PD pattern recognition is an important part of the insulation state evaluation of GIS. In order to identify PD signals accurately and efficiently, a new method of PD signal pattern recognition based on dual attention mechanism is proposed to optimize CNN. Firstly, the GIS PD test platform is built, and four typical defects are set up in GIS air room. The PD signals of different defects are collected by UHF and ultrasonic detection respectively. Then, the data preprocessing is carried out based on the data characteristics of the two, and the image features of UHF PD spectrum and the gram angle field density distribution group of ultrasonic signal are constructed Finally, the input image is extracted by convolutional neural network optimized by double attention mechanism, and the result is predicted by softmax classifier at the end of the network. The results show that the fusion algorithm can achieve about 97.57% recognition accuracy, higher than single feature recognition rate, and the convolutional neural network optimized by double attention mechanism is superior to common algorithm in recognition rate, training speed and robustness.