Research On Rail Crack Detection Method Based On CNN And LSTM

In recent decades,China's high-speed railways have developed rapidly,forming a huge railway network.High-speed railways promote the rapid development of China's economy and becomes the main mode of travel.Since the railway has experienced several speed increases,the safety of railway operation has received extensive attention.Safeguarding the safe operation of trains is the most important issue in China's railway transportation.Traditional detection methods suc h as ultrasonic and eddy currents are slow,occupy railway lines and have low detection efficiency,and have not kept pace with the development of railways.An efficient,timely method of rail crack detection must be studied.As an emerging non-destructive testing method,acoustic emission detection compensates for the shortcomings of traditional detection methods.However,the existing acoustic emission signal detection method for rail cracks is based on artificial extraction features,and the feature selection is empirical and the accuracy is low.In this paper,the following work is carried out for the research on rail crack detection method based on CNN and LSTM:Firstly,the collected sample of the rail crack acoustic emission signal is pre-processed to construct a sample set.The difference between rail crack signal and noise is analyzed based on time domain,frequency domain and time-frequency domain method,and features are extracted for classification.Secondly,in order to excavate the hidden information in the rail crack signal,try to construct different convolution network models for crack signal detection,and construct the rail crack detection model based on one-dimensional convolution and two-dimensional convolution respectively.Experiments show that the detection effect of the one-dimensional convolutional network model is better than that of the two-dimensional convolutional network model,and the network parameters related to the one-dimensional convolutional network model are optimized.Then,considering the time series information of the rail crack signal,a rail crack detection model based on the Bi-directional Long-Short Term Memory Network is constructed.Then,Experiments were performed to compare the performance with Recurrent Neural Network and Long Short Term Memory Network in rail crack detection.Finally,the rail crack detection model of Bi-directional Long-Short Term Memory Network is optimized.Rail crack detection is performed by using one-dimensional convolution extraction feature vector as the network input data.Aiming at the problem of crack acquisition and no labeling,the existing model is improved,and the Autoencoder structure and K-Means clustering algorithm are introduced to realize the rail crack detection from the noise.Using the encoder structure to extract the latent layer features of noise,perform K-Means algorithm clustering,and determine whether the signal is a crack signal by setting the threshold to realize the detection of rail cracks.