Partial discharge is the main form of early insulation failure of electrical equipment. Pattern recognition of discharge type is of great significance for the estimation of equipment insulation performance. Considering that the extreme learning machine (ELM) method has the advantages of simple structure and fast training speed, yet the initial parameter selection is random and the algorithm is unstable. A pattern recognition method based on fusion ELM algorithm for partial discharge is proposed. Considering the different judgement precisions based on variable features, the adaptive weight assignment is used to achieve the decision-level fusion of the output. In this paper, four physic discharge models are designed to simulate typical partial discharge defects. Discharge signal waveform and phase-amplitude spectrum is collected by high-frequency current transformer method, sufficient samples of experiment data are obtained to extract time-frequency domain and statistical features for classification. The result shows that the fusion ELM algorithm is superior to the traditional ELM algorithm and BP neural network in the recognition accuracy and stability without sacrificing training speed.