To address the issue of data imbalance in the hard disk failure prediction, this paper proposes an improved data augmentation method which integrates Wasserstein generative adversarial network with gradient penalty (WGAN-GP) and residual networks. The residual modules are introduced into the generator and discriminator of WGAN-GP to enhance the model’s capabilities of extracting complex data features and distinguishing between real data and generated data samples. Experiments carried out on a public dataset related to the hard disk failure demonstrate that the proposed model can effectively alleviate the problem of gradient vanishing, and generate enhanced data samples closer to the real data. Furthermore, hard disk failure prediction conducted on the model is more powerful in identifying fault samples and exhibits higher overall classification performance when compared with other models.