Fault Diagnosis Of High-speed Railway Switch Based On Semi-supervised Incremental Learning

As an important railway line connection device,the switch has a complicated structure and a poor outdoor environment,so it frequently fails.At present,the working state of the switch is mainly judged by the workers according to the electrical characteristic curve collected by the microcomputer monitoring system when the switch is in operation.In order to avoid the misjudgment and missed judgment of the switch fault caused by the lack of staff experience,this thesis takes the switch driven by ZYJ7 electro-hydraulic switch machine as the research object,and takes the switch action current curve as the research data,combined with semi-supervise learning and incremental learning,studies the fault diagnosis method of the switch,which can realize the initial diagnosis model based on a small amount of labeled data,and expand the training sample of the model by combining the unlabeled data for incremental training.The research on fault diagnosis of switch mainly involves the following problems: first,the current curve is converted into the data representation which can be processed by the diagnosis model,so that the input data of the diagnosis model can represent the data characteristics of different types of current curves.The second is to use a small amount of labeled data combined with unlabeled data to build an efficient diagnostic model,so that the model can quickly give higher accuracy diagnostic results.The third is to dynamically increase the training samples of the diagnosis model,so that the generalization ability of the diagnostic model is further improved.In this thesis,based on the principle of switch control circuit and the principle of monitoring and collecting,the types of switching current curves studied in this thesis are summarized.According to the wrench process of switch,the current curve is segmented in three ways,and the statistical characteristics of the current data in each segment are taken as the data characteristics of the current curve.Secondly,the K-Means algorithm is used to detect other fault category data that may exist in the unlabeled data set,so that the unlabeled data set is consistent with the data category of the labeled data set.In this thesis,the Naive Bayesian classification model is used to distinguish the switch fault.By using semi-supervised discriminant analysis(SDA)algorithm,the original data feature is reduced in dimension,the redundant information in the feature is eliminated.When the new feature is used as the input of the model,the model can improve the classification performance of test data.On this basis,the model is incrementally trained with unlabeled data to correct the parameters in the model and further improve the generalization ability of the modelAccording to the diagnostic model,the thesis finally developed the switch fault diagnosis system.The system realized the functions of data communication,incremental training of diagnostic models,fault diagnosis and alarm,historical fault case query and diagnostic model information display,and combined with specific data to operate the whole system.The debugging of the process indicates that the diagnostic model studied in this thesis has certain applicability.