signals, called residuals, which is an indicative of the fault presence in the IM. Finally, the residuals are evaluated to arrive at the fault detection, isolation and identification. The model based diagnosis can be more effective than other model-free approaches, if accurate model is developed. However, several assumptions must be made to develop a mathematical model to take care of the nonlinear and stochastic machine dynamics, still it is not robust enough in the presence of perturbation and noise [33]. In addition, explicit modelling may not be possible and feasible for complex systems because it is not easy to consider the disturbances and the uncertainty in the models [172]. The signal based method is more accurate and preferred than the model based method because it does not require any assumptions and complex mathematical models [47,47]. The machine fault diagnosis based on signal based and mathematical model based methods require practice engineers to have sufficient knowledge and experience. As the number of rotating machines is increasing steadily in all types of industries, so it is not practical to fulfill the demand of required expertise. The increasing economic pressure requires the development of a cost-effective maintenance system to guarantee machine operating reliability and at relatively low cost. Therefore, during the last two decades, in order to automate, improve the reliability and sensitivity, and reduce the cost of the fault monitoring and diagnosis technique, a more sophisticated technique, called intelligent fault diagnosis, is developing and growing popularity in the field of mechanical engineering [64,104]. The AI based diagnosis methods are a single reliable procedure for diagnosing any types of the mechanical and electrical faults in IMs. AI based methods are powerful and improves the fault diagnosis effectiveness and efficiency in IMs, especially during the predictive maintenance process [180,184]. The AI based diagnoses have shown improved performance over the conventional signal and modelling based approaches. This reduces the direct human–machine interaction for the diagnosis. Moreover, these are data based techniques; therefore, they do not require any detailed knowledge of the IM model and its system parameters. The AI techniques use association, reasoning and decision making processes similar to a human brain for solving diagnostic problems [12]. 4. Artificial intelligence based fault diagnosis of IMs The AI based diagnostic system consists of signal-based methods and classification tools such as the Neural Network (NN), Fuzzy Logic (FL), Genetic Algorithm (GA), hidden Markov model, Bayesian classifier, Support Vector Machine (SVM) algorithm and Deep Learning (DL) algorithm [226,222,16,111,37,39,237,70,69,167,165]. The AI based fault diagnosis, if properly established and effectively implemented, can significantly reduce maintenance cost by reducing the number of unnecessary scheduled preventive maintenance operations [102,128,130]. It comprises of five steps: data acquisition, feature extractions, feature selection, diagnosis of faults, and prognosis of systems, as shown in Fig. 20. Data can be the vibration, current, acoustic, temperature, pressure, oil analysis data, etc. The practical reliability of the diagnostics depends upon the extraction of representative features corresponding to IM health conditions [155], Tiwari [202]. The feature extraction is used to reduce the dimension of data by selecting the important features. The feature can be determined using time domain data; such as the mean, variance, standard deviation, kurtosis, skewness, and so on [63,167,165]; frequency domain data, such as the frequency center, root-mean-square frequency, amplitudes of FFT spectrum, and so on; and time–frequency domain data, such as the statistical features and coefficient of STFT, WT, and so on [98,37,39]. For an AI based fault diagnosis, 20 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 it is very crucial to choose a suitable signal processing method and then extract fault features buried within the signals. The reason is that these features are fed to the classifier and classifier’s performance depends on these. In addition, for the selection of effective features various optimization techniques, such as the principal component analysis (PCA), independent component analysis (ICA) and genetic algorithm have been used [129,143]. The diagnosis and prognosis is the final step of AI based fault diagnosis. The diagnosis deals with the fault detection, isolation and identification of already developed fault. The fault detection indicates whether something is going wrong in the machine; the fault isolation locates the machine component that is faulty; and the fault identification determines the severity of the fault when it is detected. The prognosis deals with the machine life prediction before it fails [42,209]. Obviously, the prognosis is superior to diagnosis in the sense that it can prevent faults or failures, and if impossible, be ready (with prepared spare parts and planned human resources) for the problems, and thus save extra unplanned maintenance cost. Nevertheless, the prognosis cannot completely replace the diagnosis since in practice, there are always some machine faults and failures that are not predictable. Besides, the prognosis, like any other prediction techniques, cannot be 100% sure to predict faults and failures. In the case of unsuccessful prediction, the diagnosis can be a complementary tool for providing a maintenance decision support. In addition, the diagnosis is also helpful in improving the prognosis in the way that the diagnostic information can be useful for preparing more accurate event data and hence building better CBM model for the prognosis. Furthermore, the diagnosis information can be used as useful feedback information for the system redesign [98]. In this review paper, the literatures are mainly focused on the diagnosis part, which is used for the identification of IM faults. For prognosis part, interested researchers can refer Lei et al. [122]. They reviewed machinery health prognostics, which presented a systematic review from data acquisition to remaining useful life (RUL) prediction. 4.1. ANN based diagnosis ANN is the most commonly used AI technique in the fault diagnosis of IM [89]. Chow et al. [43] successfully developed a neural network (NN) based incipient fault detector for the small and medium size IMs. It avoids the difficulties associated with traditional incipient fault detection method after using additional parameters such as the stator current and the rotor speed. Schoen et al. [175] combined the FFT and the ANN for enhancing detection of various fault condition including the BRB through the MCSA. The methodology does not need additional information about the motor parameters or load characteristics. Salles et al. [168] used NN techniques for monitoring of IM loads. In this work, extraction of the current signature is achieved by the time–frequency spectrum analysis. Finally, they showed that the technique could be used for recognition of torque and ultimately improving the interpretation of machine anomalies effects. Kolla and Varatharasa [110] applied the ANN for the detection of external faults of an IM by using RMS values of voltages and currents. A feed-forward layered neural network (FFNN) structure is applied, and trained using the back propagation algorithm. They suggested that the other signals, such as the instantaneous and symmetrical components of voltages and currents can also be tried. Finally, they showed that for effective fault detection the training sets should have a larger data with more fault situations. In addition, they suggested to test the method against any other noisy data. Ye et al. [223] successfully applied the wavelet packet based fault features and the ANN for diagnosing the BRB and the air–gap eccentricity. Likewise, Kim and Parlos [116,115] used a combination of WPT and ANN to develop a model based fault diagnostic system using the MCSA, voltage and speed for early detection of several IM faults. In other work, Parlos et al. [146] successfully performed detection of induction motor faults by combining signal-based and model-based techniques using same WPT and ANN. Subsequently, Ye et al. [224] used the MCSA with the WPT and the ANN for the detection of air–gap eccentricity and BRB. This method offers feature representation of multiple frequency resolution for IM faults. Kowalski and Orlowska-Kowalska [113] showed various IM faults, such as the stator, rotor, rolling bearings and supply asymmetry can be effectively detected based on the NN by suitable measurements and analysis of the vibration and current spectra. Yang et al. [219] presented an idea for improving the IM fault diagnosis, which is based on integrating case-based reasoning (CBR) and an adaptive resonance theory- the Kohonen neural network (ART-KNN). They showed that the proposed ART-KNN, by synthesizing the theory of ART and the learning strategy of KNN, can solve the plasticity-stability dilemma of conventional neural networks. In this work, seven different motor faults (bearing fault, rotor, stator, air gap, misalignment, mechanical unbalance, and looseness) were classified using vibration signals with the developed methodology with good accuracy. Han et al. [82] used a combination of PCA, GA and ANN for the fault detection of various IMs. In this work, the PCA is used to remove the relative features of vibration and current, and extract the principal components (PCs) from the original features. The critical fault features are then chosen using the GA. In this work, the GA is also applied for the optimization of ANN parameters. Ayhan et al. [14] compared the performance of the ANN and the multiple discriminant analysis (MDA) based detection method for the BRB using a single or multiple signature processing. These two methods are used with two different detection schemes, i.e. the monolith and the partition. The monolith scheme involves various single largescale ANN or MDA unit expressing the whole motor operating load-torque region, whereas the partition scheme involves various small scale ANN or MDA blocks, every blocks expressing specific load torque operating region. They used the MCSA for the analysis and showed that multiple signature processing performs better as compared to a single signature processing. In addition, the ANN provides a higher accuracy performance than the MDA. Lee et al. [121] developed an approach based on the Fourier and wavelet transforms, and the ANN using the MCSA to detect various IM faults. In other work, Yang and Kim [220] used the vibration and current signals along with the ANN and the Dempster-Shafer theory to detect various IM faults P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 21 including the BF, SWF, BRB, air–gap eccentricity and phase unbalance. They showed that the fault classification accuracy increases when the fusion of vibration and current is performed. Ballal et al. [19] presented the fault diagnosis of IMs based on ANFIS. They applied this system to diagnose two critical IM faults, i.e. the bearing and inter-turn insulation faults. In this work, they used five different parameters, such as the motor speed, stator current, bearing and winding temperatures, and the machine noise for the fault diagnosis. Initially, they applied the first two parameters for training and testing, and later remaining three parameters were also included. Finally, they showed that the performance of the system improves significantly with all five parameters as inputs. Bacha et al. [17] compared the current and magnetic flux monitoring in the detection of BRB and the phase unbalance based on the spectrum analysis and the ANN. They showed that the stray flux monitoring performs better than the current monitoring to detect these faults using the low frequency resolution data, especially when there is no load on the motor. Su and Chong [190] presented the BRB and the air–gap eccentricity detection using an analytical redundancy technique, which is based on the NN. The STFT is employed to process the quasi-steady vibration signals to continuous spectra that is required in the training of the NN. Result showed that the developed methodology is very effective and robust. Martins et al. [133] developed fully automatic on-line IM fault diagnostics for the SWF with the help of unsupervised Hebbian-based NN. In this work, principal components of the stator current are used. The developed method offers several advantages. First, the absence of previous training, or the incorporation of heuristic knowledge, makes it interesting for industry applications. Second, since there is no need to perform any FFT computations, it makes it simpler to implement. Third, the method is able to indicate severity of the fault, rather than only detect its presence. Huang et al. [90] used the ANN for detecting the eccentricity in IMs based on the voltage and current signals. They used the ANN to learn the complex relationship between the operating conditions and eccentricity-related harmonic amplitudes. Finally, they showed that the developed system could evaluate a threshold corresponding to an operating condition, which can then be applied for the fault detection. Sadeghian et al. [166] performed the BRB fault diagnosis using the MCSA based on the WPT and the ANN. The proposed diagnostics was shown effective even with reduced load condition. In addition, it does not require any information about the motor slip speed. Ghate and Dudul [72] introduced an approach to the fault diagnosis of an IM using the radial basis function (RBF) and the multilayer perception (MLP) cascade NN. In this work, they considered the rotor eccentricity, stator winding inter-turn short, and both these defects simultaneously for the diagnosis using current signal. Result showed that the developed classifier is robust even in case of external noise. Zarei et al. [228] explored the cascaded NN for the fault detection of bearing fault in IMs, where first the NN are used as a removing non-bearing fault component (RNFC) filter for signal denoising, and then the second NN are used this filter output for the fault detection. They showed the approach had better performance ability than the common NN even when dealing with the noisy samples. Wu et al. [216] diagnosed bearing faults in IMs based on the wavelet packet (i.e., the Daubechies 8) and the ANN with 90% detection success. Lashkari et al. [119] investigated a major drawback of the existing fault diagnosis technique to diagnose and trace the location of winding short-circuit