Research On Abnormal State Detection And Classification Method For High-Speed Rails Based On Generative Adversarial Networks

The achievements of our country’s high-speed railway construction have attracted worldwide attention,and the high-speed railway has been transformed from a large-scale construction to a long-term safe operation stage.With the growth of the operation scale and the continuous improvement of the speed,new diseases continue to emerge,which affects the smooth and comfortable operation of the train and poses a serious threat to the operation safety.The railway department conducts regular inspections on the service status of the track infrastructure through high-tech means such as rail inspection vehicles and comprehensive inspection vehicles,which ensures the safe operation of high-speed railways and accumulates a large amount of track dynamic detection data.At present,the mining and analysis of track dynamic inspection data is still in its infancy.The large scale and diversity of track dynamic inspection data,unclear and complex relationships between data attributes,noise and anomalies in the detection data,and lack of semantic annotations have caused great obstacles to the effective use of data.In this paper,the deep learning method is used to carry out anomaly analysis and research on track dynamic inspection data.The main work is as follows:(1)Aiming at the problem of efficient detection of anomaly data in track dynamic inspection data sets,a multi-dimensional data correlation anomaly detection algorithm based on generative adversarial networks is proposed.Starting from mileage attention and inter-attribute attention,the mileage information and data information are extracted through the graph attention mechanism,and a reconstruction model based on generative adversarial network and a fully connected network prediction model are respectively constructed.The dynamic time warping algorithm is used to calculate the reconstruction and prediction scores,and the anomaly scores are obtained by multiplicative combination for anomaly judgment.Experimental results on real datasets show that the F1 score of our method outperforms other anomaly detection methods.(2)In allusion to the problem of accurate classification and matching of track dynamic anomaly inspection data when the labeled anomaly sample size is small,this paper proposes a semi-supervised orbit geometric anomaly state classification algorithm based on graph attention mechanism.Relevant information is extracted from the two perspectives of mileage and attributes through graph attention mechanism,long-term temporal dependencies are extracted using gated neural network GRU,a fully connected layer-based classification model is constructed,and the model is trained in a semi-supervised manner.The comparative experimental results show that all the indicators of the model in this paper are better than the existing time series classification models.(3)In view of the problems of automatic display,anomaly detection and anomaly classification of massive high-speed railway dynamic inspection data,this paper uses the currently popular Spring boot and React frameworks to build a high-speed railway dynamic inspection data anomaly analysis system based on MVC architecture.The system mainly integrates four parts: system management,data visualization,anomaly detection and anomaly classification.On the one hand,the performance of the algorithm is actually tested,and on the other hand,it provides a sample for the development of open functions of high-speed railway related systems.