Robust Tractor Oil Leak Detection: SwinTransformer-based Deep Learning Approach

This study proposes a robust oil leak detection in tractor manufacturing, integrating SwinTransformer as the backbone of YOLOv11 to enhance global feature learning. SwinTransformer’s shifted window mechanism and hierarchical feature representation were leveraged to overcome the convolutional neural network (CNN)-based YOLO’s local feature learning limitations. The proposed model evaluated on a dataset of 2,800 oil leakage images captured under UV light conditions on a TYM T25 tractor yielded 0.832 mAP@50, outperforming YOLOv11x (0.817). Grad-CAM visualizations confirmed accurate localization of oil leaks, providing high interpretability. Although the model requires increased computational resources, the inference times remain practical for agricultural tractor manufacturing processes. Future optimization could further improve efficiency and accuracy for real-time, practical industrial deployment.