基于混合准则的软测量建模辅助变量选择方法

Mixed Criterion Based Secondary Variables Selection for Soft Sensor

摘要:

软传感器在工业中被广泛应用于预测与产品质量密切相关的关键过程变量，这些变量很难在线测量。要建立一个高精度的软传感器，选择合适的辅助变量是至关重要的。针对这个问题，本文通过耦合训练集的BIC准则以及验证集的MSE准则得到一个混合整数非线性规划问题，并将该MINLP问题分成内外两层结构，外层采用遗传算法对二元整数变量进行寻优，内层在整数变量固定之后退化成了较易于求解的非线性规划问题。在此基础上经过进一步分析提出了基于混合准则的变量选择方法，然后将所得辅助变量子集代入BP神经网络进行软测量建模。最后，通过4组案例对所提出方法进行验证。结果表明，所提出方法建立的软测量模型具有较好的预测性能。

Abstract:

Soft sensors are widely used in industry to predict key process variables that are closely related to product quality, and these variables are difficult to measure online. To build a high-precision soft sensor, it is important to choose the appropriate auxiliary variables. Aiming at this problem, this paper obtains a mixed integer nonlinear programming problem by coupling the BIC criterion of the training set and the MSE criterion of the verification set, and divides the mixed integer nonlinear programming problem into two layers, the inner and outer layers, and the outer layer uses the Genetic Algorithm (GA). The integer variable is optimized, and the inner layer degenerates into an easier to solve nonlinear programming problem (NLP) after the integer variable is fixed. Based on this analysis, a variable selection method based on hybrid criteria is proposed. Then the subset of secondary variables obtained is substituted into BP neural network for soft sensor modeling. Finally, the proposed method is validated by four actual cases. The results show that the soft-measurement model established by the proposed method has better prediction performance.