Abstract: When implementing health management for radar, it is of great significance to introduce the function of fault prediction. When modeling to carry out fault prediction by means of monitoring parameter time series which reflect health of radar system, prediction accuracy of a single model is limited. In order to improve prediction accuracy, models adapted to the failure mechanisms of radar should be considered. On the basis of the introduction of autoregressive model and radial basis function neural network model, an optimized combined prediction algorithm was promoted. Singular value decomposition filtering algorithm was employed to identify factors from different sources which influence radar performance and divide the observation sequence optimally. The simulation results show that compared with algorithm employing a single model and traditional combined prediction algorithm, the prediction accuracy index of the promoted algorithm is improved by at least an order of magnitude.