With the rapid expansion of data centers, efficient detection of disk storage system failures has become critical for ensuring data security. However, due to the extremely low occurrence rate of disk failures, conventional deep learning-based detection methods suffer from severe data imbalance, leading to poor recognition of fault events. To address this issue, we propose a clustering-enhanced generative adversarial network, termed OA-WGAN. Specifically, hierarchical agglomerative clustering and One-Class SVM are jointly employed to analyze disk failure data and divide it into fine-grained fault subcategories, enabling the model to better focus on the recognition and modeling of rare failure types. The clustering results guide the sample generation process of a Wasserstein GAN, with a dedicated classification loss introduced to significantly enhance the generator’s sensitivity to rare fault patterns. Experiments conducted on the publicly available Backblaze hard drive dataset demonstrate that OA-WGAN outperforms existing GAN variants in both distribution balance and data fidelity across fault subcategories. The model achieves an average accuracy of 0.99382 and an average recall of 0.9913 across five classifiers, exceeding the second-best model, CYCLEGAN, by approximately 0.39% in accuracy (0.98996) and 0.44% in recall (0.98694). Furthermore, ablation studies show that OA-WGAN consistently outperforms all ablation variants and baseline models, confirming the effectiveness of the proposed approach.