BDSER-InceptionNet: A Novel Method for Near-Infrared Spectroscopy Model Transfer Based on Deep Learning and Balanced...

Area of Science:

• Analytical Chemistry
• Chemometrics
• Machine Learning

Background:

• Near-Infrared Spectroscopy (NIRS) models struggle with generalization due to instrumental differences, environmental factors, and sample variety.
• Data distribution shifts and feature mismatches limit model transferability across different NIRS instruments and crop varieties.
• Existing NIRS data and models are often underutilized for new applications, hindering efficiency and accuracy.

Purpose of the Study:

• To develop a novel transfer learning framework for enhancing cross-instrument compatibility in NIRS analysis.
• To improve the generalization capability and prediction accuracy of NIRS models in industrial settings.
• To enable effective utilization of previously accumulated NIRS data for new instruments and varieties.

Main Methods:

• Proposed a transfer learning framework integrating a multi-scale network architecture (RX-Inception) with Balanced Distribution Adaptation (BDA).
• Incorporated RX-Inception structure combining depthwise separable convolution and residual connections for enhanced feature coupling.
• Integrated Squeeze-and-Excitation (SE) attention to dynamically recalibrate spectral band weights for improved feature representation.
• Systematically evaluated six transfer strategies for model adaptation performance.

Main Results:

• The proposed BDSER-InceptionNet achieved state-of-the-art performance on primary instruments for NIRS analysis.
• Method 6 demonstrated successful NIRS model sharing from primary to secondary instruments, mitigating spectral discrepancies.
• Significant improvements in transfer efficacy were observed, enabling better model adaptation across different instruments.

Conclusions:

• The developed transfer learning framework effectively addresses instrumental heterogeneity in NIRS.
• The proposed method enhances model sharing and prediction accuracy, maximizing the utility of existing NIRS data.
• This approach offers a promising solution for robust and adaptable NIRS modeling in diverse industrial applications.