Adaptive Robust Stochastic Configuration Networks for Near-Infrared Multivariate Analysis

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Background:

• Near-infrared (NIR) spectroscopy is widely used but faces challenges with high-dimensional data.
• Existing Stochastic Configuration Networks (SCNs) struggle with fast convergence and robust weight estimation in NIR analysis.
• Outliers and noise in NIR data can degrade prediction model performance.

Purpose of the Study:

• To develop an adaptive robust SCN (AR-SCN) algorithm for accelerated model construction and improved performance in high-dimensional NIR spectra analysis.
• To enhance the stability and generalization of prediction models in NIR applications.

Main Methods:

• Proposed an adaptive robust SCN (AR-SCN) algorithm.
• Implemented adaptive incremental learning based on prediction residual.
• Employed a global-local shrinkage strategy for robust output weight estimation.

Main Results:

• The AR-SCN algorithm demonstrated effectiveness on benchmark NIR datasets and a real-world gasoline blending process.
• Achieved simultaneous improvements in construction efficiency and robustness compared to state-of-the-art SCNs.
• Validated the method's ability to handle high-dimensional spectra and resist outliers/noise.

Conclusions:

• The proposed AR-SCN algorithm offers a significant advancement for NIR spectral analysis.
• AR-SCN effectively addresses limitations of existing SCNs in high-dimensional modeling.
• This method enhances both the speed and reliability of predictive modeling in NIR applications.