Scholarlink

<p>In the field of machinery fault diagnosis, data distributions vary across different working conditions, and obtaining sufficient labeled data is often costly and time-consuming in real-world scenarios. To address this problem, this study proposes a multi-representation unsupervised domain adaptation hybrid convolutional neural network (MRUDA). MRUDA integrates a novel residual deep convolutional neural network with wide first-layer kernels (RWDCNN) and a new multiple batch normalization (MBN) domain adaptation method. The RWDCNN has a hybrid structure with various branches and cross-layer connections, enhancing the multi-representation capability of the model. Meanwhile, MBN introduces multiple batch norms and applies the extracted feature knowledge to unlabeled data of a similar nature but from different domains, enabling unsupervised cross-condition fault diagnosis. Experiments on fault diagnosis benchmarks verify the superiority of MRUDA compared with other transfer methods, achieving an average test accuracy of up to 99.44 % on the CWRU dataset and demonstrating improved accuracy and generalization ability.</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

Multi-representation unsupervised domain adaptation hybrid convolutional neural network for machinery fault diagnosis

  • 期刊论文

  • 发表日期：

    2026年1月15日

查看全文
  • Wenhua Chen,
  • Jianbin Li,
  • Sixing Wu,
  • Qi Wang

摘要

In the field of machinery fault diagnosis, data distributions vary across different working conditions, and obtaining sufficient labeled data is often costly and time-consuming in real-world scenarios. To address this problem, this study proposes a multi-representation unsupervised domain adaptation hybrid convolutional neural network (MRUDA). MRUDA integrates a novel residual deep convolutional neural network with wide first-layer kernels (RWDCNN) and a new multiple batch normalization (MBN) domain adaptation method. The RWDCNN has a hybrid structure with various branches and cross-layer connections, enhancing the multi-representation capability of the model. Meanwhile, MBN introduces multiple batch norms and applies the extracted feature knowledge to unlabeled data of a similar nature but from different domains, enabling unsupervised cross-condition fault diagnosis. Experiments on fault diagnosis benchmarks verify the superiority of MRUDA compared with other transfer methods, achieving an average test accuracy of up to 99.44 % on the CWRU dataset and demonstrating improved accuracy and generalization ability.