YOLOv11s-CD: An Improved YOLOv11s Method for Catenary Dropper Fault Detection

Abstract

The catenary dropper (CD) fault detection is an important technical means to ensure the train current collection quality and operational safety. The existing you only look once (YOLO) detection algorithms need improvement in terms of accuracy, especially in the detection of small objects. To address the problem, this article proposes a CD fault detection model based on improved YOLOv11s, named YOLOv11s-CD. First, a four-detection head structure DASFFHead is designed to achieve multiscale feature fusion by integrating a small object detection layer into the neck network and combining a dynamic adaptive spatial feature fusion (DASFF) module. Subsequently, the squeeze–excitation and attention module (SEAM) attention mechanism is embedded in the neck network layer to extract more small object features in occluded areas. In addition, combining the InnerIoU and CIoU methods, the InnerCIoU loss function is designed to enhance the small object detection ability. Finally, the effectiveness and accuracy of the proposed model are validated on the dataset, which is processed by the optimized contrast-limited adaptive histogram equalization (CLAHE) algorithm to enhance the contrast and clarity of the small object defects. Experimental results show that the proposed YOLOv11s-CD has superior performance compared with several other YOLO algorithms, whose mAP@0.5 has increased to 92.3% and AP of small object detection has significantly increased to 91.3%.