The rapid development of new energy vehicles and energy storage industries has imposed higher demands on power battery health management. Accurate prediction of State of Health has become a core challenge in ensuring system safety and economic viability. Current prediction methods face three major bottlenecks: insufficient adaptability to dynamic operating conditions, difficulties in characterizing micro-level degradation mechanisms, and high implementation costs. Therefore, this paper conducts in-depth research on a multi-physics field feature collaborative extraction mechanism and a lightweight deep learning framework. The goal is to establish a novel prediction paradigm that combines real-time responsiveness with industrial-grade accuracy, providing a theoretical breakthrough direction for battery lifecycle management.