Abstract: In order to remedy the randomness and uncertainty in selection process, the parameters of the least squares support vector machines (LS-SVM) classifier are optimally selected by the Bayesian inference with three levels hierarchy which can significantly improves modeling efficiency. Combined with probability outputs of multiclass LS-SVMS, the Bayesian inference LS-SVM classification method is applied to diagnose the power transformer fault diagnosis. The experimental simulation results show that the proposed approach can identify faults successfully. Both the diagnosis accuracy and modeling efficiency are better than traditional LS-SVM method.