IPA: A deep CNN based on Inception for Petroleum Analysis

Deep learning is emerging as a challenger to chemometrics which was used for decades to predict properties of petroleum cuts using Near-Infrared spectroscopy (NIR). In this study, a new CNN ”Inception for Petroleum Analysis” (IPA) has been developed and its performance compared with both PLS and recent deep learning reference model DeepSpectra to predict CN of middle distillates. IPA, is inspired from state-of the-art computer vision architectures. It is based on several computational blocks with the ambition to meet the daily needs in NIR analysis and is especially suited for small datasets like less than 250 samples. Two datasets were considered, a first one with a reduced spectral range, using variable selection, in which saturated variables have been removed for PLS modeling, and a larger dataset, without variable selection, to evaluate the robustness of IPA. IPA demonstrates regression errors 40% smaller than PLS, and therefore much closer to measured data. Compared with a benchmark Deep Learning model, in particular DeepSpectra, IPA is 20% more accurate. On the same dataset in its full-range of 4150 features, IPA demonstrates a regression error 50% smaller than PLS, and 21% more accurate than DeepSpectra. These results are due to the ability of IPA to both extract complex representation from spectra and not being distracted by parts of the dataset who are considered saturated by experts. This new method improves the performance and reduces the dependence on complex pre-processing operations on dataset, like variable selection and digital signal processing. • Development of a new deep learning architecture for NIR analysis. • The model has been applied to fuel data. • Model robustness improvement over PLS.