Standard-free calibration transfer - An evaluation of different techniques

Abstract

The combination of spectroscopic measurements and multivariate calibration techniques (chemometrics) has become a state-of-the-art technology for process analytical chemistry. Changes, intended or unintended, in the environmental conditions, the measurement setup or of the measured substance itself can result in a calibration model that is no longer adequate for the intended purpose. In such a situation, either a new model needs to be developed or (calibration) transfer methods, can be applied to transfer models from the original (main, master) to the new (remote, slave) setting. In this contribution, we introduce, discuss and evaluate a wide-ranging subset of transfer approaches available in chemometrics and the field of machine learning, where we focus on techniques applicable in situations where transfer standards, i.e. a set of samples measured under the original as well as the new setting, cannot be provided and only few reference measurements are available for the new setting. The introduced techniques are evaluated on a public data set as well as a new industrial data set displaying three forms of transfer problems. The efficiency of the proposed transfer approaches in terms of the number of required reference measurements compared to full model recalibration can bee confirmed. Average rank maps are presented to provide guidance on a proper choice among evaluated techniques.