Abstract: In order to accurately predict the feature parameters of a hydraulic pump, a new algorithm called sequential regularized extreme learning machine (SRELM) is proposed and a prediction method based on SRELM is studied. On the basis of structural risk minimization theory, SRELM balances the empirical risk and structural risk to enhance the generalization performance of conventional extreme learning machine (ELM). In comparison with the regularized extreme learning machine (RELM), SRELM can complete the training procedure recursively without retraining when there are sequential training samples. Thus, SRELM is suitable for on-line feature parameter prediction. In the SRELM-based prediction method, feature parameters of the hydraulic pump are used to train an SRELM model. The latest feature parameter is adopted iteratively to update the prediction model and then the trained prediction model is used to predict future feature parameters. Experiments on the hydraulic pump feature parameter prediction indicate that the SRELM-based prediction method has better performance in prediction accuracy and computational cost in comparison with conventional neural-network-based prediction and support-vector-machine-based prediction.

Key words: neural networks, regularized extreme learning machine, feature parameter prediction, time series analysis, condition-based maintenance