Scholarlink

<p>In contemporary industrial environments characterized by complex operational conditions, mechanical fault diagnosis has been recognized as critical for ensuring equipment reliability and operational safety. However, the inherent nature of real-world machinery operations presents significant challenges, particularly severe class imbalance, which continues to impede diagnostic performance. To address this challenge, a dual-path transformer-based adaptive multiscale network (DTAMNet) is proposed, integrating parallel time-domain and frequency-domain pathways with adaptive multiscale convolution modules. A decoupled domain self-attention mechanism models intra-domain dependencies, while a cross-domain enhancement module facilitates effective feature interaction. To mitigate class imbalance, an adaptive focal classifier is introduced alongside a majority-guided self-distillation loss that emphasizes minority fault classes during training. Extensive experiments on three benchmark datasets under extreme imbalance conditions demonstrate that DTAMNet significantly outperforms state-of-the-art methods, achieving average accuracies of 80.57%, 83.50%, and 84.38% under severe imbalance ratios of 0.05, substantially exceeding comparative methods that ranged from 14.29% to 76.48%. Ablation studies confirm the effectiveness of each proposed component. The proposed framework offers a robust solution for real-world machinery health monitoring systems.</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

Fault Diagnosis of Bearings based on a Dual-Path Transformer-based Adaptive Multiscale Network under Imbalanced Samples

  • 期刊论文

  • 发表日期：

    2025年10月15日

查看全文
  • Shixin Li,
  • Jie Liu,
  • Hui Ma,
  • Na Yang

摘要

In contemporary industrial environments characterized by complex operational conditions, mechanical fault diagnosis has been recognized as critical for ensuring equipment reliability and operational safety. However, the inherent nature of real-world machinery operations presents significant challenges, particularly severe class imbalance, which continues to impede diagnostic performance. To address this challenge, a dual-path transformer-based adaptive multiscale network (DTAMNet) is proposed, integrating parallel time-domain and frequency-domain pathways with adaptive multiscale convolution modules. A decoupled domain self-attention mechanism models intra-domain dependencies, while a cross-domain enhancement module facilitates effective feature interaction. To mitigate class imbalance, an adaptive focal classifier is introduced alongside a majority-guided self-distillation loss that emphasizes minority fault classes during training. Extensive experiments on three benchmark datasets under extreme imbalance conditions demonstrate that DTAMNet significantly outperforms state-of-the-art methods, achieving average accuracies of 80.57%, 83.50%, and 84.38% under severe imbalance ratios of 0.05, substantially exceeding comparative methods that ranged from 14.29% to 76.48%. Ablation studies confirm the effectiveness of each proposed component. The proposed framework offers a robust solution for real-world machinery health monitoring systems.