Scholarlink

<p>Tool wear prediction remains a significant challenge because multi-sensor signals exhibit complex temporal dynamics and heterogeneous relationships across different sensing locations. However, most existing deep-learning-based methods focus primarily on either temporal patterns or local spatial features, which limits their ability to effectively model such heterogeneous spatiotemporal dependencies. To address these challenges, a novel dual-branch local-global graph convolutional network (DLGGCN) is proposed. Specifically, a dual branch architecture including global and local feature extraction and fusion modules with graph structure learning is designed to adaptively capture hidden spatiotemporal dependencies across multiple temporal scales. Then, a multi-scale cross attention (MSCA) mechanism is further introduced to fuse the global and local representations by emphasizing inter-branch semantic interactions. Experimental results on the PHM2010 dataset show that the DLGGCN model outperforms eleven state-of-the-art methods, achieving an average MAE of 3.41&#xa0;μm and RMSE of 4.47&#xa0;μm. This corresponds to a 29.98% reduction in MAE and a 22.12% reduction in RMSE compared to the best baseline.</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

A dual-branch local-global graph convolutional network for tool wear prediction

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Yu Xia,
  • Hui Liu

摘要

Tool wear prediction remains a significant challenge because multi-sensor signals exhibit complex temporal dynamics and heterogeneous relationships across different sensing locations. However, most existing deep-learning-based methods focus primarily on either temporal patterns or local spatial features, which limits their ability to effectively model such heterogeneous spatiotemporal dependencies. To address these challenges, a novel dual-branch local-global graph convolutional network (DLGGCN) is proposed. Specifically, a dual branch architecture including global and local feature extraction and fusion modules with graph structure learning is designed to adaptively capture hidden spatiotemporal dependencies across multiple temporal scales. Then, a multi-scale cross attention (MSCA) mechanism is further introduced to fuse the global and local representations by emphasizing inter-branch semantic interactions. Experimental results on the PHM2010 dataset show that the DLGGCN model outperforms eleven state-of-the-art methods, achieving an average MAE of 3.41 μm and RMSE of 4.47 μm. This corresponds to a 29.98% reduction in MAE and a 22.12% reduction in RMSE compared to the best baseline.