fault in IMs. In this work, they used three-phase shifts between the linecurrent and phase-voltage. Restricting the detection of the SWF for applying only the three-phase shifts can cause uncertainty, as the three-phase shifts are found to be sensitive to the supply voltage unbalance. An ANN based scheme was developed for detecting and discriminating between the effects of the phase unbalance and the SWF (inter-turn) in three-phase IMs. Bessam et al. [31] used the NN for the BRB fault detection especially at a low load condition. In this work, the current envelope is extracted by using the Hilbert transform. The harmonic frequencies and corresponding amplitudes are observed and used as input for the NN. This approach is capable of detecting the correct number of BRBs for various loadings. In other study, Bessam et al. [32] combined DWT and NN for IM fault diagnosis. In this, the stator current was analyzed by DWT for computing the energy associated with the stator fault in the frequency bandwidth and found effective results. In a study, Huang et al. [88] established the fault diagnostics for IMs by analyzing the vibration signal based on extension neural network (ENN). The fault frequencies were first calculated and further used as a learning data for developing the ENN model. Then after testing of the same model, it is found that developed diagnostic is capable of diagnosing the type of IM faults, accurately. In addition, it performs better than NN in terms of training time and classification accuracy. In a study, Bazan et al. [22] performed the stator fault analysis of IMs using the ANN based on measures of mutual information between the phase current signals. The results presented good classification accuracies even when the motors were subject to several conditions of the load torque and the power supply voltage unbalance. Wen et al. [211] proposed a new Convolutional Neural Network (CNN) based on LeNet-5 for diagnosing fault in IM and pumps. CNN is an effective Deep learning (DL) method. It provides an effective way to extract the features of raw data automatically. Glowacz and Glowacz [74] presented thermal imaging based fault detection of IMs with the NN, K-means, BNN. In this work, the authors developed a technique for the feature extraction based on thermal images, i.e. MoASoID (Method of Areas Selection of Image Differences). The diagnostics could be able to classify thermal images of the three-phase IM properly. However, there are some drawbacks of this method, like the cost of thermal imaging camera, because it is more expensive than analysis of other signals like vibration, current and acoustic signals. The other drawback is that it takes some time to heat up stator coils. Moreover, long durations of shorted coils can cause permanent damage to the analyzed machine. In addition, the thermal imaging camera can be set in many ways, so the training and test sets of thermal images may be improper for recognition process. In other study, Glowacz et al. [75] and Glowacz [73] explored acoustic signals in the fault diagnosis of IMs using the NN classifier, back propagation neural network (BNN) and modified algorithm using words coding. Here, two feature calculation techniques have been used, i.e. the Shortened Method of Frequencies Selection Multi-expanded 2 Groups (SMOFS-32-MULTIEXPANDED-2- 22 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 GROUPS) and the SMOFS-32-MULTIEXPANDED-1-GROUPS. Finally, the result showed that the acoustic signals are very good in recognition with the real data. 4.2. Fuzzy logic based diagnosis Fuzzy logic (FL) is another AI technique, which is explored by many researchers in fault diagnosis of IM [161,163]. In a study, Altug et al. [7] combined synergy of FL and NN for a better understanding of the heuristics underlying the motor fault diagnosis process. This study presents two neural fuzzy (NN/FZ) inference systems, namely, fuzzy adaptive learning control/ decision network (FALCON) and adaptive network based fuzzy inference system (ANFIS). Result showed that the proposed methodology was effective for incipient fault detection in IM. Lasurt et al. [120] applied the fuzzy logic procedure to develop a method for the fault detection of IMs. They used higher order statistical (HOS) analysis for preprocessing of vibration signals. Filippetti et al. [60,60] added a review on the IM fault detection based on the AI methods such as neural network and the FL. In addition, the amplitudes corresponding to the fault frequency components obtained from the FFT were considered as an input vector in AIs. Finally showed that, the fault detection is difficult when spectral leakage occurs especially under the low motor slip. Benbouzid and Nejjari [29] applied FL for the IM fault diagnosis. In this, fuzzy subsets and the corresponding membership functions describe stator current amplitudes. The IM condition is diagnosed using a compositional rule of fuzzy inference. Zidani et al. [240] showed that an effective motor fault diagnosis could be achieved based on the fuzzy logic strategy. In this work, they used the fuzzy approach using Concordia patterns of the stator current. Tan and Huo [194] developed a generic neuro-fuzzy model-based fault detection scheme for the BRB detection in a three-phase IM. The fault detector comprises the generic neuro-fuzzy model and the customized threshold levels. Variable thresholds were selected according to the difference between the output of the generic model and an empirical torque–speed relationship. These are then used to account for variations between machines. This approach overcomes a practical limitation of model-based strategies as it reduces the amount of experimental data that are needed to design the fault detector. Siddique et al. [181] presented a review of AI techniques used for the diagnosis of the SWF in IMs. They showed that till that time the research had been concerned with the development of the ANN and the fuzzy logic in conjunction with several signal processing methods, like the higher order statistics, bi-spectrum, tri-spectrum, cyclo-stationary statistics and wavelets. There is a possibility to add an intelligence diagnostic system to motors itself, providing a level of communications and diagnostic capability. Ye et al. [222] developed an Adaptive Neuro-fuzzy inference systems (ANFIS) with the wavelet packet decomposition to detect the air–gap eccentricity and the BRB in a variable speed drive system. The system could diagnose these faults with current signals even when the precise information regarding motor slip is not available. The proposed method significantly reduces scale and complexity of the system and also speed up the training process. Finally, they showed that the diagnostics was accurate in detection the BRB and the eccentricity. In other study, Ballal et al. [19] also developed ANFIS for successful detection of BRB and BF in IMs. Rodríguez et al. [162] and Rodríguez et al. [163] developed fuzzy systems with the FFT for improving the MCSA in detecting the SWF, BRB and the air–gap eccentricity. The layout has been implemented in MATLAB/ SIMULINK, with both data from a FEM motor simulation program and real measurements. The developed method has the ability to work with variable speed drives and avoids the detailed spectral analysis of the motor current. Tran et al. [204] performed the fault detection of IMs using the ANFIS based on the vibration as well as current signals. In this work, the classification and regression tree (CART) was used for selecting the useful fault features. Romero-Troncoso et al. [165] developed a fault diagnostic methodology for the identification of multiple combined faults in the IM based on the information entropy and the fuzzy logic inference. The fusion of these techniques allowed satisfactory results for this difficult task in an automatic way by investigating one phase of the steady-state current from the rotating machine. Ibrahim and Elzahab [92] implemented the fuzzy modeling system for diagnosing faults in IMs. The stator currents and time domain data were used in the fuzzy model as an input. The experimental work was carried out to verify the efficiency of the developed system. Approximately similar results were found for the simulated and experimental results. Khater et al. [106] performed the IM (inverter feeding) fault detection based on the fuzzy logic. Various inverter faults were simulated and voltage spectrum were measured, which was further used as a database for the fuzzy logic system. The methodology developed was effective for diagnosing the type and location of the faults. Liu et al. [131] replaced Proportional Integral (PI) controllers in the current loops of the rotor flux oriented control (RFOC) by FL controller for multiphase IMs. They showed that the FL controller enhanced the fault-tolerant ability of the nine-phase drive system in various kinds of faulty conditions. Amezquita-Sanchez et al. [10] successfully developed a diagnosis system for BRB by combining fractal dimension and fuzzy logic systems at start-up as well as steady-state condition of IM. De Araujo Cruz, et al. [49] developed fuzzy logic based system to diagnosis fault in IM. In this work, the signals are processed in the frequency and time domain through short time Fourier transform and wavelet multi-resolution analysis. In a study, Kumar et al. [117] used fuzzy based system to develop a direct torque control of IM and represent different faults of IM by MATLAB/SIMULINK. The estimated three-phase currents are further used as input to FL for successful fault diagnosis. 4.3. SVM based fault diagnosis Now a days, SVM have been getting popularity in the field of fault diagnosis [87,97,100,125]. Many researchers have explored this in the IM fault diagnosis also. Initially, SVM was introduced for binary classification only which handles P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 23 two class problem. However later, SVM were developed for multiclass problems to handle the multi-fault situations [38,137,206]. Mainly, three multiclass techniques were developed, i.e. One Versus One (OVO), One-Versus-All (OVA) and Direct Acyclic Graph (DAC)-SVM [85]. All these techniques were used and compared by different researchers for the fault diagnosis of IMs. In a study, Widodo and Yang [215] and Widodo et al. [214] presented the application of SVM and nonlinear feature extraction for the IM fault diagnosis. They employed the PCA and ICA techniques for the nonlinear feature extraction. Finally, they showed that the SVM performance increased significantly with the non-linear feature extraction techniques. In other work, Widodo and Yang [213] performed the IMs fault diagnosis based on the transient current CBM and the wavelet SVM. The PCA and kernel PCA were used to decrease the dimensionality of the features and to find the critical features required for effective diagnostics. In this work, they built the SVM kernel function using three different wavelets, i.e. the Haar, Symlet and Daubechies. A fault diagnostic procedure called the W-SVM was developed by introducing nonlinear kernels using wavelets in the SVM based fault diagnosis. Results showed the significant improvement in the SVM based the fault diagnosis process and performance. Nguyen and Lee [141], and Nguyen et al. [142] presented the diagnostics of the mechanical fault of IMs using the vibration based on the SVM. The study was mainly focused on the selection of useful signal features and SVM parameters. The GA and decision tree were used for selecting critical fault features as well as SVM parameters. In a study, Kurek and Osowski [118] presented fault diagnostic tools for the BRB based on the SVM. In this work, they defined and considered specific fault features from instantaneous forms of the phase current, voltage and shaft magnetic field based on the FFT. The Gaussian kernel based SVM was used for developing two different diagnostic tools. The results showed that the first diagnostic tool detected only the fault occurrence. The second tool could be used for the detection of number of bar damaged. Martínez-Morales et al. [132] performed the vibration and current data fusion for diagnosing the IM mechanical faults (bearing fault, unbalanced and misaligned rotor) based on the multi-class SVM. In this work, signatures created from frequency-domain characteristics are used. Finally, they concluded that the reliability of the developed diagnostics increased significantly by data fusion. Zhang et al. [230] performed bearing fault diagnosis using the WPT and the SVM. They compared the results with the ANN based diagnosis and showed that the SVM is better. In a study, Baccarini et al. [16] showed a practical industrial application of the SVM in diagnosing mechanical faults of the IM. In this study, they also determined the best position for the signal acquisition for vibration signals, which is very important for the maintenance task. This is valuable information to decrease the number of sensors and to reduce the maintenance costs. Konar and Chattopadhyay [111] performed the diagnosis of the bearing fault in IMs using the CWT and the SVM. They used a number of wavelet function and finally concluded that the selection of wavelet is crucial for the intelligent fault diagnosis. Bacha et al. [18] and Salem et al. [167] presented a successful fault diagnosis of IM using the Hilbert-Park transform and the SVM. Two fault signatures i.e., Hilbert modulus current space vector (HMCSV) and Hilbert phase current space vector (HPCSV) are introduced from original signature and further analyzed using classical FFT. After applying magnitude of resulting spectrum with SVM, results show its effectiveness and its robustness in both electrical and mechanical fault detection. Silva and Pederiva [183] performed the IM fault detection based on vibration spectrum analysis using three different AI techniques, i.e. the fuzzy logic, ANN and SVM, and concluded that the SVM has a good generalization capability than other techniques. Chattopadhyay and Konar [41] used features of the continuous and discrete wavelets for the BRB detection using the RBF and MLP neural networks, and the SVM. The greedy-search method with continuous wavelet transform (CWT) is used for the feature selection and found that it is far better than a widely used the DWT technique, even with the high noise. To examine the impact of wavelets on the feature extraction, four wavelets from Daubechies family were selected. Finally, they concluded that db8 gave the best classification accuracy. Also, they showed that performance of classifiers was promising even in the presence of high noise level and