面向车间人员宏观行为数字孪生模型快速构建的小目标智能检测方法

Intelligent small object detection approach for fast modeling of digital twin of global human working activities

作为制造过程中的重要影响因素,人员的活动具有很高的自主性和不确定性,因此人员行为的数字孪生模型构建已成为阻碍数字孪生车间技术发展的瓶颈之一。同时,出于火工品生产安全考虑,航天装备等许多产品的制造对车间中不同区域人员的宏观分布有严格的控制需求。鉴于此,针对人员宏观行为数字孪生模型构建问题提出一种三阶段级联卷积神经网络(3-Stage CCNN)的深度学习算法,对车间现场视频中的车间人员进行识别,以支撑人员在图像中的位置信息提取,并为后续构建人员的宏观行为数字孪生模型的实现提供基础。实验证明,该方法在国际公开数据集Caltech Pedestrian和复杂机电产品某院实际数据集上都达到了较好的执行效果和效率,具有较好的实用价值。

As an important factor in manufacturing,the human working activities are highly autonomous and stochastic,which makes it a chore in modeling the digital twins for human working activities for the development of the digital twin workshops.Besides,for production safety,some defense vendors such as certain aerospace complex mechanical and electrical products usually strictly restrict the number of workers in some key working areas,where the density of people may cause severe accidents.To well handle the workers and improve the production efficiency,a fast approach for modeling global working activities was proposed,and a 3-Stage Cascade Convolutional Neural Network(3-Stage CCNN)was constructed for fast recognition of workers in workshops,and the position of workers were useful to build the digital twin of global working activities.Experiments on open datasets Caltech Pedestrian and the real-world production dataset from certain aerospace vender demonstrated that the approach was both effective and efficient,and was of great practical impacts.