Quantitative analysis of spectral data based on stochastic configuration networks.

Area of Science:

• Chemometrics
• Machine Learning
• Spectroscopy

Background:

• Traditional linear models in spectral data analysis are fast but lack accuracy for nonlinear problems.
• Nonlinear models offer higher accuracy but can be slow and prone to local optima.
• A hybrid approach is needed to leverage the strengths of both linear and nonlinear methods.

Purpose of the Study:

• To introduce Stochastic Configuration Networks (SCNs), a type of single-hidden layer feedforward neural network, into chemometrics.
• To analyze SCN model termination parameters, including error tolerance and maximum hidden nodes.
• To determine optimal random configuration settings for efficiency and stability in SCNs.

Main Methods:

• Implementation of Stochastic Configuration Networks (SCNs) for quantitative spectral data analysis.
• Analysis and determination of key SCN parameters: error tolerance, maximum hidden nodes, and random configuration iterations.
• Validation of SCN performance on two public spectral datasets, comparing against Principal Component Regression (PCR), Partial Least Squares (PLS), Back Propagation Neural Network (BPNN), and Extreme Learning Machine (ELM).

Main Results:

• SCNs demonstrated good stability and high prediction accuracy on spectral datasets.
• The SCN method exhibited superior efficiency compared to other tested techniques.
• Optimal model termination and random configuration parameters were identified for SCNs.

Conclusions:

• Stochastic Configuration Networks (SCNs) are effective for quantitative analysis of spectral data.
• SCNs provide a robust alternative, balancing speed and accuracy in chemometric applications.
• The findings support the suitability of SCNs for complex spectral data challenges.