with a lower sampling rate of 5.12 kHz. Esfahani et al. [58] used a combination of multiclass linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA), and the SVM for the eccentricity and bearing fault diagnosis of IMs. In this work, the feature obtained using the power spectral density of the current and acoustic signals, and the HHT of vibration data are used in a hierarchical classifier. Finally, the multiclass LDA and the QDA were used for the primary classification, i.e. for the classification of fault category, and the SVM is used for the secondary classification, i.e. for the classification of fault sub-category. The result shows the minimum classification accuracy up to 95% in detection of various IM faults and severity. Seshadrinath et al. [178] performed the inter-turn fault detection in IMs using the dual-tree complex wavelet transform and the SVM. They performed the inter-turn fault detection for a balanced supply conditions, voltage imbalance and interturn fault with the voltage imbalance (both occurring at the same time). Since these faults are often failed to differentiate, this methodology is appropriate in the circumstances where supply imbalances and grid frequency changes are common. Das et al. [48] investigated performance of a load-immune classifier in the detection of minor faults in the stator winding of IMs. Park’s vector modulus (PVM) was obtained through Park’s transformation of three-phase stator line currents in all the experiment cases. Several features were extracted through the time, frequency and time–frequency analyses from the PVM. The SVM based Recursive Feature Elimination (SVM-RFE) algorithm was employed to select, rank and optimize the number of effective features. Finally, they showed the effectiveness of the present method even when the loading level is unknown. For the diagnosis of the BRB, Keskes et al. [105] used a combination of the stationary WPT, and one-versus-all (OVA) and one-versus-one (OVO) SVMs. In this work, the highest prediction accuracy was attained using the Daubechies kernel with the stationary WPT and the OVO-SVM. In other work, Keskes and Braham [103] combined pitch synchronous wavelet transform (PSWT) and the directed acyclic graph (DAG) SVM. Here, the Symlet kernel with the PSWT and DAG SVM perform with highest classification accuracy. In other study, Keskes and Braham [104] used the recursive un-decimated wavelet packet 24 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 transform (RUWPT) and the DAG-SVM. In all the studies, different wavelets, like the Haar, Symlet and Daubechies, are used for kernel functions of the SVM. The highest prediction accuracy was attained using the Symlet kernel function based on DAG-SVM. Moreover, the RUWPT improved the detection time. Vishwakarma et al. [207] proposed the bearing fault diagnosis in IMs based on the wavelet packet decomposition and the SVM, and showed that perfect predictions can be achieved using the energy feature of the 3rd level Daubechies mother wavelet. Palacios et al. [144] performed a comparative investigation of different classifiers f-naive Bayes, k-nearest neighbor, SVM, ANN, decision tree and repeated incremental pruning to produce error reduction. The bearing, stator, and rotor faults were considered for diagnosing. The time domain amplitudes of current signals were analyzed for different loading and power supply conditions, and further used as the input to the classifier. The results showed that all the classifiers were suitable for the detection of IM faults. Konar and Chattopadhyay [112] performed the IM fault diagnosis with the Hilbert and Wavelet transforms using the MLP-NN, RBF-NN and SVM. In this work, the GA is used to find the most relevant fault frequencies, which can be used to effectively diagnose various IM faults even when the noise is present. Consequently, reducing the dimension of the feature space without affecting the overall classification accuracy along with significant reduction in the computation time. In a study, Gangsar and Tiwari [63] performed the fault diagnosis for detecting mechanical faults in IM based on time domain vibration feature and SVM. In this work, the diagnosis is successfully performed for various operating speeds and loads. In other study, Gangsar and Tiwari [64] performed a comparative investigation of the vibration and current signals in order to find out which signal(s) (i.e., vibration or/and current) is/are mandatory for an effective SVM based diagnosis of the electrical and/or mechanical damages of IMs. The investigation concluded that to diagnose mechanical faults alone, the vibration signal is good and sufficient. To diagnose electrical faults alone, the current signal is sufficient; however, when the current signal is not available, the diagnosis can be performed successfully with the vibration signal alone. Moreover, when the mechanical and the electrical faults are considered simultaneously for the diagnosis, both current as well as vibration signals are required for the effective fault diagnosis. From the available literature, it is observed that the features in different domain (time, frequency or time–frequency) perform different for the same problem [62,66, 70]. In a study, Gangsar and Tiwari [65] presented a comparative study of the vibration and current features of the time, frequency and time–frequency domains for diagnosing different IM faults with the SVM. First, three fault features, such as the standard deviation, skewness and kurtosis are calculated based on all three domains. It has been observed that the SVM gives good classification accuracy with the features of all three domains. Though, the wavelet (Shannon) and the SVM with a feature called the standard deviation showed the best classification performance. In other work, Gangsar and Tiwari [66] performed the IM fault diagnosis based on the SVM and the WPT. In this work, five different wavelet functions (i.e., the Haar, Daubechies, Symlet, Coiflet, and Discrete Meyer) with fourteen different statistical features are considered in order to analyze the impact of different wavelets on the IM fault diagnosis. Results demonstrated that all the considered mother wavelets with the mean and standard deviation features could be used for the fault diagnosis of IMs. However, the discrete Meyer with same features shows the highest prediction accuracy. In one study, Gangsar and Tiwari [68] successfully combined SVM and CWT for detecting a number of mechanical and electrical IM faults. In this work, a number of wavelets were tried and found that though all the wavelets showed good prediction performance, the Shannon wavelet was found to be the most appropriate wavelet for the extraction of the significant vibration and current features. Zgarni et al. [229] developed a new extension of the multiclass SVM for the BRB fault diagnosis by embedding the Support Vector Data Description (SVDD) in the classical MSVM. In this work, for the training process, the decision boundaries are constructed with a hyper-sphere. For the feature extraction, the stationary wavelet packet transform (SWPT) with lower sampling rate is applied. Finally, it is found that the present method gives better results as compared to the MSVM. 4.4. Other hybrid AI method based diagnosis As mentioned before, the IM fault diagnosis based on ANN, FL and SVM have their own limitations and research works are continued in this field. In addition, other hybrid and new AI methods have also been explored in the IM fault diagnosis. In a study, Haji and Toliyat [80] successfully developed a fault diagnostics using Bayes minimum error for the BRB. In this work, they used features from Park’s transformation of the current signal. Nakamura et al. [138] presented an IM fault diagnosis based on the Hidden Markov Model (HMM), which is extensively preferred in the speech recognition field. Results showed that the HMM based diagnosis is quite effective to recognize faults such as SWF and BRB. In other work, Park et al. [145] performed the IM fault diagnosis using combined algorithms of the LDA (linear discriminant analysis) and the PCA. The diagnosis is performed with current signals under several noisy conditions to find the robustness of the proposed algorithm and found that it performed better than the individual LDA or PCA. Da Silva et al. [47] used the Gaussian mixture models and the Bayesian maximum model for diagnosing the BRB and the SWF. They used the envelope analysis of the current spectrum for the feature extraction. The method is capable of identifying different fault stages of the BRB and the stator winding fault. Gunal et al. [77] developed diagnostic system based on the Bayesian-Gaussian mixture model, and the Fisher’s linear discriminate analysis using the statistical time domain analysis for diagnosing the air–gap eccentricity and the BRB. This methodology detected the type of faults and load levels in the multi-dimensional feature representation based on time domain features of current. Garcia-Perez et al. [71] performed the high-resolution spectral analysis for detecting the multiple combined faults in IMs using the MUSIC algorithm. In this work they have considered BRB, BF, UR and a combination of these faults for detection based on vibration and current. The finite impulse response (FIR) filter bank with high-resolution spectral analysis was per- P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 25 formed. Finally, they showed that this method has potential to detect the multiple combined faults in IM. Basßaran [21] performed the classification of faulty motors based on three classifiers, namely the linear discriminant classifier (LDC), quadratic discriminant classifier (QDC), and Fisher’s linear discriminant analysis (LDA). In this work, features determined from wavelet packet coefficients were performed corresponding to the frequencies that showed high perturbations in the supply-side current. Ergin et al. [57] presented the fault diagnosis for the stator, rotor and bearing based on a method called the common vector approach (CVA). In this work, they have introduced wavelet energy component based features from one-Dimensional Discrete Wavelet Transform (1D-DWT) of current signal. Result shows that the satisfactory detection result for IM with the developed approach. Tran et al. [203] performed the fault diagnosis in IMs based on the Fourier–Bessel expansion and the simplified fuzzy ARTMAP, a combination of fuzzy logic and neural-network architecture based on the adaptive resonance theory (ART). In this study, generalized discrimination analysis (GDA) is employed to take care of the high dimensionality of current features. The present methodology significantly improves the classification performance. Soualhi et al. [188] developed an improved artificial ant clustering (ACC) technique for the fault diagnosis of IMs. This method is an unsupervised learning techniques which is inspired by the behavior of real ants to optimize the fault identification. Results shows its efficiency in IM fault diagnosis over other supervised learning technique. Seera et al. [177], and Seera and Lim [176] developed an IM fault detection hybrid model comprising fuzzy min–max (FMM), the classification and regression tree (CART), and the ANN. Finally, they showed that this model performs well even in noisy environments in comparison to the FMM and the CART. Zhou et al. [238] developed a new IM fault detection method based on invariant character vectors and showed that this method performs superior than SVM and BP in noisy environment. Bandyopadhyay et al. [20] performed the fault detection by using a combination of image processing and nearest neighbor algorithm tool. Three-phase line currents are acquired from a PWM inverter driving an IM under different faulty situations. The current signal then converted to Concordia patterns as the first stage of data reduction. These Concordia patters are then converted to binary images, discretized and relevant shape descriptor are derived from there. These shape descriptors are further processed by the nearest neighbor algorithm to identify and classify different faults. In AI based fault diagnosis, the selection of different useful signals, like the vibration, current, acoustic, etc. that is required for effective diagnosis of particular machine fault is very critical. It is because different machine or its fault condition produces different features or signatures in different signals. Patel and Giri [147] performed feature selection and identification of IM faults (mechanical) based on the random forest (RF) classifier. They calculated statistical features from the bearing vibration signal and fed to the RF and ANN classifier, and showed that the RF performs better than the ANN. Rajeswaran et al. [154] developed a fault diagnostic methodology using a hybrid artificial intelligence (Neuro-Genetic) of the SVPWM voltage source inverters for IMs. In this work, the IM drive is modeled and analyzed on the MATLAB Simulink. They explored the voltage, current and speed, which were further processed, estimated and converted into output signals representing the torque and the flux. Finally, they compared the results of the methodology with the fuzzy logic, BPN and neuro-fuzzy, and showed that the proposed methodology offers the best results. Singh and Naikan [187] performed the detection of half BRB in IMs based on the motor square current-multiple signal classification (MSC-MUSIC) analysis. In this work, they presented that the BRB produced at least two fault frequency components nearby the line frequency, although square of current signal produced additional frequency components. Finally, they showed that the MSC-MUSIC analysis is more effective than the MCSA for diagnosing half broken bar in rotor under different load conditions. [99] developed a method for detecting the initial short-circuit faults of IMs based on multiple linear regression models with genetic algorithms. They considered the measurement of the RMS supply voltage and the stator current for the analysis. 