Fault detection of catenary hanger based on EfficientDet and Vision Transformer

Aiming at the problems of low recognition rate and slow recognition speed of traditional detection methods in the fault state detection of overhead catenary hanging string, this paper proposed a catenary hanger state detection algorithm based on lightweight network EfficientDet and Vision Transformer network. The algorithm includedtwo parts: target location and classification detection. With the improved EfficientDet for hanger location, the located hanger was sent to the improved Vision Transformer network for fault classification detection. First, the empty convolution was used to replace ordinary convolution in the second and third layers of EfficientDet network and to expand the receptive field. The CBAMwas used instead of the original SE attention mechanism in the network to gather the high-level semantic information of the hanging string. The improved EfficientDet can efficiently locate the catenary hanging string with a relatively small size. Second, in order to reduce the number of parameters and retain a large range of feature correlation, four small convolution of 3×3 was applied to replace 16×16 convolution layer of Embedding in Vision Transformer to extract the relationship between shallow and deep features by depth. Meanwhile, when num-head values are different, the effect of attention mechanism on spatial information was analyzed to determine the optimal model of hanger fault classification and detection. Finally, compared with YOLOv3, Faster R-CNN, AlexNet, and VGG16, the accuracy of the hanger location model is 95.2% with a real-time rate of 31 frames/s, and the accuracy of the fault detection model is 98.6% with a real-time rate of 28 frames/s. Experiments show that the proposed algorithm can quickly and accurately detect the fault state of small target hanger, and effectively improve the efficiency of intelligent inspection of catenary. © 2023, Central South University Press. All rights reserved.