Using Convolutional Neural Networks for Assembly Activity Recognition in Robot Assisted Manual Production

Due to ever-shortening product life cycles and multi variant products the demand for flexible production systems that include human-robot collaboration (HRC) rises. One key factor in HRC is stress that occurs because of the unfamiliar work with the robot. To reduce stress induced strain for assembly tasks we propose an adjustment of cycle times to the human’s performance, so that the stress that is exerted on the working person by a waiting robot is minimized. For an autonomous adaptation of the cycle time, the production system should be aware of the human’s actions and assembly progress without the need to inform the system manually. Therefore, we propose an activity recognition in assembly based on a machine learning technique. A convolutional neural network is used to distinguish between different activities during the assembly by analyzing motion data of the hands of the working person. The results show that the network is suitable for distinguishing between nine different assembly activities like screwing with a screwdriver, screwing with a hexagon wrench or general assembly and further activities.