高铁接触网支持装置紧固件识别与定位的深度学习方法

高铁4C检测系统可以获取接触网的大量图像，如何利用人工智能技术检测接触网支撑装置的紧固件松动、脱落、变形等故障，是一项迫切需要攻克的技术难题．由于紧固件在整幅图像中占比非常小，解决这一问题的可行方案是先对紧固件识别定位，然后对其进行图像分割，再识别其故障类型．本文提出一种改进的Faster R-CNN算法，可以准确地实现各种紧固件的识别与定位．具体的改进策略为在深度网络中引入一种基于SE模型的注意力机制，加强各通道对有效特征的提取，在Faster R-CNN中以GA-RPN替代RPN网络．实验结果表明，本文所提出的方法对接触网紧固件识别准确率达93.4%以上．

Deep Learning Methods for Fastener Identification and Location of High Speed Railway Catenary Support Devices

The 4C detection system of high-speed railway can obtain a large number of pictures of high-speed railway catenary. How to use artificial intelligence technologies to detect the looseness, dropping, deformation and other faults of catenary support devices is an urgent technical problem to be solved. Because the fasteners occupy a very small part of the whole images, a feasible solution to the problem is to identify and locate the fasteners first, then segment them, and finally identify the fault type of the segmented fasteners. Aiming at the problem of fastener identification and location, we propose an improved Faster R-CNN algorithm, which can accurately identify and locate various fasteners. The specific improvement strategy is to introduce an attention mechanism based on the SE model into the deep network, extract effective features from each channel, and use GA-RPN instead of RPN in the Faster R-CNN. The experimental results show that the method proposed in this paper has a recognition accuracy of more than 93.4%.