Fault diagnosis method based on triple generative adversarial nets for imbalanced data

Abstract

Abstract Deep learning (DL) fault diagnosis methods require no expert knowledge and can adaptively extract fault features to realize automated diagnoses. However, factories’ limited and imbalanced data cause DL fault diagnosis methods to fail to meet data diversity requirements and perform poorly. To solve this problem, this paper proposes triple Wasserstein generative adversarial nets with classifier penalty (Triple-WGAN-CP). We first train Triple-WGAN-CP to generate samples to balance the original unbalanced dataset, then input the new balanced dataset to the fault classifier of Triple-WGAN-CP to continue training. Finally, when the numbers of consecutive sampling points in each of the nine fault classes are only 3140, 2300, and 2076, we achieve the highest prediction accuracies of 99.5%, 95.1%, and 65.1%, respectively, and the highest average accuracies for the nine environments (signal-to-noise ratio −4, −2, 0, 2, 4, 6, 8, 10, and ∞) of 96.2%, 84.1%, and 55.1%, respectively. Comparisons with other methods show that this has achieved significant improvements in accuracy and noise robustness and has broad application prospects in the field of limited and imbalanced data fault diagnosis.