Signals Mutation Point Identification Based On Time-Frequency Analysis And Machine Learning Methods

With the rapid economic growth,the requirements for railroad transportation efficiency and transportation capacity are getting higher and higher.Insulation joints,as an important part of the track circuit,achieve electrical isolation of the track signals to ensure the normal operation of the railroad.Insulation joints can be divided into two types:mechanical insulation joints and electrical insulation joints.Currently,the most common type of mechanical insulating joints used in domestic high-speed railway stations are glued insulating joints,which are usually weaker than electrical insulating joints in terms of smoothness and can cause high-frequency vibration of the vehicle when severely damaged.To facilitate a comprehensive evaluation of the condition of the glued insulation joints,this paper needs to address two issues:insulation joint positioning and insulation joint classification.In view of this,this paper explores the data characteristics of the glued insulated joints and electrical insulated joints,and identifies the location of the two types of insulated joints and distinguishes the types of insulated joints based on the track circuit data.According to the characteristic properties of data signals,this paper establishes two types of insulation joint identification methods based on wavelet transform and support vector machine(SVM).To locate the insulated nodes,firstly,the data signal of the track circuit is normalized.Secondly,with the theory that the modal maxima of wavelet transform coefficients correspond to the signal mutation points,the db3 wavelet is selected to do wavelet transform on the signal and obtain the modal maxima of wavelet transform coefficients.After that,for the situation that the wavelet transform coefficients modal maxima include many interference points at the non-insulated joints,according to the signal characteristics at the insulated joints,the interference points are filtered out by setting thresholds and other measures,and only the modal maxima with the largest value at the insulated joints are retained to achieve accurate positioning of the insulated joints.Finally,we compare and analyze the localization effect of different wavelets and the localization effect of other methods.Aiming at the classification problem of insulation joints,firstly,according to the positioning point of the insulation section and the length information of the insulation joint,the track circuit data values corresponding to the position point of the insulation joint and the effective feature points on both sides are selected,divided into a single sample,and the appropriate feature parameters are extracted to construct the feature set.Finally,the input feature set is put into classifiers such as RF,XGBoost,BP neural network,SVM,etc.for classification detection experiments.The results show that the S VM-based modeling method obtains an accuracy of 99.42%,a precision rate of 100%,a recall rate of 95.24%,and an F1 score of 97.56%,which is better than other models and has a shorter average run time.Therefore,the SVM model can more effectively distinguish between the two types of insulation joints.