On-line robust modeling of nonlinear systems using support vector regression

To improve robustness of support vector regression (SVR) in nonlinear systems on-line modeling, the relationship between outliers and the robustness of SVR is derived mathematically, and a new modeling method using SVR is proposed. The relationship indicates that the effect of outliers to SVR is decided by the training data distribution and the distance between outliers and the support vectors nearest to them. Therefore, in the method, each component of the training data is normalized into the same range, and then the components representing the system output are compressed differently to change the training data distribution to reduce the effects of the outliers. Meanwhile, a data updating criterion is presented to eliminate outliers. The method is applied to multichannel electrohydraulic force servo synchronous loading system to predict the load output, and the results show its effectiveness.