Low cost color assessment of turbid liquids using supervised learning data analysis – proof of concept

Abstract

This work reports the development of a low cost in-line color sensor for turbid liquids based on the transmission and scattering phenomena of light from RGB and IR LED sources, gathering multidimensional data. Three different methodologies to discriminate color from the turbidity influence are presented as a proof of concept approach. They are based in regression models, expectation maximization Gaussian mixtures and artificial neural networks applied to labeled measurements. Each methodology presents advantages and disadvantages which will depend on the intended implementation. Regression models revealed to be best suited for standard or occasional measurements, the EM Gaussian mixture will perform better for well-known controlled range of colors and turbidities and the neural networks have easy implementation and potential suited for real-time IoT platforms.