Adversarial training defense strategy for lithium-ion batteries state of health estimation with deep learning

Deep learning (DL) methods have great potential in the estimation of the lithium-ion batteries' states, yet those networks can be manipulated by adversarial attacks to a wrong perception which could induce severe safety to the operation of the battery energy storage system. Since this issue has not been addressed in existing research, this paper proposes a comprehensive investigation of the adversarial attacks and the corresponding adversarial training (AT) defense strategy for a trustworthy battery state of health (SOH). A residual convolutional network (RCN) is selected for normal examples (NEs) where the effect of untargeted, semi-targeted, and targeted projected gradient descent attacks on the RCN model are investigated. Since the manipulators can control the battery's estimated degradation trajectory, where the root mean square error (RMSE) of the estimated SOH can be enlarged 11.4 times. In addition, the adversarial-trained RCN (ATRCN) model for adversarial examples (AEs) with different parameters shows a good defense ability and the maximum RMSE can be reduced to 15 % of the normal-trained RCN (NTRCN) model. The proposed ATRCN model is more accurate on NEs and AEs compared to other DL models after AT and also achieves 0.303 % and 1.53 % RMSE on NEs and AEs for another battery chemistry.