Machine Learning Framework for Multi-Fault Diagnosis within Three-Phase Induction Motor in Electric Vehicle Applications

• Comprehensive Comparative Framework: Developing a rigorous benchmarking framework that evaluates and compares five distinct machine learning architectures (ANN, KNN, SVM, DT, and Ensemble) for the simultaneous diagnosis of multiple electrical and mechanical faults in induction motors. • Comprehensive Multi-Fault Diagnostic Scope: Developing a robust framework capable of identifying and classifying a wide range of both electrical and mechanical faults (including BRB, ITSC, OV, UV, and SPH), providing a more holistic diagnostic solution for EV induction motors compared to single-fault studies. • Large-Scale High-Resolution Dataset: Generating an exceptionally large dataset consisting of over 2.25 million samples recorded at a high sampling frequency of 150 kHz. This massive data scale ensures that the ML models capture subtle transient behaviors and high-frequency fault signatures that are often missed in smaller datasets. • Diverse Operational Loading Scenarios: Evaluating the diagnostic system under various load conditions specifically Full-Load (FL), Half-Load (HL), and No-Load (NL). Unlike studies restricted to constant loads, this approach ensures the model’s reliability across the dynamic power demands typical of real-world electric vehicle operation. • Extensive Comparative Benchmarking: Conducting a rigorous performance evaluation of five distinct machine learning architectures (ANN, KNN, SVM, DT, and Ensemble). This large-scale comparison identifies the most effective algorithms for high-dimensional fault data and establishes a performance baseline for future research. • Methodological Integrity and Transparency: Implementing a strict hierarchical data isolation protocol and advanced multi-domain feature engineering (Time-Frequency). By utilizing Chi-Square analysis for feature ranking, the study provides a transparent justification for the models' success and establishes a "theoretical upper-bound benchmark" for induction motor diagnostics. Induction Motors (IMs), which are known for their reliable performance, Low capital cost, and minimal operational expenditures, are a key component of Electric Vehicle (EV) powertrains. Nevertheless, they are susceptible to different electrical and mechanical faults that severely impact vehicle safety and performance. This necessitates the need for robust early fault detection systems. Although Machine Learning (ML) and Artificial Intelligence (AI) outperform traditional methods in terms of fault diagnostic features and accuracy, many existing studies remain limited by using simplistic models, focusing on single-fault investigations, a small number of input features, and a lack of comprehensive validations. To address these limitations, this paper presented a robust and extensively validated ML framework for IM fault diagnosis in EV settings. In this work, a high-resolution dataset was generated using MATLAB/Simulink at sampling rate of 150 kHz. The date set comprises many fault types, including Broken Rotor Bar (BRB), Inter-Turn Short Circuit (ITSC), Over/Under Voltage (OV/ UV), and Single Phasing (SPH) faults, under different load conditions. Five ML models, comprising Artificial Neural Network (ANN), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Ensemble Model, and Decision Tree (DT), were optimized via Grid Search. These models were evaluated with strict sequential data partitioning, which in turn reduced data leakage. The five ML techniques achieved near-optimal accuracy (≈100%) in classifying IM faults. As explicitly clarified in this study, this exceptional performance establishes a robust theoretical baseline; it is primarily attributed to the high separability of features extracted from the clean, noise-free, and high-fidelity Simulink environment. Consequently, these results provide a strong proof-of-concept for the proposed diagnostic framework under idealized conditions.