Residual networks using multi-task learning algorithm for near-infrared spectroscopy: A case study.

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• Near-infrared spectroscopy (NIRS) is a valuable non-destructive technique for chemical analysis.
• NIRS data is complex and high-dimensional, posing challenges for accurate composition prediction.
• Existing methods struggle to effectively correlate spectral data with specific chemical constituents.

Purpose of the Study:

• To develop an advanced deep learning model for enhanced chemical composition analysis using NIRS data.
• To improve the accuracy and generalization of predicting multiple chemical contents from full-dimensional NIRS spectra.
• To optimize a ResNet-18 architecture for NIRS data by reducing channels and maintaining depth to prevent overfitting.

Main Methods:

• An improved ResNet-18 deep learning model was developed.
• Multi-task learning was integrated to estimate multiple chemical contents simultaneously.
• The model architecture was optimized by reducing network channels while preserving depth.

Main Results:

• The proposed ResNet-18 model accurately predicted four chemical compositions in tobacco.
• The model demonstrated superior performance compared to traditional machine learning algorithms.
• Experimental results confirmed excellent generalization and predictive accuracy of the modified model.

Conclusions:

• The optimized ResNet-18 model with multi-task learning effectively analyzes complex NIRS data.
• This approach offers a significant advancement for predicting chemical compositions in tobacco and potentially other matrices.
• The study highlights the potential of deep learning for robust spectral data analysis.