Machine Learning Based Classification Of Adhesion Failures For Heavy Locomotive

Heavy-duty locomotives are the best way to transport bulk goods over long distances.With the development needs of the country's economic construction and the strategy for the development of the western region,“cargo heavy load” will become the new focus of China's railway construction after “high-speed passenger transportation”.The traction of heavy locomotive locomotives is determined by the adhesion between wheels and rails,which is usually characterized by stickiness.Locomotive operating conditions are complex.If a sticking fault occurs,the locomotive will cause idling/slipping and even serious traffic accidents.Based on the theory of machine learning,this dissertation studies the classification of adhesion failures of heavy-duty locomotives under conditions of dry,wet,rain and snow on heavy-duty locomotive simulation platforms.The main work is as follows:The parameters of cost-sensitive SVM were optimized by using adaptive mutation particle swarm optimization algorithm to solve the problem of unequal cost of misclassification between samples in adhesion fault classification.In the actual operation of the locomotive,the normal state data far exceeds the fault status data.In order to solve the data imbalance problem,cost-sensitive support vector machine modeling is used.An improved particle swarm optimization algorithm is proposed to optimize the penalty parameter and kernel function of cost-sensitive SVM,and the model is used to realize the classification of adhesion failure of heavy-duty locomotives.The simulation results show that the improved particle swarm optimization algorithm can improve the classification performance of cost-sensitive support vector machines.This model can accurately classify locomotive adhesion failures.This paper proposes an optimization method for the classification of adhesion states in nuclear extreme learning machines to solve the problems of the random initialization of connection weights and hidden layer neuron thresholds for extreme learning machine algorithms.Aiming at the problem of the random initialization of connection weights and hidden layer neuron thresholds by the Extreme Learning Machine(LLE)algorithm,an optimization method for the optimization of adhesion failure of nuclear extreme learning machines is proposed.An off-line parameter optimization model was established using the improved particle swarm algorithm.An online sticking fault classification model was established using a nuclear extreme learning machine.The classification of adhesion failures is accomplished jointly by offline parameter optimization and online adhesion state classification.In order to ensure the timeliness of the optimal parameters and reduce the computational complexity of the online classification,a periodic update framework for the parameters was designed.Comparison experiments with support vector machines and extreme learning machines show that this method has better classification performance.The feature extraction and classification methods based on deep learning were studied to solve the problem of difficulty in extracting the adhesive features in the noise environment of adhesion classification.The PreLU activation function is used to design a sparse noise reduction automatic coding network framework to realize network weight learning.With the lamination fusion strategy,the original pure signal of the original data is recovered,and the deep features of the initial sample data are obtained.And construct multi-classification support vector machine for its classification.Finally,the effectiveness of the method is verified by simulation experiments.