In order to improve the generalization ability of a soft-sensor model, a compositional model of SVM based on AdaBoosting algorithm is proposed. On the basis of Bayesian analysis, the penalty coefficient is initialized by using the Bayesian probability of the samples, and then the penalty weight is updated by the loss function in the regression process so that the SVM training model can highlight some important samples to improve its estimation accuracy and generalization ability. Simulation result shows that this approach can greatly improve the estimation capacity and generalization ability of the model.