A feature-dimension-adaptive physics-informed neural network for cross-operational state of health prediction of lithium-ion batteries

Accurate state of health (SOH) prediction is critical for reliable lifespan management of lithium-ion batteries. Existing data-driven approaches predominantly rely on manually engineered features or fixed-dimensional indicators, limiting adaptability to diverse operational conditions. To overcome this limitation, we propose a feature-dimension-adaptive physics-informed neural network (FDA-PINN) for cross-operational SOH prediction. The framework comprises two core modules. The adaptive feature extraction module employs a one-dimensional convolutional neural network to learn degradation-sensitive representations from raw time-series data, supported by a dual-path temporal encoder that decouples intra-cycle dynamics from inter-cycle aging trends via an orthogonality loss. The physics-informed prediction module incorporates a subnetwork with sinusoidal activations that models degradation kinetics and enforces consistency between the data-driven SOH estimate and the physically derived degradation rate through a residual loss. The feature-dimension-adaptive mechanism is implemented via an information bottleneck objective that maximizes mutual information between fused features and SOH while penalizing redundancy within the convolutional features. This design allows the model to identify the optimal feature dimension from a predefined search space during training, where prediction error and feature dimension converge jointly with substantially reduced manual tuning effort. Experiments on XJTU, NASA, and HUST datasets under four sampling strategies show that FDA-PINN consistently achieves MAE and RMSE below 0.01 in most scenarios, outperforming seven baseline models. Cross-dataset transfer learning further confirms its generalization capability, with RMSE below 0.03 across distribution shifts. These results validate that integrating adaptive feature extraction with soft physical constraints offers a robust and generalizable solution for battery health management across diverse settings.