4.5. New challenges in AI based fault diagnosis After review a number of literatures in last two decades, it is found that many researchers have explored various intelligent systems such as the ANN, fuzzy logic, neuro-fuzzy, and GA for the IM fault diagnosis; however, the use of SVM is rare in the same field [212,239]. The fault monitoring and diagnosis with the Hybrid AI techniques have lots of scope in near future [226,222,72,74–75,169,168,218]. One of the major challenge in AI based system is that the training and testing of AI methods are done with the data which are generated at the same operating conditions (i.e., load and speed) [171,173,193,233]. It is also noted that training data come from the vast history available in the industry and the testing data come from the present online measurement [160, 164,191,192]. However, there are the chances where the fault symptoms’ database may not be available at all the required operating conditions [37,39,62,66]. To take care of this problem, Gangsar and Tiwari [66] and Gangsar and Tiwari [67] developed and performed the SVM based fault diagnosis for IM fault identification for intermediate working conditions in order to handle the cases where the required data is not available. In this study, they considered three cases for the fault diagnosis, i.e. the same speed and load case, the intermediate speed case, and the intermediate load case. The identification performance was found to be effective in the first case and reasonable in last two cases. The prediction performances for the intermediate speed and load cases show the effectiveness of the SVM based diagnostics even with the limited information about operating conditions. However, this results should be further checked with other AI techniques, like ANN, FL etc. Another challenge that almost all the AI based research work that have been done for IM fault diagnosis, the diagnosis has been performed and checked using the signals, which are acquired in the laboratory itself. Result in fault diagnosis have been 26 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 Table 1 A chronological summary of papers on the AI based fault diagnosis of IM. S. No. Reference Fault considered in IM Signals/method AI methods Remarks 1 2 Chow et al. [43] Schoen et al. [175] Lasurt et al. [120] Filippetti et al. [60,60] Salles et al. [168] Kolla and Varatharasa [110] SWF and BF BRB Stator current and rotor speed MCSA/FFT/Spectrum Harmonic Vibration/Higher order statistics Instantaneous voltage and current Current/Time frequency analysis Voltage and current/RMS ANN ANN Accuracy more than 95% – Fuzzy Logic 93–100% accuracy NN, fuzzy logic and Neuro-fuzzy ANN – – ANN – Stator current/Park’s transform Current, voltage and speed/ WPT Bayes Minimum error classifier ANN – MCSA/WPT Vibration and current/Spectra 3 4 5 6 7 8 Haji and Toliyat [80] Kim and Parlos [116,115] 9 10 Ye et al. [224] Kowalski and OrlowskaKowalska [113] 11 Yang et al. [219] 12 Tan and Huo [194] 13 Han et al. [82] 14 16 Zidani et al. [240] Nakamura et al. [138] Ye et al. [222] 17 BRB, SWF, eccentricity IM loads External faults (Overloads, single phasing, unbalance supply voltage, locked rotor, ground fault, over voltage and under voltage) BRB BRB with different severity levels, SWF with different severity levels, BF, Air gap eccentricity BRB, Air gap eccentricity SWF, BRB, BF, Supply asymmetry Successfully developed mode based technology BF, Rotor Damage, SWF, Air gap, misalignment, mechanical unbalance and looseness BRB Vibration ANN NN (Multilayer perceptron network and self-organizing Kohonen network) ART-KNN Current and Torque Neuro-fuzzy BRB, BR, BF, UR, adjustable eccentricity (misalignment) and Phase unbalance SWF Vibration and current ANN/PCA and GA Successfully identified absence and presence of cracked rotor under varying load condition – Current/ Concordia patterns Fuzzy Logic – SWF and BRB Current/spectrum analysis Effective results BRB and Air gap eccentricity Stator current/Wavelet packet decomposition (WPD) Hidden Markov Model (HMM) Adaptive neuro-fuzzy inference systems (ANFIS) Ayhan et al. [14] BRB MCSA 18 Lee et al. [121] BRB, SWF, PUF MCSA/FFT and Wavelet 19 Yang and Kim [220] vibration and current signals ANN and the DempsterShafer theory 20 Ballal et al. [19] BF, SWF, BRB, air–gap eccentricity and phase unbalance Bearing and inter-turn insulation faults ANFIS Performance increases with all five parameter input 21 Bacha et al. [17] BRB and PUF ANN Stray flux monitoring performs better than the current monitoring 22 Su and Chong [190] Martins et al. [133] Huang et al. [90] BRB and the air–gap eccentricity SWF Motor speed, stator current, bearing and winding temperatures, and the machine noise current and magnetic flux monitoring/ Spectrum analysis Vibration/STFT 15 23 24 25 Widodo and Yang [213,215] Widodo et al. [214] Eccentricity BF, SWF, BRB, air–gap eccentricity and phase unbalance Stator current/Principal components Voltage and current signals/ Harmonic amplitudes vibration and current signals / PCA and ICA ANN and the multiple discriminant analysis (MDA) ANN – 93% accuracy 96.88% accuracy Significantly reduces the scale and complexity of the system and speeds up the training of the network ANN provides a higher accuracy performance than the MDA The method is able to predict impending functional failure, significantly in advance Accuracy increases with fusion of vibration and current ANN Unsupervised Hebbianbased NN ANN No need a prior identification of the system – SVM SVM perform better with nonlinear feature extraction techniques 27 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 Table 1 (continued) S. No. Reference Fault considered in IM Signals/method AI methods Remarks 26 Park et al. [145] BF, BR, BRB, static and dynamic eccentricity. Current signals LDA (linear discriminant analysis) and the PCA 27 Rodríguez et al. [162], Rodríguez et al. [163] Silva et al. (2008) [47] SWF, BRB and the air–gap eccentricity MCSA/FFT Fuzzy system Fusion of algorithm is more effective than the individual one. Used both FEM and real measurement data BRB and the SWF Current/Envelope analysis of the current spectrum UR, BF and mechanical looseness Vibration/time domain analysis Gaussian mixture models and the Bayesian maximum model SVM/GA/Decision tree BF, SWF, BRB, air–gap eccentricity and phase unbalance BF, SWF, BRB, air–gap eccentricity and phase unbalance Air-gap eccentricity and BRB Transient current/WPT SVM/PCA/ICA SVM effectively perform even with small no. of datasets Vibration as well as current signals ANFIS/CART – Current/time domain analysis – BRB MCSA/WPT Bayesian-Gaussian mixture model, and Fisher’s linear discriminate analysis ANN BRB Instantaneous forms of the phase current, voltage and shaft magnetic field/FFT Vibration and current/ Frequency domain characteristics Vibration/WPT 28 29 30 Nguyen and Lee [141], and Nguyen et al. [142] Widodo and Yang [213] 31 Tran et al. [204] 32 Gunal et al. [77] 33 Sadeghian et al. [166] Kurek and Osowski [118] 34 35 36 37 MartínezMorales et al. [132] Zhang et al. [230] Garcia-Perez et al. [71] Baccarini et al. [16] BF, MR and UR BF Reliable with data fusion ANN and SVM SVM gives better result than ANN Successfully detected the combined multiple fault Selected best position for signals and performed reliable diagnosis Selection of wavelet is crucial for fault diagnosis MUSIC algorithm Konar and Chattopadhyay [111] Basßaran [21] BF Vibration/CWT SVM BF, SWF, and BRB Current/WPT 41 Ghate and Dudul [72] Rotor eccentricity, SWF, and both these defects simultaneously Stator current/time domain features/PCA 42 RomeroTroncoso et al. [165] Bacha et al. [18] Multiple combined faults Current/Information entropy Linear discriminant classifier (LDC), quadratic discriminant classifier (QDC), and Fisher’s linear discriminant analysis (LDA). Radial basis function (RBF) and the multilayer perception (MLP) cascade NN Fuzzy logic inference Unbalanced voltage, broken rotor bar, air–gap eccentricity and outer raceway ball bearing defect. Current/Advanced Hilbert park transform SVM Salem et al. [167] Air-gap eccentricity and outer raceway ball bearing defect Current/Advanced Hilbert park transform SVM 40 43 44 Effective even at low load SVM Vibration and current/highresolution spectral analysis Vibration/FFT 39 Classification significantly improves after feature selection SVM BRB, BF, UR and a combination of all three Mechanical faults (UR, MR and mechanical looseness) 38 – SVM Successfully classified all the faults Classifier is robust even in case of noise data – Proposed two fault signatures i.e., Hilbert modulus current space vector (HMCSV) and Hilbert phase current space vector (HPCSV) shows its effectiveness and its robustness in both electrical and mechanical fault detection – (continued on next page) 28 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 Table 1 (continued) S. No. Reference Fault considered in IM Signals/method AI methods Remarks 45 Ergin et al. [57] BF, BRB, SWF Common vector approach (CVA). 46 Tran et al. [203] BR, BRB, eccentricity, BF, UR and PUF Stator current/ OneDimensional Discrete Wavelet Transform (1D-DWT) Transient current/time and frequency domain analysis/ Fourier–Bessel expansion Wavelet energy componentbased features show effective fault classification with CVA Present methodology significantly improves the classification performance 47 Soualhi et al. [188] BRB and BF Stator current/ spectrum analysis 48 Silva and Pederiva [183] Chattopadhyay and Konar [41] Zarei et al. [228] UR, MR, Mech. looseness SWF, PUF and BRB BRB Vibration/spectrum analysis 49 50 BF Continuous and discrete wavelets transform Vibration/time domain analysis A combination of fuzzy logic and neuralnetwork architecture based on the adaptive resonance theory (ART). Improved artificial ant clustering (ACC) technique Fuzzy logic, ANN and SVM RBF and MLP neural networks, and the SVM Cascade NN 51 Seera et al. [177], and Seera and Lim [176] PUF, BRB, SWF and eccentricity Stator current/Spectrum analysis 52 [58] Eccentricity and BF Current, vibration and acoustic/ power spectral density/HHT 53 Seshadrinath et al. [178] SWF, PUF, and both at the same time Current/Dual-tree complex wavelet transform 54 Zhou et al. [238] BF, BRB, SWF and rotor eccentricity Stator current/Spectrum analysis Invariant character vectors 55 Das et al. [48] SWF Current/Park’s vector modulus (PVM) 56 Keskes et al. [105] BRB Current/Stationary WPT 57 Keskes and Braham [103] Keskes and Braham [104] BRB Current/Pitch synchronous wavelet transform (PSWT) Current/Recursive undecimated wavelet packet transform (RUWPT) and VIbration/WPT SVM based Recursive Feature Elimination (SVM-RFE) One-versus-all (OVA) and one-versus-one (OVO) SVMs. Directed acyclic graph (DAG) SVM. DAG-SVM 58 BRB Hybrid model comprising fuzzy min– max (FMM), the classification and regression tree (CART), and the ANN A combination of multiclass linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA), and the SVM SVM 59 Wu et al. [216] BF 60 Vishwakarma et al. [207] BF 61 Palácios et al. [144] BRB, SWF, BF 62 Konar and Chattopadhyay [112] BRB, SWF, PUF, BF, UR and BR Vibration/Hilbert and Continuous Wavelet transforms 63 Ibrahim and Elzahab [92] Stator currents/time domain analysis Fuzzy modeling system 64 Lashkari et al. [119] Open phase, PUF, Overload fault, Over voltage fault and under voltage fault SWF and PUF Three-phase shifts between the line-current and phasevoltage ANN Vibration/time domain analysis and wavelet packet decomposition Current/Time domain amplitudes of current signals ANN SVM f-naive Bayes, k-nearest neighbor, SVM, ANN, decision tree MLP-NN, RBF-NN and SVM The resent unsupervised learning based technique perform better as compared to other supervised learning techniques SVM has a good generalization capability CWT feature performs better even with high noise Present approach shows good accuracy despite low quality (noisy) of measured vibration signal. This hybrid model performs better than individual classifier even with noise Discriminates fault and severity with minimum 95% accuracy Present method is appropriate for distinguishing PUF and SWF This method shows superior result than SVM and BP in noisy environment Present method perform good classification even when loading level is unknown OVO-SVM perform best with Daubechies wavelet kernel function DAG SVM perform best with Symlet kernel DAG-SVM perform fast, robust and effective classification with Symlet kernel 90% accuracy with daubechies wavelet – All the classifiers are suitable for the detection of IM faults. GA is successfully used for feature reduction which significantly improves the classification performance Simulated results are successfully validated with the experimental results Diagnose and trace the location of the fault 29 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 Table 1 (continued) S. No. Reference Fault considered in IM Signals/method AI methods Remarks 65 Bessam et al. [31] BRB NN Successfully detect the correct no. of BRB 66 Bandyopadhyay et al. [20] A combination of image processing and nearest neighbor algorithm tool Present method perform better than the TPS_RPM algorithm and the SVM classifier 67 Huang et al. [88] Improper contact points, poor connections and problematic solder joints in Insulated Gate Bipolar Transistor (IGBT) switching Device of IM BRB, BF,UR and MR Current/Harmonic frequencies and corresponding amplitudes/Hilbert Transform Current/ Concordia patterns/ binary images/shape descriptor vibration signal/fault frequency Extension neural network (ENN). 