SVM handles classification problem only considering samples themselves and the classification effect depends on the characteristics of the training samples but not the current information of classified problem. From the phenomena of data crossing in systems, this paper improves the classification effect of SVM by adding the prior probability item reflecting the classified problem information into the decision function, which fuses the Bayesian criterion into SVM. The detailed deducing process and realizing steps of the algorithm are put forward. It is verified through two instances. The results show that the new algorithm has better effect than the traditional SVM algorithm, and its robustness and sensitivity are all improved.