Scholarlink

<p>Fault diagnosis is often impeded by data scarcity and severe noise interference, which hinder effective feature extraction and correlation analysis. Graph neural networks (GNNs) are promising for few-shot learning, leveraging their inherent proficiency in capturing both intrinsic features and complex inter-node correlations. However, their performance relies on graph quality, and conventional metric-based graph construction methods are easily distorted by noise, limiting diagnostic reliability. To overcome this limitation, this study introduces a graph construction optimized GNN named dynamic intensity noise injection Siamese GNN (DINISiam-GNN). Firstly, Siamese networks are introduced, and the contrastive loss function is innovatively improved to optimize graph construction by clustering node features while mitigating overfitting during training. Subsequently, a DINI module is designed within the Siamese network to enhance robustness in strong noise environments. Finally, ablation and comparative experiments are conducted on rolling bearings and planetary gearboxes based on the generalized framework for fault diagnosis, demonstrating the proposed method’s superior robustness and higher diagnostic accuracy across different transmission components.</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

Graph construction optimized Siamese graph neural network for robust fault diagnosis under few-shot and strong noise conditions

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Jinhua Mi,
  • Zhiguo Wang,
  • Hao Tian,
  • Yan-Feng Li,
  • Zhen Liu

摘要

Fault diagnosis is often impeded by data scarcity and severe noise interference, which hinder effective feature extraction and correlation analysis. Graph neural networks (GNNs) are promising for few-shot learning, leveraging their inherent proficiency in capturing both intrinsic features and complex inter-node correlations. However, their performance relies on graph quality, and conventional metric-based graph construction methods are easily distorted by noise, limiting diagnostic reliability. To overcome this limitation, this study introduces a graph construction optimized GNN named dynamic intensity noise injection Siamese GNN (DINISiam-GNN). Firstly, Siamese networks are introduced, and the contrastive loss function is innovatively improved to optimize graph construction by clustering node features while mitigating overfitting during training. Subsequently, a DINI module is designed within the Siamese network to enhance robustness in strong noise environments. Finally, ablation and comparative experiments are conducted on rolling bearings and planetary gearboxes based on the generalized framework for fault diagnosis, demonstrating the proposed method’s superior robustness and higher diagnostic accuracy across different transmission components.