Design and Fault Diagnosis of Induction Motor Using ML-Based Algorithms for EV Application

The need for alternate transportation is driven by the increased fossil fuel cost and the adverse effects of climatic change. Electric vehicles (EVs) are the best option as they have less carbon footprint and reduced dependency on fossil fuels. Prodigious efforts to enhance the efficiency of EVs resulted in the development of highly efficient three-phase induction motors. Difficulties in designing highly efficient induction motors (IM) with high torque and power factors hindered the success of EV applications. Hence, our aim is to diagnosis fault in the designed IM under variable load conditions. The proposed EV motor is designed for 415V, 50Hz, and 5HP output power rating using ANSYS RMxprt simulation software. A fault detection strategy is also implemented with various machine learning (ML) techniques like Support Vector Machine (SVM), K-nearest neighbors (k-NN), ML perceptron (MLP), Random Forest (RF), Decision Tree (DT), Gradient boosting (GB), Extreme Gradient Boosting (XGBoost), and Deep Learning (DL) for both healthy and faulty conditions. Short Circuit (SC), High Resistance connection (HRC), and Open-Phase circuit (OPC) are considered as faulty states for this study. Motor performance with variable load for all the states healthy and faulty are evaluated through machine learning.


I. INTRODUCTION
Automobile sector is playing a vital part in the world's economic growth. Internal combustion vehicles used in most vehicles consume directly the fossil fuel and create a large amount of greenhouse gases affecting the entire world. This paved the way to discover new energy vehicles as an alternative to conventional vehicles, which are Electric Vehicles (EVs) [1], [2], [3]. The core part of an EV is the electric motor which converts electric energy to mechanical energy. Hence it is necessary to build an electric motor that enhances the efficiency of EV and its performance [4]. Induction Motor (IM), Brushless DC motor (BLDC), and Permanent Magnet Synchronous motor (PMSM) are the most commonly used motors for commercial purposes [5]. IM is more effective and economical than other motors due to its reliability, simple mechanical design, and effective field-weakening characteristics [6], [7]. However, the limitations such as core The associate editor coordinating the review of this manuscript and approving it for publication was R. K. Saket . loss, friction loss, and copper loss, reduces the efficiency of IM [8], [9]. To overcome these limitations and to enhance the efficiency, researchers are focusing to optimize the length of the stator as it can also reduce harmonic losses. Also, IM characteristics are influenced by geometric dimensions like core length and size of the stator and rotor slots [10]. The motor's efficiency is also determined by the materials used for manufacturing [11], [12]. Several materials have been used for the motor design for many years, and the most often used materials for core & winding are Iron (Fe), Carbon steel 1008, and steel 1010 laminated cores, Aluminium (Al), Copper (Cu), and Silver (Ag). Traditionally, Al is used as winding material for IM but the Conductivity is lower than Cu [13].
Though Cu has high conductivity and increased mechanical efficiency than Al, it is costly. Ag, which has higher conductivity than Cu is expensive with a low melting point [14]. Hence, selecting winding materials and core lamination is vital for achieving higher efficiency and effective motor operation. In addition to material selection, the fault detection (FD) system has also been used an effective approach to increase the performance of motor operation in EVs.
Researchers are developing FD strategies for IM based on two approaches. They are Model-based approaches and Data-driven approaches. Model-based approaches attempt to predict the faulty behavior by mathematically modelling the motor. The main disadvantage of this methodology is due to the machine's natural wear, because the degradation of machine components causes a difference between the actual machine and its mathematical model, when the fault magnifies. Furthermore, it is critical that the model assumes that the machine parameters are available, which does not always happen. This makes the diagnosis more challenging because it is necessary to estimate the machine parameters for the appropriate modelling of the machine [15], [16]. Data-driven approaches do not require IM model as well as the characteristics of the motor and load coupled to the machine. Furthermore, these methods have been widely used in fault diagnosis of nonlinear complex and time-varying systems and demonstrated promising outcomes in identifying faults. Machine learning (ML) is a most popular data driven approach which includes the k-Nearest Neighbor (k-NN) method, Naive Bayes (NB), Support Vector Machines (SVM), and Artificial Neural Networks (ANN) [15], [17], [18], [19], [20]. Deep Learning (DL) architectures have attracted the attention of several researchers, mainly those based on Convolutional Neural Networks (CNN). These algorithms are able to extract the failure characteristics of signals, as well as eliminate the need for a priority knowledge, allowing efficient independent fault diagnosis in electrical machines [21], [22], [23], [24], [25].
