In the field of intelligent fault diagnosis of industrial equipment, fault samples are scarce and the labeling cost is high, resulting in limited available training data and an imbalanced distribution (with more normal samples and fewer fault samples in the training data). Based on deep learning theory, this paper adopts a Siamese data augmentation strategy to address the problem of sample imbalance and proposes an improved Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) data generation method. First, the generator is redesigned with a transposed convolutional network to improve the quality of generated samples; then, a polynomial loss function and regularization techniques are introduced to alleviate the problem of mode collapse during training; finally, the Case Western Reserve University bearing dataset and the Dalian centrifugal pump bearing dataset are used for experimental verification, and iterative correction is applied to obtain generated data that better match the characteristics of real data. The results show that the proposed fault sample generation method can alleviate the data imbalance problem and effectively improve diagnostic accuracy.

deep learning,class imbalance,data augmentation,generative adversarial network