基于改进GA-PLS算法的最优辅助变量选择及其在软测量建模中的应用

Optimal Selection of Secondary Variables Based on GA-PLS Algorithm and Its Application to Soft Sensor Modeling

提出了改进的遗传算法与部分最小二乘回归相结合的最优辅助变量的选择方法。用遗传算法来选择变量时,编码方法简单,染色体的长度为候选变量的个数,每一位的取值(0或1)表示某个变量是否被选中,具有全局搜索性能的遗传算法与传统的变量选择方法相比可以比较准确地找到最(次)优解;同时部分最小二乘回归能够克服多元回归中常见的多重共线性问题,在样本点个数少于变量个数的条件下也能进行回归建模分析。用文中提出的方法建立了催化重整过程中稳定油组分的软测量模型,结果表明了本文提出的辅助变量选择方法的优越性和实用性。

A novel method based on genetic algorithm and partial least square regression is proposed to select the most suitable secondary process variables used as a soft sensor inputs. In the proposed approach, coding technique is very simple by exploiting binary code in which the numerical value （0 or 1 ） of each bit represents whether one variable is selected or not. Genetic algorithm is able to find optimal （or near） solution faster than typical schemes, such as stepwise regression, especially in high dimensional input spaces. Partial least square regression can be employed to overcome difficulties encountered with the existing collinearity in multivariable regression and is also suitable for situation with low observation/variable ratio. The proposed method has been applied to develop a soft-sensor for measuring the stabilizing oil components of catalytic reform reactor and the results demonstrate that the method is effective and suitable for practical application.