68 Patel and Giri [147] Bazan et al. [22] BF vibration signal SWF Khater et al. [106] Rajeswaran et al. [154] Open and close circuit Current/measures of mutual information between the phase current signals Voltage/spectrum analysis Random forest (RF) and ANN classifer ANN Short circuit fault Voltage, current and speed/ A hybrid artificial intelligence (NeuroGenetic) of the SVPWM 72 Singh and Naikan [187] Half BRB Current/spectrum analysis 73 [99] Initial short-circuit faults 74 Gangsar and Tiwari [63] Gangsar and Tiwari [70,69] Mechanical faults (BF, UR, MR, BR) BRB, SWF (two severity levels), PUF (two severity levels), BF, UR, BR, and MR Supply voltage and the stator current/measurement of the RMS Current/time domain analysis Motor square currentmultiple signal classification (MSCMUSIC) analysis. Multiple linear regression models with genetic algorithms SVM ENN perform better than NN in training time and classification accuracy The RF performs better than the ANN. Effective even in case of variable loading and phase unbalance Effective for diagnosing the type and location of the faults This methodology offers the best result as compared to fuzzy logic, BPN and neurofuzzy, - 76 Glowacz and Glowacz [74] BRB and faulty ring 77 Glowacz et al. [75] and Glowacz [73] SWF and BF 78 Zgarni, et al. [229] BRB 79 Gangsar and Tiwari [66] 80 Gangsar and Tiwari [67] BRB, SWF (two severity levels), PUF (two severity levels), BF, UR, BR, and MR BRB, SWF (two severity levels), PUF (two severity levels), BF, UR, BR, and MR 81 Gangsar and Tiwari [68] 82 Gangsar and Tiwari [69] 69 70 71 75 BRB, SWF (two severity levels), PUF (two severity levels), BF, UR, BR, and MR BRB, SWF (two severity levels), PUF (two severity levels), BF, UR, BR, and MR Fuzzy logic Current and vibration/ Time domain analysis SVM Thermal imaging/ MoASoID (Method of Areas Selection of Image Differences). Acoustic signals/ Shortened Method of Frequencies Selection Multi-expanded 2 Groups (SMOFS-32MULTIEXPANDED-2-GROUPS) and the SMOFS-32MULTIEXPANDED-1-GROUPS. SWPT NN, K-means, BNN Vibration and current/WPT NN classifier, back propagation neural network (BNN) and modified algorithm using words coding Support Vector Data Description (SVDD) with classical MSVM SVM Vibration and current/Time domain analysis SVM Vibration and current/CWT SVM Vibration and current/ Frequency domain analysis SVM – Successful fault detection even at intermediate speed case Both current and vibration are required for fault diagnosis of mechanical and electrical fault simultaneously Successfully classified fault but expensive and time taking method Acoustic signals are very effective for fault diagnosis Present method gives better results as compared to the MSVM All wavelet with specific feature perform good for fault diagnosis Performance is very effective for the same speed and load case,, and very encouraging for intermediate speed and load cases 30 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 found to be effective. The reason may be that, in these studies various statistical features are extracted from the signals that are generated in controlled environment and finally most suitable features are selected that represent each IM fault effectively. Further, these selected features make separate clusters in the feature space for each faults and that gives better fault diagnosis. However, it is noted that, in these cases the training and testing data are collected from the same source and in the same controlled environment. If the same fault diagnosis problem is considered for an industrial setting, there is almost always some difference between the training and testing distributions, even in the large-sample limit. These differences can be due to variations in operating conditions, changes in environment conditions, variations of fault types of same class, sensor noise, etc. However, even if there are the uncertain signal observations (may be due to additional extraneous inputs and variation of parametric condition), the basic features of all the faults would remain same in dynamics or electrodynamics. Therefore, it is expected that, using the suitable features the algorithm would be able to successfully classify the IM faults. Generalization is one of the challenge in the AI systems to produce outputs for inputs not encountered during training for the cases where variance is high and information is low. It gives ability to produce results as near as real outputs. To make an AI system more generic, the pre-processing of the signal is one of the most necessary task to do. The present study suggests that having an appropriate feature or input set is more effective for a fault diagnosis problem, which improves the diagnosis accuracy and reduces the computational burden. Therefore, in order to solve the same problem in an industrial setting, identification of the crucial feature(s) for a particular system is compulsory. Moreover, normalization of the crucial features can also help to make a system more generalize. In addition, input parameters of the machine learning system have to be optimized before building the optimal system for the final fault prediction. To inspect the potency of the classifier, the fault identification could also be executed at different operating or working conditions. The problem of generalization can be solved by making a learning system more adaptive. More the learning system understands about the real situation, the better it is able to obtain learning signals, perhaps with fewer samples. It is possible by throwing enough diversity of data at the diagnosing problem, so that the learning system can be pushed to develop a generalized model. This paper focused on the recent development of condition monitoring techniques for the fault detection and diagnosis of IMs which also includes AI based techniques. The summary of all the papers is also tabulated in Table 1. In the next section, the outcomes of the present literature review are summarized. 5. Observations, research gaps and ideas for future research The literature which mainly focuses on the fault detection and diagnosis of IMs has been reviewed. From the literature reviewed in previous sections, following important observations have been drawn: IMs are the workhorses and critical machines of any industries as they maintain and accelerate the manufacturing processes. Accordingly, industries are ready to make a great effort for the monitoring, and early detection and diagnosis of the induced and incipient faults in IMs, especially in its critical applications, before it causes the unscheduled maintenance and in the most of cases a complete breakdown of IMs. Several types of IM faults, for example the bearing fault, stator winding fault, broken rotor faults, rotor related faults, and phase unbalance have been considered. However, combined study of these faults and their severity is still rare. Various techniques for the IM fault diagnosis have been applied by using the MCSA, vibration analysis, electromagnetic torque measurement, acoustic analysis, and thermal analysis. However, the most popular techniques are vibration signature analysis and the MCSA because they are easily measurable, highly accurate and reliable. Generally, research papers have considered maximum two to three types of faults that are simulated in different IMs, however, it is very essential to consider all the possible faults occurring in the mechanical and electrical components of IMs, and simultaneously diagnose them so that the chances of further damage or complete motor failure due to occurrence of any specific fault can be reduced. In practical applications, different severity levels of faults may be developed, so it is very important to consider faults under progression to detect the faults at an incipient stage, which is very rare in the literature. A diagnosing system that can detect faults as well as their severity simultaneously is still less explored. This work is of practical significance because if the methodology is robust enough to classify a number of different fault conditions including different severity levels, then it can be easily used to classify any one fault condition and healthy motor condition. In addition, it is very useful when a number of IMs are working in the industry need to be diagnosed simultaneously based on a single developed methodology. A single fault detection system that can detect faults occurring simultaneously in various IMs of same kind is not available in the literature. As the condition monitoring techniques for fault detection and diagnosis of rotating machines has been improvised from conventional methods to AI methods, so there is much scopes of research in this field. The AI based diagnostic systems still have several challenging tasks to accomplish in regards to its efficiency, reliability, computational time, sufficient database, and robustness. Nowadays, the SVM is frequently used method for diagnosing faults in various machines due to its better classification accuracy and less computational time. The diagnosis of multi-faults in IMs based on the SVM is still uncommon. P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 31 In aforementioned works, a higher sampling frequency has been adopted to acquire the signals in AI based fault detection, however, it is very significant to choose adjusted sampling frequency and data points (or frequency resolution) that can improve fault diagnostic and detection time, and reduce the cost of implementation. The AI, for the purpose of pattern recognition or fault diagnosis, include a large collection of very different types of mathematical tools, namely, the pre-processing, extraction and selection of suitable statistical features, selection of AI parameters and final recognition. In many cases, it is difficult to say which features are significant and what kind of tool would be best for a particular problem and machines. In order to extract useful fault features, a number of signal processing methods are applied, for example the time, frequency, or more advanced time–frequency approach, like wavelets. The selection of domain and representative features is governed by the nature of signals and needed information from the machine. From the available literature, it could be summarized that the wavelets have still an immense prospect and still very few people have tried it for the AI based fault diagnosis of IMs. In addition, as wavelet having immense information, a combined study of wavelet and comparatively new AI methods such as the SVM and deep learning particularly for the IM fault identification can be further explored. Based on existing literatures, it could be summarized that several mother wavelets are available that can be used with the AI based fault diagnostic. In spite of a lot of study on wavelets, the choice of the mother wavelet and their features, which is considerable issue of intelligent fault diagnosis, is still open for discussions. The research observation so far is confined to diagnose the faults at specific operational condition of IMs and the observations shows that it is very difficult to detect faults in IMs for light loads. In addition, the accuracy of fault diagnosis may reduce due to occurrence of fluctuation of rotor speeds during data acquisition under different loadings of IMs. Thus, considering the influence of motor operational conditions on the AI based fault diagnosis is still an open challenge. Usually, AI systems are trained using a symptoms database available by the measurement in the industry. However, it is not always possible to have a database at all IM operational conditions. The AI based fault diagnosis is difficult when test data do not match the database used for the training. The fault diagnosis of IMs, when the training and testing data available at different operating conditions used for the diagnosis, is hardly available in the literature. The performance of AI based fault diagnostics can be checked when the training of algorithm is performed with data obtained from a particular IM and the testing is performed with data obtained from a different configuration of IMs. It would be helpful for the generalization of the problem by normalizing signal features. Furthermore, it will be interesting to use the training data that are generated by numerical models of IMs based on multiphysics and the testing data from signals, which are acquired using an experimental measurement. However, it will be a great challenge to make an accurate numerical model for each possible fault in IMs. The effect of simultaneous fault occurrence in IMs on the performance of the AI based fault diagnostics can also be studied. The fault diagnosis of different systems can be tried simultaneously for the combined rotating systems such as IM attached with a pump, IM attached with a gearbox, IM attached with the rotor system, etc., using measurement at one location. Practical constraints of availability of limited system signal to be addresses by applying the AI based methodology in real industry setup. 6. Concluding Remarks In this paper, an attempt has been made to review the researches and developments of the present fault detection and diagnosis techniques of IMs for a number of electrical and mechanical faults. Capabilities and restrictions of these techniques have also been discussed in details. AI based fault diagnosis techniques of IMs, which are being more popular due to their effectiveness and so many advantages over conventional signal based techniques, is added here. As reported by discussed literatures, AI based condition monitoring and fault diagnosis techniques for IMs still require more encouragement and attention in the future, due to lack of research in this field. The generalization is one of the crucial thing in the machine learning to produce outputs for inputs not encountered during training for the cases where variance is high and information is low. References [1] M. Abd-el-Malek, A.K. Abdelsalam, O.E. Hassan, Induction motor broken rotor bar fault location detection through envelope analysis of start-up current using Hilbert transform, Mech. Syst. Sig. Process. 93 (2017) 332–350. [2] G.G. Acosta, C.J. Verucchi, E.R. Gelso, A current monitoring system for diagnosing electrical failures in induction motors, Mech. Syst. Sig. Process. 20 (4) (2006) 953–965. [3] M. Akar, H.S. Gercekcioglu, Instantaneous power factor signature analysis for efficient fault diagnosis in inverter fed three phased induction motors, Int. J. Hydrogen Energy 42 (12) (2017) 8338–8345. [4] Akcay, H., & Germen, E. (2013, September). Identification of acoustic spectra for fault detection in induction motors. In AFRICON, 2013 (pp. 1-5). IEEE. [5] P.F. Albrecht, J.C. Appiarius, E.P. Cornell, D.W. Houghtaling, R.M. McCoy, E.L. Owen, D.K. Sharma, Assessment of the reliability of motors in utility applications, IEEE Trans. Energy Convers. 3 (1987) 396–406. [6] M.A. Alsaedi, Fault diagnosis of three-phase induction, Motor: A Reviewk. Optics. Special Issue: Applied Optics and Signal Process. 4 (1–1) (2015) 1–8. 32 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 [7] S. Altug, M.Y. Chen, H.J. Trussell, Fuzzy inference systems implemented on neural architectures for motor fault detection and diagnosis, IEEE Trans. Ind. Electron. 46 (6) (1999) 1069–1079. [8] Ameid, T., Menacer, A., Talhaoui, H., & Azzoug, Y. (2018). Discrete wavelet transform and energy eigen value for rotor bars fault detection in variable speed field-oriented control of induction motor drive. ISA Transactions. [9] T. Ameid, A. Menacer, H. Talhaoui, I. Harzelli, Rotor resistance estimation using Extended Kalman filter and spectral analysis for rotor bar fault diagnosis of sensorless vector control induction motor, Measurement 111 (2017) 243–259. [10] J.P. Amezquita-Sanchez, M. Valtierra-Rodriguez, C.A. Perez-Ramirez, D. Camarena-Martinez, A. Garcia-Perez, R.J. Romero-Troncoso, Fractal dimension and fuzzy logic systems for broken rotor bar detection in induction motors at start-up and steady-state regimes, Meas. Sci. Technol. 28 (7) (2017) 075001. [11] J. Antonino-Daviu, P. Jover, M. Riera, A. Arkkio, J. Roger-Folch, DWT analysis of numerical and experimental data for the diagnosis of dynamic eccentricities in induction motors, Mech. Syst. Sig. Process. 