Abstract: In order to improve the classification performance of Extreme Learning Machine (ELM) and retain its advantage of the training speed, after a detailed analysis of feature level fusion and decision level fusion, fusion of ELM is proposed. To implement decision level fusion ELM, Probabilistic ELM (PELM) is proposed, which transforms the numeric outputs of ELM to the probabilistic outputs and unifies the outputs in a fixed range. On this basis, an adaptive weighted feature fusion method is introduced, which considers fully the difference accuracy rates of different features without the prior knowledge and subjective definition. Simulation experiments verify the correctness and the validity of the method, thus achieving a higher recognition rate compared to the Support Vector Machine (SVM) and ELM, and a good performace in terms of training time.