Intelligent assembly operations monitoring with the ability to detect non-value-added activities as out-of-distribution (OOD) instances

Abstract

Recognition and localization of actions in manufacturing assembly operations improves productivity and product quality by identifying bottlenecks and assembly errors. In our previous work, we developed an approach that can recognize and localize the assembly standard operating procedures (SOP) steps in real-time using vision cameras. In this work, we augment the previous study with the ability to detect objects corresponding to the step being performed. Additionally, identifying non-value-added (NVA) activities in an assembly operation is challenging, hence, in this study, we propose an approach to detecting NVA activities by considering the out-of-distribution for deep learning models.