For the low signal to noise ratio, nonlinear and nonstationary ch ar acteristics of engine vibration signals, a modeling and forecasting technique is put forward by integrating the empirical mode decomposition(EMD) denoising, t h e theory of phase space reconstruction and support vector machine(SVM). Firstly, the vibration signals were decomposed by using the EMD method and the intrinsic mode function (IMF) components related to main interference factors were elimin ated. Then, the true vibration signal was obtained by reconstructing the remaini ng IMF components, and the timeseries embedding space about the denoised vibra t ion signals were rebuilt after the embedding dimension and time delay had been d etermined based on the theory of chaos. Finally, the modeling prediction was car ried out by SVM and compared with radialbasisfunction (RBF) neural network. Th e experimental results show that the proposed method can predict the trend of en gine vibration signals and its performance surpasses the traditional techniques and has better project practicability.