Enhanced fault detection in lithium-ion battery energy storage systems via transfer learning-based conditional GAN under limited data

The accuracy of fault detection in large-scale lithium-ion battery-based energy storage system is limited due to the scarce and low-quality fault dataset. This study proposes a data augmentation technique integrating transfer learning with a conditional generative adversarial network (TL-cGAN) to generate high-quality synthetic fault dataset, thereby enhancing fault diagnosis performance. By embedding fault information as a condition during retraining, this approach enhances the dataset across different fault scenarios. Moreover, the proposed conditional inverse normalization dynamically adjusts normalization parameters based on fault conditions, ensuring physical plausibility. It also refines the kernel density estimation distribution of the generated dataset, thereby preserving key fault signatures. A robust evaluation framework validates the proposed method across static, dynamic, and practical dimensions. Experimental results show that TL-cGAN reduces KL divergence by up to 90 % in key fault types and significantly enhances recall and F1-score in few-shot and low-visibility fault detection tasks, confirming the effectiveness of the method under data-limited conditions.