Scholarlink

<p>This paper addresses the challenge of predicting spindle vibration performance in winding machine assembly. In multi-stage manufacturing, error propagation and accumulation complicate prediction, while existing methods struggle with dynamic coupling and structural complexity. To address this, the paper proposes a multi-scale topology-semantic attention graph neural network (MS-TSAGNN). The model employs a multi-scale convolution module to extract local and global features, and a topology-semantic collaborative attention mechanism to capture dependencies between assembly nodes, integrating both physical connections and nonlinear characteristics. A kernel density estimation method is further introduced to mitigate data imbalance. Experiments show that MS-TSAGNN outperforms graph convolutional and graph attention networks, reducing Mean Absolute Error and Root Mean Square Error by about 11.32% and 10.75%, respectively. Validation on real industrial data confirms the model’s accuracy and robustness, with on-site testing of 50 BWA40-II1800 spindle assemblies achieving an RMSE below 0.034&#xa0;mm, demonstrating its engineering applicability. This work provides practical guidance for vibration performance optimization and digital-twin-based quality prediction in spindle assembly and other complex manufacturing systems.</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

An improved GNN integrating multi-scale topology-semantic attention for performance prediction in spindle assembly

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Siyi Ding,
  • Yan Meng,
  • Yongfeng Li,
  • Jie Zhang,
  • Xinhua Mao

摘要

This paper addresses the challenge of predicting spindle vibration performance in winding machine assembly. In multi-stage manufacturing, error propagation and accumulation complicate prediction, while existing methods struggle with dynamic coupling and structural complexity. To address this, the paper proposes a multi-scale topology-semantic attention graph neural network (MS-TSAGNN). The model employs a multi-scale convolution module to extract local and global features, and a topology-semantic collaborative attention mechanism to capture dependencies between assembly nodes, integrating both physical connections and nonlinear characteristics. A kernel density estimation method is further introduced to mitigate data imbalance. Experiments show that MS-TSAGNN outperforms graph convolutional and graph attention networks, reducing Mean Absolute Error and Root Mean Square Error by about 11.32% and 10.75%, respectively. Validation on real industrial data confirms the model’s accuracy and robustness, with on-site testing of 50 BWA40-II1800 spindle assemblies achieving an RMSE below 0.034 mm, demonstrating its engineering applicability. This work provides practical guidance for vibration performance optimization and digital-twin-based quality prediction in spindle assembly and other complex manufacturing systems.