Reinforcement learning-driven adaptive stealthy attacks manipulating battery state of health estimation

The accurate estimation of the state of health (SOH) of lithium-ion batteries is paramount for the reliability and safety of the battery management system (BMS). However, the increasing connectivity of BMS exposes it to cyber threats, particularly advanced false data injection attacks (FDIA) with adaptive capabilities, which can precisely manipulate SOH estimation and jeopardize system security. To further investigate the mechanisms of such attacks and provide effective tools for the robustness evaluation of SOH models, this paper proposes a deep reinforcement learning-driven adaptive stealthy attack method (ReLASA), which models FDIA as a Markov decision process with a multi-source state space and continuous action space for voltage-current perturbations, which enables precise, smooth control over long-term SOH degradation trajectories. The reward function jointly optimizes attack effectiveness, multi-modal stealthiness, and physical constraint compliance. The agent trained based on the Soft Actor-Critic algorithm with entropy regularization can generate optimal attack sequences in real time. Experimental results demonstrate that ReLASA reliably manipulates SOH estimation to track an attacker-defined trajectory with high accuracy (R² > 0.97), while maintaining low detection scores for stealth. This multi-dimensional performance significantly surpasses both GAN-based and PGD attacks, thereby providing an effective tool for adversarial testing and system security analysis.