生物发酵过程中VIP优化神经网络逆系统的软测量方法  

VIP optimal neural network inverse system soft sensing method in bio-fermentation process

作　　者：于霜[1] 刘国海[2] 梅从立[2] 程锦翔[2] 

机构地区：[1]苏州工业职业技术学院机电工程系,江苏苏州215104 [2]江苏大学电气信息工程学院,江苏镇江212013

出　　处：《南京理工大学学报》2015年第4期447-451,共5页Journal of Nanjing University of Science and Technology

基　　金：国家中小型企业创新基金(6612C26213202207);苏州市科技基础设施建设计划项目(SZP201303)

摘　　要：针对生物发酵过程中生化变量难以在线检测的问题,提出一种基于变量投影重要性(Variable importance in the project,VIP)方法优化的神经网络逆系统软测量模型。根据逆系统理论建立发酵过程生化变量的软测量模型,由于发酵系统的复杂性,逆系统软测量模型具有不惟一性,且难以得到精确的表达式。文中提出采用VIP方法对逆系统软测量模型的辅助变量进行优选,以对主变量贡献率较高的变量作为软测量模型的辅助变量,离线采集发酵过程各变量值,训练神经网络近似逆系统软测量模型,得到优化的神经网络逆系统软测量模型,实现发酵过程中菌体浓度和基质浓度的在线估计。利用Pensim平台采集数据,对所提方法做了仿真实验,结果表明:经过优化辅助变量的神经网络逆系统软测量方法具有更高的估计精度和泛化能力。To solve the online measuring of biochemical variables in the fermentation process,a neural network inverse soft sensing method which is optimized using the variable importance in the project( VIP) is proposed. According to the inverse system theory,a soft sensing model of biochemical variables is constructed. Due to the complexity of the fermentation process,the soft sensing model is not unique and not exact. This paper proposes that secondary variables should be optimized using the VIP method. The variables which have greater contribution to key variables are selected as the secondary variables of the soft sensing model. This paper collectes the fermentation process data offline and trains neural network approximating complex soft sensing model. The optimal neural network inverse system soft sensing model is obtained. It can estimate the mycelium concentration and substrate concentration online. Numerical simulations based on the Pensim data platform show that the optimal soft sensing model has higher estimation accuracy and stronger generation ability.

关 键 词：变量投影重要性 神经网络逆系统 软测量 发酵过程 内含传感器 变量筛选 在线检测 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]