Different approaches to multivariate calibration of nonlinear sensor data.

Area of Science:

• Analytical Chemistry
• Chemometrics
• Sensor Technology

Background:

• Multivariate calibration is essential for analyzing complex mixtures like refrigerant vapors.
• Traditional linear methods like Partial Least-Squares regression (PLS) fail with nonlinear sensor data.
• Nonlinear relationships between sensor signals and analyte concentrations necessitate advanced calibration techniques.

Purpose of the Study:

• To evaluate various nonlinear multivariate calibration methods for analyzing refrigerant vapors.
• To identify the most effective algorithms for handling nonlinearities in time-resolved sensor data.
• To determine the optimal balance between calibration accuracy and algorithmic complexity.

Main Methods:

• Box-Cox transformation for nonlinear dependent variables.
• Implicit nonlinear PLS regression incorporating squared independent variables.
• Quadratic PLS (QPLS) utilizing a nonlinear quadratic inner relationship.
• Tree algorithms for splitting nonlinear problems into linear subproblems.
• Neural networks (including genetic algorithms and growing neural networks) for modeling complex relationships and preventing overfitting.

Main Results:

• Quadratic PLS (QPLS) demonstrated good performance among simpler algorithms.
• Various neural network implementations achieved excellent calibration results.
• Sophisticated algorithms like growing neural networks provided the best performance, despite higher computational demands.

Conclusions:

• Nonlinear calibration methods are superior to linear approaches for the analyzed refrigerant vapor data.
• Neural networks, particularly advanced implementations, offer the highest accuracy.
• The choice of the optimal method involves a trade-off between calibration quality and algorithmic complexity.