Finding out the partial discharge in time is an important means to avoid fault， improving the quality of transformer and ensuring safe operation of power grid. In order to identify defect type of transformer insulation pressboard quickly for removal of fault， the partial discharge experiments of the three kinds of insulation defects common in transformer insulation pressboard was carried out in laboratory. Results show that the partial discharge of the three samples has unique characteristics. The partial discharge of flawless pressboard specimens occurs mostly at the phases of 0 to 120 degrees and 180 degrees to 300 degrees， while the air gap defect samples are discharged at the phase of 60 to 160 degrees， 240 degrees to 330 degrees， and the metal particle defect samples are pulsed at 80 to 160 degrees and 260 degrees to 340 degrees. On certain condition， it is need to extracted the text data secondarily， for some partial discharge experimenter can only export picture data. Based on the experimental PRPD（Phase Resolved Partial Discharge） graph， a data set was built， with 80% as a training set， 20% as a test set. A convolutional neural networks (CNN) was adopted to analyse the data set， compared with K-Nearest Neighbors(KNN)， Support Vector Machine(SVM) and Error Back Propagation Training. The convolutional neural network， which is constructed and optimized， obtains the correct rate of about 96.5% and 89.9% respectively in the training set and the test set

showing that convolutional neural networks are suitable for local discharge recognition based on PRPD images.