Data-driven FD techniques using ML and DL algorithms are being the recent focus by researches in diagnosing the motor faults. Many research findings are reported to detect the stator faults through monitoring the vibration and current signals extracted from the stator [26], [27], [28]. However, the detection of motor wiring or connection failures of IM has received the least attention. It is important to note that catastrophic failure modes, such as open-phase circuit (OPC), High Resistance Connections (HRC) in IM connections, are common. These failure modes are rare and their maintenance cost is also high. Furthermore, in most cases, these faults are caused due to human error during manufacturing. Various methods have been used to identify HRC. Thermal imaging is a valuable and effective approach for manual inspections. However, the process is expensive and complex [19]. Online detection techniques have been developed since past few decades, primarily based on resistance estimation or current sequence analysis. In a study [29], resistance is calculated by injecting voltage pulses with the inverter fixed in the drive. But, it requires the measurement of voltage between the motor's neutral and the DC-negative link terminal. The main disadvantage of this method is the requirement for an additional sensor and also the neutral point is accessed frequently.
In [30] and [31], induction motor model is designed with the consideration of HRC and short circuits (SC). These models are used to predict the negative sequence current and voltage resulting from unbalance to determine the type of failure (HRC or SC). In a similar technique is presented in [32], but the drive control strategy is employed to calculate the negative sequence current and voltage to design faulttolerant control. Few studies discuss the detection of OPC fault, when it occurs at the motor connection. Park's Vector Approach is one of the most popular approach for this OPC fault. There are also model-based methods, such as the one described in [19], where a model is suggested and validated for open-circuit defects in the phases and wiring. As stated in [33], data driven fault detection and diagnosis (FDD) strategies based on ML or DL have become a feasible alternative fault detection technique for electric motors. Several reports has been published demonstrating the use of ML or DL to identify stator defects [34], [35], [36].
Hence, this paper presents data driven ML for FDD of IM and its design using Finite Element Analysis (FEA) techniques using ANSYS software. A data-driven approach is employed for SC fault, HRC fault, and OPC fault diagnosis in IM. A 4Pole, 415V, 5HP, 50Hz IM is designed using ANSYS. The data required for FDD is extracted from the motor designed, for the healthy and faulty conditions. The data extracted from various conditions (SC, HRC, and OPC) are given to the ML algorithms to identify the performance of IM. The different algorithms used to diagnosis the healthy and faulty status of the IM are Support Vector Machine (SVM), K-nearest neighbors (k-NN), ML perceptron (MLP), Random Forest (RF), Decision Tree (DT), Gradient boosting (GB), Extreme Gradient Boosting (XGBoost), and Deep Learning (DL). In [17] and [18], the authors used instantaneous output signals such as vibration and current signals of IM for FD, whereas this article considers RMS values of current, torque, slip, and efficiency. In addition, this manuscript considers multiple faults namely SC, HRC, and OPC which are not addressed in many of the articles. Also, FD of multiple faults requires multiclass labeling and hence RMS data is considered. Moreover, RMS data can reduce the computational complexity when compared to instantaneous data. This paper is divided into five sections: Section II explain the design of Induction motor. Section III gives Finite Element analysis of Motor. Section IV examines the different ML techniques using the data extracted from the motor. Section V presents the conclusion of the article and the potential future work.

