Abstract: In view of the disadvantages of the traditional time-varying parameters modeling algorithm about nonstationary random vibration signal of a spacecraft with closed spaced modal frequency, a multicomponent process neural network (PNN) autoregressive model was proposed, which was based on the empirical mode decomposition (EMD). The EMD was utilized to decompose the original time series into several local orthogonal intrinsic mode functions (IMF) with different time scale. The PNN was established for anyone of these IMF, and obtained time-varying power spectrum density (PSD). The time-varying PSD of the original signal was reconstituted by superposing. The simulation and example analyzed results suggest that this method avoids the correlative estimation calculation and reduces the calculation complexity. There is no cross-term interference, so it improves the time-frequency distributing characteristic of the signals. This method has a higher ability of extracting signal characteristic and can be used to analyze the vibration signals of spacecraft under various work conditions.