The performance of Support vector machine (SVM) is affected mainly by model parameters, but there is no special method for the model parameter selection of SVM based incremental learning. A new method is proposed in this paper, i.e. taking robustness as criteria for performance evaluation of incremented learning, the range of solution space is designed by fitting error and scale factor, then the gradient descent algorithm is used to search the parameters. The experiments with this new method are made on the Logistic model regressing and aero engine vibration monitoring, and the comparison of this new method with the genetic algorithm and the gradient descent algorithm is made. The result indicates that the use of the proposed method can take full advantage of the results of historical learning, thus the solution space is narrowed, and iteration steps are reduced.