Scholarlink

In a cloud-edge collaborative environment, diverse equipment types and operating conditions yield disparate data distribution. This negatively impacts fault diagnosis model accuracy and generalization. To tackle this challenge, domain adaptation is proposed as an effective solution. However, traditional methods have limitations in capturing complex industrial equipment data patterns and discriminating domain features. Therefore, this paper proposes a novel domain adaptation method based on adversarial training called TFADAMA. Firstly, a feature extraction network is designed by integrating Convolutional Neural Network (CNN) and Transformer to capture comprehensive time-frequency feature representations with domain invariance. Additionally, the construction of a multi-scale domain discriminator enhances the discriminative power of domain features and alleviates negative transfer effects. Experimental results demonstrate that TFADAMA achieves a significant improvement of approximately 10% compared to baseline methods, with an average diagnostic accuracy of around 97%.

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

Multiscale Adversarial Domain Adaptation Approach for Cloud-Edge Collaborative Fault Diagnosis of Industrial Equipment

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yuanyuan Yang,
  • Liang Zhao,
  • Ningjiang Chen

摘要

In a cloud-edge collaborative environment, diverse equipment types and operating conditions yield disparate data distribution. This negatively impacts fault diagnosis model accuracy and generalization. To tackle this challenge, domain adaptation is proposed as an effective solution. However, traditional methods have limitations in capturing complex industrial equipment data patterns and discriminating domain features. Therefore, this paper proposes a novel domain adaptation method based on adversarial training called TFADAMA. Firstly, a feature extraction network is designed by integrating Convolutional Neural Network (CNN) and Transformer to capture comprehensive time-frequency feature representations with domain invariance. Additionally, the construction of a multi-scale domain discriminator enhances the discriminative power of domain features and alleviates negative transfer effects. Experimental results demonstrate that TFADAMA achieves a significant improvement of approximately 10% compared to baseline methods, with an average diagnostic accuracy of around 97%.