Online sequential extreme learning machine learns samples one by one or chunk by chunk, so it is appropriate to analyze online process data and to detect system faults. To improve the detection accuracy and rapidity, a new online extreme learning machine algorithm with varying weights and decision level fusion has been proposed, which increases the weights of the new samples predicted wrongly by current data monitoring model in learning, and introduced decision level fusion to improve integrated decision-making ability of the model. Performance comparisons of the method are presented using UCI datasets and Tennessee Eastman process. The results show that the proposed algorithm produces comparable or better performance with higher accuracies and lower training time.