Machine learning applied as an in-situ monitoring technique for the water content in oil recovered by means of NIR spectroscopy

Abstract

The present work proposes a study on the combination of artificial neural network based model and multivariate calibration techniques, as principais component analysis (PCA) and partial least squares (PLS) combined with near infrared spectrophotometer (NIR) to formulate online water content monitoring models for recovered fuel oils after effluent treatment of a Brazilian thermoelectric plant. The database for adjustments of these models was built using oil samples supplied by the SUAPE II thermoelectric plant, where they were characterized in laboratories and analyzed via NIR. 450 spectra were used to construct the PLS model for predictive model calibration using the PLS technique and 118 spectra for the model based on artificiais neurais networks. In the obtained PLS model, it was possible to obtain correlations around 0.97, cross-validation error (RMSECV) of 0.0906 and test prediction error (RMSEP) of 0.0651. The RNA model presented coefficient of determination R² of 0.99 for training and R² of 0.98 for test with prediction error of 0.062