Lightweight video analytics for cycle time detection in manufacturing

In manufacturing, manual processes along an assembly line can have considerable productivity fluctuations. Cycle time, or the length of a certain process, can be used as a diagnostic tool in understanding the performance degradation of a line. In this paper we propose a multi-task classification model that leverages RGB camera data from the production line to solve the problem of cycle time detection, as well as giving additional information that can explain variations in the cycle time. The design of the IoT architecture required for our video-based solution is also considered. We conduct an experiment at a soldering inspection process in a factory by building a computer vision ML module that can estimate the cycle time. Our results show a competitive advantage over a benchmark of YOLO that can also work well in spite of occlusion within the environment and is capable of maintaining strong performance over time.