Research On State Identification Of Catenary Dropper Based On Improved YOLO Algorithm

In recent years,the national high-speed railway industry has developed rapidly.Monitoring and maintaining important facilities in the railway system is an important guarantee for the safe and stable operation of trains.The dropper in the catenary system supports and suspends the contact line.The abnormal working state of the dropper will change the elasticity of the contact line,affect the current collection quality of the pantograph,shorten the service life of the contact line and the pantograph,and affect the operation quality of the locomotive.Therefore,the fast and accurate identification of the catenary dropper state is a problem of great concern.The particularity of dropper data form leads to the low recognition accuracy of existing detection algorithms and network models,which is difficult to meet the actual needs.Aiming at the problem of low accuracy of the traditional algorithm in catenary dropper detection,this paper adopts a two-stage detection method with positioning and state recognition.Based on the morphological characteristics of the dropper,the detection method of the target detection algorithm is improved and optimized to solve the problem of insufficient positioning accuracy of the dropper;Based on the sample analysis and existing technology,a dropper state identification network is established to solve the problem of low accuracy of dropper state identification;Aiming at the problem of few fault samples,an anomaly detection network based on convolutional autoencoder is designed to avoid the poor training effect of classification model caused by data imbalance.The main work contents are as follows:(1)When dealing with the dropper object with an inclined angle,the horizontal detection frame of the traditional target detection algorithm will contain a lot of redundant information.This inaccurate positioning method increases the difficulty of subsequent dropper state recognition.To solve this problem,taking the YOLOv5 algorithm as an example,this paper improves the traditional object detection algorithm,improves the horizontal detection frame to the rotation detection frame,and optimizes the data processing method.The improved algorithm can rotational positioning along the inclined direction of the dropper.The experimental results show that its positioning accuracy is higher than that of the traditional detection algorithm.It provides high-quality data for subsequent dropper classification network and convolution autoencoder,and reduces the difficulty of dropper state recognition.(2)Aiming at the problem that the dropper state is difficult to identify,this paper designs the classification network structure based on the morphological characteristics of the dropper,proposes a VRNet classification network suitable for identifying the dropper state and uses the fault simulation to balance the data set for the training of the classification network.The core of VRNet is the wide residual structure embedded in the SE attention mechanism,which is used as a feature extraction module to replace the general convolution in VGG-16.Ghost mechanism is used to lighten the model,reduce the number of parameters and computation of the model,and improve the processing efficiency.The accuracy of the VRNet classification network in the dropper fault classification experiment is 97%,which is better than other traditional classification networks.(3)Aiming at the problem of poor training effect of classification network caused by unbalanced data and lack of various forms of fault samples,an anomaly detection method based on autoencoder is adopted in this paper.According to the characteristics of the dropper,a convolutional autoencoder based on a memory enhancement module is proposed.The encoder and decoder are symmetrical to each other,and a memory module is embedded between them.This module plays a significant role in distinguishing the reconstruction loss of fault samples and normal samples.With the help of lightweight modules,the number of parameters and computation of the network are reduced.In the experiment of dropper fault detection,the network has ideal recognition accuracy,avoids the problem of failure training due to the scarcity of fault samples,and saves the work of fault data processing.