| The internal damage of the rail can cause the rail to break and cause serious accidents.Therefore,timely detection of internal cracks in the rail is of great significance for railway maintenance.At present,the rail flaw detection vehicles used at the railway site at home and abroad are all detected during the skylight time,and there is a blind spot in the bottom corner of the rail,which cannot realize the online monitoring of the full cross-section of the rail.Ultrasonic guided waves have low attenuation and are suitable for long-distance detection.However,due to the relatively complex geometric structure of the rail,the inherent multi-modality and dispersion of ultrasonic guided waves are more significant,and different areas of the rail have different sensitivities to different modes,It is impossible to use a single mode for monitoring.Based on the analysis of the propagation characteristics of ultrasonic guided waves in the rail,this thesis studies the rail crack signal extraction method and the crack location algorithm in the case of modal aliasing,and provides theoretical support for realizing the online monitoring of the full section of the rail.Aiming at the problem of rail crack signal extraction,a crack signal extraction method suitable for complex environments is proposed.Since guided waves will generate echoes when they encounter structural changes such as cracks,the received echo signals include complex and changeable environmental noise,structural acoustic noise and other interference signals and echo signals caused by cracks.This thesis proposes a method based on The methods of singular value decomposition and independent component analysis have realized the effective extraction of crack echo signals,laying the foundation for the precise location of cracks in the next step.Aiming at the problem of identifying the regions where rail cracks occur(rail head,rail waist,rail bottom),the extracted crack echo signals are used as input,and traditional machine learning and deep learning networks are used for classification and recognition respectively.When using traditional classifiers,the effects of cracks in different regions on different receiving points of the same cross-section are used as features,and multiple classifiers such as SVM and Random Forests are used for recognition,and the recognition accuracy of more than 98% is achieved on the simulation data set;When using the deep learning network,on the basis of CNN,one-dimensional and two-dimensional CNN are used to supervise and learn the crack echo signals.After comparison,the optimal network parameters are determined,and the best rail crack characteristics are obtained.The accuracy rate is over 99%.Aiming at the problem of distance location of rail cracks,this thesis determines to use the first wave packet to achieve accurate location of cracks along the longitudinal direction of the rail.Due to the different propagation characteristics of ultrasonic guided waves at different positions of the rail section,the crack echo velocity in different areas is different,so the rail cross section is refined into three parts: rail head,rail waist,and rail bottom;After preprocessing the crack echo signals such as time-frequency analysis,filtering,and Hilbert transform,comprehensively compare the positioning methods of inflection point,stagnation point,and extreme point to realize the fitting of the first wave packet velocity in different areas.And in the case of simulation,comparing different excitation methods,verifying on rail models with different distances of cracks,the results show that the relative error of crack location is within 3%,which meets the requirements of practical applications.Finally,experiments were carried out on 3.5m short rails and 125 m long rails in the laboratory to verify the feasibility of the above crack signal extraction and location identification methods,and provide a theoretical basis for the subsequent development of long-distance,rail full-section online monitoring equipment. |
PDF Full Text Request