Research On Track Disease Extraction And Recognition Technology Based On Distributed Acoustic Sensing System

Real-time monitoring of railway track diseases is of great significance for maintaining railway safety and safeguarding the national economy.Distributed Acoustic Sensing(DAS)system is widely used in structural monitoring,pipeline security,perimeter pipelines,and other fields due to its tremendous advantages of resistance to electromagnetic radiation,corrosion resistance,long life,and low power consumption.The application of DAS system in real-time monitoring of railway track diseases has a very broad application prospect.However,in the actual monitoring process of railway diseases by the DAS system,the interference of noise can affect the effective extraction of train vibration signals.The data collection of up and down train signals simultaneously requires separate monitoring,and the effective distinction and recognition between different track disease signals greatly restricts the application of the DAS system in the field of railway track disease monitoring.Aiming at the problems encountered by the DAS system in the field of railway track disease monitoring,thesis focuses on the research of track disease extraction and recognition technology for the DAS monitoring system based on existing trackside communication optical cables.Thesis proposes algorithms including preprocessing and denoising of train vibration signals,distinguishing between up and down trains,multidimensional fusion feature extraction,and track disease recognition,to achieve feature extraction and recognition of track disease signals.The main research work of this thesis includes:(1)A train signal preprocessing method based on time frequency spatial multi domain analysis is proposed.A fusion train signal denoising algorithm based on timefrequency feature filtering and spectral subtraction is used to denoise train vibration signals,improving the signal-to-noise ratio.The spatiotemporal response image is enhanced based on image binarization and morphological processing methods.The line detection algorithm based on Hough transform and coordinate mapping are used to distinguish the up and down directions of trains.(2)A track disease recognition algorithm based on multi-dimensional feature extraction and SVM classification recognition is proposed.The feature extraction section is a multi-dimensional fusion feature vector composed of four single time-frequency feature values and normalized wavelet energy feature vectors.In the classification and recognition part,a Bayesian optimized Support Vector Machine(SVM)model is constructed,and feature data set is constructed for training and testing,and compared with ANN and KNN models.The final test results show that the average recognition rate of the trained SVM model reaches 94%,and the sample recognition time is 1.9 ms.(3)A deep learning model is used to classify and recognize track disease signals,and a comparative study is conducted on the recognition effect and performance of the model.The CNN,VGG,and Res Net models were optimized for the data characteristics of railway track disease signals,and 1D-CNN,1D-VGG,and 1D-Res Net models with different convolution depths were constructed for testing.The impact of different convolution depths on model recognition performance was analyzed.The experimental results show that 1D-CNN3,1D-VGG16＿bn and 1D-Res Net34 models are the optimal models for the corresponding types of models,with recall rates of 93.0%,96.1%,and96.3%,respectively,and recognition times of 1.20 ms,4.45 ms,and 8.62 ms.