Physics-inspired degradation modeling guided by dynamic orbits for battery health prognostics

Remaining useful life (RUL) prediction must be reliable for optimal operation, maintenance scheduling and lifecycle management of battery-powered systems. In practical applications, batteries are frequently subjected to fluctuating power demands and irregular operating schedules, resulting in condition-dependent capacity variations that can obscure the underlying slow degradation trend and challenge conventional RUL prediction methods. Building on the experimentally observed orbit-like degradation behavior under dynamic operating conditions, this study develops a physics-inspired degradation modeling framework, termed PhyOrbitNet, for battery RUL prediction. PhyOrbitNet jointly represents intra-cycle degradation manifestation and cross-cycle degradation propagation. In the feature learning stage, a mathematical framework is derived to describe the coupled evolution of fast-varying observables and slow-varying degradation variables in a high-dimensional phase space, and is reformulated as a cross-attention module within the Transformer architecture. Furthermore, considering the influence of historical degradation information, the dual form of a state-space model is derived and formulated as structured masked attention. A knowledge distillation mechanism is further introduced to transfer degradation representation capability to a student model driven solely by operating conditions, enabling long-term RUL prediction. Experimental results demonstrate that PhyOrbitNet provides accurate RUL prediction across different cells and lifetime stages under dynamic operating conditions.