Closed-Loop Identification Issues in the Process Industry

摘要

Control system design in the process industry often involves model-basedcontrollers. In this thesis it is investigated whether a model can be identified from closed-loop data, such that a (re)designed controller yields a higher performance of the closed-loop system. Taking into account the identification objective, which is the design or improvement of controllers, and the application area, which is the process industry with its specific properties, several issues, related to the stated problem, can be distinguished. The first issue is closed-loop identification. This is concerned with the problem of how to obtain a model from closed-loop data. The second issue that is addressed is the parametrization problem. The parametrization or model structure of a black-box model is arbitrary, but it determines the conditioning of the parameter estimation problem, and therefore it determines the results of the identification method. The third problem at issue concerns control-relevant identification. The fourth issue that is considered is input design. The bias of a model is shaped by applying appropriate filters to the data, and the variance of the model is reduced by injecting an appropriately designed excitation signal. Finally, an industrial experiment is conducted. An excitation signal is applied to a multivariable distillation column operating in closed loop, and the resulting data is used to identify models.