II. DESIGN OF INDUCTION MOTOR
Designing the motor based on the application can provide efficient operation. Also, selection of motor size is one of the most challenging step in the design process. In this paper, IM is designed for the specifications given in Table. (1). RMxprt tool is used to design the rotor and stator geometry. The rotor and stator diameters, number of slots, iron core length, insulation material, and winding material are considered as the parameters to design the IM.   In general, the motor with a larger diameter produces more torque with less speed, whereas the motor with a small diameter produces less torque with high speed. This is due to the increase in flux density resulting in the decreased motor size. Increased flux density above the rated specification reduces the motor diameter, magnetic core becomes saturated and this saturation will cause overheating and may result in the failure of the motor operation. To overcome this limitation, the length of the motor, operating temperature, size of the wires, torque, and speed must be adequately balanced. Importantly, the design of the motor relies on its length and diameter. Moreover, the length and diameter are determined by the application [37].
The materials used for manufacturing motor and the geometrical factors like core and winding dimensions are the most important factors in designing the motor. Moreover, proper selection of material can improve motor efficiency. The proposed IM design [38], is chosen for diagnosis the fault condition in efficient IM [39], [40]. Various motor parameters considered for the motor design is depicted in Fig. (1).
The geometrical specification and design of IM are explained in the following subsections.

A. STATOR GEOMETRY
The stator geometry is designed based on the core depth, slot depth, teeth width, slot width, and slot height. These parameter equations are given as follows,

1) STATOR CORE DEPTH (d cs )
The maximum flux density (B cs max ), occurs at the stator core depth(d cs min ) and it is given by, where N s is the number of conductors per phase and (φ) is the per phase magnetic flux given by, where f , K d , E sph are frequency, distribution factor, and induced EMF of stator phase respectively. where, where, p, S s , m are the poles, Stator slots, and number of phases. Then, the motor's core length L i is obtained as, where, L and S f are active length and stacking factor.

3) STATOR TEETH WIDTH (W ts )
The maximum flux density (B ts max ), that occurs at the stator teeth that influence both the Width of the stator teeth (W ts ) and width of the stator slot (B ss ), is given as, where S s and S r are stator slots and rotor slots. Here, substituting (2) and (5) in (8).
where, VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.

4) STATOR SLOT WIDTH (B ss )
Stator slot width at teeth, at opening and at the end is denoted as B ss1 , B ss2 , B ss3 . They are expressed as,

5) STATOR SLOT HEIGHT (h s )
Stator slot height at teeth, at opening and at the end is denoted as h s0 , h s1 , h s2 and they are expressed as,

B. ROTOR GEOMETRY
Rotor geometry parameters like rotor diameter, core depth, teeth width and bar cross-sectional area are used for designing the motor and the following equation is utilized.

1) ROTOR DIAMETER (D r )
The rotor diameter can be obtained as follows, 2) ROTOR CORE DEPTH (d cr ) The maximum flux density (B max cr ), that occurs at the rotor core depth is given as, Substituting (13) and (14) in (15) rotor slot depth is given by,

3) ROTOR COPPER LOSS
The maximum shaft diameter can be specified in order to limit the saturation level of the rotor core. Rotor resistance (R r ) is designed to reduce rotor copper losses (P r cu ) as follows, where, In (18), (N s eff /N r eff ) is the ratio of the effective stator and rotor turns. Here, the rotor bar and rotor ring resistances are expressed as r br , r rr , then rotor bar and rotor ring currents are denoted as I br and I rr , respectively given by the following equations, Here, Rotor bar resistance (r br ), Rotor ring resistance (r rr ), Rotor bar current (I br ), where the rotor slot width is,  Fig. (2) represents IM stator and rotor Slot and the Geometry specifications designed using the above equations is given in Table. (2). VOLUME 11, 2023

