Mathematical methods for regression in Raman spectroscopy applied to natural gas analysis

The Raman spectroscope developed by Pietro Fiorentini aims to extract component compositions from natural gas mixtures and subsequently determine various thermodynamic properties. This work investigates and tests several techniques for processing the instrument's output data to compute composition and thermodynamic variables. The proposed machine learning approach employs a Convolutional Neural Network trained on mixture data from two units combined with data augmentation, and demonstrates strong generalization capability when tested on mixture data from previously unseen units, despite limited data availability. Unlike the current linear combination method, this approach does not require pure spectra extraction for each individual unit, as it is trained, validated, and tested directly on mixture spectra, resulting in a reduced total cost and cycle time per unit.