Detection of partial discharge by UHF signal has become an ordinary operation and maintenance detection method for High Voltage GIS, transformers, and other equipment. However, this method indirectly detects PD intensity by measuring electromagnetic waves, so the accuracy of PD pattern recognition technology based on this method is not high, and has not been well solved. This paper introduces the cut-off frequency of TEm1 mode as the division basis of the time-frequency distribution of the PD UHF signal. The PD UHF signal excited by a single insulation defect is used for signal preprocessing, time-frequency distribution segmentation, feature parameter extraction and selection, PSO-ELM recognition, and other operations. Finally, the PD signal type can be accurately determined and the identification accuracy of insulation defect types in GIS can be improved. This segmentation method enhances the spatial distribution information of image features, and it is of great practical value to improve the accuracy of the UHF method in PD identification.