Detection and Classification of Automobile Wheel Hub Surface Defects Using You Only Look Once Version 8

To address the challenges of high misdetection rates, false positives, low classification efficiency, and limited adaptability to automated production lines for automobile wheel hubs, we developed a defect detection and classification algorithm based on the You Only Look Once Version 8 (YOLOv8) algorithm.The developed algorithm was deployed in a desktop-grade industrial collaborative robot, the MG400 robot (Shenzhen Dobot Corp Ltd., China), which contains a combination of 2D and 3D cameras with a calibrated ring light source to ensure a defect-to-background contrast ratio greater than 30:1.We collected images from a production line and constructed a dataset comprising microcracks, scratches, dents, protrusions, stains, and deformations.Using contrast enhancement, rotation, and flipping, the original 958 images were augmented to 1870.To improve sensor-driven defect detection, we modified the YOLOv8 algorithm by integrating the coordinate attention mechanism into the backbone module to enhance the spatial localization of defects, particularly small microcracks under low-contrast conditions.Additionally, a windowed self-attention structure from the Swin Transformer was embedded in the neck module to extract multiscale features and improve their integration.Focal loss was used in the classification loss function to mitigate model bias arising from class imbalance in the dataset.The improved YOLOv8 model showed an average detection accuracy of 98.33%, a recall of 98.89%, and a mean average precision at the intersection over union threshold of 0.5-0.95 of 99.1% on the test dataset, demonstrating significant improvements over the original YOLOv8 algorithm.The processing time per image was 45 ms, satisfying the production line requirements.In the production line verification for 1000 hubs, the classification accuracy rate of the robot with the improved algorithm reached 98.7%, with a missed inspection rate of 1.3%, and the average classification time was 0.92 seconds per product.The robot system's performance was stable despite environmental fluctuations, significantly improving the sensitivity for microcrack detection compared with traditional manual inspection methods.