Kernel Method-Based Fault Diagnosis For ACS600Frequency Converter IGBT

With the development of both science and technology and human progress, a variety of large-scale electrical equipment is showing a trend of complication and intelligence. Although humans have created endless wealth through automated systems, the security issue has always been inevitable. Once large-scale systems produce accidents, they will bring in serious loss to both human and finance, even affecting the stable development of the country. Therefore, fault diagnosis for automated systems has always been considered a research focus and priority.In this thesis, IGBT (Insulated-Gate Bipolar Transistor) device in ACS600frequency converter is the goal of research. With the data collected in rolling process of different conditions, the types of fault were specified into two kinds, on the one hand, IGBT was damaged by the overcurrent in the boot process, on the other hand, IGBT was damaged by the overcurrent of sudden load. The specific method for IGBT fault diagnosis was chosen as follows:In terms of fault detection and feature extaction, it compared the pros and cons of PC A (Principal Component Analysis) and KPCA (Kernel Principal Component Analysis). Through simulation analysis, it shows the advantage of KPCA. In terms of fault identification, based on the characteristic data extracted through PCA and KPCA respectively, it trained three SVM, and then compared the correct classification rate of one against one SVM (Support Vector Machine) method and DAG (Directed Acyclic Graph) SVM method, which were both multi-class classification methods. Through simulation analysis, it shows the advantage of classification performance for DAGSVM.On the basis of KPCA and DAGSVM, it focused on improved methods to improve the classification performance of IGBT fault classifier. In view of the pivotal position of SVM kernel function and its parameters, which influence classification performance, firstly, Quantum Particle Swarm Optimization strategy was chosen for the optimization of kernel function parameters. A lot of optimization simulations for the selection of the contraction-expansion factor show that the convergence of optimization algorithm is fast and the correct classification rate is high when the interval [1.0,0.5] of the factor is chosen. Secondly, in terms of the choice of kernel function, it selected different kernel functions for three SVM in DAGSVM. Through simulation analysis, it chooses one group of kernel functions to achieve high classification correct rate. At last, in terms of the SVM trained with the data containing two types of fault, it did some simulation analysis with mixed kernel function, which consisted of polynomial kernel function and RBF kernel function. The result shows that the classification performance is enhanced with mixed kernel function.