Assessment of in-line spectroscopy and CNNs for CDW fine fraction classification

Abstract

Effective management of construction and demolition waste (CDW), particularly its fine fraction (<4 mm), faces technical challenges due to its heterogeneity and limited recyclability. Conventional characterisation and sorting technologies are ill-suited to handle this fraction, as its fine particle size and compositional heterogeneity exceed their operational limits, resulting in widespread landfill disposal. Exploiting this fraction as a secondary raw material requires accurate, real-time classification or regression methods. Unlike previous studies that used individual spectroscopic instruments, this study leverages a novel multi-sensor system capable of in-line analysis using four complementary techniques: laser-induced breakdown spectroscopy (LIBS), Raman, ultraviolet–visible (UV–Vis) and near-infrared (NIR) on fine CDW. The system was tested on 12 standard materials, including cementitious materials, recycled aggregates, and organic and polymeric compounds, resulting in a total of 49 152 spectra across the four modalities. One-dimensional (1D) convolutional neural networks (CNNs) were trained for multiclass classification. LIBS achieved the highest micro F1 scores (0.999), followed by UV–Vis (0.990), Raman (0.989) and NIR (0.869). This work demonstrates that the multi-sensor system captures material-specific spectral features that, coupled with 1D-CNN classifiers, allow reliable discrimination among fine CDW classes. The system’s ability to acquire data continuously and in-line provides a viable pathway towards scalable, automated characterisation in industrial settings.