Trustworthy data augmentation for fault detection of lithium-ion batteries in energy storage systems

Fault detection in battery energy storage systems (ESS) is challenging due to the limited quantity of real-world datasets. This study proposes a transfer learning generative adversarial network (TL-GAN) approach for augmenting fault data, effectively enhancing few-shot learning for the detection of slight overcharging. To address the dataset's inherent quality constraints, the Wasserstein distance and gradient penalty are employed to enhance the stability of GAN training. To generate a high-fidelity synthetic dataset, a physically informed parameter projection (PIPP) method is introduced to evaluate key parameters from a battery electrochemical model, including ohmic, activation, and concentration overpotentials, as well as the open-circuit voltage. Validation results indicate that the synthetic dataset closely matches critical features observed in the actual battery. The kernel density estimation, t-distributed stochastic neighbor embedding, and Kullback-Leibler divergence also confirm the statistical similarity between the synthetic and original datasets. In addition, the performance of a recurrent neural network-based fault detection can be significantly improved using the augmented dataset from the proposed method.