Research Initiation Award: Recurrent Neural Networks as Representation of Nonlinear Dynamical Systems

9309057在系统控制中，过程的数学描述通常是控制器设计的先决条件。由于非线性系统本身的复杂性和辨识算法的困难，线性系统的系统辨识技术是人们熟知的新的非线性系统。由于现实世界应用中不可避免的噪声，这些困难在建立基于经验的模型时被切断。本项目将使用递归神经网络进行非线性系统辨识。定义网络的参数通过输入-输出数据训练来确定。培训的方法基本上是基于梯度的。稳定性方法也将被检验，原因有两个。它提供了一种识别性能的定性度量，并为改进梯度方法的效率提供了一些启示。这些方法本身并不独特，除非符合实际的应用标准(较低的错误率、不受噪声影响等)。都被考虑在内了。为了考虑来自经验数据的噪声影响，稳健范数将被视为参数预测误差的度量。将采用基于自适应系统理论的理论分析和统计分析来对所提出的问题进行定性的理解。将进行数值模拟，以检验所提出的识别程序。理论和经验方法的目标都是相同的：揭示神经网络在现实世界问题中未被探索的方面。提出的研究受到启发，并将通过一个工业规模的过程辨识问题进行检验。特别是，实验将在霍尼韦尔TDC-3000实时工厂信息和控制系统上进行，该系统连接到甲醇-异丙醇蒸馏塔。***

9309057 Si In system control, a mathematical description of a process is often a prerequisite to the controller design. The system identification techniques for linear systems are new well understood nonlinear systems due to the inherent complexity of nonlinear systems and the difficulty of deriving identification algorithms. These difficulties are severed in building empirically based models because of unavoidable noise in real-world applications. This project will use recurrent neural networks to perform nonlinear system identification. The parameters defining the network are determined through input-output data training. The methodology for training is essentially gradient- based. Stability methods will also be examined for two reasons. It provides a qualitative measure of identification performance and provides some insight as t improve the efficiency of gradient methods. These approaches themselves are not unique until real-world application criteria (lower error rates, immune to noise etc.) are taken into consideration. To account for noise effect from empirical data, robust norm will be considered as a measure of parameter prediction error. Theoretical analysis based on adaptive systems theory and statistical analysis will be employed to gain qualitative understanding of the proposed problem. Numerical simulation will be carried out to test proposed identification procedures. The goal of both theoretical and empirical approaches are the same: reveal unexplored aspects of neural networks for real world problems. The proposed research is inspired and will be tested by an industrial scale process identification problem. In particular the experiments will be conducted at a Honeywell TDC-3000 real time plant information and control system interfaced to a methanol- isopropanil distillation column. ***