Automated Inspection of Flywheel Metallic Objects Using Deep Learning-Based Defect Detection

In the current industrial era, automated inspection is crucial for maintaining the quality and reliability of manufactured components. Accurate and efficient inspection processes are essential to meet industry standards. This research presents a deep learning-based model integrated with automated inspection hardware for detecting defects in flywheel workpieces. The flywheels, with dimensions of 22.75 cm outer radius, 16.75 cm inner radius, and 6 cm thickness, were inspected to detect holes, scratches, and pitted defects. The model was trained on 1900 images from 38 flywheels and tested on 7, with 40 ground truth defects identified by professional QA from industries. To meet the requirement of inspecting each flywheel within 20 seconds, we designed a hardware system incorporating a stepper motor, camera, and lighting. The deep learning model is based on YOLOv4Tiny, with modifications to the backbone and feature fusion through an improved Feature Pyramid Network (FPN). We evaluated the model using a confusion matrix to determine precision, recall, and accuracy. Our model detected defects as small as 0.5 mm and achieved the highest accuracy at 95% and 40 fps, outperforming YOLOv5, YOLOv6, YOLOv7, and DETR. The optimal motor speed, determined by real-world observation, was 200 RPM, enabling the system to complete the inspection within 20 seconds.