Research On Slipping Detection Methods Of Electric Locomotive Based On Feature Extraction And Machine Learning

With the continuous advancement of socialist modernization with Chinese characteristics,heavy-haul rail transportation,as the backbone of China’s comprehensive transportation system,has also achieved rapid development.In the field of heavy-haul rail transportation,from the perspective of wheel-rail safety and economy,it is of great significance to detect the wheel-rail slipping of electric locomotives efficiently,quickly and accurately.Thus,based on the signal-based slipping detection methods,this thesis studies the electric locomotive slipping detection method based on feature extraction and machine learning.By studying the signal feature extraction method and machine learning-based classification/clustering algorithms and corresponding optimization algorithms for locomotive wheelset velocity data.And in-depth exploration and optimization are carried out from the three perspectives of high precision technology realization,industrial real-time application and unsupervised adaptive application.And then three types of slipping detection methods,namely off-line high precision,real-time online high accuracy and unsupervised self-adaptation,are designed and proposed successively.Firstly,aiming at the accuracy problem of the existing methods and taking the large amount of running data stored by locomotives into account,this thesis focuses on improving the accuracy of slipping detection,and takes the analysis of the large amount of data as the guide,and proposes an accurate slipping detection method based on fuzzy entropy and support vector machine.By using the proposed feature extraction method based on empirical wavelet transform and fuzzy entropy to extract the features of the speed signal of locomotive wheelset velocity signals,and inputting features into the support vector machine model for feature classification to achieve high-precision slipping detection.At the same time,the genetic particle swarm algorithm is proposed to solve the superparameter selection problem of the support vector machine and optimize the model.The experimental results show that this method has excellent slipping recognition accuracy.Secondly,in order to meet the feasibility and real-time requirements of engineering practice,and solve the problem that the precise slipping detection method based on fuzzy entropy and support vector machine can identify slipping with high precision but cannot be applied to millisecond real-time detection.This thesis focuses on improving the speed of feature extraction and slipping detection,and is guided by data-driven and maintaining highprecision detection,and further constructs and proposes an online slipping detection method based on detrended analysis and kernel extreme learning machine.A multi-scale detrended fluctuation analysis algorithm is constructed by combining the idea of multi-scale analysis and detrended fluctuation analysis to realize fast high-dimensional feature extraction of locomotive wheelset velocity signals,and the kernel extreme learning machine model optimized by differential evolution algorithm is used to quickly classify features for fast and accurate slipping detection.Experimental results show that this method has excellent performance in feature extraction speed and real-time slipping detection while maintaining high precision,and can be effectively applied to real-time slipping detection in engineering practice.Finally,considering that the existing methods and the proposed slipping detection methods are based on supervised learning methods,and model training requires a large number of labeled samples,this thesis further studies the unsupervised slipping detection method based on autoencoder and cluster analysis,oriented by adopting unsupervised learning algorithm to save the sample labeling time.By using the proposed deep sparse contractive autoencoder network to mine the deep feature of locomotive data,and then using K-means clustering algorithm to cluster and divide the features of slipping and adhesion data to achieve unsupervised slipping detection.Experimental results show that the proposed method not only maintains relatively good detection accuracy,but also has the convenience of omitting sample labeling and the excellent modeling and recognition speed.