In this paper, 7 different artificial partial discharge (PD) models of power transformer defect types are built in a high voltage laboratory, and then PD signals are recorded in the whole power frequency cycle to develop a preliminary PD knowledge base. A specific technique is used to extract single PD events from the recorded knowledge base. By application of wavelet transform, different decomposition levels of extracted PD signals are investigated. Feature extraction is performed by application of statistic moments on the autocorrelation function of PD signal decompositions. "Best" features are selected based on the F-test. To evaluate the performance of features, a naive Bayes classifier is applied to selected features. Results show that nearly 100 percent accuracy in PD discrimination is achieved using multivariate-multinomial distribution estimation of the feature space. A case study is carried out to show the contribution of the proposed approach. The combined PD detection-classification system proposed in this paper can serve as a complement to conventional PD monitoring systems.