Performance evaluation of kernel adaptive filtering algorithm in soft sensor modeling

Jun Li ； XueMei Wang

《淡江理工學刊》 24卷1期 (2021/02) Pp. 9-20

https://doi.org/10.6180/jase.202102_24(1).0002

摘要

Aiming at the problems of strong nonlinearity and complicated mechanism for chemical processes, a class of soft sensor modeling methods based on kernel adaptive filtering (KAF) algorithms are proposed, including sliding-window kernel recursive least-squares (SW-KRLS), fixed-budget kernel recursive least-squares (FBKRLS) and quantization kernel least mean squares (Q-KLMS). The major idea of the KAF algorithms is to use a linear adaptive filtering algorithm in feature space to solve nonlinear problem, and by introduction of ＂SW criterion＂, ＂fixed-budget (FB)＂ and quantization (Q) technical criterion to deal with the problem of difficulty in kernel matrix operation due to the increase of training data. In order to verify the effectiveness of the employed algorithms, they are applied to the soft sensor modeling instances including the key product output prediction of industrial fluid catalytic cracking unit (FCCU) which is the core unit of the oil secondary operation and product quality (melt index) estimation of sequential reactor-multi-grade (SRMG) industrial process. Compared with the employed KAF algorithm using the different kernel functions and the basic KAF algorithms as well as other kernel learning methods under the same conditions, experimental results confirm that the employed KAF algorithms can obtain better modeling accuracy, which can improve the generalization ability of the model and are suitable for training of large-scale data sets. In addition, FB-KRLS and Q-KLMS algorithms show similar modeling effects using the gaussian kernel.

關鍵字

Soft sensor ； Modeling ； Kernel recursive least squares ； Kernel least mean square ； Chemical processes