Research On The High-Speed Train Running Gear Fault Diagnosis With Mutual Information And RBF Neural Network

High-speed train is a complex nonlinear system. With the development of high-speed train, the requirements of high-speed train fault diagnosis system tend to be more and more rigid. Not only high accuracy but also good real-time performances demanded by the fault diagnosis system of high-speed train are stricter. In addition, higher request of classifier design came up with mass characteristics data of high-speed trains’diagnosing. Two aspects of high-speed train running gears fault diagnosis method were studied in this thesis. One is the classifier design and the other is feature dimension reduction. The simulation using the experimental data was accomplished as well.At the beginning of classifier design part, four kinds of running state of high-speed train, experimental data source, and several commonly used feature extraction methods were introduced in detail. Then introduction of the high-speed train running gears diagnosis model with RBF network based on gradient descent were referred. On this basis, PSO algorithm was used to adjust and optimize the parameters of RBF network. To a certain extent, this way can avoid the RBF network parameters training into local optimum thus improving the precision of neural network classification. By combining PSO algorithm and chaos local search (CLS), the local search ability of PSO algorithm were improved. Then, a new heuristic intelligent algorithm, namely of the Bat algorithm (BA), were used to optimize parameters of RBF network for further improvement of the classification accuracy. With high-speed train running gears experimental data, the simulation confirms the application of BA RBF network for high-speed train running gears fault diagnosis can increase the recognition speed of fault detection and improve the recognition accuracy.Althghou the performance of the diagnosis system were improved, the accuracy of high speed train running gears fault diagnosis system using single sensor is not good enough, result in misjudgment between similar fault states. Fault diagnosis method based on the multi-sensor fusion was used in this thesis. Fault characteristics were collected from more than one sensors. Then UFS-MI, short for a unsupervised feature selection method based on mutual information, is used to eliminate redundancy and interference between the fault features. Experimental data simulation using the selected fault characteristics confirms that UFS-MI and BA-RBFNN can improve the diagnosis speed and classification accuracy successfully.