Abstract: Partial discharge (PD) is the main cause of insulation degradation in electrical equipment. Compared with traditional detection methods, fiber optic acoustic emission detection technology has advantages such as not affecting equipment operation mode and strong resistance to electromagnetic interference. This paper designed and produced four typical partial discharge models: tip discharge, bubble discharge, suspended discharge, and surface discharge. A partial discharge acoustic emission signal acquisition experimental platform based on fiber optic acoustic emission was built, and the partial discharge acoustic emission signals under gradient applied voltage were collected and analyzed for parameter statistics. A recognition model was proposed that took the denoised signal time-frequency map as the feature input of a convolutional neural network. The results showed that the recognition model had good recognition performance for weak partial discharge signals submerged in noise, with a recognition accuracy of 98%.