Raw on-line monitoring PD data for transformers consist of true PD signals as well as several kinds of interference signals too. Among the interference signals, pulse type interference signals are the most difficult to eliminate. In this paper, we propose a set of practical pulse type interference signals elimination method for an on-line UHF PD monitoring system of a 500 kV three-phase split type power transformer. That is noise sensor method and inter-phase antitheses method to exclude external pulse type interferences. The characteristic parameters from the pre-processed data in single cycle are extracted. The data are then classified with fuzzy clustering and grey clustering method to make further discrimination of all kinds of coincident pulse type interferences with true PD signals. At last, the classified PD signals are statistically analyzed with PRPD method, the evaluation on PD activities are then made based on the clustered PD data.