基于改进的PCA-RBFNN过程变量软测量建模及应用  

Modeling and application of soft measurement for process variables based on improved PCA-RBFNN model

作　　者：朱荷蕾 高慧敏[1] ZHU Helei;GAO Huimin(College of Information Science and Engineering, Jiaxing University, Jiaxing 314001, China)

机构地区：[1]嘉兴学院信息科学与工程学院,浙江嘉兴314001

出　　处：《计算机集成制造系统》2022年第5期1306-1313,共8页Computer Integrated Manufacturing Systems

基　　金：嘉兴市公益性研究计划资助项目(2019AD32009,2020AY10012);教育部产学合作协同育人资助项目(202101014039);浙江省高等学校国内访问学者教师专业发展资助项目(FX2020037)。

摘　　要：针对过程控制系统中关键变量的软测量建模及应用问题,结合主成分分析法(PCA)和径向基(RBF)神经网络法(RBFNN),提出了改进的PCA-RBFNN软测量建模方法。首先利用PCA分析变量筛选法从过程变量集合中找到对系统过程特性具备最佳解释能力的过程变量子集;然后将该过程变量子集作为输入、被估计变量作为输出构建PCA-RBFNN模型,并使用K-means聚类和最小均方误差法初始化RBF神经网络的数据中心、扩展系数和连接权值;最后采用梯度下降法训练、校正所建模型。以某纺织原料生产过程为实例,对所建模型进行了验证和输出性能对比分析。结果表明,该模型可以实现过程变量在线预测,比原模型具有更好的泛化能力、预测能力和输出精度,能够提高过程控制系统的稳定性和可靠性。Aiming at the on-line prediction and monitoring of key process variables in process control system,by combining Principal Component Analysis(PCA)and Radial Basis Function Neural Network(RBFNN),an improved PCA-RBFNN soft measurement modeling method was proposed.The PCA analysis variable selecting method was used to find the process variable subset with the best interpretation ability for the system process characteristics.Then,the process variable subset was taken as the input and the estimated variable as the output to establish the improved PCA-RBFNN model.K-means clustering method and the least mean square error method were used to initialize the data center,expansion coefficient and connection weight of RBFNN.The gradient descent method was used to train and correct the model.Taking the production process of a textile raw material reactor as an example,the improved PCA-RBFNN model was verified and the output performance was compared and analyzed.The results showed that the improved PCA-RBFNN model could realize online prediction of process variables,improve the stability and reliability of the process control system,which had better generalization ability,prediction ability and output accuracy.

关 键 词：过程控制 主主成分分析 径向基神经网络 软测量 在线预测 

分 类 号：TP278[自动化与计算机技术—检测技术与自动化装置]