An on-line Least Square Support Vector Regression (LS-SVR) algorithm is given in this paper. Based on this algorithm the method of on-line fault prediction by temporal pattern estimation is proposed. The method needs neither the model to approximate the true system nor the fault training data and primary knowledge. It can study and predict while system’s running, and it is believed with fast speed, fewer amounts of calculation and better real-time capability. The result of simulation on CSTR proved the efficiency.