• Introduce a physics-informed dual-stream network for battery SOH estimation. • Separate macroscopic aging trends from microscopic electrochemical features. • Integrate convolutional feature extraction with attention-based temporal fusion. • Enforce thermodynamic monotonicity through degradation-consistency regularization. • Demonstrate 0.299% MAPE with stable cross-domain transfer performance. Accurate state-of-health (SOH) estimation is critical for the reliability and maintenance of lithium-ion batteries in electric vehicles. However, existing data-driven methods often suffer from poor generalization under varying operating conditions and a lack of physical interpretability. To address these limitations, this study proposes a Physics-Informed Dual-Stream Network (PID-Net) for robust SOH estimation. The architecture uniquely integrates a dual-stream mechanism: a macroscopic stream capturing global degradation trends from raw charging profiles, and a microscopic stream extracting electrochemical differential features to identify aging mechanisms. These heterogeneous features are dynamically fused via a multi-head self-attention mechanism. To mitigate prediction volatility caused by measurement noise, a degradation consistency regularization (DCR) strategy is introduced, enforcing physical monotonicity constraints on the degradation trajectory. Furthermore, a transfer learning paradigm is employed to enhance adaptability to data-scarce scenarios. Experimental validation on public datasets demonstrates that the proposed method achieves superior accuracy and robustness, yielding a root mean square error (RMSE) of 3.39 mAh, a mean absolute percentage error (MAPE) of 0.299%, and a coefficient of determination ( R 2 ) of 0.987. These results confirm that PID-Net offers a highly accurate, physically plausible, and transferable solution for real-time battery health monitoring.
Yang et al. (Sun,) studied this question.