EXPRESS: Attention-based Multi-level Fusion for NIRS Modeling

Area of Science:

• Analytical Chemistry
• Chemometrics
• Process Analytical Technology (PAT)

Background:

• Near-Infrared Spectroscopy (NIRS) offers direct molecular-level insights into product quality (e.g., concentration, viscosity).
• NIRS models are highly susceptible to environmental noise, limiting their real-world applicability.
• Robust modeling requires effective integration of process variables to mitigate noise interference.

Purpose of the Study:

• To develop a robust multi-level fusion model for enhancing NIRS-based quality prediction.
• To leverage Attention Mechanism (AM) for effective feature extraction and fusion.
• To address the sensitivity of NIRS to environmental noise through process variable integration.

Main Methods:

• A two-branch Convolutional Neural Network (CNN) was employed for NIRS feature extraction.
• A novel fusion block incorporating cross-attention, residual connection, and Multi-layer Perceptron (MLP) was designed.
• Two fusion strategies (central and shared) were explored to optimize feature integration across different CNN levels.

Main Results:

• The proposed multi-level fusion model demonstrated superior performance compared to existing methods.
• The Attention Mechanism effectively weighted and integrated NIRS and process variable features.
• Experimental validation on 2,6-dimethylphenol distillation data confirmed the model's robustness and accuracy.

Conclusions:

• The developed Attention Mechanism-based multi-level fusion model significantly enhances the robustness of NIRS quality prediction.
• This approach offers a promising solution for overcoming environmental noise challenges in NIRS applications.
• The findings support the broader adoption of advanced chemometric techniques in industrial process monitoring.