Scholarlink

<p>Tool wear state directly influences CNC machining quality and efficiency. Some non-universal sensors closely related with the tool wear state are only existed in the laboratory environment rather than the industrial scenes. Differences in the characteristics of acquired signals between the industrial and laboratorial scenes restrict to establish an industrial tool wear monitoring(TWM) model with the high-accuracy wear prediction from the laboratorial TWM model with the uncommon sensors. To this end, a cascaded two-stage CNN-Transformer-BiGRU (CT2S-CTBiGRU) model is proposed. Firstly, a cascaded two-stage training scheme is developed to transfer the laboratorial TWM model with the intermediate supervisory labels to the industrial TWM model. By involving the supervisory labels from cutting force and sound signals, the correlation relationship between sensor signals used in industrial applications and tool wear conditions can be sufficiently explored, and then the industrial TWM model can be quickly constructed. Secondly, a PHC module coupled with a BiGRU regression network is designed to extract latent features from multi-source signals. The implicit information of the multi-sensor signals about tool wear condition is sufficiently explored. Thirdly, a divide-and-conquer Transformer strategy dynamically re-weights the features of sensor signals to enforce the TWM model to concentrate on task-relevant information. By fusing multi-scale spatial features and temporal dependencies, the TWM model reveals the hierarchical latent patterns about tool wear states. Experimental results with varying cutting conditions validate that the proposed CT2S-CTBiGRU model maintains the robust adaptability to complex monitoring scenarios while preserves monitoring accuracy.</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

Multi-sensor signal fusion for tool wear condition monitoring based on a cascaded two-stage CNN-Transformer-BiGRU network

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Huang Xiaoyong,
  • Wu Yulong,
  • Liu Chuangchuang,
  • Dong Fangfang,
  • Tian Xiaoqing,
  • Xiao Qunbao,
  • Mei Xuesong

摘要

Tool wear state directly influences CNC machining quality and efficiency. Some non-universal sensors closely related with the tool wear state are only existed in the laboratory environment rather than the industrial scenes. Differences in the characteristics of acquired signals between the industrial and laboratorial scenes restrict to establish an industrial tool wear monitoring(TWM) model with the high-accuracy wear prediction from the laboratorial TWM model with the uncommon sensors. To this end, a cascaded two-stage CNN-Transformer-BiGRU (CT2S-CTBiGRU) model is proposed. Firstly, a cascaded two-stage training scheme is developed to transfer the laboratorial TWM model with the intermediate supervisory labels to the industrial TWM model. By involving the supervisory labels from cutting force and sound signals, the correlation relationship between sensor signals used in industrial applications and tool wear conditions can be sufficiently explored, and then the industrial TWM model can be quickly constructed. Secondly, a PHC module coupled with a BiGRU regression network is designed to extract latent features from multi-source signals. The implicit information of the multi-sensor signals about tool wear condition is sufficiently explored. Thirdly, a divide-and-conquer Transformer strategy dynamically re-weights the features of sensor signals to enforce the TWM model to concentrate on task-relevant information. By fusing multi-scale spatial features and temporal dependencies, the TWM model reveals the hierarchical latent patterns about tool wear states. Experimental results with varying cutting conditions validate that the proposed CT2S-CTBiGRU model maintains the robust adaptability to complex monitoring scenarios while preserves monitoring accuracy.