Research On Track Circuit Fault Detection Based On Improved Deep Belief Network

In recent years,high-speed railroads have developed rapidly,and they have become a popular approach of travel,meeting the requirements of daily travel and operational efficiency.As an important part of the railroad system,track circuit may malfunction and cause unexpected accidents during operation because of its complex structure,the working environment which is susceptible to external influences.This will not only bring economic losses,but also threaten our life safety.At present,there is a single method of fault detection for track circuits in the railway field,which is equipment-specific and requires manual intervention;they rely heavily on manual experience,increasing the labor intensity of maintenance personnel,which may lead to a decrease in diagnostic efficiency and affect train operation;in addition,the large amount of fault data monitored by the microcomputer leads to long fault processing time and poor real-time performance.Therefore,the research on the fault detection method of track circuits has great application value.In this thesis,we propose the following research for the current status of rail circuit fault detection,based on the existing methods:（1）Taking the ZPW-2000 track circuit as the research object,this thesis analyzed the working principle of the track circuit system based on the equipment composition and functions of the main equipment;by referring to the technical specification and combining with the actual situation,the influences of"red light strip"and"bad shunting"faults on the train running state are summarized;divided the faults according to the transmitting channel,receiving channel and rail line,and summarized 14 common fault types,which laid the foundation for the subsequent research on the detection and extraction of fault characteristics of the track circuit.（2）A feature extraction method with improve intrinsic timescale decomposition and sample entropy was proposed.Firstly,the Intrinsic Timescale Decomposition（ITD）algorithm is improved to effectively improve the ITD signal decomposition endpoint effect and waveform easy distortion.The IITD method is used to decompose various types of signals,screen the Proper Rotation（PR）components containing the main feature information to simplify the data volume and shorten the data processing time,and obtain PR₁ and PR₂ as the key PR components.Next,the sample entropy value of each signal is calculated as the feature value to complete the feature extraction.Aiming at the problems of one-sidedness of single feature extraction,poor extraction effect,and difficulty in meeting the demand of high precision detection of track circuit faults,the multi-feature fusion based on entropy distance method is proposed.The energy entropy,sample entropy and permutation entropy of the training set samples are extracted for features,and the entropy center is calculated as the basis for subsequent classification;the extracted features are evaluated by inter-class divergence and intra-class divergence indexes;the entropy distance method is used for feature fusion to form clusters in the three-dimensional entropy space to complete the initial division of each fault type,while the features of three types of entropy values are extracted as the data for subsequent fault detection to solve the problem of inadequate feature extraction.（3）A track circuit fault detection method based on Deep Belief Network（DBN）is proposed to construct a deep belief network with the dimensionality of input samples and fault classes,while forward unsupervised layer-by-layer training and backward fine-tuning of network parameters are performed using Contrastive Divergence（CD）algorithm and Back Propagation（BP）algorithm.The Restricted Boltzmann Machine（RBM）is improved by using Gauss-Bernoulli Restricted Boltzmann Machine to improve the binary constraint problem of RBM nodes and effectively extract the distribution features in fault data.Multiple hidden layer and hidden layer nodes are set and the extracted feature values are input into the network for simulation analysis.The experimental results show that the best fault detection performance is achieved with 3 hidden layer layers and 12 hidden layer nodes,which determines the internal structure of the DBN model.Finally,the fused eigenvalues are input into the improved DBN model for fault detection,and the detection accuracy reaches 98.25%,which is improved in different degrees compared with the fault detection results of single energy entropy,sample entropy and permutation entropy.（4）For the problem that the network structure needs to be manually adjusted and the training efficiency is low during the training of DBN model,a fault detection algorithm based on improved GA-DBN is proposed based on the improved DBN model,and the number of hidden layer nodes is optimized using Genetic Algorithm（GA）to obtain the optimal combination as 3-17-21-12-14.The simulation experimental results show that the optimized deep belief network can achieve 98.96%fault recognition rate,and the average accuracy is improved by 8.86%,6.26%,5.56%and 0.71%compared with BP neural network,convolutional neural network,DBN and DBN not optimized by GA algorithm,respectively.Finally,the comprehensive performance of the model is evaluated using recall,average accuracy,F-Measure.It is also compared with the algorithm before improvement,and it is proved that the algorithm can achieve better detection results.In summary,this thesis combines the track circuit history data in the microcomputer monitoring system of a Railway Administration and uses multi-feature fusion for fault feature extraction.As well as the research on track circuit fault detection using improved GA-DBN model to quickly determine the type of track circuit faults based on the monitored data,it solves the problems of high operation and maintenance costs,high workload and low detection accuracy.Computerized detection instead of manual experience solves the disadvantage that centralized signal monitoring cannot make discriminations on the monitored data and fault detection still requires the judgment of technicians.It is conducive to rapid fault detection,can meet the actual needs of railroad sites,and has good application prospects.