21 (6) (2007) 2575–2589. [12] M.A. Awadallah, M.M. Morcos, Application of AI tools in fault diagnosis of electrical machines and drives-an overview, IEEE Trans. Energy Convers. 18 (2) (2003) 245–251. [13] M.A. Awadallah, M.M. Morcos, Automatic diagnosis and location of open-switch fault in brushless DC motor drives using wavelets and neuro-fuzzy systems, IEEE Trans. Energy Convers. 21 (1) (2006) 104–111. [14] B. Ayhan, M.Y. Chow, M.H. Song, Multiple discriminant analysis and neural-network-based monolith and partition fault-detection schemes for broken rotor bar in induction motors, IEEE Trans. Ind. Electron. 53 (4) (2006) 1298–1308. [16] L.M.R. Baccarini, V.V.R. e Silva, B.R. De Menezes, W.M. Caminhas, SVM practical industrial application for mechanical faults diagnostic, Expert Syst. Appl. 38 (6) (2011) 6980–6984. [17] K. Bacha, H. Henaob, M. Gossa, G.A. Capolino, Induction machine fault detection using stray flux EMF measurement and neural network-based decision, Electr. Power Syst. Res. 78 (7) (2007) 1247–1255. [18] K. Bacha, S. Ben Salem, A. Chaari, An improved combination of Hilbert and Park transforms for fault detection and identification in three-phase induction motors, Int. J. Electr. Power Energy Syst. 43 (1) (2012) 1006–1016. [19] M.S. Ballal, Z.J. Khan, H.M. Suryawanshi, R.L. Sonolikar, Adaptive neural fuzzy inference system for the detection of inter-turn insulation and bearing wear faults in induction motor, IEEE Trans. Ind. Electron. 54 (1) (2007) 250–258. [20] I. Bandyopadhyay, P. Purkait, C. Koley, A combined image processing and Nearest Neighbor Algorithm tool for classification of incipient faults in induction motor drives, Comput. Electr. Eng. 54 (2016) 296–312. [21] M. Basßaran, Condition monitoring of speed controlled induction motors using wavelet packets and discriminant analysis, Expert Syst. Appl. 38 (7) (2011) 8079–8086. [22] G.H. Bazan, P.R. Scalassara, W. Endo, A. Goedtel, W.F. Godoy, R.H.C. Palácios, Stator fault analysis of three-phase induction motors using information measures and artificial neural networks, Electr. Power Syst. Res. 143 (2017) 347–356. [23] G.H. Bazan, P.R. Scalassara, W. Endo, A. Goedtel, R.H.C. Palacios, W.F. Godoy, Stator short circuit diagnosis in induction motors using mutual information and intelligent systems, IEEE Trans. Ind. Electron. (2018). [24] A.M. Bazzi, P.T. Krein, Review of methods for real-time loss minimization in induction machines, IEEE Trans. Ind. Appl. 46 (6) (2010) 2319–2328. [25] A. Bellini, F. Filippetti, G. Franceschini, C. Tassoni, G.B. Kliman, Quantitative evaluation of induction motor broken bars by means of electrical signature analysis, IEEE Trans. Ind. Appl. 37 (5) (2001) 1248–1255. [26] A. Bellini, F. Filippetti, C. Tassoni, G.A. Capolino, Advances in diagnostic techniques for induction machines, IEEE Trans. Ind. Electron. 55 (12) (2008) 4109–4126. [27] M.E.H. Benbouzid, G.B. Kliman, What stator current processing-based technique to use for induction motor rotor faults diagnosis?, IEEE Trans. Energy Convers. 18 (2) (2003) 238–244. [28] M.E.H. Benbouzid, A review of induction motors signature analysis as a medium for faults detection, IEEE Trans. Ind. Electron. 47 (5) (2000) 984–993. [29] Benbouzid, M. E. H., & Nejjari, H. (2001, June). A simple fuzzy logic approach for induction motors stator condition monitoring. In IEMDC 2001. IEEE International Electric Machines and Drives Conference (Cat. No. 01EX485) (pp. 634–639). IEEE. [30] M.E.H. Benbouzid, M. Vieira, C. Theys, Induction motors’ faults detection and localization using stator current advanced signal processing techniques, IEEE Trans. Power Electron. 14 (1) (1999) 14–22. [31] B. Bessam, A. Menacer, M. Boumehraz, H. Cherif, Detection of broken rotor bar faults in induction motor at low load using neural network, ISA Trans. 64 (2016) 241–246. [32] B. Bessam, A. Menacer, M. Boumehraz, H. Cherif, Wavelet transform and neural network techniques for inter-turn short circuit diagnosis and location in induction motor, Int. J. Syst. Assurance Eng. Manage. 8 (1) (2017) 478–488. [33] P. Bilski, Application of Support Vector Machines to the induction motor parameters identification, Measurement 51 (2014) 377–386. [34] M. Blodt, P. Granjon, B. Raison, G. Rostaing, Models for bearing damage detection in induction motors using stator current monitoring, IEEE Trans. Ind. Electron. 55 (4) (2008) 1813–1822. [35] A.H. Bonnett, G.C. Soukup, Analysis of rotor failures in squirrel-cage induction motors, IEEE Trans. Ind. Appl. 24 (6) (1988) 1124–1130. [37] D.J. Bordoloi, R. Tiwari, Support vector machine based optimization of multi-fault classification of gears with evolutionary algorithms from time– frequency vibration data, Measurement 55 (2014) 1–14. [38] C.J. Burges, A tutorial on support vector machines for pattern recognition, Data Min. Knowl. Discovery 2 (2) (1998) 121–167. [39] S.C. Chang, R. Yacamini, Experimental study of the vibrational behaviour of machine stators, IEE Proc. Electric Power Appl. 143 (3) (1996) 242–250. [41] P. Chattopadhyay, P. Konar, Feature extraction using wavelet transform for multi-class fault detection of induction motor, J. Inst. Eng. (India): Ser. B (2014) 1–9. [42] F. Cheng, L. Qu, W. Qiao, Fault prognosis and remaining useful life prediction of wind turbine gearboxes using current signal analysis, IEEE Trans. Sustain. Energy 9 (1) (2018) 157–167. [43] M.Y. Chow, P.M. Mangum, S.O. Yee, A neural network approach to real-time condition monitoring of induction motors, IEEE Trans. Ind. Electron. 38 (6) (1991) 448–453. [44] T.W.S. Chow, G. Fei, Three phase induction machines asymmetrical faults identification using bispectrum, IEEE Trans. Energy Convers. 10 (4) (1995) 688–693. [45] B. Corne, B. Vervisch, S. Derammelaere, J. Knockaert, J. Desmet, The reflection of evolving bearing faults in the stator current’s extended park vector approach for induction machines, Mech. Syst. Sig. Process. 107 (2018) 168–182. [46] I.M. Culbert, W. Rhodes, Notice of violation of IEEE publication principles using current signature analysis technology to reliably detect cage winding defects in squirrel-cage induction motors, IEEE Trans. Ind. Appl. 43 (2) (2007) 422–428. [47] A.M. Da Silva, R.J. Povinelli, N.A. Demerdash, Induction machine broken bar and stator short-circuit fault diagnostics based on three-phase stator current envelopes, IEEE Trans. Ind. Electron. 55 (3) (2008) 1310–1318. [48] S. Das, P. Purkait, C. Koley, S. Chakravorti, Performance of a load-immune classifier for robust identification of minor faults in induction motor stator winding, IEEE Trans. Dielectr. Electr. Insul. 21 (1) (2014) 33–44. [49] A.G. De Araujo Cruz, R.D. Gomes, F.A. Belo, A.C. Lima Filho, A hybrid system based on fuzzy logic to failure diagnosis in induction motors, IEEE Lat. Am. Trans. 15 (8) (2017) 1480–1489. [50] P.A. Delgado-Arredondo, D. Morinigo-Sotelo, R.A. Osornio-Rios, J.G. Avina-Cervantes, H. Rostro-Gonzalez, R. de Jesus Romero-Troncoso, Methodology for fault detection in induction motors via sound and vibration signals, Mech. Syst. Sig. Process. 83 (2017) 568–589. [51] G. Didier, E. Ternisien, O. Caspary, H. Razik, A new approach to detect broken rotor bars in induction machines by current spectrum analysis, Mech. Syst. Sig. Process. 21 (2) (2007) 1127–1142. P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 33 [52] H. Douglas, P. Pillay, A.K. Ziarani, A new algorithm for transient motor current signature analysis using wavelets, IEEE Trans. Ind. Appl. 40 (5) (2004) 1361–1368. [53] M. Drif, A.J.M. Cardoso, Rotor cage fault diagnostics in three phase induction motors, by the instantaneous non-active power signature analysis, Proc. IEEE Int. Symp. Ind. Electron. (2007) 1050–1055. [54] B.M. Ebrahimi, J. Faiz, S. Lotfi-Fard, P. Pillay, Novel indices for broken rotor bars fault diagnosis in induction motors using wavelet transform, Mech. Syst. Sig. Process. 30 (2012) 131–145. [55] E. Elbouchikhi, V. Choqueuse, M. Benbouzid, Induction machine bearing faults detection based on a multi-dimensional MUSIC algorithm and maximum likelihood estimation, ISA Trans. 63 (2016) 413–424. [56] El-Shafei, A., & Rieger, N. (2003). Automated diagnostics of rotating machinery. ASME Turbo Expo, vol. 4, Atlanta, GA, USA, pp. 491–498. [57] S. Ergin, A. Uzuntas, M.B. Gulmezoglu, Detection of stator, bearing and rotor faults in induction motors, Procedia Eng. 30 (2012) 1103–1109. [58] E.T. Esfahani, S. Wang, V. Sundararajan, Multisensor wireless system for eccentricity and bearing fault detection in induction motors, IEEE/ASME Trans. Mechatron. 19 (3) (2014) 818–826. [59] J. Faiz, M. Ojaghi, Different indexes for eccentricity faults diagnosis in three-phase squirrel-cage induction motors: a review, Mechatronics 19 (1) (2009) 2–13. [60] F. Filippetti, G. Franceschini, C. Tassoni, P. Vas, Recent developments of induction motor drives fault diagnosis using AI techniques, IEEE Trans. Ind. Electron. 47 (5) (2000) 994–1004. [61] R. Fiser, S. Ferkolj, Application of a finite element method to predict damaged induction motor performance, IEEE Trans. Magn. 37 (5) (2001) 3635– 3639. [62] P. Gangsar, R. Tiwari, Multiclass fault taxonomy in rolling bearings at interpolated and extrapolated speeds based on time domain vibration data by SVM algorithms, J. Fail. Anal. Prev. 14 (6) (2014) 826–837. [63] P. Gangsar, R. Tiwari, Taxonomy of induction-motor mechanical-fault based on time-domain vibration signals by multiclass SVM classifiers, Intell. Ind. Syst. 2 (3) (2016) 269–281. [64] P. Gangsar, R. Tiwari, Comparative investigation of vibration and current monitoring for prediction of mechanical and electrical faults in induction motor based on multiclass-support vector machine algorithms, Mech. Syst. Sig. Process. 94 (2017) 464–481. [65] Gangsar, P., & Tiwari, R. (2017, December). Analysis of Time, Frequency and Wavelet Based Features of Vibration and Current Signals for Fault Diagnosis of Induction Motors Using SVM. In ASME 2017 Gas Turbine India Conference (pp. V002T05A027-V002T05A027). American Society of Mechanical Engineers. [66] P. Gangsar, R. Tiwari, Multifault diagnosis of induction motor at intermediate operating conditions using wavelet packet transform and support vector machine, J. Dyn. Syst. Meas. Contr. 140 (8) (2018) 081014. [67] P. Gangsar, R. Tiwari, A support vector machine based fault diagnostics of Induction motors for practical situation of multi-sensor limited data case, Measurement 135 (2019) 694–711. [68] P. Gangsar, R. Tiwari, Diagnostics of mechanical and electrical faults in induction motors using wavelet-based features of vibration and current through support vector machine algorithms for various operating conditions, J. Braz. Soc. Mech. Sci. Eng. 41 (2) (2019) 71. [69] P. Gangsar, R. Tiwari, Online diagnostics of mechanical and electrical faults in induction motor using multiclass support vector machine algorithms based on frequency domain vibration and current signals, ASCE-ASME J. Risk nd Uncertainty in Eng. Syst. Part B Mech. Eng. 5 (3) (2019) 031001. [70] R.X. Gao, R. Yan, Non-stationary signal processing for bearing health monitoring, Int. J. Manuf. Res. 1 (1) (2006) 18–40. [71] A. Garcia-Perez, R. de Jesus Romero-Troncoso, E. Cabal-Yepez, R.A. Osornio-Rios, The application of high-resolution spectral analysis for identifying multiple combined faults in induction motors, Ind. Electron. IEEE Trans. 58 (5) (2011) 2002–2010. [72] V.N. Ghate, S.V. Dudul, Cascade neural-network-based fault classifier for three-phase induction motor, IEEE Trans. Ind. Electron. 58 (5) (2011) 1555– 1563. [73] A. Glowacz, Acoustic based fault diagnosis of three-phase induction motor, Appl. Acoust. 137 (2018) 82–89. [74] A. Glowacz, Z. Glowacz, Diagnosis of the three-phase induction motor using thermal imaging, Infrared Phys. Technol. 81 (2017) 7–16. [75] A. Glowacz, W. Glowacz, Z. Glowacz, J. Kozik, Early fault diagnosis of bearing and stator faults of the single-phase induction motor using acoustic signals, Measurement 113 (2018) 1–9. [76] S. Grubic, J.M. Aller, B. Lu, T.G. Habetler, A survey on testing and monitoring methods for stator insulation systems of low-voltage induction machines focusing on turn insulation problems, IEEE Trans. Ind. Electron. 55 (12) (2008) 4127–4136. [77] S. Gunal, D. Gokhan Ece, O. Nezih Gerek, Induction machine condition monitoring using notch-filtered motor current, Mech. Syst. Sig. Process. 23 (8) (2009) 2658–2670. [78] K.N. Gyftakis, A.J.M. Cardoso, J.A. Antonino-Daviu, Introducing the filtered park’s and filtered extended park’s vector approach to detect broken rotor bars in induction motors independently from the rotor slots number, Mech. Syst. Sig. Process. 93 (2017) 30–50. [79] K.N. Gyftakis, D.V. Spyropoulos, J.C. Kappatou, E.D. Mitronikas, A novel approach for broken bar fault diagnosis in induction motors through torque monitoring, IEEE Trans. Energy Convers. 28 (2) (2013) 267–277. [80] M. Haji, H.A. Toliyat, Pattern recognition-a technique for induction machines rotor broken bar detection, IEEE Trans. Energy Convers. 16 (4) (2001) 312–317. [81] Halme, J. (2002, December). Condition monitoring of oil lubricated ball bearing using wear debris and vibration analysis. In Proceedings of the International Tribology Conference (AUSTRIB’02), Frontiers in tribology, Perth, University of Western Australia (pp. 2-5). [82] Han, T., Yang, B. S., & Lee, J. M. (2005, May). A new condition monitoring and fault diagnosis system of induction motors using artificial intelligence algorithms. In Electric Machines and Drives, 2005 IEEE International Conference on (pp. 1967–1974). IEEE. [83] H. Henao, G.A. Capolino, M. Fernandez-Cabanas, F. Filippetti, C. Bruzzese, E. Strangas, S. Hedayati-Kia, Trends in fault diagnosis for electrical machines: a review of diagnostic techniques, IEEE Ind. Electron. Mag. 8 (2) (2014) 31–42. [84] H. Henao, C. Demian, G.A. Capolino, A frequency-domain detection of stator winding faults in induction machines using an external flux sensor, IEEE Trans. Ind. Appl. 39 (5) (2003) 1272–1279. [85] C.W. Hsu, C.J. Lin, A comparison of methods for multiclass support vector machines, IEEE Trans. Neural Netw. 13 (2) (2002) 415–425. [86] J.S. Hsu, Monitoring of defects in induction motors through air-gap torque observation, IEEE Trans. Ind. Appl. 31 (5) (1995) 1016–1021. [87] Q. Hu, Z. He, Z. Zhang, Y. Zi, Fault diagnosis of rotating machinery based on improved wavelet package transform and SVMs ensemble, Mech. Syst. Sig. Process. 21 (2) (2007) 688–705. [88] S.R. Huang, K.H. Huang, K.H. Chao, W.T. Chiang, Fault analysis and diagnosis system for induction motors, Comput. Electr. Eng. 54 (2016) 195–209. [89] X. Huang, T.G. Habetler, R.G. Harley, Detection of rotor eccentricity faults in a closed-loop drive-connected induction motor using an artificial neural network, IEEE Trans. Power Electron. 