2019 Volume 52 Issue 6 Pages 554-561
Just-in-time modeling methods, such as locally weighted regression, construct a local model using samples stored in a database each time when the output estimation is required. To reduce the computational burden of online estimation, the number of samples stored in the database should be limited. Thus, a database management method that selects an appropriate set of samples from all the historical samples is required. We propose a new database management method that takes into account the strength of the nonlinearity as well as the sample density in order to realize the systematic sample selection. Locally weighted linear regression models with different degrees of localization are used to evaluate the strength of the nonlinearity. We compared the proposed method and conventional methods, such as first-in first-out methods, through a numerical example and a case study of an industrial distillation process. It was confirmed that, using the proposed method, 7 to 48% less estimation error is accomplished when the number of samples in the database is the same.
