Vibration diagnosis method based on wavelet analysis and neural network for turbine-generator

Abstract

The turbine-generator plays a crucial rule in modern industrial plant. The risk of turbine-generator set failure can be remarkably reduced if normal service condition can be arranged in advance. An effective approach based on wavelet neural network is presented for vibration signal analysis and fault diagnosis. The wavelet transform exhibits not only more comprehensive results, but also delivers a variety of possible explanations to the investigated problem. The main advantage of wavelet transform for signal analysis is that the wavelet coefficients are obtained by correlating vibration signal with the wavelet basis functions so that all possible fault patterns can be displayed by time-scale results. The feature vector obtained from wavelet transform coefficients are presented as input vector for neural network. The improved training algorithm is used to fulfill network training process and parameter initialization. From the output values of the neural network, the fault pattern is identified in accordance with the predefined fault feature vectors, which are obtained from practical experience. At the meantime, the convergence property of wavelet network for fault diagnosis is discussed. The experiment results demonstrate that the proposed method is effective and accurate.