Comparison of machine learning approaches for spectroscopy applications

Abstract

In energy production the characterization of the fuel is a key aspect for modelling and optimizing the operation of a power plant. Near-infrared spectroscopy is a wellestablished method for characterization of different fuels and is widely used both in laboratory environments and in power plants for real-time results. It can provide a fast and accurate estimate of key parameters of the fuel, which for the case of biomass can include moisture content, heating value, and ash content. These instruments provide a chemical fingerprint of the samples and require a calibration model to relate that to the parameters of interest.

A near-infrared spectrometer can provide point data whereas a hyperspectral imaging camera allows the simultaneous acquisition of spatial and spectral information from an object. As a result, an installation above a conveyor belt can provide a distribution of the spectral data on a plane. This results in a large amount of data that is difficult to handle with traditional statistical analysis. Furthermore, storage of the data becomes a key issue, therefore a model to predict the parameters of interest should be able to be updated continuously in an automated way. This makes hyperspectral imaging data a prime candidate for the application of machine learning techniques. This paper discusses the modelling approach for hyperspectral imaging, focusing on data analysis and assessment of machine learning approaches for the development of calibration models.