[An Improved ELM Algorithm for Near Infrared Spectral Quantitative Analysis]

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Background:

• Extreme Learning Machine (ELM) is a chemometric method for spectral analysis.
• Original ELM faces challenges with high-dimensional and collinear spectral data.
• Existing solutions can be ill-conditional due to matrix properties.

Purpose of the Study:

• To propose an improved Extreme Learning Machine (iELM) for spectral regression.
• To address the limitations of original ELM in handling high-dimensional spectral data.
• To enhance the precision of spectral analysis models.

Main Methods:

• Utilizing Extreme Learning Machine (ELM) to link spectral variables and response variables.
• Treating hidden node outputs as new variables.
• Employing VIP-SPLS (Variable Importance in Projection - Stacked Partial Least Squares) for regression modeling on these new variables.

Main Results:

• The proposed iELM effectively handles high dimensionality and collinearity in spectral data.
• VIP-SPLS leverages hidden node outputs and offers model ensemble advantages.
• iELM demonstrated a 29.06% precision improvement over PLS and 27.47% over original ELM on benchmark NIR data.

Conclusions:

• The integration of VIP-SPLS with ELM (iELM) significantly improves spectral regression model performance.
• iELM offers a robust solution for analyzing high-dimensional and collinear spectral data.
• This enhanced chemometric approach provides greater precision in spectral analysis.