The signal of online partial discharge（PD）detection of XLPE cable may come from the cable

its end joint

or the switchgear connected with it. PD from different source does different harm to electric equipment

and is recognized with different criteria. Therefore

it is necessary to recognize the source of PD. Based on wavelet packet decomposition

four eigenvectors

i.e. energy spectrum

Shannon entropy

logarithmic energy entropy

and 1.5-order norm entropy

are extracted in this paper. These eigenvectors are taken as the input of BP neural network to recognize PDs. The conclusions are as follows: 1)the recognition rates of BP neural network with respective four eigenvectors are all around 90%; 2)for surface discharge

the recognition rate with energy spectrum

Shannon entropy or logarithmic energy entropy is relatively low

while the recognition rate with 1.5-order norm entropy is high

however it is relatively low for cable PD. An improved method is proposed by taking the combined eigenvector of energy spectrum and 1.5-order norm entropy as the input of BP neural network

thus the recognition rate reaches 97%.