In this paper, a new method for vibration fault diagnosis of large steam turbine-generation is proposed based on theory of adaptive wavelet network(AWN). A special adaptive wavelet network classifier for arbitrary nonlinear system is developed, algorithm of back propagation type is proposed. By means of computer simulation and with applying to fault diagnosis, and the performance of the AWN are compared with that of the conventional multilayer perception. The results show that this method has better diagnostic, faster speed of learning properties and it is an efficient new method. And then some important related elements, which affect the performance of classification AWN is analyzed, conclusions that are of great importance in real application, are obtained.