III. DESIGN AND ANALYSIS OF IM USING FEA
Finite Element Analysis (FEA) is used to analyze the material for designing the 5HP motor with better efficiency for EV application. The selection of material relies on its mass and cost of it. The structure of IM is designed using ANSYS with optimized specifications like dimensions, core, and winding material [41]. Carbon steel 1008 as core and Cu as winding material provided better efficiency, torque, power factor, and slip performance in our previous study [38]. Hence, these materials are used for the design of IM in RMxprt and 2D Maxwell in ANSYS platform. The cross-sectional view of the motor designed is shown in Fig. (3), and the motor specifications are given in Table. (1).
The per phase equivalent circuit of IM is given in Fig. (4). Here, R 1 , R 2 are stator and rotor resistances respectively, X 1 , X 2 are stator and rotor leakage reactance respectively, X m is magnetizing reactance, and s is slip [42]. The electrical parameters of the proposed IM obtained from ANSYS design is R 1 = 1.64 , R 2 = 1.97 , X 1 = 8.41 , X 2 = 7.17 , X m = 114.19 , and s = 0.05.
The per phase impedance from Fig. (4) is where Z 2 , Z m , and R ′ 2 are given by, (21) can be expressed as, Z ph = R ph + jX ph (22) where, These per phase values are calculated from the designed IM's stator and rotor parameters as R ph = 33. 19 and X ph = 25.40 . For the proposed design the core loss is 0.002 W, and the stator and rotor copper losses are 238.7 W and 236.3 W, respectively.
The pitch factor is one of the important design parameter that enhances the performance of IM [43].
Pitch factor (K p ) = cos α 2 (23) where, α is the short pitch angle given by, Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. For a pitch short angle α of 12 • , the pitch factor (K p ) = 0.99, distribution factor (K d ) = 0.95 and hence the winding factor (K W ) is 0.94. Double layer stator winding pattern is used in the proposed IM model and the winding pattern designed using ANSYS is shown in Fig. (5).
Based on the design consideration, Healthy and faulty conditions are created by integrating RMxprt with the Simplorer model in ANSYS, and the developed model is used for FDD. Load variation is given to the Healthy and Faulty motor model. The motor faults are created in the stator winding as a SC fault [44], HRC, and OPC fault as depicted in Fig. (6). The performance of the designed IM is analyzed with various loads. The RMS data of motor parameter such as stator currents (I1, I2, I3), and torque is extracted from the simulated motor designed, for healthy and faulty conditions at different load settings from Simplorer. A sample of data obtained for 80kg (0.0245 kgm 2 ) load is shown in Fig. (8). The healthy motor produced an efficiency of about 85% at a speed of 1437.122 rpm with torque 9.8 Nm. The faulty motor data is obtained for two different cases. In case I, fault is created at t = 0 and the fault is applied throughout the simulation. In case II, motor is run as a healthy motor from t = 0s to 0.5s and further the fault is created at t = 0.5s. The Short circuit fault is created between the phase A & B, B & C, and C & A. Similarly, HRC and Open phase circuit are created at each phases A, B, and C of the motor supply terminal. Machine learning is used to predict the condition of the motor (healthy or faulty) from the collected data.

IV. FAULT DIAGNOSIS OF IM USING ML
FD strategy is implemented through data driven ML approach to predict the three phase supply connection failures in IM. The Data is obtained from the Simplorer platform. The healthy and various faulty conditions of IM are analysed in this study. Stator winding failure is one among the most common failure modes in IM. HRC faults can be caused due to human error during the motor assembly. Additionally, SC connection and OPC fault are also found as the reason for motor failure. Hence, industries are more interested in diagnosing these failures to overcome human errors [19], [45].

A. CLASSIFICATION OF DATA FOR THE HEALTHY AND FAULTY CONDITIONS OF IM AND LABEL GENERATION
Data sets obtained from the simulation platform are classified as healthy and faulty using ML classification algorithms and the list of labels used in this study is shown in Table. (3). From the total 4000 data, 250 data are classified as healthy and the remaining data are classified as faulty. Under faulty condition, data is further classified into 19 different classes and each class contains 250 sample data. As, HRC and OPC showed similar wave pattern for the motor parameters current, torque, slip, and efficiency in Case II (Healthy & Faulty) condition in simulation and they are grouped as same cluster, named as imbalanced fault. The total number of classes are 16 and their details are given in the Table. (3). The classified data is fed to ML to differentiate healthy and faulty condition of the motor.

