Robust oil leak detection for tractor hydraulic fittings using multiscale feature learning and attention

Oil leak inspection in tractor hydraulic fittings is a critical quality control process, but it still relies largely on manual visual inspection, causing inconsistent detection and reduced efficiency. Although ultraviolet (UV) fluorescence imaging has been explored, existing methods remain sensitive to environmental variability and are rarely validated under production-line conditions. This study presents a manufacturing-oriented automated inspection system that integrates an enhanced You Only Look Once (YOLO)-based detector with a manufacturing execution system (MES) for real-time oil leak inspection. To address low illumination, noise, and irregular leak morphology in UV inspection, Swin Transformer V2 is integrated into the YOLOv11 backbone for global contextual learning, and the Convolutional Block Attention Module (CBAM) is placed before the three multiscale detection heads for channel- and spatial-selective refinement. A dataset of 1340 images from the TYM T25 production line was used to train and evaluate the model, which achieved an mAP@50 of 0.838 with an inference time of 15.6 ms. The deployed system was evaluated during a nine-month on-site trial comprising 2000 controlled inspection trials; all 750 reproduced-leakage trials and 1250 defect-free trials were correctly classified. The automated system reduced the average inspection time by 22.71% compared with manual inspection. These results demonstrate the practical applicability of the proposed MES-linked system for standardized and traceable final quality inspection with reduced dependence on operator experience.