An Automated Inspection System for Hose Assembly of Engine Main Control Valves Using Deep Learning-based Object Detection

In many manufacturing sites, the completion of the process is examined by a manual visual inspection, which is costly and prone to human errors. To overcome this, we propose an automated inspection system for hose assembly of engine main control valves using deep learning-based object detection. CCTVs capture the operator's real-time work process, and the hose connection points and work gears are automatically recognized to determine whether the assembly is completed at the end of the process. We trained a deep learning model based on the Normalized Gaussian Wasserstein Distance and YOLOv5 to detect small objects accurately. The images collected from multiple cameras at different angles were processed in parallel using multi-camera processing. The experimental results prove the applicability of our proposed method, as the object detection model detects all classes correctly.