基于大数据建模的玉米淀粉糖工艺参数预测系统开发  

System Development for the Prediction of Key Parameters in Corn Starch Sugar Process Based on Big Data Modeling

作　　者：李义 刘颖慰 都健[3] 赵优 赵国兴 佟毅 Li Yi;Liu Yingwei;Du Jian;Zhao You;Zhao Guoxing;Tong Yi(COFCO Biotechnology Limited Liability Company,Anhui,233000;COFCO Nutrition and Health Research Institute Co.,Ltd,Beijing,102200;College of Chemical Engineering,Dalian University of Technology,Liaoning,116024)

机构地区：[1]中粮生物科技股份有限公司,安徽233000 [2]中粮营养健康研究院有限公司,北京102200 [3]大连理工大学化工学院,辽宁116024

出　　处：《当代化工研究》2022年第14期145-152,共8页Modern Chemical Research

摘　　要：近年来,传统制造业面临着数字化挑战。建立基于大数据的智能工厂能够通过模型有效识别,将隐性和碎片化的工业问题显性化,形成新的知识积累。本文根据玉米深加工制备淀粉糖工艺中记录的实时数据,首先基于大数据技术对原始数据进行数据清洗,随后提出了三种数据降维与人工神经网络(ANN)相结合的策略对玉米深加工工艺的关键参数进行识别,并建立了过滤工段压差的预测模型,包括两步神经网络预测(T-ANN)法、聚类分析结合神经网络预测(C-ANN)法、Lasso回归结合神经网络预测(L-ANN)法。结果表明,三种方法均能准确地关联玉米淀粉上游工段的工艺条件与过滤工段压差的关系,复相关系数R均在0.99以上。通过以上三种模型,依据Olden权值排序法筛选出了权值较大的20个控制位点,从而大幅降低输入变量的维数,达到筛选关键控制位点的目的。通过对三种方法进行机理分析、敏感性分析以及超参数分析,发现L-ANN法对数据集的敏感性最低,预测精度较高,且能筛选出更多的具有机理可解释性的关键位点。Corn starch sugar industry have long faced the risks of high energy consumption and thin profits.Due to its long and complicated process,it is hard to be upgraded or optimized based on mechanism unit operation models.Fortunately,big data technology provides a promising solution due to its ability to turn huge amounts of data into insights of informed business and operational decisions.Therefore,big data technology can be used to establish the process model of the corn starch deep processing.In this paper,real-time data is recorded and analyzed.The collected data is first cleaned and then used to train the big-data model.Artificial neural network(ANN)is considered to build digital models and identify key parameters.For the purpose of dimensional reduction,three ANN-based methods are proposed,namely two-step-ANN(T-ANN)prediction method,cluster-analysis-ANN(C-ANN)prediction method and Lasso-regression-ANN(L-ANN)prediction method.The proposed methods are used to establish the prediction model of the filtering section,which has a great influence on production efficiency.The results show that the three proposed methods can accurately correlate the relationship between the process conditions in the upstream section of corn starch and the pressure difference of the filtration section,the correlation coefficient R is all above 0.99.Through each method,20 control sites with larger weights are selected and analyzed.Thereby,the dimension of input variables can be greatly reduced,thus the key control sites is identified.Through the comparison of mechanism analysis,sensitivity analysis and hyperparameter analysis,it is found that the L-ANN method has the lowest sensitivity to the data set and highest prediction accuracy.

关 键 词：神经网络 玉米淀粉糖 聚类分析 Lasso分析 参数预测 大数据 

分 类 号：TQ021.8[化学工程]