A YOLOv3 and ODIN Based State Detection Method for High-speed Railway Catenary Dropper

Abstract

The dropper is one of the core equipment of high-speed railway catenary, and its working state affects the power supply stability of pantograph catenary system. In this paper, we propose an effective detection method of catenary dropper state based on target detection algorithm You Only Look Once (YOLOv3) and Out-of-Distribution Detector for Neural Networks (ODIN). This method uses YOLOv3 as dropper locating network to detect the dropper area in catenary. The designed dropper state classification model based on ODIN is trained by augmented dropper area images of various states, and then is applied to analyze the specific state of dropper area from the location area images which is output by dropper location network. The extensive experimental results show that YOLOv3 can accurately detect dropper. The ODIN can effectively eliminate the interference of locating errors on the classification of dropper state, and the detection performance of the dropper state classification model is significantly improved by data augmentation. On the testing set, the accuracy of dropper locating network is more than 94.1%, and the precision of dropper state classification model achieve 97.97%.