Research On Track Circuit Fault Diagnosis And Prediction For Predictive Maintenance

Track circuit is the fundamental infrastructure of China train control system.It plays an vital role in checking occupancy,ground-to-vehicle information transmission,and checking broken rail.Track circuits are one of the core technical equipment to ensure the safe and efficient operation of trains.With the gradual expansion of China’s railway network,the number of track circuits also increases,and the railway transportation has put forward higher and higher requirements for the reliability,availability,maintainability and safety of track circuits.In this context,it has become rigid demands to ensure track circuits are always in the safe operation state,to improve the quality of service and to reduce the cost of operation.The traditional maintenance of track circuit is mainly breakdown maintenance and scheduled maintenance,which leads to the problems of under-maintenance and over-maintenance,and can not meet the demands of railway development.In contrast,predictive maintenance is based on the operating status of equipment and can accurately locate and predict faults through mathematical models,which can greatly improve maintenance efficiency and reduce maintenance costs.It is an inevitable development trend of track circuit maintenance.A favorable condition for the transformation of track circuit maintenance to predictive maintenance is that the Railway Bureaus have widely used ground static monitoring and on-board dynamic monitoring to obtain important parameters that reflect the working state of track circuits.The two monitoring methods complement each other in terms of monitoring purpose and scope,and provide data support for the intelligent transformation of maintenance work.How to make full use of existing monitoring resources to realize more comprehensive state awareness,more accurate fault diagnosis and location,and more effective trend analysis and fault prediction of track circuits are the basis of formulating predictive maintenance scheme and also the urgent problem to be solved in current track circuit maintenance.Therefore,this paper systematically analyzes the characteristics of track circuit,fault type,monitoring means and monitoring data,constructs the predictive maintenance framework for track circuits based on onboard and ground monitoring data,formulates the fault processing process and functions.On this basis,the challenges and solutions in monitoring data processing,state detection,fault diagnosis and fault prediction are studied in depth.The main innovations of this dissertation are as follows:(1)Aiming at the problem that the on-board monitoring data(induced voltage curves)will not correctly reflect the working state of track surface equipment due to noise interference,a denoising method of track circuit signals based on denoising convolutional autoencoder is proposed.This study analyzes the problem that the full connection layer cannot characterize the high and low frequency signals at the same time,and proposes to replace the full connection layer with a convolutional layer with local connectivity,which improves the characterization ability of timevarying track circuit signals.Based on the ability of the autoencoder to learn the essential characteristics and structure of the signal,a time domain denoising model based on the denoising convolutional autoencoder is constructed.The experimental results show that this method can effectively reduce the impact of noise that can invade the working frequency band,and provide data basis for the subsequent diagnosis and prediction process based on the induced voltage curves.(2)Aiming at the problem of track circuit state detection under large-scale unlabeled data,an unsupervised anomaly detection method combining relative mass isolation forest and sparse autoencoder is proposed.In this method,a pseudo-label training strategy is formulated to solve the problem that the sparse autoencoder is prone to be biased by anomalies in the training process,which leads to a high error rate.Aiming at the problem that relative mass isolation forest cannot flexibly detect cluster anomalies,a nonlinear fusion strategy is developed.The combination method can integrate the advantages of two kinds of composition methods,overcome the shortcomings of a single detection method,effectively improve the recall rate and reduce the false positive rate.This method can separate the rare fault state from the massive normal state and reduce data volume for subsequent fault diagnosis and prediction,and improve maintenance efficiency.(3)Aiming at the problem that the current track circuit fault diagnosis methods are vulnerable to the influence of data imbalance and do not consider the correlation of monitoring and measurement,a hybrid resampling strategy is proposed to reduce the influence of unbalanced data,and a fault diagnosis model based on onedimensional convolution neural network and feature fusion method to learn spatial correlation is proposed,which can achieve the purpose of fault feature extraction and fault pattern recognition at the same time.The proposed two strategies improve the fault detection rate and comprehensive performance of the diagnosis algorithm.However,this method cannot solve the problem of accurate fault location of rail surface equipment,and as a supplement,a Matrix profile-based inductive voltage curve processing method is proposed,and different fault diagnosis strategies are designed for electrical separation joints and compensation capacitors.This method avoids the additional feature extraction step and directly calculates the similarity of the sequence data through the accelerated algorithm,so as to accurately detect anomalies on the induced voltage curve.It provides new ideas for further excavating the information contained in the induced voltage curve.(4)Aiming at the problem that the degree of fault deterioration is difficult to accurately characterize,a remaining useful life prediction method of track circuit based on spatio-temporal feature extraction is proposed.On the basis of the spatial features extracted by one-dimensional convolution,the long short-term memory neural network is introduced to perceive the changes of monitoring data in the time dimension,and the regression mapping relationship between monitoring data and remaining useful life is further established.Taking the capacitance degradation of 4 700μF electrolytic capacitors as an example,the simulation results show that the prediction method is accurate and stable,and the spatial and temporal characteristics play an important role in the prediction process.Similarly,in order to make up for the deficiency of ground monitoring data when facing rail surface equipment faults,a detection method for deterioration sequence of induced voltage curve based on time series chain is proposed to depict the deterioration trend of rail surface equipment.This method can support maintenance personnel to make qualitative judgment on the deterioration trend,and can also calculate the deterioration degree and speed quantitatively,which provides the basis for fault prediction.