Ball Bearing Fault Diagnosis Based on Hybrid Adversarial Learning

Abstract

Ball bearings are prone to faults in their inner and outer rings and rolling elements. Timely detection of these faults is crucial, especially when adversarial perturbations are present, as deep learning-based fault diagnosis models may misclassify these faults. To address this issue, this study proposes a hybrid adversarial learning method that combines convolutional neural networks with a generative adversarial network framework. In this method, the generator introduces perturbations and adaptively adjusts them based on their magnitude and gradient information. The discriminator was used to verify the effectiveness of adversarial perturbations. The goal of this hybrid adversarial learning method is to improve the fault recognition accuracy of a model when subjected to perturbation attacks. The experimental results show that under adversarial perturbation attacks, the proposed method outperforms other deep learning models and defence methods, demonstrating the effectiveness of this approach.