Abstract: To address the challenges of frequent condition variations, limited fault acoustic signal samples, and imbalanced data in pumped storage unit anomaly detection, this paper proposes an improved Vision Transformer (ViT)-based method for detecting abnormal acoustic signals. First, the Mel spectrogram algorithm is employed to convert one-dimensional acoustic signals into two-dimensional spectrograms, enhancing the information content of fault samples. The generated spectrograms are then input into the ViT network, where the interaction mechanism between self-attention layers and image features is leveraged to learn invariant features across multiple working conditions. Furthermore, a domain prompt and prompt adaptation module is introduced to predict the unit status in the target domain based on feature similarities between the source and target domains. Results show that the proposed method achieves an average accuracy of 90.0%, a recall of 87.9%, and an F1-score of 88.7% on a real-world dataset. On the MIMII dataset, it outperforms other methods by 8.7%, 6.92%, and 4.52% in accuracy, recall, and F1-score, respectively. Therefore, the proposed model effectively addresses anomaly detection tasks under multi-condition and limited-sample scenarios.