MI-SOH: a multi-indicator feature dependency model for lithium-ion battery state-of-health Estimation

Abstract

Abstract Accurate state-of-health (SOH) estimation is vital for guaranteeing the safety and longevity of lithium-ion batteries. However, most existing methods employ static feature fusion strategies that fail to account for temporal evolution of indicator correlations throughout battery degradation, leading to compromised estimation accuracy under complex, non-stationary aging patterns. To address this gap, this study proposes MI-SOH, a multi-indicator SOH estimation model that dynamically adapts to evolving feature importance across battery lifecycle stages. MI-SOH primarily consists of four core components: (1) Multi-indicator Feature Weighting Block that employs dual-correlation analysis to adaptively prioritize health factors based on correlation patterns that reflect multi-stage degradation characteristics; (2) Temporal Pattern Extraction Block that processes these weighted features through dilated convolutions to capture multi-scale degradation dynamics; (3) Cross-Variable Dependency Modeling Block that utilizes inverted transformers to learn complex interdependencies among different health indicators throughout battery aging; and (4) Adaptive Hyperparameter Optimization Block that automatically configures model hyperparameters for optimal performance across diverse battery conditions. Extensive experiments on benchmark National Aeronautics and Space Administration (NASA) and Center for Advanced Life Cycle Engineering (CALCE) datasets demonstrate that MI-SOH outperforms current mainstream prediction approaches across diverse battery chemistries and lifecycles, achieving average Root Mean Squared Error (RMSE) of 0.00312 and 0.01126 respectively. This research advances intelligent battery management systems (BMS) by providing a practical SOH monitoring framework critical for electric vehicle safety and energy storage reliability.