22 (4) (2007) 1552–1559. [90] X. Huang, T.G. Habetler, R.G. Harley, E.J. Wiedenbrug, Using a surge tester to detect rotor eccentricity faults in induction motors, IEEE Trans. Ind. Appl. 43 (5) (2007) 1183–1190. [92] S.O. Ibrahim, E.A. Elzahab, Implementation of fuzzy modeling system for faults detection and diagnosis in three phase induction motor drive system, J. Electr. Syst. Inf. Technol. 2 (1) (2015) 27–46. [93] J. Ilonen, J.K. Kamarainen, T. Lindh, J. Ahola, H. Kalviainen, J. Partanen, Diagnosis tool for motor condition monitoring, IEEE Trans. Ind. Appl. 41 (4) (2005) 963–971. [94] F. Immovilli, A. Bellini, R. Rubini, C. Tassoni, Diagnosis of bearing faults in induction machines by vibration or current signals: a critical comparison, IEEE Trans. Ind. Appl. 46 (4) (2010) 1350–1359. [95] F. Immovilli, C. Bianchini, M. Cocconcelli, A. Bellini, R. Rubini, Bearing fault model for induction motor with externally induced vibration, IEEE Trans. Ind. Electron. 60 (8) (2013) 3408–3418. 34 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 [96] M. Iorgulescu, R. Beloiu, Vibration and current monitoring for fault’s diagnosis of induction motors, Ann. Univ. Craiova, Electr. Eng. Ser. 32 (2008) 102– 107. [97] L.B. Jack, A.K. Nandi, Fault detection using support vector machines and artificial neural networks, augmented by genetic algorithms, Mech. Syst. Sig. Process. 16 (2–3) (2002) 373–390. [98] A.K. Jardine, D. Lin, D. Banjevic, A review on machinery diagnostics and prognostics implementing condition-based maintenance, Mech. Syst. Sig. Process. 20 (7) (2006) 1483–1510. [99] A.M. Júnior, V.V. Silva, L.M. Baccarini, L.F. Mendes, The design of multiple linear regression models using a genetic algorithm to diagnose initial shortcircuit faults in 3-phase induction motors, Appl. Soft Comput. 63 (2018) 50–58. [100] P.K. Kankar, S.C. Sharma, S.P. Harsha, Fault diagnosis of ball bearings using continuous wavelet transform, Appl. Soft Comput. 11 (2) (2011) 2300– 2312. [101] Kanovic, Z., Matic, D., Jelicic, Z., Rapaic, M., Jakovljevic, B., & Kapetina, M. (2013, August). Induction motor broken rotor bar detection using vibration analysis—A case study. In Diagnostics for Electric Machines, Power Electronics and Drives (SDEMPED), 2013 9th IEEE International Symposium on (pp. 64–68). IEEE. [102] Kaviarasan, M., TamilSelvan, A., & Venugopal, E. (2016, February). Fault diagnosis of three phase squirrel cage induction motor due to bearing by using artificial intelligence. In Emerging Trends in Engineering, Technology and Science (ICETETS), International Conference on (pp. 1-4). IEEE. [103] Keskes, H., & Braham, A. (2014, April). DAG SVM and pitch synchronous wavelet transform for induction motor diagnosis. In Power Electronics, Machines and Drives (PEMD 2014), 7th IET International Conference on (pp. 1-6). IET. [104] H. Keskes, A. Braham, Recursive undecimated wavelet packet transform and DAG SVM for induction motor diagnosis, IEEE Trans. Ind. Inf. 11 (5) (2015) 1059–1066. [105] H. Keskes, A. Braham, Z. Lachiri, Broken rotor bar diagnosis in induction machines through stationary wavelet packet transform and multiclass wavelet SVM, Electr. Power Syst. Res. 97 (2013) 151–157. [106] F. Khater, M.I.A. El-Sebah, M. Osama, Fault diagnostics in an inverter feeding an induction motor using fuzzy logic, J. Electr. Syst. Inf. Technol. 4 (1) (2017) 10–17. [107] K. Kim, A.G. Parlos, Induction motor fault diagnosis based on neuropredictors and wavelet signal processing, IEEE/ASME Trans. Mechatron. 7 (2) (2002) 201–219. [109] G.B. Kliman, R.A. Koegl, J. Stein, R.D. Endicott, M.W. Madden, Noninvasive detection of broken rotor bars in operating induction motors, IEEE Trans. Energy Convers. 3 (4) (1988) 873–879. [110] S. Kolla, L. Varatharasa, Identifying three-phase induction motor faults using artificial neural networks, ISA Trans. 39 (4) (2000) 433–439. [111] P. Konar, P. Chattopadhyay, Bearing fault detection of induction motor using wavelet and Support Vector Machines (SVMs), Appl. Soft Comput. 11 (6) (2011) 4203–4211. [112] P. Konar, P. Chattopadhyay, Multi-class fault diagnosis of induction motor using Hilbert and Wavelet Transform, Appl. Soft Comput. 30 (2015) 341– 352. [113] C.T. Kowalski, T. Orlowska-Kowalska, Neural networks application for induction motor faults diagnosis, Math. Comput. Simul 63 (3–5) (2003) 435– 448. [114] C. Kral, T.G. Habetler, R.G. Harley, Detection of mechanical imbalances of induction machines without spectral analysis of time-domain signals, IEEE Trans. Ind. Appl. 40 (4) (2004) 1101–1106. [115] Kral, C., Habetler, T. G., Harley, R. G., Pirker, F., Pascoli, G., Oberguggenberger, H., & Fenz, C. J. M. (2003, August). A comparison of rotor fault detection techniques with respect to the assessment of fault severity. In Diagnostics for Electric Machines, Power Electronics and Drives, 2003. SDEMPED 2003. 4th IEEE International Symposium on (pp. 265–270). IEEE. [116] C. Kral, F. Pirker, G. Pascoli, Model-based detection of rotor faults without rotor position sensor-the sensorless Vienna monitoring method, IEEE Trans. Ind. Appl. 41 (3) (2005) 784–789. [117] Kumar, R. S., Raj, L. G. C., & Abarna, J. (2018, February). Analysis of Fuzzy Logic Based Fault Detection for Three Phase Induction Motor Drive System. In 2018 4th International Conference on Electrical Energy Systems (ICEES) (pp. 700–705). IEEE. [118] J. Kurek, S. Osowski, Support vector machine for fault diagnosis of the broken rotor bars of squirrel-cage induction motor, Neural Comput. Appl. 19 (4) (2010) 557–564. [119] N. Lashkari, J. Poshtan, H.F. Azgomi, Simulative and experimental investigation on stator winding turn and unbalanced supply voltage fault diagnosis in induction motors using Artificial Neural Networks, ISA Trans. 59 (2015) 334–342. [120] Lasurt, I., Stronach, A. F., & Penman, J. (2000). A fuzzy logic approach to the interpretation of higher order spectra applied to fault diagnosis in electrical machines. In Fuzzy Information Processing Society, 2000. NAFIPS. 19th International Conference of the North American (pp. 158–162). IEEE. [121] S. Lee, M.D. Bryant, L. Karlapalem, Model-and information theory-based diagnostic method for induction motors, J. Dyn. Syst. Meas. Contr. 128 (3) (2006) 584–591. [122] Y. Lei, N. Li, L. Guo, N. Li, T. Yan, J. Lin, Machinery health prognostics: a systematic review from data acquisition to RUL prediction, Mech. Syst. Sig. Process. 104 (2018) 799–834. [124] B. Li, M.Y. Chow, Y. Tipsuwan, J.C. Hung, Neural-network-based motor rolling bearing fault diagnosis, IEEE Trans. Ind. Electron. 47 (5) (2000) 1060– 1069. [125] N. Li, R. Zhou, Q. Hu, X. Liu, Mechanical fault diagnosis based on redundant second generation wavelet packet transform, neighborhood rough set and support vector machine, Mech. Syst. Sig. Process. 28 (2012) 608–621. [126] W. Li, C.K. Mechefske, Detection of induction motor faults: a comparison of stator current, vibration and acoustic methods, J. Vib. Control 12 (2) (2006) 165–188. [127] G.M. Lim, D.M. Bae, J.H. Kim, Fault diagnosis of rotating machine by thermography method on support vector machine, J. Mech. Sci. Technol. 28 (8) (2014) 2947–2952. [128] T.I. Liu, J.H. Singonahalli, N.R. Iyer, Detection of roller bearing defects using expert system and fuzzy logic, Mech. Syst. Sig. Process. 10 (5) (1996) 595– 614. [129] Y. Liu, A.M. Bazzi, A review and comparison of fault detection and diagnosis methods for squirrel-cage induction motors: State of the art, ISA Trans. 70 (2017) 400–409. [130] Z. Liu, H. Cao, X. Chen, Z. He, Z. Shen, Multi-fault classification based on wavelet SVM with PSO algorithm to analyze vibration signals from rolling element bearings, Neurocomputing 99 (2013) 399–410. [131] Z. Liu, Z. Zheng, Y. Li, Enhancing fault-tolerant ability of a nine-phase induction motor drive system using fuzzy logic current controllers, IEEE Trans. Energy Convers. 32 (2) (2017) 759–769. [132] Martínez-Morales, J. D., Palacios, E., & Campos-Delgado, D. U. (2010, September). Data fusion for multiple mechanical fault diagnosis in induction motors at variable operating conditions. In Electrical Engineering Computing Science and Automatic Control (CCE), 2010 7th International Conference on (pp. 176–181). IEEE. [133] J.F. Martins, V.F. Pires, A.J. Pires, Unsupervised neural-network-based algorithm for an on-line diagnosis of three-phase induction motor stator fault, IEEE Trans. Ind. Electron. 54 (1) (2007) 259–264. [134] Maruthi, G. S. & Vittal, K. P. (2005, November). Electrical fault detection in three phase squirrel cage induction motor by vibration analysis using MEMS accelerometer. In 2005 International Conference on Power Electronics and Drives Systems (Vol. 2, pp. 838–843). IEEE. [135] C.K. Mechefske, Objective machinery fault diagnosis using fuzzy logic, Mech. Syst. Sig. Process. 12 (6) (1998) 855–862. [136] M.R. Mehrjou, N. Mariun, M.H. Marhaban, N. Misron, Rotor fault condition monitoring techniques for squirrel-cage induction machine-A review, Mech. Syst. Sig. Process. 25 (8) (2011) 2827–2848. P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 35 [137] Meng, L., Miao, W., & Chunguang, W. (2010, May). Research on SVM classification performance in rolling bearing diagnosis. In Intelligent Computation Technology and Automation (ICICTA), 2010 International Conference on (Vol. 3, pp. 132–135). IEEE. [138] Nakamura, H., Yamamoto, Y., & Mizuno, Y. (2006). Diagnosis of electrical and mechanical faults of induction motor. In Electrical Insulation and Dielectric Phenomena, 2006 IEEE Conference, (pp. 521–524). IEEE. [139] S. Nandi, H.A. Toliyat, X. Li, Condition monitoring and fault diagnosis of electrical motors-A review, IEEE Trans. Energy Convers. 20 (4) (2005) 719– 729. [140] H. Nejjari, M.E.H. Benbouzid, Monitoring and diagnosis of induction motors electrical faults using a current Park’s vector pattern learning approach, IEEE Trans. Ind. Appl. 36 (3) (2000) 730–735. [141] N.T. Nguyen, H.H. Lee, An application of support vector machines for induction motor fault diagnosis with using genetic algorithm, in: Advanced Intelligent Computing Theories and Applications. With Aspects of Artificial Intelligence, Springer, Berlin, Heidelberg, 2008, pp. 190–200. [142] N.T. Nguyen, H.H. Lee, J.M. Kwon, Optimal feature selection using genetic algorithm for mechanical fault detection of induction motor, J. Mech. Sci. Technol. 22 (3) (2008) 490–496. [143] A. Nikranjbar, M. Ebrahimi, A.S. Wood, Model-based fault diagnosis of induction motor eccentricity using particle swarm optimization, Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci. 223 (3) (2009) 607–615. [144] R.H.C. Palácios, I.N. da Silva, A. Goedtel, W.F. Godoy, A comprehensive evaluation of intelligent classifiers for fault identification in three-phase induction motors, Electr. Power Syst. Res. 127 (2015) 249–258. [145] W.J. Park, S.H. Lee, W.K. Joo, J.I. Song, A mixed algorithm of PCA and LDA for fault diagnosis of induction motor, in: Advanced Intelligent Computing Theories and Applications. With Aspects of Artificial Intelligence, Springer, Berlin, Heidelberg, 2007, pp. 934–942. [146] Parlos, A. G., Kim, K., & Bharadwaj, R. (2002, May). Detection of induction motor faults-combining signal-based and model-based techniques. In American Control Conference, 2002. Proceedings of the 2002 (Vol. 6, pp. 4531–4536). IEEE. [147] R.K. Patel, V.K. Giri, Feature selection and classification of mechanical fault of an induction motor using random forest classifier, Perspect. Sci. 8 (2016) 334–337. [148] Z.K. Peng, F.L. Chu, Application of the wavelet transform in machine condition monitoring and fault diagnostics: a review with bibliography, Mech. Syst. Sig. Process. 18 (2) (2004) 199–221. [149] J. Penman, H.G. Sedding, B.A. Lloyd, W.T. Fink, Detection and location of interturn short circuits in the stator windings of operating motors, IEEE Trans. Energy Convers. 9 (4) (1994) 652–658. [152] R. Puche-Panadero, M. Pineda-Sanchez, M. Riera-Guasp, J. Roger-Folch, E. Hurtado-Perez, J. Perez-Cruz, Improved resolution of the MCSA method via Hilbert transform, enabling the diagnosis of rotor asymmetries at very low slip, IEEE Trans. Energy Convers. 24 (1) (2009) 52–59. [154] N. Rajeswaran, M.L. Swarupa, T.S. Rao, K. Chetaswi, Hybrid artificial intelligence based fault diagnosis of SVPWM voltage source inverters for induction motor, Mater. Today. Proc. 5 (1) (2018) 565–571. [155] R.B. Randall, Vibration-based condition monitoring: industrial, aerospace and automotive applications, Willey-Blackwell, Hoboken, New Jersey, USA, 2011. [156] J. Rangel-Magdaleno, H. Peregrina-Barreto, J. Ramirez-Cortes, I. Cruz-Vega, Hilbert spectrum analysis of induction motors for the detection of incipient broken rotor bars, Measurement 109 (2017) 247–255. [158] Rapur, J. S., & Tiwari, R. (2017, December). A Compliant Algorithm to Diagnose Multiple Centrifugal Pump Faults With Corrupted Vibration and Current Signatures in Time-Domain. In ASME 2017 Gas Turbine India Conference (pp. V002T05A007-V002T05A007). American Society of Mechanical Engineers. [159] J.S. Rapur, R. Tiwari, Experimental fault diagnosis for known and unseen operating conditions of centrifugal pumps using MSVM and WPT based analyses, Measurement (2019). [160] J.S. Rapur, R. Tiwari, On-line time domain vibration and current signals based multi-fault diagnosis of centrifugal pumps using support vector machines, J. Nondestr. Eval. 38 (1) (2019) 6. [161] P.V.J. Rodríguez, A. Arkkio, Detection of stator winding fault in induction motor using fuzzy logic, Appl. Soft Comput. 8 (2) (2008) 1112–1120. [162] P.V.J. Rodríguez, A. Belahcen, A. Arkkio, A. Laiho, J.A. Antonino-Daviu, Air-gap force distribution and vibration pattern of induction motors under dynamic eccentricity, Electr. Eng. 90 (3) (2008) 209–218. [163] P.V.J. Rodríguez, M. Negrea, A. Arkkio, A simplified scheme for induction motor condition monitoring, Mech. Syst. Sig. Process. 22 (5) (2008) 1216– 1236. [164] Rojas, A., & Nandi, A. K. (2005, September). Detection and classification of rolling-element bearing faults using support vector machines. In Machine Learning for Signal Processing, 2005 IEEE Workshop on (pp. 153–158). IEEE. [165] R.J. Romero-Troncoso, R. Saucedo-Gallaga, E. Cabal-Yepez, A. Garcia-Perez, R.A. Osornio-Rios, R. Alvarez-Salas, N. Huber, FPGA-based online detection of multiple combined faults in induction motors through information entropy and fuzzy inference, IEEE Trans. Ind. Electron. 