Research On Anomaly Detection Algorithm Of Dynamic Inspection Data Based On Absolute Mileage Calibration

With the rapid development of high-speed railway construction in our country,the technical requirements for identifying and assessing the state of high-speed railway infrastructure are becoming more and more urgent.The dynamic inspection data is a special multi-dimensional time series data obtained by regularly detecting the condition of the track line by the comprehensive track inspection train,which can reflect the actual track service status in a timely and authentic manner.Based on the theory of data mining,the mileage dimension calibration of the large-scale dynamic inspection data and the construction of an effective track state anomaly detection model are significant for ensuring the safety of high-speed rail operations.At the same time,as a special time series data,dynamic inspection data also have important research value in the field of data mining.Based on the theoretical basis of machine learning,this paper studies the mileage dimension calibration method for the mileage deviation of the dynamic inspection data,studies the anomaly detection model of dynamic inspection data,and develops the mileage calibration and anomaly detection system of dynamic inspection data from the application requirements.The main work and research contents of this paper include:(1)Aiming at the problem of unevenly distributed mileage deviation along the mileage caused by internal and external factors in the process of comprehensive inspection train detection,this paper proposes an account-assisted dynamic inspection data absolute mileage calibration method.The method firstly uses the D-DTW(Derivative Dynamic Time Warping)algorithm to locate the feature points of the dynamic inspection data after sampling and automatically segment the curve based on the feature points of the account Cant’s curve.Then,perform high-precision calibration based on the shape-DTW(shape Dynamic Time Warping)algorithm on the curve segments of the account Cant’s data and the inspection data,and the inspection data curve feature points are located by combining the calibration path and the prior feature points of the account curve.The absolute mileage calibration of the dynamic inspection data is realized by calibrating the feature points of the inspection data curve to the mileage position of the feature points of the prior curve in the account.Experiments on the measured dynamic inspection data of a high-speed railway line show that the calibration method is better than the comparison method in terms of TAM(Time Alignment Measurement)value and calibration speed and the average error of the detected curve length is only 0.36%.In addition,the standard deviation of the gauge of the calibrated dynamic inspection data is reduced by 23% on average,and the Pearson correlation coefficient,DTW distance and mean absolute error of the Lprf data series are better than those before the calibration,which meets the engineering application requirements.(2)In view of the characteristics of dynamic inspection data with many types of anomalies,uncertain distribution and deficient proportion of anomalies,this paper proposes a method combining graph attention and graph bias learning on the basis of mileage calibration of dynamic inspection data.The dynamic inspection data anomaly detection model(GATDN,Graph Deviation Network based on Graph Attention Network)realizes the unsupervised anomaly detection of dynamic inspection data.The model captures attention scores in multiple dimensions,learns the interrelationships among the Top K features with the most significant correlation between data features through graph bias,and makes anomaly judgments based on graph bias scores.Experiments on the highspeed rail line real-time dynamic inspection data show that the recall rate and F1 score of this model are increased by 4.9% and 1.1%,respectively,compared with the comparison model.(3)According to the application requirements of dynamic inspection data mileage calibration and anomaly detection,a high-speed rail dynamic inspection data mileage calibration and anomaly detection system has been developed.This system can realize the mileage calibration function based on the calibration method proposed in this paper and the anomaly detection function of dynamic detection data based on GATDN.The system provides strong technical support for the intelligent processing and analysis of dynamic inspection data for relevant scientific research and engineering personnel in the industry.