XSOH-NET: An Explainable Deep Learning Framework for Battery State of Health Prediction in Electric Vehicles

Abstract

The assessment of battery State of Health (SOH) is vital for enhancing electric vehicles' safety, reliability and energy efficiency. Deep learning models achieve high prediction accuracy, but their black-box nature renders them unsuitable for implementation in Battery Management Systems (BMS). This paper proposes an explainable deep learning framework based on CNNs (XSOH-Net) that uses CNNs for degradation features, the Bidirectional Long Short-Term Memory (BiLSTM) networks for analysis of temporal battery aging, and SHapley Additive exPlanations (SHAP) design for a good understanding of the model. The efficacy of the proposed framework was assessed on the NASA lithium-ion battery dataset and it provides MAE $(0.007-0.010)$, MSE (0.00019 -0.00031), RMSE (0.0138-0.0176), and R2 (0.993-0.996). According to the comparative results, XSOH-Net offers superior prediction accuracy, as well as transparent decision-making capabilities, compared to existing deep learning methods. Hence, it is a reliable solution for intelligent battery health monitoring and prognosis in EVs.