Abstract: Production safety is the basic requirement of human-oriented intelligent manufacturing.To meet the real-time detection and edge deployment requirements of unsafe behaviors of workers in intelligent workshops,a lightweight YOLOv5s-based method for detecting unsafe behaviors of workers in intelligent workshops was proposed.Structural optimizations were deleted on the feature fusion network and output layer of YOLOv5s.The resulting model files after improvement were subjected to structured pruning.The knowledge distillation was applied to fine-tune the pruned network model.Experimental results demonstrated that the improved YOLOv5s algorithm achieved a mAP@0.5 of up to 97.8%,a 108% improvement in FPS and requires computational powers decreases by 69.0%.The proposed YOLOv5s-2Detect network and lightweight design scheme manifested high accuracy,real-time performance,and robustness in detecting unsafe behaviors of workers,thereby satisfying the detection needs of unsafe behaviors of workers in the practical environment of intelligent workshops.

Key words: intelligent workshop, unsafe behavior, YOLOv5s, structural pruning, real-time performance

摘要： 生产安全是以人为本的智能制造基本要求,为满足智能车间工人不安全行为的实时性检测和边缘端部署需求,提出一种基于轻量化的YOLOv5s的工人不安全行为检测方法。首先,对YOLOv5s特征融合网络以及输出层进行删除;其次,对改进后网络训练得到的模型文件进行结构化剪枝;最后,使用知识蒸馏对剪枝后的网络模型进行微调。实验结果表明,改进后YOLOv5s算法的mAP@0.5高达97.8%,刷新率提升108%,所需算力下降了69.0%。所提出的YOLOv5s-2Detect网络及轻量化设计方案对智能车间工人不安全行为检测具有较高的精度,实时性与鲁棒性能够满足智能车间实际环境中工人不安全行为的检测需求。

关键词: 智能车间, 不安全行为, YOLOv5s, 结构化剪枝, 实时检测