化工学报 ›› 2022, Vol. 73 ›› Issue (7): 3120-3130.DOI: 10.11949/0438-1157.20220096

• 过程系统工程 • 上一篇    下一篇

基于IDPC-RVM的多模态间歇过程质量变量在线预测

周新杰(),王建林(),艾兴聪,随恩光,王汝童   

  1. 北京化工大学信息科学与技术学院,北京 100029
  • 收稿日期:2022-01-16 修回日期:2022-04-14 出版日期:2022-07-05 发布日期:2022-08-01
  • 通讯作者: 王建林
  • 作者简介:周新杰(1995—),男,博士研究生,zhouxj@mail.buct.edu.cn
  • 基金资助:
    国家自然科学基金项目(61973025)

IDPC-RVM based online prediction of quality variables for multimode batch processes

Xinjie ZHOU(),Jianlin WANG(),Xingcong AI,Enguang SUI,Rutong WANG   

  1. College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China
  • Received:2022-01-16 Revised:2022-04-14 Online:2022-07-05 Published:2022-08-01
  • Contact: Jianlin WANG

摘要:

间歇过程具有多模态特性,现有的间歇过程模态划分方法中过程数据高维特征和模态中心的选取直接影响模态划分结果的合理性,进而影响间歇过程质量变量在线预测的精度。为提高间歇过程质量变量在线预测的精度,提出了一种基于改进密度峰值聚类相关向量机(improved density peaks clustering-relevance vector machine,IDPC-RVM)的间歇过程质量变量在线预测方法。首先,在密度峰值聚类算法基础上,考虑过程数据的高维特征进行样本相似性度量,并通过样本密度不平衡下的模态中心选取策略准确获取间歇过程模态中心;其次,利用模态划分指标在无须先验知识的情况下获取间歇过程最优模态数目,并识别过渡模态完成间歇过程的模态划分;最后,建立各模态数据的RVM预测模型,实现间歇过程质量变量的在线预测。青霉素发酵过程的实验结果表明,与RVM、SCFCM-RVM和DPC-RVM方法相比,对青霉素浓度预测的均方根误差(RMSE)降低至0.0093,判定系数(R2)提升至0.9995,有效地提高了预测精度。

关键词: 间歇式, 改进密度峰值聚类, 模态划分, 模型, 预测

Abstract:

Batch processes have multimode characteristics. The high-dimensional characteristics of process data and the selection of mode center in the existing batch process mode partitioning method directly affects the rationality of the mode partitioning results, which in turn affects the accuracy of online prediction of batch process quality variables. To cope with this challenge, an online prediction method of batch processes quality variables based on improved density peaks clustering relevance vector machine (IDPC-RVM) is proposed. First, based on the DPC algorithm, the sample similarity is measured by considering the high-dimensional characteristics of the process data, and the mode centers of the batch process are accurately obtained by the mode centers selection strategy with the unbalanced sample density. Then, the optimal number of modes is obtained without prior knowledge according to the mode partitioning index (MPI), and the transition modes are identified to complete the mode partitioning of batch processes. Finally, the RVM prediction model of each mode data is established to realize the online prediction of batch process quality variables. The experimental results of penicillin fermentation process show that compared with RVM, SCFCM-RVM and DPC-RVM methods, the RMSE of the proposed method for the prediction of penicillin concentration was reduced to 0.0093, and the R2 was increased to 0.9995, which effectively improving the prediction accuracy.

Key words: batchwise, improved density peaks clustering, mode partitioning, model, prediction

中图分类号: