Crosslinked polyethylene (XLPE) insulated power cable has a wide application in flexible DC transmission areas. PD measurement is considered as an effective tool to detect and even identify insulation defects for XLPE cable. In this paper, four kinds of insulation defects were constructed, i.e. inner semi-conductive layer breakage, internal air cavity defect, insulation surface scratch defect and outer semi-conductive layer creepage, and PD aging experiments for each type of defect were conducted at DC voltage. It was found that PD characters significantly varied with different aging phases, leading to the fluctuation of PD fingerprint parameters. Without taking the time-sequence feature of PD data into consideration, the recognition rate was only 72.93% in the usage of fingerprint parameters. For the aim to improve recognition effect, a model based on bidirectional recurrent neural networks (BRNN) algorithm was proposed. By dividing the PD process into several stages, the model takes both the fingerprint parameter and the stage information as input, in which the output of each stage is not only related to itself, but also influenced by the preceding and subsequent inputs. Therefore, the model can reflect the time-sequence characteristic of PD. With the dataset acquired by four insulation defects testing, the recognition rate of the BRNN model is 93.71%. It is proved that the BRNN-based recognition model effectively eliminate the influence of insulation aging on the PD fingerprint of XLPE cable, and improve the identification efficiency to a certain extent.