Measurement of particle size in suspension based on VIS-NIR-RGB sensors and gradient-boosted regression tree

Abstract

To better utilize the spectral features of near-infrared and visible light and address issues in traditional suspension particle size detection, such as long measurement cycles, high costs, and complex operation, we propose a new method that combines multispectral sensor characterization technology with machine learning algorithms for rapid particle size measurement. By integrating a lab-designed light source driving circuit and spectral data acquisition device, we obtained 15 spectral data points for visible light, near-infrared, and red–blue–green (RGB) values to build a prediction model. The near-infrared-based model obtained a coefficient of determination (R2) of 0.9130, whereas models combining RGB-near-infrared and visible-near-infrared features performed best, with R2 values of 0.9670 and 0.9845. These models improved performance by 5.91% and 7.83%, effectively overcoming the limitations of single spectral features in different suspensions. This combination of visible and near-infrared light with machine learning proves effective for detecting average particle size in suspensions.