Based On The Target Detection And Deep Learning Of The Catenary Supporting Device Small Size Parts Defect Condition Monitoring

With the rapid development of China’s high-speed railway in recent years,the total mileage of operation is increasing day by day.Catenary is the only device to provide electricity for high-speed railway trains.Only when the catenary is in a stable working state,the trains can be guaranteed to provide continuous and stable electricity.In this paper,picture data of parts and components of catenary support device are collected by4 C detection truck.Based on target detection and deep learning technology,quick and accurate state detection is carried out for small-size parts and components of catenary support device.The main work is as follows:First of all,the YOLOv3-W target optimization algorithm was proposed in oder to solve the problems of lower efficiency of positioning and higher residual rate for detection of catenary support system small size parts.Based on the four aspects improving,such as network backbone to planning,introducing RFB module,adjusting the anchors,optimization the loss function a,the precise positioning of cotter pin bolts and other small size parts were completed.Based on the actual operation data of Beijing-Guangzhou railway line,the training data set of small size parts cogging pin and jack-bolt based on Labelimg image annotation tool was given,and the positioning simulation experiment and method analysis of various positioning algorithms were completed.Then,the first round of expansion was carried out by using image processing technology to solve the problems of poor universality of the detection model to different detection scenarios and over-fitting caused by the disharmony between defect samples of small size parts and normal samples.The DCGAN was improved by optimizing the loss function and adjusting the network structure,and then the defect samples were expanded in the second round based on the improved DCGAN-W algorithm.Based on machine learning framework Tensor Flow and deep learning Keras library,DCGAN-W was used to expand the defect samples of split pins and jack-bolts using the actual operation database of Beijing-Guangzhou Railway line,and the validity and universality of DCGAN-W were verified.Finally,the classification algorithm VGG16 was improved in four aspects,such as optimization activation function and so on,to achieve fast and accurate detection of the defect state.At the same time,a defect detection model of small size parts of catenary support device based on cascade architecture was designed to realize automatic detection of small size parts of catenary support device.Based on the optimized algorithm VGG16-W,the importance of each link of the cascade architecture was verified under the machine learning framework Tensor Flow,and the experimental results of state detection of coving pins and buttoning bolts on the Beijing-Guangzhou railway line prove the superiority and portability of the model proposed in this paper.