Fault detection in electric vehicle lithium batteries is a crucial task to ensure the safety and reliability of electric vehicles. However, the scarcity of labeled data is a major challenge in achieving high-accuracy fault detection. This paper proposes a novel training method to address the issue of limited labeled data. The method begins with unsupervised pre-training of a feature extractor using a large amount of unlabeled data, followed by training a base classifier on a small set of labeled data. Then, through progressive training with self-generated labels, the base classifier predicts the unlabeled data to generate high-confidence labels, iteratively enhancing model performance. We validated the effectiveness of this method on the task of fault detection in electric vehicle lithium batteries. Experimental results demonstrate that the proposed method significantly outperforms traditional methods in handling the problem of labeled data scarcity.