核主元回归模型间歇过程加权迭代学习的控制  

A weighted iterative learning control for batch process using kernel principal component regression model

作　　者：李赣平[1] 黄涛[1] 

机构地区：[1]南昌大学信息工程学院,江西南昌330031

出　　处：《计算机与应用化学》2013年第10期1097-1101,共5页Computers and Applied Chemistry

基　　金：国家自然科学基金资助项目(61064004)

摘　　要：间歇过程在生产中起到重要作用。针对间歇过程的控制提出了很多方法,迭代学习控制是其中一种。迭代学习控制需要合理的模型,目前数据驱动的建模方法受到重视。由于间歇过程通常为复杂的非线性过程,过程数据具有非线性相关性的特点。为了消除数据的非线性相关性,本文采用核主元回归方法对间歇过程进行建模,即在间歇过程的控制变量和终点质量之间建立间歇过程的模型。在此模型上,通过围绕标称轨迹线性化核主元回归模型,并最小化与终点质量相关的二次型目标函数,导出迭代学习控制算法从而计算控制策略。为了克服过程变化和扰动的不利影响,本文提出在批次间将最早的数据从训练数据集移除并加入最新的数据对核主元回归模型进行更新。由于迭代学习律中的增益矩阵反映的是过程的梯度信息,易使迭代学习控制过早收敛或偏离实际工况,为了获得更好的收敛效果,可对学习增益矩阵进行加权。通过对一个间歇聚合反应仿真过程的应用,加权迭代学习控制有良好的控制性能并显示出对过程变化和扰动的适应能力。该方法比基于主元回归模型的迭代学习控制方法具有更好的性能,因此基于核主元回归模型的加权迭代学习控制是一种有效的间歇过程控制方法。Batch process plays an important role in manufacturing. Many methods have been proposed for batch process control, one of them is iterative learning control （ILC）. Iterative learning control needs a reasonable model. By now, data-driven modeling methods have drawn great attention. Since batch process generally is a complicated nonlinear process, the process data have the character of nonlinear correlation. In order to eliminate the nonlinear correlation of the data, kernel principal component regression （KPCR） method is employed to model a batch process in the paper, by which a batch process is modeled between control variables and end-point qualities. Based on the model, the ILC algorithm is derived to calculate the control policy by linearizing the KPCR model around the nominal trajectories and minimising a quadratic objective function concerning the end-point product quality. To overcome the detrimental effects of uncertain process variations or disturbances, it is proposed in the paper that the KPCR model should be updated in a batchwise manner by removing the earliest batch data from the training data set and adding the latest batch data to the training data set. Because the gain matrix of the iterative learning law reflects the gradient information of a process, which will make the ILC algorithm converge early or deviate from the actual working conditions, the learning gain matrix is weighted to acquire a better convergence. The weighted iterative learning control （WILC） has good control performance and shows adaptability for process variations or disturbances when applied to a simulated batch polymerization process. The method has better performance than the ILC based on principal component regression （PCR） model, hence the WILC based on KPCR model is an effective method for batch process control.

关 键 词：加权迭代学习的控制 核主元回归 间歇过程 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]