Points Fault Prognosis And Diagnosis Based On Neural Network

With the rapid development of Chinese railway in recent 10 years,intensity and density of railway transportation improve continually.Points devices whose performance is closely related to the safety and efficiency of railway transportation are basic component of railway signal system.Maintenance and repair work of points devices in railway site is completed by the staff through browsing the current and power curves of point machine recorded by signal centralized monitoring system.Failure is processed manually according to experience,which is inefficient.And this method carried out after failure happens,cannot provide guidance for routine maintenance and generate systematic maintenance and repair schedule.In view of the above questions,this thesis proposes points fault prognosis and diagnosis methods based on neural network.Focused on the research of points soft fault,points soft fault prognosis model based on PSO-ELM neural network algorithm is set up through establishing degradation performance index by studying normal non-failure data.Meanwhile,points fault diagnosis model based on ELM algorithm is set up by analyzing features of various failure data.In order to realize intelligence of points devices fault prognosis and diagnosis and provide theoretical support for site maintenance and repair work,the research of this thesis is as follows:(1)Failure mode analysis of points devices.Working process and monitoring theory of points is studied.Formation mechanism of points power curve is researched.From the perspective of points power curves,points failure sample database is established by analyzing historical data and summarizing points failure modes and their forming reasons.(2)Data processing and feature processing research.Feature extraction is achieved for points power data by using statistic method.Aimed at imbalance among various failure data,SMOTE algorithm is used for data alignment and feature dataset is composed.Feature dimensionality reduction is carried out by introducing Fisher linear discriminant analysis and t-SNE algorithm.Feature data which has appropriate dimensions is obtained as the input of the fault diagnosis model.(3)Points soft fault prognosis method research.In allusion to normal non-failure data,the relationship between degradation state and various failure modes is established by using feature processing methods.Degradation performance candidate indexes are built.Degradation performance index and its failure threshold which can represent degradation state obviously are set up by KPCA algorithm.At last,PSO-ELM model is set up to predict the changing tendency of degradation performance index.Compared with the prognosis results of BP,ELM and SVR models,it can be found that when the number of PSO-ELM model hidden-layer neurons are 30,it emerges optimal prognosis performance.(4)points fault diagnosis research.In allusion to failure data,points fault diagnosis is achieved by setting up t-SNE-ELM neural network model.Compared with the diagnosis results of PCA-ELM and t-SNE-SVM model,it can be found that when the input data is 5 dimensions,the recognition accuracy rate of t-SNE-ELM model reaches 95.8%.Degradation performance index prognosis of soft fault is conducted in this thesis based on normal data and fault location and recognition based on failure data by studying points normal and failure power data,so as to realize the health management of points devices and provide theoretical support for maintenance and repair work of railway site.Figure 52,Table 12,Reference 72.