Towards an general unknown actuator fault of the satellite attitude control system, this paper propose a new method based on SVM observer which use two LS-SVM regression models to identify the unknown fault and the nonlinear term respectively, because the training datasets of an unknown fault cannot be acquired before it occurred, we combine offline training with online incremental learning method to reduce the approximation error. In the last part, this method is applied to detect a kind of actuator fault, the simulation shows that the method proposed by this paper is effective.