Research On Rail Defect Detection Based On Graph Spectrum

Cracks on the rail surface will cause serious defects such as rail head nuclear injury and peeling,which will seriously threaten the safe operation of the railway.Therefore,timely and effective detection of rail surface cracks is an important guarantee for railway operation safety.This topic focuses on improving the accuracy of rail crack recognition,and researches on feature extraction,feature optimization,and multi-sensor decision fusion in rail crack recognition based on Magnetic Flux Leakage(MFL)signals.The main contents include:Considering the consistency between the MFL signal of the rail crack and the structure of the path graph,the MFL signal of the rail crack was modeled by the Graph Signal Processing(GSP)technology to obtain the MFL graph signal of the rail crack.From the time domain and frequency domain perspectives of the MFL graph signal,the features of the MFL graph signal represented by the Laplacian matrix eigenvalues,pseudo-Laplacian energy,frequency mean and frequency center of gravity are extracted.In order to solve the problem of the validity of the MFL graph features of rail cracks extracted by using different adjacency matrix parameters during the modeling of MFL signals.Based on the KNN classifier’s recognition rate of rail cracks,the optimal adjacency matrix parameters for modeling MFL signals of different channels in different directions were determined.At the same time,the relationship between the features of the MFL graph signal of rail cracks extracted from the optimal adjacency matrix parameters and the parameters of rail cracks was analyzed.Aiming at the redundant problem between the extracted rail crack MFL graph signal features,an SFS + KNN based optimization algorithm for rail crack MFL signal graph features was designed.The algorithm takes the recognition rate of rail cracks by the KNN classifier as the feature evaluation index.Each time a feature is selected from the original feature set so that the combination with the selected feature has the greatest recognition accuracy to form an optimal feature subset.The experimental results show that compared with the use of unoptimized features,the use of optimized features for rail crack identification has a better effect.At the same time,the adaptability of the optimized features in the SVM classifier was verified through experiments.In order to make full use of the information diversity in MFL signals in different directions in different channels,a new multi-sensor rail crack recognition algorithm based on entropy weight DS fusion is proposed.The algorithm calculates the information entropy for each SVM classifier to determine the classifier weight,adjusts the initial BPA,and then uses the classic DS fusion rule to calculate the final BPA of the test sample.The experimental results show that compared with the singlechannel and unidirectional MFL signals,the multi-sensor rail crack recognition algorithm based on entropy weight DS fusion has a higher recognition rate,reaching 99.49%.