The extrinsic Fabry-Perot interferometer (EFPI) optical fiber ultrasonic sensor can be used for the detection and pattern recognition of the partial discharge ultrasonic signal inside the gas-insulated switchgear (GIS). Compared with the traditional piezoelectric sensor, it has many advantages such as high sensitivity and strong anti-interference ability. Based on this, four typical partial discharge models of tip, metal particles, suspension and quay are set in the GIS cavity filled with 0.4 MPa SF₆ gas. The EFPI sensor is used to detect the discharge ultrasonic signal. The waveform characteristics of a single ultrasonic pulse signal are extracted to form a characteristic parameter database, and the probabilistic neural network (PNN) algorithm and the support vector machine (SVM) algorithm are respectively used for pattern recognition. The recognition results of the two algorithms are compared and analyzed. The ultrasonic signals detected by the EFPI sensor have outstanding features. Based on the extraction of feature parameters, the two pattern recognition algorithms can achieve an average recognition rate of over 85%, and the recognition rate of SVM is higher than that of PNN.