Database Management Method Based on Density and Non-Linearity for Locally Weighted Linear Regression Model

Soft-sensors have been widely used for predicting difficult-to-measure variables by using easy-to-measure variables in many industries, and they have successfully contributed to operation cost reduction in real industrial processes. However, there are still some problems to solve to enhance the estimation performance of soft-sensors and make their use more widespread. One of the problems is that researches on database management have been not actively conducted although the estimation performance of soft-sensors significantly depends on quality and quantity of the samples in a database, and the samples are selected arbitrary in most research. To solve this problem, in the past research, database management methods based on the density of the samples, in which samples that are close to other samples are merged, were proposed. In addition to the density, the strength of non-linearity should also be taken into account when locally weighted regression models are used since the density required to construct an accurate model varies with the strength of non-linearity. Thus, this research proposes a database management method that takes into account process non-linearity as well as density of the process data. It was confirmed that, by using the proposed method, high prediction accuracy was realized with limited number of samples in a numerical example.