Using Nguyen-Widrow method initializes the variable parameters of the L-M algorithm, based on the application of Levenberg-Marquardt algorithm in multilayer feed-forward neural network training, reduce the number of the adjustment of the threshold value and the weight When the network training, speed up the convergence speed of BP algorithm, which ensures the global minimum of the error of plane, and apply the N-W method combined with L-M algorithm for fault diagnosis of steam turbine. The theory and practice has proved that: this method has improved the accuracy and reliability of fault diagnosis, which provides effective method for the turbo generator fault diagnosis.