Hydropower Generating Unit's Fault Diagnosis Based On Improved SVM And Information Fusion

With high-speeded growth of Chinese economy, demand for electricity increased and requirement for reliable electriciy became stronger. In the past twenty years, water resource and hydropower was exploited widely and capacity of hydro-electricity played a more and more important role in capacity of power system. Development of manufacture technology promoted hugemazation of Hydropower Generating Unit (HGU). Security and reliability of HGU became one of the most attentional objects in research and application of hydropower science.HGU is a complex nonlinear system with strong coupling between partions. It is difficult to comment condition of HGU with single and linear method. By information fusion, all kinds of HGU's information with different distruction and from different resource are integrated. And Support Vector Machine (SVM) with kernel results HGU's nonlinear problem with mapping input data into feature space. It is valuable for HGU's fault diagnosis and condition-based mentanence to apply these methods.Fault diagnosis methods are generally in the condition of stabilization. Vibration signal in starting process is so complex that it is not used to fault diagnosis in starting process. In fact, more fault information could be got in starting process. The idea of fault ditection and analysis in starting process is expounded and the method of fault dignosis by multi-feature information fusion is proposed in this thesis. HGU starting process includes rising-speed phase, rising-excitation phase and rising-load phase, which meets condition of HGU vibration experiments (altering-speed experiment, altering-excitation experiment and altering-load experiment). Frenquency feature, temporal feature and spatial feature of vibration could be gained in HGU starting process. It could raise veracity of fault diagnosis with multi-feature information fusion.Shaft orbit is the result of HGU running information fusion in time domain of data level. It is an open line. Linear moment invarint is a kind of graphics feature of shaft orbit and could reflect condition of running HGU. However, discrete linear moment invariant is change in zoom. Improved linear moment invariant is presented in this thesis. Experiment result shows improved linear moments are invariant in movement, rotation and zoom. In order to map sample from input space to feature sapce effectively, geometry of RBF kernel used in SVM is analyzed and data dependent method based on Riemannian geometry is used in this thesis. Results of analysis and experiments show delecting redandent support vectors is one reason for raising fault diagnosis speed and veracity of SVM. In addition, results of analysis and experiments demonstrate SVM has strong generalization capability and learning capability from little sample set, which meets condition of HGU fault diagnosis with a little prior knownledge. A fault diagnosis example exhibit improved SVM could raise speed and veracity of HGU fault diagnosis.For the problem of combination explosion in inforemation fusion, exchange theorem and conjunction theorem in combination process amonge evidences are researched in this thesis. Results of analysis on calculate complex of combine among evidences show information fusion by D-S theory could be reduced with exchange theorem and conjunction theorem. Results of experiments demonstrate these theorems. In addition, information fusion is applied in HGU fault diagnosis. By fusing frequency featrure information, temporal feature information and spatial feature information, HGU is diagnosed with lower uncertainty and higher reliability.At last, a HGU condition monitoring and fault diagnosis system is designed. Hardware and software design of each module is finished based on general design of the system. At the same time, Fault Diagnosis Expert Subsystem and SVM Assistant Decision Subsystem are built. The system could satisfy many requirements of HGU condition monitoring and fault diagnosis.