Deep learning-based prediction of particle size distributions in construction and demolition waste recycling using convolutional neural networks on 3D laser triangulation data

Abstract

To enhance sustainability in the construction industry, substituting primary with recycled aggregates from construction and demolition waste (CDW) is essential. However, the necessary quality assessment of recycled aggregates, especially their particle size distribution (PSD), through sampling and manual sieving is time-consuming and prone to sampling errors due to the heterogeneity of CDW waste and fluctuating material flows combined with small sampling and manual sieving volumes. Here, we introduce a novel inline monitoring approach using convolutional neural networks (CNNs) to estimate PSDs from inline 3D laser triangulation (3DLT) sensor data of both primary and recycled aggregate particles. Analyzing 174,220 particles across nine size classes with a dual camera 3DLT sensor, a customized VGG-inspired CNN model outperformed other architectures, achieving accuracies of 80.8 % and 75.0 % for primary and recycled aggregates at particle level, respectively. Most errors were near-miss classifications, yielding a mean absolute error of 1.0 vol% in PSD predictions at material flow level. Explainable artificial intelligence techniques confirmed the reliance of CNNs on particle contours for robust classification. Our findings offer a pathway for inline PSD monitoring in processing of both primary and recycled aggregates, contributing to a more quality-orientated, circular, and sustainable construction industry. • Sensor-based prediction of particle size distributions (PSDs) investigated. • Novel for PSD prediction using convolutional neural networks (CNNs) presented. • PSDs could be predicted within 1.0 vol% mean absolute error. • XAI confirmed reliance of CNN on particle contours for robust classification. • Insights enable inline PSD monitoring in construction & demolition waste recycling.

Topics & Concepts