Study On Void Identification Method Of Ballastless Track Mortar Layer Based On Convolution Neural Network

CA mortar layer is the elastic adjustment layer of CRTSI、CRTSII type ballastless track.Its state directly affects the durability of ballastless track and the comfort and safety of train running.The damage of the mortar layer has the characteristics of concealment which can not be observed directly.Therefore,it is very important for early detection and treatment to study the hidden damage nondestructive testing technology of CA mortar layer.The traditional diagnosis method can not achieve deep feature extraction and the accuracy of elastic wave detection is limited.This paper focuses on the feature extraction method,feature signal processing method and classification recognition model of CA mortar layer elastic wave detection signals.Based on Hilbert Huang transform and deep learning,a defect identification model of CA mortar layer,HHT-CNN model,is established.The main research results are as follows:(1)The time domain,frequency domain and Hilbert Huang(HHT)algorithm are used to extract the feature of different defect condition signals.According to the characteristics of time domain,amplitude and spectrum of response waveform,the feasibility and limitation of feature extraction based on time and frequency analysis are discussed,and some effective identification parameters are proposed.At the same time,according to the advantages of Hilbert spectrum in instantaneous frequency time energy extraction,a method of defect feature extraction of echo signal based on Hilbert Huang transform is proposed.(2)Considering the influence of image preprocessing on the performance of deep learning network.In this paper,the HHT spectral image is processed by clipping,graying,compression and other normalization,and the processed data is labeled to establish the data set.According to the requirement of data set,the software of signal preprocessing for nondestructive testing is designed based on MATLAB,which realizes the fast and accurate extraction of signal characteristics.(3)Feature image input convolution neural network completes signal classification.In this paper,DE-CNN-1 model and DS-CNN-2 model are trained according to the needs of void recognition to judge whether or not and the size of void respectively.The experimental results show that the recognition accuracy is more than 98%.The CA mortar layer defect detection model named HHT-CNN model designed in this paper has a good defect recognition ability,and the automatic recognition accuracy of defects is high.