B. DATA PROCESSING AND ML ALGORITHM
From the simulated model, 16 different healthy and faulty conditions are considered for data collection as given in Table. (3). The input features considered for data collection are, the phase currents, torque, slip, and efficiency. Label encoder is used to normalize the labeled data without any reduction in dimension, since there are minimal feature set. Finally, the normalized feature vectors are given as input to various ML algorithms.
The objective of all the ML algorithms is to distinguish between healthy and faulty classes, and the model is trained using a set of data with healthy and faulty classes that have been labeled already. This is done for all eight ML algorithms namely Support vector Machine (SVM), K-nearest neighbors (k-NN), Multi-layer perceptron (MLP), Random Forest (RF), Decision Tree (DT), Gradient boosting (GB), Extreme Gradient Boosting (XGBoost), and Deep Learning (DL) are used for classification of IM conditions. The data is split into 80:20 as training and test set. Grid search approach was used to select the ideal hyper parameters. Table. (4) provides the model settings and hyper parameters used for machine learning methods that can provide good precision    VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.   Fig. (10). These evaluation metrics provide the performance VOLUME 11, 2023 34193 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. of the model and it is used to compare with different ML algorithm.
The accuracy of different algorithms has been descried in the Table. (5). Due to the hyperparameter tuning, the SVM approach gives 65% accuracy. Fig. (9a), shows that, some of the test data/labels are misclassified to class 3 of predicted labels, and the true class labels of 5 and 10 are inappropriately assigned to the predicted class label of 8 and 7, respectively. The MLP algorithm provides 64% accuracy due to the misclassification of six classes of test data as dipicted in Fig. (9c). VOLUME 11, 2023 34195 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.  In this, true labels of 6, 7, 8, 11, 12, and 13 are inappropriately assigned to the predicted labels of 9, 10, 5, 2, 3 and 4 respectively. The probability of misclassification is depicted in Fig. (11a, 11b), which shows that the misclassification varies from '0 to 1'; where '1' denotes 100% misclassification. Due to the minimal misclassification of test data, the k-NN is provide an accuracy of 98% as given in Fig. (9b). Apart from this, other algorithms such as RF, DT, GB, XGBoost, and DL produced the best value of accuracy (100%) in healthy and faulty conditions as given in Fig. (9d, 9e, 9f, 9g,9h). Therefore, these ML models are selected at the end of the training/testing process for the FDD of IM EV.

V. CONCLUSION
This article describes the ML-based FDD strategy for IMs under Healthy and Faulty condition. Data is generated using simulation-based models in ANSYS Simplorer for the proposed strategy. The data generated is used for training, validation, and testing various algorithms such as Support vector Machine (SVM), K-nearest neighbors (k-NN), Multilayer perceptron (MLP), Random Forest (RF), Decision Tree (DT), Gradient boosting (GB), Extreme Gradient Boosting (XGBoost), and Deep Learning (DL). To enhance the efficiency of the ML based FDD, feature extraction and selection methods are utilized for this model. By optimizing the K-nearest neighbors (k-NN), Random Forest (RF), Decision Tree (DT), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), and Deep Learning (DL) algorithms provides superior results with an accuracy of 98% to 100%. Hence, these technique can diagnose the healthy and faulty conditions accurately. Also, the deployment of ML algorithms for FDD in EV application has the ability to extract important data features automatically which leads to flexibility and versatility. Moreover, the real time data analysis enables early fault identification, reduces downtime, lowers the maintenance cost, and provides better motor performance. In addition, the FD information can be used to design Fault Tolerant Control (FTC), that can provide better reliability and safety for the EV application.
M. AISHWARYA received the bachelor's degree in electrical engineering from Anna University, in 2017, and the master's degree in power electronics and drives from the Vellore Institute of Technology (VIT), Vellore, India, in 2019, where she is currently pursuing the Ph.D. degree with the School of Electrical Engineering.
Her research interests include the design of electrical machines (ANSYS), transformer, fault diagnosis, and machine learning. She is currently an Associate Professor with the School of Electrical Engineering, Vellore Institute of Technology, Vellore, India. Her research interests include nonlinear systems and control, the control of drives and robotic systems, and fault diagnosis and control.