Fault Diagnosis For Switch Based On Improved Convolutional Neural Network

In the past ten years,the construction of the Chinese high-speed railway has made a qualitative leap.The railway network is moving from "four vertical and four horizontal networks" to "eight vertical and eight horizontal networks".Since the end of 2018,the total length of Chinese railways has exceeded to 131,000 kilometers,of which the total length of high-speed railways is 29,000 kilometers,accounting for 66%of the world's total mileage.The rapid development of high-speed railway not only shortens the travel time but also promotes the coordinated development of the regional economy.As an important branch of the railway transportation system,signal systems are responsible for the train operation and maintenance.The switch is the key equipment for the railway to make sure the safety of train operation,especially in high-speed trains.The switch machine is a basic component which is used to change the traveling track during the train operation.At present,the analysis of the switch operation and the fault identification mainly relies on maintenance personnel.This kind of analysis method has poor real-time performance.Also,this method is both inefficient and prone to misjudgment,resulting in some accidents.As the speed of passenger trains goes up,the existing fault diagnosis methods for turnouts are not suitable for the development.More and more intelligent methods of turnout fault diagnosis are needed to help signal workers diagnose the causes of turnouts.In view of the above background,this thesis chooses ZYJ7 AC electro-hydraulic switch machine as the research object,selects the convolutional neural network as the research method.The aim of this paper is to classify the faults of the switch machine and realize the intelligentization of the fault diagnosis and prediction.The main findings of this thesis are as follows:Firstly,the composition and action process of ZYJ7 AC electro-hydraulic switch machine is introduced in detail,including the working principle and normal action current curves.Based on some current curves and expert experience from the on-site investigation of the computer monitoring data in the Lanzhou Electric Power Station Microcomputer Monitoring Center,the author summarizes five common failure modes.At the same time,the corresponding action current curve forms are summarized,and the causes of various failure modes are analyzed as well.Further,because current curves of the ZYJ7 switch machine are much complicated,the author proposes three methods to extract and select characteristic of current parameters based on the current curve sample data:(1)directly selecting the original curves of the action currents after processing simply,(2)selecting current parameters based on the different states,(3)selecting current parameters based on the different time periods.The extraction methods of current parameters are good foundations for the following research on the fault diagnosis method based on convolutional neural network.Afterward,some basic structures of the convolutional neural network are analyzed.The key points are the local connection and weight sharing characteristics.The author elaborates the network training process of forwarding conduction and backpropagation in the way of mathematical formula deduction,which is the basis of the improved convolutional neural network.Next,the core algorithm of adaptive enhanced convolutional neural network(AE-CNN)is elaborated,that is,an adaptive module is added between the forward conduction and backpropagation processes of the original convolutional neural network.AE-CNN can be used to analyze the original classification results and execute feature extracted.Based on the last iteration number,the accuracy of feature extraction and the recognition result,AE-CNN can adjust the feature error adaptively using automatic enhancement coefficients,in order to achieve adaptive enhancement of feature residuals.A model based on AE-CNN can be used for fault diagnosis on ZYJ7 switch machines.Finally,comparative experiments of three parameter extraction methods and two convolutional neural networks are carried out.The conclusions are as followed:(1)the first method of extracting current parameters has a good expression on switch fault characteristics,(2)the convergence rate and the stability of the adaptive enhanced convolutional neural network for fault diagnosis are better than that of the original convolutional neural network.Especially,the recognition rate is increased by about 5%.Since the generalization,convergence and recognition rates of the selected current curves are better,this method of using the improved convolutional neural network to diagnose faults of switch current curves can be further extended.