Scholarlink

<p>In manufacturing industries, monitoring the assembly operation performed in an assembly workstation is essential to improve product quality. Assembly monitoring enables meticulous measurement of step and cycle time and facilitates the identification of anomalies such as sequence breaks and missed steps. In this work, we propose an approach to model the actions performed by an operator in an assembly workstation, captured using vision cameras, as graphs. The spatial and temporal information from the video data was captured by modeling the interaction between objects within a video frame and between video frames, respectively. By classifying the graphs constructed through these interactions, we discern different action types, enabling a comprehensive understanding of assembly physics. Our approach was evaluated using two assembly operation datasets collected from industries, demonstrating its ability to identify assembly steps with a macro-average F1 score of 0.70 and 0.9052, respectively. This work aims to develop a real-time monitoring system capable of accurately detecting and localizing actions with an assembly cycle in an assembly workstation while maintaining computational efficiency and precision in the assembly physics modeling process.</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

Spatial-Temporal Graph Learning for Action Detection in Manual Assembly Operations

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Vignesh Selvaraj,
  • Monami Bhuyan,
  • Sangkee Min

摘要

In manufacturing industries, monitoring the assembly operation performed in an assembly workstation is essential to improve product quality. Assembly monitoring enables meticulous measurement of step and cycle time and facilitates the identification of anomalies such as sequence breaks and missed steps. In this work, we propose an approach to model the actions performed by an operator in an assembly workstation, captured using vision cameras, as graphs. The spatial and temporal information from the video data was captured by modeling the interaction between objects within a video frame and between video frames, respectively. By classifying the graphs constructed through these interactions, we discern different action types, enabling a comprehensive understanding of assembly physics. Our approach was evaluated using two assembly operation datasets collected from industries, demonstrating its ability to identify assembly steps with a macro-average F1 score of 0.70 and 0.9052, respectively. This work aims to develop a real-time monitoring system capable of accurately detecting and localizing actions with an assembly cycle in an assembly workstation while maintaining computational efficiency and precision in the assembly physics modeling process.