基于迁移成分分析的发酵过程集成软测量建模

Integrated Soft Sensor Modeling of Fermentation Process Based on Transfer Component Analysis

青霉素发酵过程具有不确定性和多阶段等特点,不同批次发酵过程间存在工况差异,过程数据的分布不一定相同,使传统软测量模型性能下降。结合迁移学习策略和高斯混合模型,提出一种基于迁移成分分析的多模型集成软测量建模方法。该方法使用迁移成分分析求解样本间共享特征映射矩阵,适配建模过程数据与待测数据的边缘概率分布;并基于高斯混合模型对建模数据进行聚类划分,与偏最小二乘算法结合建立子模型的集成模型,完成对主导变量的预测。基于青霉素平台数据的仿真结果表明,所提方法不仅能够有效提高青霉素发酵过程软测量模型的精度,而且适应于变工况下青霉素浓度的预测。

The Penicillin fermentation process is an uncertain and multi-stage process.There are different working conditions among different batch fermentation processes,and the distribution of process data is not necessarily the same,which degrades the performance of the traditional soft sensing model.Combined with the transfer learning strategy and Gaussian mixture model,a multi-model ensemble soft sensor modeling method based on transfer component analysis is proposed.In this method,the transfer component analysis is used to get the shared feature mapping matrix between samples,and adapt the edge probability distribution of labeled dataset and unlabeled dataset;the modeling data are clustered based on Gaussian mixture model,then use partial least squares algorithm to establish an ensemble model to predict dominant variables.The simulation results based on penicillin platform data show that the proposed method can not only improve the accuracy of the soft sensor model of penicillin fermentation process effectively,but also adapt to the prediction of penicillin concentration under variable working conditions.