Research On Registration And Anomaly Detection Algorithm Of Track Dynamic Inspection Data Of High-speed Railway

In recent years,China’s high-speed rail construction has developed vigorously,and the state detection and evaluation technology of high-speed rail infrastructure has also been advancing with the times.Dynamic inspection data obtained by the comprehensive track inspection train regularly inspecting the track line status is the basic data for evaluating the service status of the track.Based on the machine learning theory,discover the track state transition law and the internal correlation relationship between the inspection data from the large-scale dynamic inspection data,establish an effective track state anomaly detection model,and provide auxiliary decision-making for high-speed rail line state evaluation and line maintenance.It has important practical significance for the quality assessment and maintenance of high-speed railway tracks,and also has important research value in the field of data mining technology.In this paper,the data mining research is carried out on the dynamic inspection data of high-speed railway.First,the registration study is carried out on the mileage offset that is common in the track dynamic inspection data,and then the dynamic inspection data anomaly detection research is carried out for the typical abnormal data of the dynamic inspection data,and finally combined with the actual engineering application requirements developed a comprehensive application system for dynamic inspection data.The main work and research contents of this paper include:(1)Aiming at the ubiquitous mileage offset and uncertainty problems in the motion inspection data detected at different time periods,a sliding window dynamic time warping mileage registration algorithm based on curve feature points(CFP-SW-DTW)is proposed.The algorithm first extracts the characteristic points of the track curve by means of ultrahigh linear fitting,and rough registration according to the extracted characteristic points,reduces the mileage deviation within the error range of 22 m,and then uses the sliding window dynamic time warping mileage algorithm(SW-DTW),the algorithm performs fine registration,which reduces the singularity phenomenon in sequence registration.This paper constructs a simulation data set of movement inspection data mileage migration to carry out a quantitative registration experiment analysis of the proposed algorithm,and verifies the performance advantages of this algorithm in TAM indicators and time consumption on the real dynamic inspection data set,which is a follow-up data mining research provides the basis for registration data.(2)Aiming at the problem of multiple types of abnormal data in track dynamic detection data,which is distributed uncertainly,this paper proposes an unsupervised anomaly detection multivariate time series based on graph attention network(GATUSAD).The graph attention network in the model extracts the data feature relationship of the dynamic detection data,uses an adversarial trained autoencoder to construct the reconstruction error between the reconstruction result and the original input,and judges the abnormal data through the reconstruction error.This paper constructs a simulation data set of typical abnormalities in dynamic detection data.Experiments on this data set verify the advantage of data feature relationship extraction and points’ relationship in neighborhood window extraction in GAT-USAD.Compared with the baseline network USAD model,GAT-USAD achieves an increase of about 1.2% in the comprehensive performance index F1 of motion detection data anomaly detection.(3)Combining the actual needs in the field of dynamic inspection data research and engineering application,this paper designs and develops the high-speed track dynamic inspection data comprehensive application system.The system integrates functional modules such as dynamic inspection data retrieval,data visualization,data preprocessing,mileage registration,and anomaly detection.It has the characteristics of convenient management and intelligent analysis,and provides scientific researchers and engineering technicians engaged in dynamic inspection data mining and research with a professional data management and analysis system.