EXPERIMENTAL-DESIGN CONSIDERATIONS FOR DYNAMIC-SYSTEMS

Although experimental designs have been extensively applied for steady -state modeling of process behavior, their use in studying dynamic sys tems has been more limited. Optimal designs were primarily focused on minimizing uncertainty in the transfer function parameter estimates, u ntil Ljung pointed out the importance of minimizing bias in the assume d form of the transfer function. This paper reviews existing methodolo gies for designing input test signals and draws comparisons among them through the use of simulation examples. Results indicate that input d esigns which emphasize relevant frequencies for the controller applica tion give superior performance when either an adequate process transfe r function model has been specified or bias in its form exists. In the latter case, minimization of the frequency domain bias is found to be the most important factor in ensuring reliable controller performance . Distributional results for the closed-loop performance criterion pro vide important insights into the dominant source of model uncertainty.