For a class of uncertain systems based on nonlinear TS fuzzy model, the steady state error and dynamic quality of sliding mode control algorithm are related to the accuracy of the description model of TS fuzzy algorithm. Using the least squares support vector machine (LSSVM) learning algorithm of TS fuzzy model can approximate the actual system well. However, due to the LSSVM algorithm has a certain amount of data requirements, and learning speed is relatively slow. The requirements of the algorithm is not suitable for higher dynamic response or less memory. In this paper, an improved conjugate gradient online learning algorithm is proposed to study TS model, which can approach the actual model in real time, and the algorithm can achieve the asymptotic stability of the control system. In this paper, the control algorithm is simulated under different error conditions. In the random error is less than 100, the steadystate error is 0.01, and the system showed stable characteristics of fast with time varying error, which showed the robustness of the control algorithm is strong. Finally, the experimental results whose random error amplitude is 500 show that the TS model has the ability to de noise the random error of the system.