Optimal Input Design for Dynamic Model Prediction Accuracy

Abstract

A new optimality criterion for the optimal input design for discrete dynamic nonlinear system identification is presented. The criterion weights the parameter covariances by the magnitude of regressors in order to reduce the prediction error. An iterative optimization procedure to the optimal signal design is proposed. The effectiveness of the optimal test signal design is demonstrated by the system identification and validation of a nonlinear multiple-input single-output (MISO) automotive engine fuelling model. The output-error simulation accuracy of the resulting model is compared with those of models identified by means of commonly used non-optimal test signals and optimal signals designed by alternative optimality criteria. The proposed test signal design method is shown to provide superior outcomes in output prediction fit and also to allow the application of input and output constraints as required for experimental industrial applications.