58 (11) (2011) 5263– 5270. [166] A. Sadeghian, Z. Ye, B. Wu, Online detection of broken rotor bars in induction motors by wavelet packet decomposition and artificial neural networks, IEEE Trans. Instrum. Meas. 58 (7) (2009) 2253–2263. [167] S.B. Salem, K. Bacha, A. Chaari, Support vector machine based decision for mechanical fault condition monitoring in induction motor using an advanced Hilbert-Park transform, ISA Trans. 51 (5) (2012) 566–572. [168] G. Salles, F. Filippetti, C. Tassoni, G. Crellet, G. Franceschini, Monitoring of induction motor load by neural network techniques, IEEE Trans. Power Electron. 15 (4) (2000) 762–768. [169] B. Samanta, Gear fault detection using artificial neural networks and support vector machines with genetic algorithms, Mech. Syst. Sig. Process. 18 (3) (2004) 625–644. [170] A.K. Samanta, A. Naha, A. Routray, A.K. Deb, Fast and accurate spectral estimation for online detection of partial broken bar in induction motors, Mech. Syst. Sig. Process. 98 (2018) 63–77. [171] B. Samanta, K.R. Al-Balushi, Artificial neural network based fault diagnostics of rolling element bearings using time-domain features, Mech. Syst. Sig. Process. 17 (2) (2003) 317–328. [172] A. Sapena-Bano, J. Martinez-Roman, R. Puche-Panadero, M. Pineda-Sanchez, J. Perez-Cruz, M. Riera-Guasp, Induction machine model with space harmonics for fault diagnosis based on the convolution theorem, Int. J. Electr. Power Energy Syst. 100 (2018) 463–481. [173] N. Saravanan, K.I. Ramachandran, Incipient gear box fault diagnosis using discrete wavelet transform (DWT) for feature extraction and classification using artificial neural network (ANN), Expert Syst. Appl. 37 (6) (2010) 4168–4181. [174] A.Y.B. Sasi, F. Gu, Y. Li, A.D. Ball, A validated model for the prediction of rotor bar failure in squirrel-cage motors using instantaneous angular speed, Mech. Syst. Sig. Process. 20 (7) (2006) 1572–1589. [175] R.R. Schoen, B.K. Lin, T.G. Habetler, J.H. Schlag, S. Farag, An unsupervised, on-line system for induction motor fault detection using stator current monitoring, IEEE Trans. Ind. Appl. 31 (6) (1995) 1280–1286. [176] M. Seera, C.P. Lim, Online motor fault detection and diagnosis using a hybrid FMM-CART model, IEEE Trans. Neural Networks Learn. Syst. 25 (4) (2014) 806–812. [177] M. Seera, C.P. Lim, D. Ishak, H. Singh, Fault detection and diagnosis of induction motors using motor current signature analysis and a hybrid FMM– CART model, IEEE Trans. Neural Netw. Learn. Syst. 23 (1) (2012) 97–108. [178] J. Seshadrinath, B. Singh, B.K. Panigrahi, Incipient turn fault detection and condition monitoring of induction machine using analytical wavelet transform, IEEE Trans. Ind. Appl. 50 (3) (2014) 2235–2242. [179] J.M.B. Siau, A.L. Graff, W.L. Soong, N. Ertugrul, Broken bar detection in induction motors using current and flux spectral analysis, Aust. J. Electr. Electron. Eng. 1 (3) (2004) 171–178. 36 P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 [180] Siddique, A., Yadava, G. S., & Singh, B. (2003, August). Applications of artificial intelligence techniques for induction machine stator fault diagnostics: review. In Diagnostics for Electric Machines, Power Electronics and Drives, 2003. SDEMPED 2003. 4th IEEE International Symposium on (pp. 29–34). IEEE. [181] A. Siddique, G.S. Yadava, B. Singh, A review of stator fault monitoring techniques of induction motors, IEEE Trans. Energy Convers. 20 (1) (2005) 106– 114. [182] K.M. Siddiqui, K. Sahay, V.K. Giri, Health monitoring and fault diagnosis in induction motor-a review, Int. J. Adv. Res. Electr. Electron. Instrument. Eng. 3 (1) (2014) 6549–6565. [183] Silva, V. A. D. & Pederiva, R. (2013). Fault detection in induction motors based on artificial intelligence. Surveillance 7, International Conference October 29-30, 2013, Institute of Technology of Chartres, France. [184] G.K. Singh, Induction machine drive condition monitoring and diagnostic research—a survey, Electr. Power Syst. Res. 64 (2) (2003) 145–158. [185] G.K. Singh, S.A.K.S.A. Ahmed, Vibration signal analysis using wavelet transform for isolation and identification of electrical faults in induction machine, Electr. Power Syst. Res. 68 (2) (2004) 119–136. [186] G. Singh, V.N.A. Naikan, Infrared thermography based diagnosis of inter-turn fault and cooling system failure in three phase induction motor, Infrared Phys. Technol. 87 (2017) 134–138. [187] G. Singh, V.N.A. Naikan, Detection of half broken rotor bar fault in VFD driven induction motor drive using motor square current MUSIC analysis, Mech. Syst. Sig. Process. 110 (2018) 333–348. [188] A. Soualhi, G. Clerc, H. Razik, Detection and diagnosis of faults in induction motor using an improved artificial ant clustering technique, IEEE Trans. Ind. Electron. 60 (9) (2013) 4053–4062. [189] G.C. Stone, H.G. Sedding, M.J. Costello, Application of partial discharge testing to motor and generator stator winding maintenance, IEEE Trans. Ind. Appl. 32 (2) (1996) 459–464. [190] H. Su, K.T. Chong, Induction machine condition monitoring using neural network modeling, IEEE Trans. Ind. Electron. 54 (1) (2007) 241–249. [191] V. Sugumaran, K. Ramachandran, Effect of number of features on classification of roller bearing faults using svm and psvm, Expert Syst. Appl. 38 (4) (2011) 4088–4096. [192] V. Sugumaran, G.R. Sabareesh, K.I. Ramachandran, Fault diagnostics of roller bearing using kernel based neighborhood score multi-class support vector machine, Expert Syst. Appl. Int. J. 34 (4) (2008) 3090–3098. [193] Sui, W.T., Zhang, D., (2009). Rolling element bearings fault classification based on svm and feature evaluation. Machine Learning and Cybernetics, International Conference on, IEEE, 1, 450–453. [194] W.W. Tan, H. Huo, A generic neurofuzzy model-based approach for detecting faults in induction motors, IEEE Trans. Ind. Electron. 52 (5) (2005) 1420– 1427. [195] Tavner, P. J., Gaydon, B. G., & Ward, D. M. (1986, May). Monitoring generators and large motors. In IEE Proceedings B (Electric Power Applications) (Vol. 133, No. 3, pp. 169-180). IET Digital Library. [196] S.M. Tetrault, G.C. Stone, H.G. Sedding, Monitoring partial discharges on 4-kV motor windings, IEEE Trans. Ind. Appl. 35 (3) (1999) 682–688. [197] Thomson, W. T. & Orpin, P. (2002, September). Current and vibration monitoring for fault diagnosis and root cause analysis of induction motor drives. In Proceedings of the thirty-first turbomachinery symposium (pp. 61–67). [198] Thomson, W. T. (2001). On-line MCSA to diagnose shorted turns in low voltage stator windings of 3-phase induction motors prior to failure. In Electric Machines and Drives Conference, 2001. IEMDC 2001. IEEE International (pp. 891–898). IEEE. [199] W.T. Thomson, M. Fenger, Current signature analysis to detect induction motor faults, IEEE Ind. Appl. Mag. 7 (4) (2001) 26–34. [200] Thomson, W. T., & Gilmore, R. J. (2003). Motor Current Signature Analysis To Detect Faults In Induction Motor Drives-Fundamentals, Data Interpretation, And Industrial Case Histories. In Proceedings of the 32nd Turbomachinery Symposium. Texas A&M University. Turbomachinery Laboratories. [201] M. Timusk, M. Lipsett, C.K. Mechefske, Fault detection using transient machine signals, Mech. Syst. Sig. Process. 22 (7) (2008) 1724–1749. [202] R. Tiwari, Rotor Systems: Analysis and Identification, CRC Press, Boca Raton, FL, 2017. [203] V.T. Tran, F. Althobiani, A. Ball, B. Choi, Expert Systems with Applications An application to transient current signal based induction motor fault diagnosis of Fourier – Bessel expansion and simplified fuzzy ARTMAP, Expert Syst. Appl. 40 (13) (2013) 5372–5384. [204] V.T. Tran, B. Yang, M. Oh, A. Chit, C. Tan, Fault diagnosis of induction motor based on decision trees and adaptive neuro-fuzzy inference, Expert Syst. Appl. 36 (2) (2009) 1840–1849. [205] F.C. Trutt, J. Sottile, J.L. Kohler, Online condition monitoring of induction motors, IEEE Trans. Ind. Appl. 38 (6) (2002) 1627–1632. [206] V.N. Vapnik, An overview of statistical learning theory, IEEE Trans. Neural Netw. 10 (5) (1999) 988–999. [207] Vishwakarma, H. O., Sajan, K. S., Maheshwari, B., & Dhiman, Y. D. (2015, August). Intelligent bearing fault monitoring system using support vector machine and wavelet packet decomposition for induction motors. In Power and Advanced Control Engineering (ICPACE), 2015 International Conference on (pp. 339–343). IEEE. [208] C. Wang, R.X. Gao, Sensor Placement Strategy for In-Situ Bearing Defect Detection Vol. 3 (2000) 1463–1467. [209] D. Wang, K.L. Tsui, Q. Miao, Prognostics and health management: a review of vibration based bearing and gear health indicators, IEEE Access 6 (2018) 665–676. [210] Wang, Z., & Chang, C. S. (2011, June). Online fault detection of induction motors using frequency domain independent components analysis. In Industrial Electronics (ISIE), 2011 IEEE International Symposium on (pp. 2132-2137). IEEE. [211] L. Wen, X. Li, L. Gao, Y. Zhang, A new convolutional neural network-based data-driven fault diagnosis method, IEEE Trans. Ind. Electron. 65 (7) (2017) 5990–5998. [212] A. Widodo, B.S. Yang, Support vector machine in machine condition monitoring and fault diagnosis, Mech. Syst. Sig. Process. 21 (6) (2007) 2560–2574. [213] A. Widodo, B.S. Yang, Wavelet support vector machine for induction machine fault diagnosis based on transient current signal, Expert Syst. Appl. 35 (1) (2008) 307–316. [214] A. Widodo, B.S. Yang, T. Han, Combination of independent component analysis and support vector machines for intelligent faults diagnosis of induction motors, Expert Syst. Appl. 32 (2) (2007) 299–312. [215] A. Widodo, B.S. Yang, Application of nonlinear feature extraction and support vector machines for fault diagnosis of induction motors, Expert Syst. Appl. 33 (1) (2007) 241–250. [216] C. Wu, T. Chen, R. Jiang, L. Ning, Z. Jiang, ANN Based multi-classification using various signal processing techniques for bearing fault diagnosis, Int. J. Control Automat. 8 (7) (2015) 113–124. [217] F. Wu, Y. Hao, J. Zhao, Y. Liu, Current similarity based open-circuit fault diagnosis for induction motor drives with discrete wavelet transform, Microelectron. Reliab. 75 (2017) 309–316. [218] G.M. Xian, B.Q. Zeng, An intelligent fault diagnosis method based on wavelet packer analysis and hybrid support vector machines, Expert Syst. Appl. 36 (10) (2009) 12131–12136. [219] B.S. Yang, T. Han, Y.S. Kim, Integration of ART-Kohonen neural network and case-based reasoning for intelligent fault diagnosis, Expert Syst. Appl. 26 (3) (2004) 387–395. [220] B.S. Yang, K.J. Kim, Application of Dempster-Shafer theory in fault diagnosis of induction motors using vibration and current signals, Mech. Syst. Sig. Process. 20 (2) (2006) 403–420. [221] Y. Yang, D. Yu, J. Cheng, A fault diagnosis approach for roller bearing based on imf envelope spectrum and svm, Measurement 40 (9) (2007) 943–950. [222] Z. Ye, A. Sadeghian, B. Wu, Mechanical fault diagnostics for induction motor with variable speed drives using Adaptive Neuro-fuzzy Inference System, Electr. Power Syst. Res. 76 (9–10) (2006) 742–752. P. Gangsar, R. Tiwari / Mechanical Systems and Signal Processing 144 (2020) 106908 37 [223] Ye, Z., Wu, B., & Zargari, N. (2000). Online mechanical fault diagnosis of induction motor by wavelet artificial neural network using stator current. In Industrial Electronics Society, 2000. IECON 2000. 26th Annual Conference of the IEEE (Vol. 2, pp. 1183-1188). IEEE. [224] Z. Ye, B. Wu, A. Sadeghian, Current signature analysis of induction motor mechanical faults by wavelet packet decomposition, IEEE Transactions on Industry Electronics 50 (6) (2003) 1217–1228. [225] G.G. Yen, K.C. Lin, Wavelet packet feature extraction for vibration monitoring, IEEE Trans. Ind. Electron. 47 (3) (2000) 650–667. [226] Younus, A. M., & Yang, B. S. (2010, January). Wavelet co-efficient of thermal image analysis for machine fault diagnosis. In Prognostics and Health Management Conference, 2010. PHM’10. (pp. 1-6). IEEE. [227] J. Zarei, J. Poshtan, Bearing fault detection using wavelet packet transform of induction motor stator current, Tribol. Int. 40 (5) (2007) 763–769. [228] J. Zarei, M.A. Tajeddini, H.R. Karimi, Vibration analysis for bearing fault detection and classification using an intelligent filter, Mechatronics 24 (2) (2014) 151–157. [229] S. Zgarni, H. Keskes, A. Braham, Nested SVDD in DAG SVM for induction motor condition monitoring, Eng. Appl. Artif. Intell. 71 (2018) 210–215. [230] L. Zhang, G. Xiong, H. Liu, H. Zou, W. Guo, Fault diagnosis based on optimized node entropy using lifting wavelet packet transform and genetic algorithms, Proc. Inst. Mech. Eng. Part I: J. Syst. Control Eng. 224 (5) (2010) 557–573. [231] P. Zhang, Y. Du, T.G. Habetler, B. Lu, A survey of condition monitoring and protection methods for medium-voltage induction motors, IEEE Trans. Ind. Appl. 47 (1) (2011) 34–46. [232] R. Zhang, X. Wang, On-line broken-bar fault diagnosis system of induction motor, Trans. Tianjin Univ. 14 (2) (2008) 144–147. [233] S. Zhang, T. Asakura, X. Xu, B. Xu, Fault diagnosis system for rotary machine based on fuzzy neural networks, JSME Int J. Ser. C 46 (3) (2003) 1035– 1041. [234] Zheng, H., Zhou, L., (2012). Rolling element bearing fault diagnosis based on support vector machine. Consumer Electronics, Communications and Networks (CECNet), 2nd International Conference on. IEEE, 544–547. [235] Zhitong, C., Hongping, C., Guoguang, H., & Ritchie, E. (2001). Rotor fault diagnosis of induction motor based on wavelet reconstruction. In Electrical Machines and Systems, 2001. ICEMS 2001. Proceedings of the Fifth International Conference on (Vol. 1, pp. 374-377). IEEE. [236] Zhongming, Y., & Bin, W. (2000). A review on induction motor online fault diagnosis. In Power Electronics and Motion Control Conference, 2000. Proceedings. IPEMC 2000. The Third International (Vol. 3, pp. 1353-1358). IEEE. [237] H. Zhou, J. Chen, G. Dong, R. Wang, Detection and diagnosis of bearing faults using shift-invariant dictionary learning and hidden Markov model, Mech. Syst. Sig. Process. 72 (2016) 65–79. [238] Z. Zhou, J. Zhao, F. Cao, A novel approach for fault diagnosis of induction motor with invariant character vectors, Inf. Sci. 281 (2014) 496–506. [239] K. Zhu, X. Song, A roller bearing fault diagnosis method based on hierarchical entropy and support vector machine with particle swarm optimization algorithm, Measurement 47 (2014) 669–767. [240] F. Zidani, M.E.H. Benbouzid, D. Diallo, M.S. Naït-Saïd, Induction motor stator faults diagnosis by a current Concordia pattern-based fuzzy decision system, IEEE Trans. Energy Convers. 18 (4) (2003) 469–475. [241] Zolfaghari, S., Noor, S. B. M., Mariun, N., Marhaban, M. H., Mehrjou, M. R., & Karami, M. (2014, December). Broken rotor bar detection of induction machine using wavelet packet coefficient-related features. In Research and Development (SCOReD), 2014 IEEE Student Conference on (pp. 1-5). IEEE.