Scholarlink

<p>In response to the challenges of feature extraction and model security faced by convolutional neural networks in bearing fault diagnosis, this study proposes a sample segmentation method based on the characteristics of the fault signal, combined with an analysis of the envelope spectrum to generate training samples. Adversarial training is employed to enhance both diagnostic accuracy and model robustness. The proposed method adaptively determines the sample length based on fault feature frequencies, replacing traditional experience-based segmentation strategies, and utilizes color-mapping techniques to visually present the feature information from the envelope spectrum. To further improve the model’s stability under small perturbations, an adversarial layer based on the FGSM-RP algorithm is introduced by applying random perturbations during training. This significantly enhances the model’s ability to defend against such attacks. The experimental results show that the proposed method not only achieves high-precision fault classification but also effectively defends against adversarial attacks from various algorithms. These findings validate the effectiveness of the proposed method and provide new insights for improving roll bearing fault diagnosis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adversarial Training-Based Convolutional Neural Network Fault Diagnosis for Bearings with Random Perturbation

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Xiaofeng Bai,
  • Qazi Mazhar ul Haq,
  • Aftab Alam Khan,
  • Chaohui Li,
  • Chengwei Li

摘要

In response to the challenges of feature extraction and model security faced by convolutional neural networks in bearing fault diagnosis, this study proposes a sample segmentation method based on the characteristics of the fault signal, combined with an analysis of the envelope spectrum to generate training samples. Adversarial training is employed to enhance both diagnostic accuracy and model robustness. The proposed method adaptively determines the sample length based on fault feature frequencies, replacing traditional experience-based segmentation strategies, and utilizes color-mapping techniques to visually present the feature information from the envelope spectrum. To further improve the model’s stability under small perturbations, an adversarial layer based on the FGSM-RP algorithm is introduced by applying random perturbations during training. This significantly enhances the model’s ability to defend against such attacks. The experimental results show that the proposed method not only achieves high-precision fault classification but also effectively defends against adversarial attacks from various algorithms. These findings validate the effectiveness of the proposed method and provide new insights for improving roll bearing fault diagnosis.