Research On Deep Learning Based Insulators Inspection Algorithm In High-speed Railway Overhead Contact System

The rapid development of China's high-speed railway system has put forward stricter requirements on the safety of contact networks.In order to adapt to the increasing complexity of railway administration,CR has proposed an overall goal of 6C system construction.Since the inspection of the high-speed rail catenary 4C system plays a key role on ensuring the safety of railway contact system,the railway management department has set strict standards for the inspection of the catenary suspensions.As the main equipment of railway catenary suspensions,accurate inspection of insulators is essential to ensure the stable operation of contact system.On the basis of summarizing relevant research around the world,this paper proposes to divide the insulators inspection algorithm in high-speed railway contact system into two stages of object detection and defect recognition.Specifically,this paper designs an insulators object detection algorithm in railway contact system and an insulator defects recognition algorithm based on deep-learning algorithm.The main results achieved in this paper are as follows:1)An improved Faster R-CNN algorithm for insulators object detection in railway contact system is proposed.To meet the multi-scale object detection requirements of contact insulators inspection,this paper introduces feature pyramid(FPN)into original Faster R-CNN,and adjusts model structure according to characteristics of FPN.Experimental results show that the improved Faster R-CNN algorithm in this paper has a comprehensive advantage over Faster R-CNN in terms of insulators detection accuracy,especially the accuracy of small objects detection.2)An insulator defect recognition algorithm based on fine-grained classification is proposed.Aiming at the problem of fewer defect samples and the insignificant defect features relative to the whole insulator,this paper innovatively proposes a fine-grained classification method for insulator defects recognition.Specifically,a NTS-Net model including a multi-agent cooperative self-supervision mechanism is utilized to achieve accurate locating of the insulator defect areas.Then the model recognize insulator defects by fusing local features and overall features.Experimental results show that the NTS-Net model significantly improves the accuracy of insulator defect recognition compared to NS-Net model and ResNet-50 model.