Abstract:

An intelligent modeling approach was developed for batch reactor to solve modeling difficulties of non-linear characteristics，dynamic process and with partially unmeasurable states. Base on separability of reaction，the topology of model was established via structure approaching hybrid neural networks（SAHNN）. Compared with normal neural networks，SAHNN has optimized structure and more nodes which represent actual states. This modeling approach utilized virtual supervisor-artificial immune algorithm（VS-AIA）to solve problems of training neural network with unmeasurable states. It firstly initialized information of unmeasurable states during reaction with only a little mechanism information（ascending or descending curve）to make them act as virtual supervisors. Then it improved the precision of modeling by optimizing virtual supervisor and training network weights at the same time. During training，immune population was used to explore new better solutions and avoid wrong direction. Detailed process of modeling and training neural networks were described in dynamical modeling of batch condensation reaction of producing promotor for vulcanizing rubber. The simulation result proved that the approach was effective.

Key words:

结构逼近式神经网络, 假想教师, 人工免疫, 间歇反应器, 建模

摘要：

提出一种智能化的神经网络建模方法，建立状态变量部分不可测的间歇反应器模型。针对间歇反应是一个非线性、非稳态过程，根据化学反应的非线性分离特性，采用结构逼近式神经网络构建模型的拓扑结构。利用反应的先验知识优化网络结构，赋予网络节点实际的物理意义，并完善网络训练过程，使建模过程灰箱化；通过假想教师-人工免疫训练算法，解决不可测变量影响常规网络训练的问题；通过并行优化假想教师和网络权值，提高建模精度。以实际橡胶硫化促进剂制备的间歇缩合反应过程为实验对象，详细论述了建模和网络训练的过程，证明了方法的有效性。

关键词:

结构逼近式神经网络, 假想教师, 人工免疫, 间歇反应器, 建模