Scholarlink

<p>Diagnosing faults across different machines–known as cross-machine fault diagnosis–is a critical yet challenging task in industrial environments. To effectively address this issue, domain adaptation techniques have been explored to transfer diagnostic knowledge from labeled source domains to unlabeled or sparsely labeled target domains. However, cross-machine adaptation remains difficult due to limited labeled data and significant heterogeneity in machine types, operating conditions, and environmental settings. To overcome these challenges, we propose a semi-supervised domain adaptation framework called Domain Adversarial Perturbation (DAP). The proposed method leverages spectrogram encoding to effectively handle vibration signals, making it especially suitable for scenarios with limited labeled data. Moreover, domain adversarial perturbations are introduced to reduce domain discrepancies between machines while preserving intra-machine class distinctions. Additionally, a pseudo-labeling strategy based on spectrogram similarity further enhances the utilization of unlabeled data from the target domain. Extensive experiments conducted on the CWRU, IMS, and HUST benchmark datasets demonstrate that our proposed DAP method consistently achieves significant performance improvements and outperforms state-of-the-art methods across various cross-machine adaptation scenarios.</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

Improving rotating machinery fault diagnosis across heterogeneous machines with domain adversarial perturbation in semi-supervised domain adaptation scenario

  • 期刊论文

  • 发表日期：

    2025年6月13日

查看全文
  • Yujun Yang,
  • Jun Kataoka,
  • Sang-Hyuk Yun,
  • Hyunsoo Yoon

摘要

Diagnosing faults across different machines–known as cross-machine fault diagnosis–is a critical yet challenging task in industrial environments. To effectively address this issue, domain adaptation techniques have been explored to transfer diagnostic knowledge from labeled source domains to unlabeled or sparsely labeled target domains. However, cross-machine adaptation remains difficult due to limited labeled data and significant heterogeneity in machine types, operating conditions, and environmental settings. To overcome these challenges, we propose a semi-supervised domain adaptation framework called Domain Adversarial Perturbation (DAP). The proposed method leverages spectrogram encoding to effectively handle vibration signals, making it especially suitable for scenarios with limited labeled data. Moreover, domain adversarial perturbations are introduced to reduce domain discrepancies between machines while preserving intra-machine class distinctions. Additionally, a pseudo-labeling strategy based on spectrogram similarity further enhances the utilization of unlabeled data from the target domain. Extensive experiments conducted on the CWRU, IMS, and HUST benchmark datasets demonstrate that our proposed DAP method consistently achieves significant performance improvements and outperforms state-of-the-art methods across various cross-machine adaptation scenarios.