A Physics-Informed Neural Network With Attention-Enhanced Multisource Feature Fusion for Lithium-Ion Battery SOH Estimation

Accurate lithium-ion battery state of health estimation is crucial for its safe and reliable operation. However, existing purely data-driven methods lack interpretability due to their black-box nature. Physics-informed neural networks (PINNs) provide a paradigm synergizing physics and data, but conventional PINNs are limited by static physical feature weight allocation or their embedded physical constraints fail to capture the underlying mechanisms of battery aging, hindering generalization throughout complex and varied aging stages. To address these limitations, this article proposes a novel PINN model with multisource feature fusion and attention enhancement (PSO-APINN). In terms of feature representation, we offer a multisource feature fusion approach. The approach synergistically integrates data features of the charging process, physical features extracted from the relaxation process, and intrinsic battery degradation mechanism indicators [loss of lithium inventory (LLI) and loss of active material (LAM)] quantified from incremental capacity curves, to achieve physics-informed data augmentation. At the model architecture and optimization level, an APINN structure capable of dynamically assigning feature weights is designed. It introduces a feature attention mechanism, which can dynamically adjust the contribution weights of features at different aging stages, thereby overcoming the weak generalization of static global features. Subsequently, to ensure estimations strictly comply with electrochemical laws, LLI and LAM mechanistic constraints are transformed into gradient-based penalty terms embedded within a composite loss function, suppressing physically inconsistent results by enforcing monotonicity. Furthermore, particle swarm optimization is utilized for global hyperparameter search to achieve an optimal balance between accuracy and physical consistency. Finally, comprehensive experiments on various chemical systems demonstrate that PSO-APINN significantly outperforms benchmark and hybrid models, confirming its superior performance.