•  
  •  
 

Terra Joule Journal

Abstract

Maintaining the dependability and operational continuity of electric vehicle (EV) drivetrains has become crucial as the world's shift to electrified transportation picks up speed. This is a major obstacle for sustainable mobility. Using a high-capacity Artificial Neural Network (ANN), this study presents a data-driven multi-sensor framework for advanced fault diagnosis of induction motors in EV applications. The publicly available University of Ottawa dataset that encompasses nine distinct fault categories and 180 test samples was utilized to replicate both mechanical failure modes (e.g., broken rotor bars, bearing defects, rotor misalignment) and electrical anomalies (e.g., stator winding faults, voltage imbalance) under constant and variable speed regimes. A hybrid time–frequency feature extraction strategy was employed, which yielded a diagnostic accuracy of 97.6%, with precision at 97.1%, recall at 97.5%, and an F1-score of 97.3%. In comparison, conventional classifiers such as Support Vector Machines (accuracy 92.3%), Random Forest (94.2%), and k-NN (89.7%) underperformed by margins ranging from 3.4% to 8.1%. For high-risk fault classes, including broken rotor bars and bearing wear, the ANN achieved class-specific F1-scores exceeding 96%, demonstrating exceptional sensitivity to critical failure signatures. Explainable AI (XAI) methods that provide a quantitative understanding of feature contributions and decision pathways, such as SHAP and Layer-wise Relevance Propagation (LRP), were integrated to improve the interpretability and reliability of the model. The suggested framework lowers operating costs, allows for real-time fault detection and predictive maintenance, and has the potential to cut unscheduled downtime by more than 30% and increase the lifespan of EV motors by up to 25%.

Download

Share

COinS