Based on the traditional method of analytical redundancy fault diagnosis, the advanced machine learning technology is combined with the model-based fault diagnosis so as to form a new intelligent approach to the fault diagnosis for satellite control systems. The support vector regression technique in statistical learning theory is employed to model the control system with a little sampling data firstly. Then the feasibility of detecting and identifying faults for the satellite attitude control system with the Mahalanobis distance is analyzed in detail. Finally a set of fault-detection observers are designed and implemented based on the residual evaluation. The simulation result indicates that the diagnosing method proposed in this paper is characterized with light computation burden and good real-time performance.