Control of Nonstationary Systems Using Information Preserving Neural Networks

基于模式识别的控制方法可以在一系列应用程序中产生稳定和健壮的性能，并且可以设计这些方法，以便工业从业者在实现方面需要相对较少的努力或专业知识。人工神经网络（ANNs）非常适合模式识别应用。人工神经网络的一个优点是它们具有插值和泛化模式分类的能力，因此它们有望成为基于模式的控制器设计的改进框架。例如，人工神经网络可以被训练来学习使用历史工厂数据的过程输入-输出关系，然后可以作为自适应控制器的内部模型在线安装。另一个例子是直接在过程的误差反馈关系中训练人工神经网络，然后使用人工神经网络作为控制器本身。为了保持准确的描述，基于人工神经网络的控制器架构必须提供足够的自由度来确保可塑性，从而允许控制器以前从未经历过的系统行为得到适当处理。该研究项目涉及使用模式识别来识别和表示当前的过程特征，同时满足这些需求。这项工作的重点是研究控制系统(1)在设计、实施和适应方面的自主决策，(2)在广泛的应用中提供稳定和鲁棒的性能，以及(3)成功处理非线性、非平稳系统。pi的方法是研究一个由几个人工神经网络组成的体系结构，每个人工神经网络执行特定的功能，并将它们集成到一个统一的设计中。这些功能包括：对被操纵输入和被控制输出的近期历史中所显示的模式进行分类，以形成当前过程特征的图形模型，并将该图形模型转换为适当的算法控制器参数，用于许多流行的在线控制器算法；基于模式的控制器性能评估显示在最近的历史上的控制器误差，并使用这种性能评估在无监督训练的图像到上述算法转换，以提高性能和保持稳定性在非线性和非平稳应用。该控制器将在一系列单输入单输出过程模拟中进行测试。过程将显示一系列特征，包括过程顺序的变化，过程增益的线性程度，主导时间常数的可变性，信噪比的大小，以及模拟过程特征中缓慢和突然变化的非平稳变化。

Pattern recognition based control methods can produce stable and robust performance in a range of applications and these methods can be designed such that relatively little effort or expertise is required on the part of an industrial practitioner for implementation. Artificial neural nets (ANNs) are well suited for pattern recognition applications. A strength of ANNs are their ability to interpolate and generalize pattern classifications, consequently they hold promise as an improved framework for pattern based controller design. For example, ANNs can be trained to learn a process input - output relationship using historical plant data, and can then be installed online in the role of the internal model of an adaptive controller. Another example is training ANNs directly in the error feedback relationship of a process and then using the ANN as the controller itself. To maintain an accurate description, ANN based controller architectures must provide sufficient degrees of freedom to ensure plasticity, thus permitting system behaviors never previously experienced by the controller to be properly addressed. The research project involves the use of pattern recognition to identify and represent current process characteristics while satisfying these requirements. The focus of this work is the study of control systems that (1) are autonomous in their decision making for design, implementation and adaptation, (2) provide stable and robust performance over a wide range of applications, and (3) succeed in dealing with nonlinear, nonstationary systems. The PIs' method of approach is to study an architecture comprised of several ANNs that each perform specific functions and integrate them into a unified design. These functions include: classification of the patterns exhibited in the recent histories of the manipulated input and controlled output to form a pictorial model of the current process character, translation of this pictorial model into appropriate algorithmic controller parameters for use in any of a number of popular on-line controller algorithms, controller performance evaluation based on the patterns exhibited in the recent history of the controller error, and use of this performance evaluation in the unsupervised training of the pictorial to algorithmic translation mentioned above to improve performance and maintain stability in nonlinear and nonstationary applications. The controller will be tested on a range of single-input single-output process simulations. The processes will display a range of characteristics, including variations in process order, the degree of linearity of the process gain, the variability of the dominant time constant, the size of the signal to noise ratio, and with both slowly and suddenly changing nonstationary changes in the simulated process character.