Apprentissage profond et proche infrarouge pour intensifier l’usage des méthodes spectroscopiques

Near Infrared spectroscopy (NIR) combined with chemometrics has been used for years in various fields to predict properties, quantify species, classify samples, etc. At IFPEN, this approach is daily used in the laboratory and on-line on pilot plant units to predict properties of interest of petroleum cuts. Today, NIR and chemometrics are under development for biomass characterization and plastic recycling processes. One of the major difficulties is determining a consensus on which spectral preprocessing method and optimal settings to use for the chosen chemometric method (regression, classification, etc.). The common practice is a time-consuming trial-and-error experimentation that has to be reiterated when changing the device, for a change in acquisition conditions, etc. In this context, the proposed thesis is to guarantee the prediction quality using deep learning approaches. 
The use of deep learning on spectroscopic data has been on the rise over the past 5 years, especially Convolutional Neural Networks (CNNs). It is proposed to work on PIR data already available at IFPEN for more than 20 years, acquired on different spectrometers, in the laboratory as well as online. Thus, the relevance of using deep learning on spectral data will be assessed on a global approach, from laboratory calibration to transfer to online analysis. The richness of existing data at IFPEN will make it possible to develop approaches to facilitate the maintenance and transfer of calibrations from a deep network, without losing performance and knowledge

Keywords: Deep Learning, near infrared spectroscopy, chemometrics, neural networks, transfer learning