基于即时学习的不完整高维数据非线性过程软测量

Nonlinear Process Soft Sensors of Incomplete High-Dimensional Data Based on Just-in-Time Learning

由于现代工业过程中数据存在高维性、强非线性、时变性和不完整性等特性,限制了局部加权偏最小二乘算法(LW-PLS)的预测精度。因此,本工作在即时学习建模算法的基础上,将偏最小二乘算法(PLS)应用于相似性样本的选择中,并研究比较了基于主成分分析法(PCA)无监督降维的即时学习算法和基于PLS有监督降维的即时学习算法。概率主成分分析法(PPCA)可以有效的解决数据不完整性的问题;基于PCA和PLS算法的处理,有效的解决了数据高维性问题。基于即时学习的局部加权建模,可以有效解决数据的时变性和非线性问题。所研究算法的有效性在一个数值例子和脱丁烷塔实例中得到了验证。

The locally weighted partial least squares algorithm(LW-PLS)has been successfully applied to industrial processes and has received extensive attention.However,due to high dimensionality,strong nonlinearity,time-varying and incompleteness of data in practical industrial processes,the prediction accuracy of locally weighted partial least squares algorithm is limited.Therefore,based on the just-in-time learning method,this paper proposes to use Partial Least Squares(PLS)to select similarity samples for removing noise and improving sample relevance to the output,as well as studies and compares performance of the just-in-time learning algorithm between PCA based unsupervised dimensionality reduction and PLS based supervised dimensionality reduction.In the proposed approach,the probabilistic principal component analysis method is utilized to effectively solve the problem of data incompleteness.Based on the PCA or PLS algorithm,the dimensionality of the data is effectively reduced for sample selections.Local weighted modeling based on just-in-time learning effectively solves the time-varying and nonlinear modeling problems.The effectiveness of the proposed algorithm is verified through a numerical example and a debutanizer example.