AI-driven diagnosis and health management of autonomous electric vehicle powertrains: An empirical data-driven approach

The reliability and operational continuity of electric vehicle (EV) powertrains critically depend on the health of traction motors, which are subject to dynamic loading conditions and mechanical degradation. Conventional diagnostic techniques often fall short in providing accurate, real-time fault identification under such variability. To address this challenge, this paper proposes a novel and integrated data-driven methodology for fault diagnosis of induction motors (IMs) in EV powertrain applications. The approach leverages vibration signals acquired from accelerometers and employs a hybrid machine learning (ML) framework. The study focuses on identifying the most informative features from time, frequency, and wavelet domains, followed by dimensionality reduction using Principal Component Analysis (PCA) and Correlation Analysis (CA) to enhance classification performance, reduce complexity, and improve model interpretability. A suite of supervised ML classifiers, decision trees, gradient-boosted trees, random forests, and artificial neural networks are evaluated using experimental vibration data collected from bench-mounted IMs operating under three distinct conditions: healthy, bearing fault, and static eccentricity. Among the models, the optimized neural network combined with CA-selected features achieved the most consistent diagnostic performance, supported by low root mean square error and balanced evaluation metrics. The novelty of this work lies in the empirical benchmarking of reduced feature sets across diverse classifier families and the end-to-end validation of diagnostic robustness using real vibration signals under controlled EV-relevant fault scenarios. The findings highlight the potential of hybrid feature selection and ML optimization to enable real-time, high-fidelity health monitoring of intelligent EV powertrains.