By means of fuzzy C-means (FCM) algorithm, this paper research the problem of distinguishing characteristic vectors of partial discharge (PD) ultra-high frequency (UHF) signals from different defects in oil-paper insulation. According to the internal insulation defects in transformer, this paper design 4 kinds of PD models characterizing typical defects of oil-paper insulation. The multi-scale wavelet packet grid dimensions and energy parameters making up the characteristic vectors are both extracted from UHF signals of PD models. So this paper get comprehensive characteristic recognition matrixes, cluster data and recognize defects from it. Using fuzzy C-means algorithm, the two matrixes are clustered and recognized respectively with and without the wavelet de-noising. Both the clustering results and characteristics show that it is available to distinguish the difference between PD models characterized by the wavelet packet multi-scale UHF grid dimensions and energy parameters; Wavelet de-noising method could effectively enhance the correctness ratios, minimum ratios, recognition stability, stability of the algorithm and astringency; and verified the Fuzzy C-means algorithm applied to the insulation defect recognition.