软测量模型的变量选择方法研究

发布时间：2018-03-10 04:13

本文选题：软测量　切入点：辅助变量选择　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来软测量建模技术在化工生产过程中得到了广泛应用。软测量技术根据某一最优准则,选择一组与主导变量相关的且易测量的辅助变量,构造以辅助变量为输入,主导变量为输出的数学模型,实现对主导变量的在线估计。虽然软测量的核心是建模,但是只有选择与主导变量密切相关的辅助变量才能建立一个有效的软测量模型。变量选择就是在软测量模型基础之上,从一系列预先给定的自变量集合中,确定一个在某种准则下可以对主导变量进行最佳描述的变量子集。假定有p个候选辅助变量,总共能产生2p-1个候选模型,即使p不是很大的时候,也能陷入组合爆炸的困境。因此,研究如何快速高效的变量选择方法,在保证模型预测性能的前提下,尽可能地减少冗余变量,是很有必要的。针对该问题,本论文开展了较为系统化的软测量变量选择方法研究。本文的主要研究内容和成果如下:1.将蒙特卡洛无信息变量消除算法(MC-UVE)、遗传算法和偏最小二乘(GA-PLS)三者相结合,首先利用MC-UVE算法剔除无信息变量,在MC-UVE所选出的有信息变量的基础上,使用GA算法进一步精选变量子集,提出了MC-UVE-GA-PLS变量选择方法。最后,通过UCI数据对算法的可行性、有效性、模型预测性能及模型复杂度等方面进行验证,并与A11-PLS模型和GA-PLS模型进行了对比,结果表明了算法的可行性与可靠性。2.考虑到变量选择本质上是数学优化问题,以多元线性回归(MLR)模型为基础,通过引入0-1决策变量,利用BIC准则,将变量选择问题描述成一个嵌套的混合整数二次规划(MIQP)问题,并提出了嵌套式MIQP-MLR变量选择方法,同时实现特征变量的选择与预测模型的建立。最后,通过UCI数据对所提出的方法进行验证,并与传统的逐步回归(Stepwise)变量选择方法进行对比,结果验证了所提出方法的有效性和实用性。3.在基于嵌套式MIQP-MLR变量选择方法基础上,进一步将模型结构从MLR拓展至鲁棒性更强的支持向量回归(SVR)模型,并利用改进的MSE准则,将变量选择描述成一个混合整数线性规划(MILP)问题,提出了 MILP-SVR变量选择方法。所提出的方法不仅不需要事先指定模型中的变量个数,避免惩罚因子的调节,而且求解效率更高。此外,SVR模型可以利用核技巧,实现非线性函数的拟合。最后通过UCI数据测试了算法的可行性与有效性,并与A11-SVR、Pearson-SVR和RFE-SVR对比,验证了方法的可靠性。4.将上述变量选择方法应用至某一工业精馏塔间苯二胺纯度的软测量建模中,为间苯二胺纯度软测量模型选择辅助变量,并建立相应的软测量模型。实际的工业数据的仿真研究结果表明所提出方法的可靠性与高性能。
[Abstract]:In recent years, soft sensor modeling technology has been widely used in chemical production process. According to an optimal criterion, soft sensing technology selects a set of auxiliary variables related to dominant variables and easy to measure, and constructs auxiliary variables as input. The dominant variable is an output mathematical model, which realizes the on-line estimation of the dominant variable. Although the core of soft sensing is modeling, But only by selecting auxiliary variables closely related to dominant variables can an effective soft-sensor model be established. Variable selection is based on the soft-sensing model and from a set of predefined independent variables. Determine a subset of variables that can best describe the dominant variable under certain criteria. Assuming that there are p candidate auxiliary variables, a total of 2p-1 candidate models can be generated, even if p is not very large. Therefore, it is necessary to study how to select variables quickly and efficiently and reduce the redundant variables as much as possible while ensuring the prediction performance of the model. The main contents and achievements of this paper are as follows: 1. Combining Monte Carlo algorithm with MC-UVEG, genetic algorithm and partial least squares GA-PLS. First of all, the MC-UVE algorithm is used to eliminate the information variable, and on the basis of the information variable selected by MC-UVE, the MC-UVE-GA-PLS variable selection method is proposed by using GA algorithm to further select the subset of variables. Finally, the feasibility and validity of the algorithm are obtained through the UCI data. The model prediction performance and model complexity are verified and compared with A11-PLS model and GA-PLS model. The results show that the algorithm is feasible and reliable. Based on the multivariate linear regression model, by introducing 0-1 decision variables and using BIC criterion, the variable selection problem is described as a nested mixed integer quadratic programming (MIQP) problem, and a nested MIQP-MLR variable selection method is proposed. At the same time, the selection of feature variables and the establishment of prediction model are realized. Finally, the proposed method is verified by UCI data, and compared with the traditional stepwise regression method. The results verify the validity and practicability of the proposed method. Based on the nested MIQP-MLR variable selection method, the model structure is further extended from MLR to a more robust support vector regression (SVR) model, and the improved MSE criterion is used. The variable selection is described as a mixed integer linear programming (MILP) problem, and the MILP-SVR variable selection method is proposed. The proposed method not only does not need to specify the number of variables in the model in advance, but also avoids the adjustment of the penalty factor. In addition, the kernel technique can be used to fit nonlinear functions. Finally, the feasibility and effectiveness of the algorithm are tested by UCI data, and compared with A11-SVRSn-Pearson-SVR and RFE-SVR. The reliability of the method is verified. 4. The above variable selection method is applied to the soft sensor modeling of the purity of resorcinenediamine in a certain industrial distillation column, and the auxiliary variable is selected for the soft sensing model of the purity of resorcinenediamine. The simulation results of practical industrial data show the reliability and high performance of the proposed method.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TQ021.8;TP18

【参考文献】