Monitoring of multivariate calibration models in the absence of new reference values: The regression case

Diaz, Valeria Fonseca; Aernouts, Ben; Bobelyn, Els; Nicolaı̈, Bart; Saeys, Wouter

doi:10.1016/j.chemolab.2023.104884

Cited by 3 publications

(2 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In short, new spectra are projected into the space of a PCA or a PLS, where the quality of their projection is evaluated with the so-called Hotelling statistic and residuals (T 2 and Q). A new method has recently been proposed for estimating a regression model's prediction bias and covariate shift in real-time, without using target reference values [37]. A set of current spectra, acquired during process monitoring is used to re-estimate the intercept of the calibration model.…”

Section: Discussion and Open Problemsmentioning

confidence: 99%

See 1 more Smart Citation

Cloning instruments, model maintenance and calibration transfer

Roger,

Fonseca-Diaz,

Nikzad-Langerodi

2024

Preprint

View full text Add to dashboard Cite

Most literature on the application of Non-Destructive Spectral Sensors (NDSS) reports proofs of concept limited to model calculation (calibration) and its application on a so-called independent data set (validation, or test). However, developing NDSS also requires proving that the performance obtained during this first validation remains valid when conditions change. This generic problem is referred to as robustness in chemometrics. When the measurement conditions change, the measured spectrum is subject to a deviation. The reproducibility of the model, and thus of the sensor, with respect to this deviation, defines its robustness. The application of NDSS involves a large number of processes, and thus deviation sources. Instrument cloning, between laboratory instruments or from a benchtop to an online device, is certainly the most concerning issue for deploying NDSS-based applications. This problem has been studied for many years in chemometrics, under the paradigm of calibration transfer, through geometric corrections of spectra, spectral spaces, or calibration model corrections. The same problem has been addressed in the machine learning community under the domain adaptation paradigm. Although all these issues have been addressed separately over the last twenty years, they all fall under the same topic, i.e., model maintenance under dataset shift. This paper aims to provide a vocabulary of concepts for formalizing the calibration model maintenance problem, reviewing recent developments on the subject, and categorizing prior work according to the proposed concepts.

show abstract

Section: Discussion and Open Problemsmentioning

confidence: 99%

“…the domains [7,35,36,37]. Figure 4 shows a decision tree that can help select a domain transfer method, depending on the characteristics of the domain in which the model is applied.…”

Section: How To Manage the Robustnessmentioning

confidence: 99%