Catenary Inspection Image Anomaly Detection Based On Semi Supervised CNN

With China becoming the most rapidly developing and technologically comprehensive country in the world's high speed railway,there are many kinds of problems in the electrified railway.The stability and safety of traction power supply system is not only related to the safety of electrified railway,but also to the efficiency and cost of operation and maintenance.Catenary safety status monitoring as a important issue in the present study,for the video and big data environment of automation and intelligent condition monitoring is an urgent and important research areas,as an important early step in the recognition and maintain work,which has an important significance to study the anomaly detection of critical areas of the parts of catenary.The research work of this paper is depended on the technical specification of the catenary security inspection device in the 6C system specification.This algorithm with C2 camera shooting device of catenary pole number data and catenary insulator as foundation,through image processing techniques and methods of deep learning anomalies of two kinds of data in the related research,the validity of the proposed method is verified by experiment.The main work of this paper is as follows:1?Aiming at the catenary image data obtained from continuous frames,through the sample selection principle determine the normal and abnormal class samples of the data set to be detected,intercept the data containing the objects to be detected,and produce related data sets.In view of the fact that there is no negative sample in the application of insulator data,the negative sample of insulator is simulated by image simulation software to get certain data.Through enhance the two kinds of data,the amount of training data is increased effectively,and the data base is prepared for the follow-up experiment.2?The abnormal detection of the serial number of the Catenary Based on HOG features is studied.By analyzing the HOG algorithm and the feature extraction principle,we designed different parameters and feature dimensions,and found the best feature dimension suitable for catenary pillar number,and served as the input of SVDD classifier later.In order to find the best combination of parameters,design a method of parameter optimization in SVDD algorithm,and through grid search optimize the parameters.Finally,a variety of evaluation indexes are applied to evaluate the experimental results.Through experimental surface: the method of parameter adjustment can find the best parameter range and parameter combination reasonably and quickly,and has a high accuracy rate for anomaly detection of catenary pillar number.3?An anomaly detection method based on CNN transfer learning and SVDD is proposed.In this paper,we analyze the non-balance of data and the sample size of data,and the model migration of Lenet-5 network is realized by using the idea of semi supervised learning.According to the two classification problem of anomaly detection in the actual need of the improvement of the traditional Lenet-5 network model,the convolution kernel size and number of layers in the fine-tuning of volume and the impact on the number of different convolution kernel fully connected layer on the accuracy of the experimental training,the feature dimension for feature contact net image data were determined.As the input of the classifier.The experiment proves the rationality of the network;finally through the SVDD training and testing,the anomaly detection of catenary pillar number image accuracy is more than 97%,the method has certain engineering value.