Fault Diagnosis Of Phase-shifted Full-bridge Converter In Locomotive Control System Power Supply

With the rapid of the development of China’s railway industry, Its kinds of the technical indicators have reached the world leader, But the requirements of the safety and reliability for railway are gradually increased. Now, according to the literatures indicate that the main fault of the train is from the failure of computer systems. However the locomotive control power supply has been seen as a core component of the locomotive control, and it’s safe and reliable operation directly impact on the normal operation of the computer system. Given the importance and the special nature of the locomotive control power, and it’s circuit structure for the process of running, The fault of the locomotive control power should diagnosed reasonably, effectively and rapidly in order to reduce the dangerous from the locomotive control power.In the thesis conduct research for the core circuit of the locomotive control power phase-shifted full-bridge converter. Firstly introduce the structure and the development of locomotive control power and the reasons for failure. According to the phase-shifted full-bridge converter circuit structure and it’s works, and the parameters for the locomotive control power, simulate the circuit. Because in the actual operation process the circuit will inevitably produce a lot of noise and harmonic voltage. So in order to meet the actual situation, need the noise and harmonics in the simulation. Then use the diagnostic methods for "energy feature for wavelet packet and neural networks" to simulation test. Pointed out that traditional wavelet packet transform in fault diagnosis have two problems. First, if the wavelet packet decomposition level for fault signal is much less, then the error message can not be fully presented. Secondly if the decomposition level is too large, the feature dimension is much more.Ultimately it affects the training time and the diagnosis accuracy as well as the speed of the BP neural network. Based on the above two problems, present new solutions. Firstly by the db3 wavelet packet, analysis the fault signal with multilayer at low and high frequency. Extract eigenvalue to completely find fault information. Due to the number of implied information nodes at the bottom of wavelet packet decomposition tree. With the increase of the number of decomposition levels, that will be in the form of the exponential 2n. If these feature vectors set as the input of neural network, which will inevitably lead to a large number of neurons, complex and difficultly to converge, leading to "curse of dimensionality".Therefore, the use of the Laplacian Eigenmap of manifold learning algorithms reduce dimensionality for data. According the standard of manifold learning algorithm for enabling high-dimensional data dimensionality reduction to "visualize", so reduce the dimensionality of the data to three-dimensional. Effectively solve the problem for the "curse of dimensionality". However, excessively high dimensional data dimensionality reduction will lost seriously some faults, affecting the accuracy of fault diagnosis.According to this shortcoming for excessive dimensionality reduction, improve the laplace feature mapping algorithm. Propose based on the Mahalanobis distance Laplace mapping algorithm, Solve neighborhood problems to make k of the LE algorithm adaptive. By studying the correlation dimension of the fractal theory, obtain estimates of high-dimensional feature vectors intrinsic fault dimension, solve d selected difficulty in the manifold learning algorithm, to achieve exactly high dimensional data dimensionality reduction purposes.Finally, verify the validity and accuracy of the proposed method for the diagnosis through the Matlab/Simulink software.