Activity recognition in manual manufacturing: Detecting screwing processes from sensor data

Abstract

Knowledge about the duration of manufacturing processes and operation times is essential for production planning and control. But data acquisition is often difficult and especially challenging if production requires manual activities. This paper presents different data analysis and machine learning approaches to detect manual manufacturing processes from sensor data. As human activity recognition approaches are not necessarily applicable in industrial environments, all sensors are attached to tools, in this case screwdrivers. A dataset covering different tool movements, sensor types and mounting options is created and analyzed. The results are evaluated in terms of feasibility of the approach.