Near-Infrared Spectroscopy and Image Classification of Refuse Derived Fuels to Increase Cement Production Quality

Abstract

Refuse derived fuels (RDF), produced from municipal and industrial waste, provide an alternative to fossil fuels like coal or lignite in the cement production, thereby reducing the significant CO_2 emissions typically associated with cement production. The composition of RDF is often unknown, which limits the substitution rate, since otherwise the risk of impacting cement quality would increase. In this contribution, both near-infrared spectroscopy (NIRS) and RGB images were used to analyze RDF in an at-line measurement on a conveyor belt setup. The goal was to classify individual RDF particles in one of six fractions (paper, foils, 3D plastic, rubber, foams, textiles), since the fractions differ in combustion and flight behavior and therefore influence cement quality. For this, training, validation, and test data were obtained from 11,526 manually sorted RDF particles, sampled from various German cement plants and processed using an at-line conveyor belt setup. The NIRS data were processed using a small convolutional neural network (CNN) to provide the respective fraction, yielding an accuracy of 99.5%. The images were processed with different CNNs with transfer learning, yielding an accuracy of 96.7%. In a second phase, both NIRS and image predictions were combined by soft voting, yielding an accuracy of 99.7%. This validates the method under lab conditions and lays the groundwork for an application in a cement plant.