Research On Fault Diagnosis For Hydro-turbine Generating Units Based On Support Vector Machines

With the fast development of modern industrialization, industrial equipments are improving with high precision, complication and highly integrated automation and the safety and stability of equipments is very important to enterprise production. Hydroturbine generating units (HGU) are the key equipments in the hydropower station, which health condition will influence the safety and economical benefits of the station. At present, the maintenance system of hydropower station is transferring to the new system—condtion based maintenance (CBM). It is useful for the station production and CBM system that the correct condition analysis and efficient fault mode classification of the HGU can be done with the condition monitoring in the station. According to the small sample problems in the HGU fault diagnois and the nonlinear and nonstability characteristics in the HGU vibration, the support vector machines (SVM) method is introduced in HGU fault diagnosis research in this thesis. The research with SVM on fault pattern classification, condition trend prediction and parameter selection in HGU fault diagnosis is done systematically.Firstly, the basic theory of SVM is introduced and the algorithms of classical SVM in classification and regression are analysed. Least squares support vector machines (LS-SVM) method, an extension version of classical SVM, is presented with the algorithms in classification and regression. Then the differences between SVM and LS-SVM are analysed. All of the algorithms and comparisons are the foundations for the application of LS-SVM in HGU fault diagnosis.Fault diagnosis can be regarded as pattern classification essentially. With the excellent performance of SVM in pattern classification, a vibration fault pattern classification model is proposed based on wavelet decomposition and LS-SVM. With the energy feature vectors extracted from the HGU vibration signals, the multi-fault pattern classification model is built on the LS-SVM, which validity is testified by samples.The asymptotic characteristics of equipment running condion are analysed. The feasibility of prediction of HGU condition is proposed in detail. The general prediction methods are summarized. It is indicated that the condition trend evolution of HGU is a process from quantitative change to qualitative change. The asymptotic characteristics of HGU provide convincing refernce for the prediction of HGU condition trend with prediction methods.According to the vibration signal series as the representation of HGU condtion, the phase space reconstruction, embedding dimensions and time lags problems are discussed. Estimation rules and functions of prediction models are summarized. Based on the feasibility analysis of HGU condition prediction, a hybrid prediction model based on wavelet decomposition and LS-SVM is proposed. In this hybrid model, vibration signal series are divided into serval subseries with obvious tendency characteristics through wavelet decomposition. Then the tendency of these subseries is forecated with LS-SVM respectively. Finally, these prediction outputs are summed up as the prediction results of original vibration signal series. Compared with other different prediction models, the hybrid model gains higher prediction accuracy and predicts the peak value change in the signal, which is useful to the HGU condition precaution.Parameter selection is an important part in SVM research; proper parameters of the SVM model only gain the perfect performance. With the full search ability of genetic algorithm (GA) in complex problems optimization, a parameter selection method of a LS-SVM prediction model based on GA is given, which is applied in HGU condition prediction research and provides a feasible and valid method for parameter selecetion of support vector machines based model in HGU fault diagnosis.Support vector machines are based on fully developed theory. But the theory studies and applications in fault diagnosis of support vector machines are still in developing. So the research of support vector machines in HGU fault diagnosis need to be further studied.