IMPLEMENTATION OF A VIOLATION BEHAVIOR IDENTIFICATION AND EARLY WARNING SYSTEM FOR INDUSTRIAL WORKSHOPS

Abstract

Aiming at the pain points in industrial workshops, such as the low efficiency of manual monitoring, poor computing power adaptability of existing vision-based systems, and the lack of a closed-loop alarm mechanism, this paper designs and implements a violation behavior identification and early warning system. First, based on the YOLOv8n model developed by the Ultralytics team, the system incorporates depthwise separable convolutions and attention mechanisms for lightweight optimization. Combined with TensorRT quantization and compression, this ensures low-latency deployment on edge devices. Second, a four-layer architecture is constructed, comprising the Perception Layer, Edge Computing Layer, Intelligent Decision Layer, and Application Layer. The introduction of an intelligent decision mechanism enables an end-to-end automated closed-loop of identification-alarm-archiving. The system is integrated with DingTalk/WeCom and MES (Manufacturing Execution System) to facilitate real-time alarm notifications and data logging. Finally, the system's performance was validated through experiments in real-world industrial workshop scenarios. The results indicate that the average inference latency at the edge is 39 ms, the identification accuracy for violations reaches 92.3%, and the false alarm rate is reduced to 3.2%. All metrics meet the real-time monitoring requirements of industrial sites.