Development of a robust calibration model for nonlinear in-line process data

Area of Science:

• Chemical Engineering
• Analytical Chemistry
• Data Science

Background:

• Industrial distillation processes present challenges for accurate water content prediction due to curved effects from temperature changes and batch variations.
• Near-infrared (NIR) analysis offers in-line measurement capabilities for process monitoring.
• Developing robust calibration models is crucial for process control and optimization.

Purpose of the Study:

• To compare the performance of global linear (partial least squares), local linear (locally weighted regression), and nonlinear (neural networks) calibration models for predicting water content in a distillation reactor.
• To investigate the impact of spectral range selection and data preprocessing on model robustness.
• To develop and validate a new model selection criterion to mitigate overfitting in neural network models.

Main Methods:

• Comparative analysis of partial least squares (PLS), locally weighted regression (LWR), and neural network (NN) calibration models.
• Utilized in-line near-infrared (NIR) spectral data from an industrial distillation process.
• Explored spectral range selection and various data preprocessing techniques.
• Developed a model selection criterion based on the median of monitoring error over replicate trials for neural networks.

Main Results:

• Neural network models, particularly back-propagation networks, demonstrated superior performance in predicting water content on independent test data compared to PLS and LWR.
• The proposed model selection criterion effectively addressed overfitting issues in neural network calibration.
• Spectral range selection and data preprocessing significantly influenced model accuracy and robustness across all methods.

Conclusions:

• Back-propagation neural network models, selected using the novel median error criterion, provide the most accurate and robust calibration for predicting water content in challenging industrial distillation environments.
• The developed model selection strategy enhances the reliability of neural network-based process analytical technology (PAT).
• This study highlights the importance of advanced chemometric techniques and rigorous model validation for effective process monitoring.