An Improved Bearing Defect Detection Algorithm Based on Yolo

Abstract

Bearings are essential and have been widely used in mechanical equipment. However, surface defects on bearings can seriously affect their performance and service life, and manual inspection requires high labor costs. Automatic visual inspection approaches for bearing defect detection become important to ensure equipment safety. Given the difficulty of different forms of defects on bearing surfaces, we propose to obtain defect images with multi-angle lights and to detect defects based on an object detection algorithm with several variations as follows. We eliminate data imbalance by using a focal loss function. We improve network nonlinearity by using a smooth activation function. Experimental results show that the proposed surface defect detection algorithm can achieve an accuracy of 97.56%, much better than traditional approaches. Ablation experiments verify the effectiveness of our variations.