Research On Operation State Identification Method Of Metro Vehicle Bogie Frame Under Different Speeds

Bogie frame is the supporting part of bogie,which can maintain the structural form of bogie and is an important part of bogie.The research object of this paper is the bogie frame of a certain type of subway vehicle.It explores the recognition method of the operation state of the bogie frame of different subway vehicle speeds and different vehicle types.The main steps of state recognition include feature extraction and pattern recognition.A method of bogie frame operation state recognition based on time-frequency map texture feature and two-dimensional convolutional neural networks(2d-cnn)migration learning is proposed.Using the time-frequency analysis method of wavelet transform,the time-frequency map of wavelet transform is drawn,and the energy texture feature of the time-frequency map is taken as the image sensitive feature;Input it into the squeezenet model of 2d-cnn migration learning pattern recognition method to realize the state recognition of the running state of bogie frame.This paper analyzes the bogie and its structure of metro vehicles,and takes the bogie frame of metro vehicles(motor car and trailer)at different speeds(40km / h and 90 km / h)as the research object,that is,four kinds of bogie frame operation states: motor car 40 km / h,motor car 90 km / h,trailer 40 km / h and trailer 90 km / h.Design the test scheme of bogie frame vibration signal acquisition.Firstly,select the acceleration sensor and install it near the end where the frame vibration is the most intense,then use the data collector to collect the data collected by the sensor,and finally screen and store the vibration data at the signal processing PC.For the processed vibration data,conventional sensitive feature extraction methods such as time domain analysis and time-frequency domain analysis are used to analyze and process the vibration data.The time domain analysis method includes sensitive characteristic indexes such as peak to peak,mean,variance,kurtosis and root mean square.After analyzing these indexes,the characteristic indexes of variance,kurtosis and root mean square are selected.Principal component analysis(PCA)is used to realize time-frequency domain feature extraction,wavelet packet decomposition of vibration signal and calculation of energy percentage feature;PCA is used to analyze the energy characteristics of wavelet packet decomposition,so as to achieve the goal of time-frequency domain feature reduction and feature extraction.Based on the analysis of the shortcomings of time domain analysis method and time-frequency domain analysis method,the time-frequency analysis method of wavelet transform is used to draw the time-frequency diagram of wavelet transform based on the time-frequency analysis method;The energy parameter index of texture feature is analyzed,and the energy texture feature is taken as the image sensitive feature.The 1d-cnn pattern recognition method is used to realize the recognition of the running state of the bogie frame.The energy features of wavelet packet decomposition extracted by PCA analysis,joint time-domain features such as variance,kurtosis and root mean square;Input into BP(back propagation)neural network pattern recognition method for frame state recognition,that is,PCA analysis wavelet packet BP(add time-domain feature)state recognition method;Experiments are carried out to verify the feasibility of this method.Compare the state recognition methods such as PCA analysis wavelet packet KNN,PCA analysis wavelet packet BP and PCA analysis wavelet packet KNN(add time-domain features);The advanced method of 1D-CNN state recognition method is verified by comparing the above methods through experiments.Based on the comprehensive analysis of the advantages and disadvantages of the above feature extraction methods and pattern recognition methods,combined with the intuitiveness of time-frequency image texture features and the application requirements of image recognition technology,a model recognition method based on the combination of time-frequency image texture features and 2d-cnn transfer learning is proposed.The 2d-cnn transfer learning method has two application models,googlenet and squeezenet.The central idea of googlenet model is to approximately optimize the structure of the recognition model with dense structure modules;Squeezenet model compression optimizes the feature extraction layer structure of Google net model.Experiments verify the efficiency of squeezenet model in identifying the running state of bogie frame and the superiority of the proposed method.