CWBLS network and its application in portable spectral measurement

Area of Science:

• Spectroscopy
• Machine Learning
• Chemometrics

Background:

• Portable near-infrared spectroscopy (NIRS) faces challenges with low signal-to-noise ratio and poor reproducibility.
• Existing regression models often struggle with accuracy and stability in real-world NIRS applications.

Purpose of the Study:

• To introduce the novel D-CWBLS network for improved regression analysis in portable NIRS.
• To enhance the accuracy, stability, and robustness of NIRS models using low-quality data.

Main Methods:

• The D-CWBLS network expands upon the BLS network by incorporating Near-Infrared characteristic spectral band data.
• A Dropout layer is added vertically to deepen the network, optimize structure, and reduce redundancy.
• Optimized and enhanced feature node weight matrices are combined to minimize training randomness and improve robustness.

Main Results:

• The D-CWBLS model demonstrated superior performance in reproducibility, accuracy, and robustness.
• It outperformed traditional machine learning (PLSR, BP-ANN, ELM) and deep learning (MLP, CNN, RNN) models.
• Performance was also better than basic BLS and CWBLS models, showing significant advancements.

Conclusions:

• The D-CWBLS network offers a significant advancement for portable NIRS data analysis.
• It addresses key challenges in outdoor NIRS applications, providing higher reliability and applicability.
• This novel approach paves the way for more robust and accurate spectroscopic measurements.