Study On Application Of SVM In Malfunction Classification Of Traction Converter

As power conversion device, traction converter is one of the key parts in locomotive.Traction converter is continuously on work during the process of train operation, whichwill result in faults inevitably. Also, traction converter is of high energy density, and thesemiconductor devices are relatively ‘weak’, making traction converter be thehigh-frequency-faulted part in traction system. Traction converter’s failure will directlydislocate railway services and even cause casualties. As a result, malfunction diagnosis fortraction converter is crucial to meet the demand for maintenance and guarantee the safeoperation of the locomotive.The state-of-the-art in malfunction diagnosis is firstly reviewed in this thesis. SupportVector Machine (SVM), a data-based malfunction diagnosis approach, is recommendedfor traction converter’s malfunction diagnosis.There are two main factors affecting the classification performance of SVM: kernelfunction and parameters optimization. Radial Bias Function is selected to be the SVM’skernel function after comparing the impact of different kinds of kernel functions used inSVM for binary classification performance; grid-search based on cross-validation isrecommended being SVM’s parameter optimization algorithm after comparing the impactof different kinds of parameter optimization algorithms used in SVM formulti-classification performance. C-SVM with selected kernel function and parameteroptimization algorithm is utilized for traction converter’s malfunction classification, andalso its classification result is compared with that of using least square SVM (LS-SVM).Study and simulationresult shows that, it’s entirely effective to apply SVM with RBFkernel function and cross-validation based grid-search optimization to traction converter’smalfunction classification, and its generalization performance is much better than that ofLS-SVM under relatively great sample data. This paper establishes the foundation for further study of malfunction diagnosis technology for traction converter, and also beinstructive and meaningful to further research of SVM method.