Adaptive ensemble of extreme learning machines and application to fault diagnosis

A modified extreme learning machine method with adaptive ensemble strategy is proposed to improve the classification accuracy, robustness and generalization performance on small sample of online sequential extreme learning machine. The classification accuracy and weight vector connecting the hidden layer and output layer of incremental learning phase is calculated based on the learning method. The input weight vector, the hidden layer biases and voting weights of the component neural network are adaptively updated. The validity of the method is verified with the UCI data sets and bearing fault data. One hundred trials show that the proposed method can obtain an increase of 1% on the classification accuracy and a decline from 0.1% to 1.2% on standard deviation. Finally, the application of the diesel engine fault diagnosis based on the combination of wavelet package and the proposed method shows that the correct ratio is 91.16%.