Fuzzy-YOLO Model for Rail Anomaly Detection: Robustness Under Limited Sample and Interference Conditions

Accurate detection of surface anomalies in railway tracks is critical for ensuring train operation safety and enabling intelligent railway management. However, the scarcity and pronounced imbalance of anomaly samples significantly constrain model training and generalisation. Moreover, complex environmental factors such as illumination variability, sensor noise, and motion blur pose additional challenges to model robustness in real-world applications. This study presents a Fuzzy-YOLO model tailored for limited sample datasets. Built upon YOLOv11, Fuzzy-YOLO incorporates a fuzzy-non-maximum suppression (NMS) mechanism and integrates a lightweight fuzzy residual neural network (RFNN-Res) module based on fuzzy logic for anomaly classification. The final anomaly type is determined via a weighted voting strategy. Experimental evaluations demonstrate that Fuzzy-YOLO achieves a mean average precision (mAP) of 98.90%, exhibiting notably enhanced stability compared to YOLOv11 under conditions of varying illumination, noise, and motion-induced blur.