Lightweight deep learning framework for real-time fault diagnosis in EV lithium-ion batteries

Abstract

With the rapid use of electric vehicles (EVs), the need for intelligent and real-time fault diagnosis systems for lithium-ion batteries (LIBs) has become critical to ensure safety, efficiency, and extended operational life. Numerous existing fault diagnosis approaches often require complex computations and fail to adapt to nonlinear, dynamic EV conditions. In addition, data-driven models lack real-time capability, exhibit high computational overhead, and are rarely optimized for embedded deployment. This research presents a hybrid DL-based diagnostic framework named Optimized Elman Spotted Hyena Tuned NeuroNet (OESH-Net) for accurate State of Charge (SOC) estimation and fault detection in EV battery systems. The proposed model integrates a Thevenin-equivalent circuit-based LIB model simulated in MATLAB with a recurrent Elman Neural Network (ENN) optimized using the Spotted Hyena Optimizer (SHO) in TensorFlow/Keras. High-frequency sensor data, including voltage, current, temperature, motor speed, and hall signals, are collected and preprocessed using low-pass filtering, sliding window segmentation, and min-max normalization, to create structured feature vectors. Then, the preprocessed data are fed into the OESH-Net for SOC prediction and health classification into Normal, Warning, and Fault states based on residual SOC error thresholds. Experimental results shows that OESH-Net outperforms models’ sin both charging and discharging phases, achieving 98.86% accuracy, 99.22% F1-score, 0.00078 MSE, and 4.85 ms inference time. It maintains robust performance with a mean absolute SOC estimation error of 2.06%. The proposed OESH-Net framework enables predictive maintenance and early fault detection, offering a scalable, efficient, and real-time solution for smart EV battery management systems.