摘要
针对过程数据中存在的噪声干扰及动态特性,提出了一种基于慢特征重构与改进DPLS的软测量建模方法。该方法首先利用慢特征分析提取变化缓慢的成分,并用于重构原始输入,同时提出一种重构相似性指标来评价重构效果,实现用尽可能少的成分刻画数据的原有趋势,减少噪声干扰;然后梳理改进DPLS方法的完整流程,并用于分析重构输入与原输出间的关系,获得的模型更符合数据间的动态关系。本方法的有效性在TE过程及脱丁烷塔过程的软测量模型中得到了验证。
A novel soft sensor method based on slow feature reconstruction and enhancing dynamic partial least square （DPLS） is presented for process data with noise and dynamic characteristics.This method firstly extracts the components with slowly varying dynamics by a rising unsupervised learning method called slow feature analysis（SFA）, which aims to extract invariant features and contain significant information from the high-dimensional signals. The extracted slow features are applied to reconstruct the raw input variables. In order to evaluate the reconstruction result, the reconstruction similarity index is proposed, which consists of the correlation between the original input and output and the similarity between reconstructed input and the original input, and the index realizes the purpose that original trend of data can be described by as few components as possible and the process noise is removed. Then, the reconstructed input variables are used for regression modeling. The traditional DPLS methods are built based on PLS and dynamic extension, which extend the input variables with time-delay inputs to describe the dynamic characteristics of process data. Although this method is easy to implement, the extended dynamic model may cause the curse of dimensionality and make the mapping matrix more complex and difficult to interpret when the dimension of input variables or time delay is too large. Therefore, the enhancing DPLS（EDPLS） is proposed by considering the importance of different time-delay input variables, and the complete flowchart of EDPLS, which consists of training and testing sections, is summarized, and the model built by EDPLS is more consistent with the dynamic relationship of process data than the general DPLS. Finally, soft sensing application of TE process and debutanizer column process have been carried out to test the effectiveness and feasibility of the proposed method, and the proposed method, SFAr-EDPLS, shows the better performance than the traditional dynamic regression models suc
作者
蒋昕祎
李绍军
金宇辉
JIANG Xin-yi;LI Shao-jun;JIN Yu-hui(Key Laboratory of Advanced Control and Optimization for Chemical Processes,Ministry of Education,East China University of Science and Technology,Shanghai 200237,Chin)
出处
《华东理工大学学报（自然科学版）》
CAS
CSCD
北大核心
2018年第4期535-542,共8页
Journal of East China University of Science and Technology
基金
国家自然科学基金(21676086
21406064)
关键词
软测量
偏最小二乘
慢特征分析
数据重构
TE过程
soft senor
partial least square
slow feature analysis
data reconstruction
TE process
