基于典型相关分析和数据自回归处理的BP神经网络在聚丙烯熔融指数预报中的应用

Melt Index Prediction by BP Neural Networks Based on Canonical Correlation Analysis and Data Auto-regression

由于聚丙烯生产是一个大量参数相互耦合的强非线性过程,使得传统的机理建模受到一定的限制。提出基于典型相关分析和数据自回归处理的BP神经网络软测量建模,通过可测变量来推知聚丙烯熔融指数。应用典型相关分析选择与输出熔融指数关系较大的独立输入变量,数据自回归处理校正一系列带有误差的量测数据,而BP神经网络用来刻画过程的非线性特征。最后,将提出的算法应用到聚丙烯大型生产工艺中进行熔融指数的预报建模并进行实例仿真,仿真结果表明该算法有较强的建模精度。

Propylene polymerization is a highly nonlinear process with a lot of variables correlated, which limits the use of traditional first principle modeling. A soft-sensor architecture based on back propagation （BP）neural networks combining canonical correlation analysis as well as data auto-regression was proposed to infer melt index （MI） from other given process variables. Canonical correlation analysis was carried out to select the independent variables which have much contact with MI, data auto-regression was introduced to acquire corrected data, and BP networks were used to characterize the nonlinearity. Finally, the algorithm is applied to the production process of polypropylene and the results of emulator indicate the excellent model accuracy of the algorithm.