In the new energy vehicle industry, fault diagnosis of lithium batteries is becoming increasingly important. However, current methods for detecting faults in lithium batteries are typically based on physical models and require the establishment of complex mathematical models. These methods have low accuracy, high latency, and low adaptability. To address this issue, we developed a time-series fault diagnosis model based on deep transformer technology called TranAD. This paper introduces the training algorithm and fault diagnosis procedure for the TranAD model and presents experimental data to validate the effectiveness of our approach. Our extensive experimental results demonstrate that the method we propose can improve the accuracy of lithium battery fault diagnosis compared to existing methods. The method achieves an average F1 score of 98%, which exceeds that of the autoencoder (AE) method on this dataset. The experiments show that this class of methods using adversarial training and transformer architecture is beneficial for fault detection under battery timing data and provides a new idea for the field of fault detection.