Imbalanced Fault Diagnosis of Harmonic Reducers Using Vibration Signals Based on an Auxiliary Classifier WGAN-GP with Spectral Normalization

Abstract

Class imbalance is common in vibration-based fault diagnosis because normal-condition data are generally more abundant than fault data. This study proposes an auxiliary-classifier Wasserstein generative adversarial network with gradient penalty and spectral normalization, termed ACWGAN-SG, for fault-sample generation and progressive dataset augmentation. The method combines class-conditioned generation, Wasserstein adversarial learning, gradient penalty, spectral normalization, and PCC-CS-based sample screening. Experiments were conducted on the public CWRU bearing dataset and a self-built 12-class harmonic-reducer dataset, with the downstream diagnostic experiments covering balance ratios from 1:100 to 1:1. The CWRU and harmonic-reducer experiments were independently repeated five and three times, respectively. Under the balanced condition, ACWGAN-SG achieved mean diagnostic accuracies of 98.40% and 97.627% on the two datasets. On the harmonic-reducer dataset at BR = 1:2, the method obtained a Macro-F1 of 94.298%, a balanced accuracy of 94.333%, and an MCC of 0.9383. Repeated-run statistical analyses showed significant overall differences among the evaluated methods across the tested balance ratios. These results indicate that the proposed generation and progressive-augmentation procedure improves downstream diagnostic performance under the reported experimental settings.