A design of experiment approach incorporating layered designs for choosing the right calibration model

Choosing a useful calibration model is a complex optimisation problem which can be solved by experimental designs. In this paper, layered designs are introduced, and it is shown how they can be used to simplify the complicated designs that choosing a useful calibration model demands. The simplification is obtained by operating with several designs in different layers, or sequentially, and the optimised result from one layer is brought forward to the next layer. In the top layer, the main design is located, and this is used to actually choose the calibration model. Optimised results from earlier layers are tested as two-level variables in the top layer, where the low level is a non-optimised solution and the high level is the optimised solution. There are also traditional variables in the top layer in addition to the test variables for the hidden layer optimised solution. The concept is demonstrated on a NIR data set describing the feed into a naphtha pretreatment distillation column. The parameters considered in the search for a useful calibration model are type of regression method, calibration set selection, variable subset selection, outlier identification, Box-Cox transformation, differentiation and number of components.