Improved dynamic recurrent-based ELM neural network for fault prediction

With the increase of the complexity of industrial system, the demand of real time of accuracy of fault prediction is higher. Therefore, an improved dynamic recurrent-based extreme learning machine (DR-ELM) neural network is proposed for fault prediction. For the network structure, a feedback layer is added to memorize the hidden output. And the trend characters of data variation is extracted from the feedback information, so as to update the output weight of feedback layer dynamically. Through the prediction of the next-time output for the nonlinear dynamic system, the diagnosis is made for the prediction output, then the fault prediction is realized. A numerical study(Sinc test) and the complicated Tennessee Eastman(TE) benchmark process show the superiority of the proposed approach not only in prediction accuracy but also in dynamic adaptability, and the better prediction ability for non-linear sequential systems.