The traditional feedforward neural networks has a slow training speed and a low generalization performance, which may produce the local minimum and can not handle the change flow of information. Unlike these conventional implement ations, a fast fault diagnosis method is proposed based on online sequential extreme learning machine. As complexity and lack of fault samples of rotating machinery，the testing data obtained in the forecast process are added to the training samples, which is known as information during the next forecast process to establish the online sequential extreme learning machine classifier model, and it enhances the precision of fault diagnosis at the largest extent. The experimental results show that, it is easier to select the parameters and the learning rate increased 200 times when online sequential extreme learning machine has similar classification accuracy compared with support vector machine(SVM).