Adversarial defense for battery state-of-health prediction models

Abstract

This study addresses the challenge of state of health (SOH) estimation for lithium-ion batteries using a generative graphical approach under adversarial conditions. We analyze the impact of adversarial data poisoning attacks on SOH prediction models, specifically employing the fast gradient sign method (FGSM) and iterative fast gradient sign method (IFGSM). To enhance model robustness, we propose a two-defense strategy against such attacks. The effectiveness of these defenses is evaluated using error metrics such as root-mean-square error (RMSE), mean absolute error (MAE), and mean-square error (MSE). Results indicate that the proposed strategy significantly improves the model’s ability to accurately predict SOH, even in the presence of malicious data.