Research On Fault Diagnosis Of Turbogenerator Set Based On Decomposition-Ensemble Method

With the development of automation and increasing capacity of turbogenerator set, higher requirements have been put forward to high speed, fully loaded, continuous state and reliable operation of the equipments as well as the on-line monitoring and fault diagnosis technology of large steam turbo-generator unit. Hilbert-Huang transform and neural network ensemble technology provide effective methods for the fault diagnosis of steam turbine under increasingly complex conditions.The following studies are carried out around some key technical issues of Hilbert-Huang transform and neural network ensemble learning:(1) A new method is proposed according to the shortage of traditional steam turbine fault diagnosis methods. Firstly, the characteristics of the steam turbine vibration signals are analysed using the EMD adaptive frequency demultiplication method combining with spectrum analysis. Secondly, composition of the vibration signal is obtained based on the physical significance of the IMF component. Finally, the characteristic frequency of fault is obtained by drawing the time-frequency chart and then the fault type can be diagnosed by the characteristic frequency.(2) Feature extraction technology based on kernal method is used for fault diagnosis of turbogenerator set. The basic principle of kernal method is described in detail. It shows that kernal feature extraction technique is effective in order to further improve the accuracy of fault diagnosis of turbogenerator set. According to the characteristics of steam turbine fault data, focusing on the kernel principal component analysis and kernel independent component analysis, two effective non-linear kernal feature extraction methods are introduced and combined with practical problems.(3) A new method based on kernel principal components analysis (KPCA) and neural network ensemble using kernel fuzzy c-means clustering (KFCM) is proposed. The fault data is first analyzed using KPCA to extract main features from high dimension patterns. Not only is the diagnosing efficiency improved but also the diagnosing accuracy is ensured. Consequently, the KFCM algorithm is used to classify the individual neural network training independent. Lastly, the individual neural network whose generalization error is minimum in its category will be selected and the predictions of the component networks are combined through majority voting. Experiments show that the proposed approach has higher accuracy and stability, compared with other methods. (4) A new method for fault diagnosis of turbogenerator set based on kernel independent component analysis (KICA) and dynamic selective neural network ensemble (DSNNE) is proposed. The KICA algorithm is used to eliminate the correlation of high-order data and remove redundant features. Thus the classifiers obtained using this method have very strong generalization ability. Then, a dynamic selective neural network ensemble method is used(DSNNE). The DSNNE method can dynamically select the corresponding neural network ensemble which tits the corresponding testing sample best from all neural networks at any time. That is to say, the concept of "concrete analysis of concrete problems" is realized by the method proposed.(5) A method of fault diagnosis of turbo-generator set based on optimized neural network ensemble is proposed. Two aspects including improving the accuracy of neural networks and reducing the degree of correlation between the member networks are considered. Besides, a method of combineing the Bagging and Boosting alogorithm and dynamically determineing the weights of individual neural networks is given. Turbo-generator fault classification experimental results show the feasibility of the method.