Cost-efficient strategies for building, transferring, and maintaining multivariate calibration models for spectroscopic sensors

The process of quality control for the inspection of products in industries requires methods that quantify the concentration of chemical constituents before products are distributed to consumers. Nowadays, industries are transforming these processes by adopting methods that are more efficient, waste fewer resources, and are non-destructive for the material that is inspected. Such methods rely on spectroscopy technology in which devices are used to acquire signals from a product and those signals can be used to make a prediction about the chemical composition of the product. The combination of spectroscopy and multivariate calibration has been one of the biggest breakthroughs in order to transforming the chemical inspection of products from such costly and destructive processes to fast non-destructive instant quantification. The chemical quantification of products based on spectroscopy signals has been a process based on the use of multivariate calibration models which are trained to be further used in quantification through prediction. Nonetheless, while the chemical relationship between spectral signals and constituents fundamentally exists, obtaining a satisfactory model that can be used in different environments of prediction and maintained in the long term has represented a major challenge since the transformation of chemical quantification was redirected, becoming a subject of study at the interface between analytical chemistry and statistics. The aim of this work was to develop an efficient methodology to enable the building, transferring, and maintenance of spectroscopy sensors through calibration models in order to keep models valid in the long term with minimum effort. The strategies here presented discuss the conditions for effective model building, satisfactory and efficient calibration transfer, and effective model maintenance. Firstly, effective model building is presented by providing answers on the conditions of sample selection to collect the reference analyses to build the models. Secondly, the conceptualization of calibration transfer has been re-structured and new algorithms have been designed to enable more efficient calibration transfer. Lastly, a complementary paradigm for calibration maintenance was developed with specific metrics for model drift quantification and model update strategies have been studied.

Publication year:2022

Accessibility:Open