A semi physical simulation system for fault prediction of wind turbines was designed. The electric signals of the system were used for the prediction. Based on MATLAB simulation and examples of real data processing, the working principle, hardware structure and software design method of the system were described. The feasibility and effectiveness of the Fast Fourier transform, wavelet transform and other traditional and new methods on nonstationary electric power data processing for fault prediction were studied, with wind turbines operating under the condition of uncontrollable natural winds. The results proved that the system was a significant method of studying fault prediction, while it was difficult to obtain the data of wind turbines.