Across different instruments about tobacco quantitative analysis model of NIR spectroscopy based on transfer learning

Area of Science:

• Analytical Chemistry
• Machine Learning

Background:

• Near-infrared (NIR) spectroscopy is crucial for quantitative analysis.
• Traditional calibration transfer algorithms often require large, similar datasets.
• Transfer learning offers a solution for calibration transfer with limited samples.

Purpose of the Study:

• To develop an instance transfer learning algorithm for NIR quantitative analysis across different instruments.
• To improve model generalization and prediction performance in tobacco analysis.

Main Methods:

• Preprocessing spectral data using Standard Normal Variate (SNV) and Principal Component Analysis (PCA).
• Comparing Support Vector Machine (SVM), weighted Extreme Learning Machine (ELM), and weighted ELM-AdaBoost models.
• Implementing a weighted ELM-TrAdaBoost model for instance transfer learning between domains.

Main Results:

• The weighted ELM-TrAdaBoost model achieved high prediction accuracy for nicotine (R²=0.9426), Cl (R²=0.8147), K (R²=0.7548), and total nitrogen (R²=0.6980).
• Ensemble learning and source domain samples significantly improved model generalization.
• The proposed method demonstrates effectiveness for small sample size modeling and fast learning.

Conclusions:

• The weighted ELM-TrAdaBoost algorithm is superior for NIR calibration transfer between instruments.
• This approach enhances model performance and generalization ability, particularly with limited data.
• The method is advantageous for practical applications requiring rapid and accurate quantitative analysis.