Deep Learning based Parts Classification in a Cognitive Robotic Cell System

Advanced manufacturing systems increasingly rely on intelligent algorithms to discriminate, model and predict system behaviours that lead to increased productivity. Edge intelligence allows the industrial systems to collect, compute and act based on process data while reducing the latency and cost associated to an hierarchical control system in which complex decisions are generated in the upper layers of the automation hierarchy. Greater local computing capabilities allow the online operation of such algorithms while accounting for increased performance requirements and lower sampling periods of the control loops. In this work we present the concept of a cognitive robotic cell that collects, stores and processes data in situ for enabling the control of a robotic arm in a production setting. The main features that characterise the robotic cell are embedded computing, open interfaces, and standards-based industrial communication with hardware peripherals and digital twin models for validation. An application of part classification is presented that uses the YOLOv4 image processing algorithm for real-time and online assessment that guides the control of an ABB IRB120C robotic arm. Results illustrate the feasibility and robustness of the approach in a real application. Quantitative evaluation underlines the performance of the implemented system.