基于核熵投影的CPLS间歇过程监测及质量预测

Monitoring and Quality Prediction of CPLS Batch Process Based on Kernel Entropy Projection

针对间歇过程的非线性、多模态特性以及质量预测问题,提出多向高斯混合模型-并发核熵潜结构投影(multi-way Gaussian mixture model-concurrent kernel entropy projection to latent structures, MGMM-CKEPLS)算法。首先根据核熵投影将低维非线性数据映射到高维核特征空间,通过Renyi熵贡献的大小选取主元,降低了主元数,克服了传统核方法计算负荷大的问题;然后通过高斯混合模型(GMM)获取每一模态数据,分别对不同模态的样本数建立并发潜结构投影(CPLS)模型,由于考虑了各个模态过程的不同特征,更符合实际间歇过程数据特性;最后通过各模态权值系数集成统一的监控统计量,实现在线监测和质量预测。通过青霉素发酵过程验证了所提算法比MKPLS算法具有更好的在线监控效果,质量预测精度更高。

A ＂multi-way Gaussian mixture model-concurrent kernel entropy projection to latent structures＂（MGMM-CKEPLS） algorithm was proposed to solve problems of nonlinearity, multimode and quality prediction of batch process. Low-dimension nonlinear data was first projected into high-dimensional kernel feature space by kernel entropy projection. Principal components were obtained by the size of Renyi entropy contribution, which can reduce principal component numbers and overcome the problem of computational complexity of traditional kernel methods. Data of each mode was then obtained by GMM, and CPLS models were established for different modes, which was more consistent with actual batch processes by considering the difference between each process. Finally, unified monitoring statistics were integrated to achieve online monitoring and quality prediction by modal weight coefficients. The model was verified in penicillin fermentation process and the results show that the proposed algorithm has better online monitoring effectiveness and higher accuracy in quality prediction than MKPLS algorithm.