Scholarlink

<p>Tool wear monitoring has gained wide attention in both research and industry, yet most models are trained with classification losses that emphasize inter-class separability on seen data while neglecting intra-class compactness and margin robustness. This work proposes a Dual Siamese Neural Network (DSNN) tailored for tool-wear monitoring and a matching loss function, the Normalised Dual-Pair Loss (NDPL). NDPL is introduced as a structurally new loss rather than a modification or combination of existing losses. Taking into account the monotonic progression of tool wear, a training sample construction strategy for the DSNN was developed. Leveraging the proposed NDPL in conjunction with traditional classification loss, effective tool wear monitoring was realized. Experiments show that DSNN learns more discriminative representations: the mean inter-class distance in the learned feature space reaches 6.9, compared with 1.2 for a standard Siamese network and 2.0 for a triplet network. In monitoring performance, DSNN attains 99.1% accuracy, outperforming the Siamese Neural Network (98.1%) and triplet Neural Network (98.3%). Further comparison with representative discriminative feature-learning methods shows that the test accuracies of Center Loss, Circle Loss, CosFace, CurricularFace, ArcFace, and AdaFace are all lower than those achieved by the proposed method. These results indicate improved margin robustness and higher classification accuracy, providing a solid basis for high-precision tool-wear monitoring in practical applications.</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

Dual Siamese neural network for discriminative tool-wear feature learning: a metric-learning approach

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Bo Qin,
  • Qinghua Song,
  • Xiaoliang Liang,
  • Feng Guo,
  • Haifeng Ma,
  • Zhanqiang Liu

摘要

Tool wear monitoring has gained wide attention in both research and industry, yet most models are trained with classification losses that emphasize inter-class separability on seen data while neglecting intra-class compactness and margin robustness. This work proposes a Dual Siamese Neural Network (DSNN) tailored for tool-wear monitoring and a matching loss function, the Normalised Dual-Pair Loss (NDPL). NDPL is introduced as a structurally new loss rather than a modification or combination of existing losses. Taking into account the monotonic progression of tool wear, a training sample construction strategy for the DSNN was developed. Leveraging the proposed NDPL in conjunction with traditional classification loss, effective tool wear monitoring was realized. Experiments show that DSNN learns more discriminative representations: the mean inter-class distance in the learned feature space reaches 6.9, compared with 1.2 for a standard Siamese network and 2.0 for a triplet network. In monitoring performance, DSNN attains 99.1% accuracy, outperforming the Siamese Neural Network (98.1%) and triplet Neural Network (98.3%). Further comparison with representative discriminative feature-learning methods shows that the test accuracies of Center Loss, Circle Loss, CosFace, CurricularFace, ArcFace, and AdaFace are all lower than those achieved by the proposed method. These results indicate improved margin robustness and higher classification accuracy, providing a solid basis for high-precision tool-wear monitoring in practical applications.