Development of Defect Detection System for BLDC Motor Stator Pin Insertion Using Deep Learning

In the manufacturing industry, vision inspection technology based on deep learning exhibits excellent performance in detecting unstructured defects, but in the actual process, it has been difficult to secure sufficient learning data owing to the absolute lack of defect data and class imbalance. In this study, we propose a semi-supervised learning system based on the FixMatch algorithm to solve this problem and accurately detect fine defects in the brushless direct current motor stator pin insertion process for automobiles. The proposed system streamlined the preprocessing process by extracting the region of interest using a jig based on fixed coordinates and maximized the generalization performance of the model by combining weak and strong augmentation. The experiment was performed using data collected from the actual manufacturing line, and only a part of the entire learning data was used as label data. According to the experimental findings, the proposed FixMatch model exhibited superior performance compared with that of the supervised learning-based baseline model under the same conditions. In particular, the proposed model achieved a high accuracy rate even in pins with high detection difficulty due to lighting interference and confirmed the robustness of field variables.