Domain Adaptation for In-Line Allergen Classification of Agri-Food Powders Using Near-Infrared Spectroscopy.

Area of Science:

• Food Science and Technology
• Spectroscopy
• Machine Learning

Background:

• Human error in adding agri-food powders poses a significant food safety risk, particularly concerning allergen cross-contamination.
• Current detection methods may be insufficient for real-time monitoring on dynamic production lines.

Purpose of the Study:

• To develop an early detection system for incorrect agri-food powder additions using near-infrared (NIR) spectroscopy and machine learning.
• To adapt machine learning models for use with moving samples on a production line, reducing the need for extensive labelled data.

Main Methods:

• Employed near-infrared spectroscopy (NIR) for sample analysis.
• Utilized domain adaptation techniques, including domain-adversarial neural networks (DAN) and semisupervised generative adversarial networks (SS-GAN).
• Combined measurements from two NIR sensors with different wavelength ranges and applied ensemble methods for enhanced accuracy and uncertainty quantification.

Main Results:

• Achieved up to 96.0% accuracy without labelled data from moving samples.
• Reached 99.68% accuracy when incorporating a single labelled data instance per material.
• Combined domain adaptation methods and multi-sensor NIR data yielded the highest average prediction accuracies.

Conclusions:

• Domain adaptation effectively transfers NIR spectroscopic models from stationary to moving samples in food manufacturing.
• The developed system offers a robust solution for early detection of incorrect ingredient additions, enhancing food safety.
• The combination of advanced machine learning, multi-sensor NIR, and ensemble techniques provides accurate and interpretable results.