FTIR, 1H and 13C NMR data fusion to predict crude oils properties

Data fusion from different analytical sources can be a feasible way to estimate physicochemical properties of petroleum when compared to using a single analytical technique. This occurs because outputs of different instrumental techniques can carry complementary information and act synergistically during calibration. In this paper, we investigate the potential of data fusion strategies for estimating seven crude oil properties: sulphur content (S), total nitrogen content (TN), basic nitrogen content (BN), total acid number (TAN), saturated (SAT), aromatic (ARO) and polar (POL) contents. We used 127 crude oil samples split into 70% for calibration and 30% for prediction. Partial least squares (PLS) regression models were constructed from Fourier transform mid-infrared (FTIR) and 1H and 13C nuclear magnetic resonance (NMR) spectroscopy. Data fusion models were built: fused at low and mid-level in different combinations. While mid-level fusion usually increased the accuracy of models, low-level fusion caused insignificant improvements. Using PLS mid-level fusion, we estimated S, TN, BN, TAN, SAT, ARO and POL contents with average prediction errors of 0.064 wt%, 0.049 wt%, 0.0070 wt%, 0.16 mgKOH·g−1, 5.34 wt%, 3.66 wt% and 6.58 wt%, respectively, with coefficients of determination equal to 0.87, 0.78, 0.98, 0.91, 0.79, 0.67 and 0.63 for the prediction set and using 4, 3, 3, 3, 2, 4 and 2 latent variables, respectively. Although promising results were obtained, mid-level fusion demonstrates to be the best strategy usually improving accuracy of models.