GOALI: Nonlinear Identification and Control Strategies

基于Seborg goal模型的过程控制策略利用非线性动态模型可以提供传统PID和基于线性动态模型的控制策略的改进。非线性控制系统成功应用的一个因素是合理准确的动态模型的可用性。对于一些工业控制应用，物理模型可以从非稳态质量和能量平衡等第一性原理开发。然而，由于模型复杂性、未知模型参数或缺乏关键测量，使用物理模型作为在线控制计算的一部分可能是不可行的。在许多应用程序中，由于各种原因无法获得精确的物理模型，这些原因包括过程复杂性、缺乏过程理解、有限的测量以及开发它们所需的时间和精力。对于这些情况，一个有吸引力的替代方案是开发一个经验非线性模型，该模型与可用的先验物理信息一致，使用非线性识别技术。在这个联合学术-工业研究项目中，加州大学圣巴巴拉分校和杜邦公司的过程控制研究人员将解决非线性识别中的一些问题，这些问题限制了现有方法的有效性及其广泛应用。这些问题包括：(1)输入序列的选择，以确保适当的过程激励；(2)纳入可用的先验工艺知识；(3)采样周期、采集数据量、模型阶数、模型时延等关键设计参数的选择；(4)模型不确定性表征；(5)在线模型更新与自适应。该研究产生的识别方法将在一些模拟和实验研究中进行评估，包括工业应用。案例研究将提供替代识别技术的比较，部分基于它们是否可用于开发有效的基于模型的控制系统。

Abstract - Seborg GOALI Model-based process control strategies utilizing nonlinear dynamic models can provide improvements over conventional PID and control strategies based on linear dynamic models. A factor for successful application of nonlinear control systems is the availability of a reasonably accurate, dynamic model. For some industrial control applications, physical models can be developed from first principles such as unsteady-state mass and energy balances. However, it may not be feasible to use a physical model as part of the on-line control calculations due to model complexity, unknown model parameters, or the lack of key measurements. In many applications, accurate physical models are not available for a variety of reasons, which include process complexity, lack of process understanding, limited measurements, and the time and effort required to develop them. For these situations, an attractive alternative is to develop an empirical nonlinear model which is consistent with available a priori physical information using nonlinear identification techniques. In this joint academic-industrial research project, process control researchers at the University of California at Santa Barbara and at DuPont will address a number of issues in nonlinear identification which limit the effectiveness of existing methods and their widespread application. These issues include: (1) selection of the input sequence to ensure appropriate process excitation; (2) incorporation of available a priori process knowledge; (3) selection of key design parameters such as sampling period, amount of data to be acquired, model orders and model time delay; (4) characterization of model uncertainty; and (5) on-line model updating and adaptation. The identification methods resulting from the research will be evaluated in a number of simulation and experimental studies, including industrial applications. A case study will provide a comparison of alternative identification techniques based, in part, on whether they can be used to develop effective model-based control systems.