Scholarlink

<p>Few-shot learning-based deep neural networks have emerged as a promising solution for bearing fault diagnosis, particularly under limited labeled data conditions. However, existing approaches often rely on a single linear comparison module at the end of the feature extraction pipeline and commonly utilize pre-trained models from other domains, which might not generalize well to specific datasets such as those involving bearing faults. In this study, we propose a novel few-shot learning model, namely the Multiple Local-Global Correlation-based Deep Neural Network (MLGCN), for bearing fault diagnosis. The model is designed by simultaneously learning nonlinear data distributions and embedding representations in an end-to-end fashion. Inspired by human perceptual mechanisms that focus attention on task-relevant information, we introduce a Selective Kernel Attention Feature Extractor module to dynamically capture diverse and critical signal features, which is especially effective in limited data conditions. In addition, we present the Multiple Dual Local-Global Cross Attention (Dual LGCA) module to leverage the full feature hierarchy for enhanced similarity learning, enabling hierarchical similarity reasoning—akin to human comparison strategies that integrate multi-scale contextual cues. The proposed model is assessed using two popular bearing fault benchmarks, including the Case Western Reserve University (CWRU) and Paderborn University (PU) datasets, in distinct experiments to assess its accuracy under limited data and its generalization to real bearing damages. Our code will be released at <a href="https://github.com/ZQuang2202/MLGCN-FewshotBearingFault">https://github.com/ZQuang2202/MLGCN-FewshotBearingFault</a>.</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

Multiple Local-Global Correlation-Based Deep Neural Network with Selective Kernel Attention for Bearing Fault Diagnosis

  • 期刊论文

  • 发表日期：

    2025年9月16日

查看全文
  • Van-Quang Nguyen,
  • Thi-Thao Tran,
  • Manh-Hung Vu,
  • Van-Truong Pham,
  • Men-Tzung Lo

摘要

Few-shot learning-based deep neural networks have emerged as a promising solution for bearing fault diagnosis, particularly under limited labeled data conditions. However, existing approaches often rely on a single linear comparison module at the end of the feature extraction pipeline and commonly utilize pre-trained models from other domains, which might not generalize well to specific datasets such as those involving bearing faults. In this study, we propose a novel few-shot learning model, namely the Multiple Local-Global Correlation-based Deep Neural Network (MLGCN), for bearing fault diagnosis. The model is designed by simultaneously learning nonlinear data distributions and embedding representations in an end-to-end fashion. Inspired by human perceptual mechanisms that focus attention on task-relevant information, we introduce a Selective Kernel Attention Feature Extractor module to dynamically capture diverse and critical signal features, which is especially effective in limited data conditions. In addition, we present the Multiple Dual Local-Global Cross Attention (Dual LGCA) module to leverage the full feature hierarchy for enhanced similarity learning, enabling hierarchical similarity reasoning—akin to human comparison strategies that integrate multi-scale contextual cues. The proposed model is assessed using two popular bearing fault benchmarks, including the Case Western Reserve University (CWRU) and Paderborn University (PU) datasets, in distinct experiments to assess its accuracy under limited data and its generalization to real bearing damages. Our code will be released at https://github.com/ZQuang2202/MLGCN-FewshotBearingFault.