A representation-first framework for laser-induced breakdown spectroscopy (LIBS) based quality assurance in recycled aggregate processing
Cheng Chang (TU Delft - Civil Engineering & Geosciences, Macau University of Science and Technology)
Siwei Peng (Southeast University)
Xiaorong Wang (Ltd.)
Shuai Zong (Wuhan University)
Wei Hu (Nanjing University of Information Science and Technology, Nanyang Technological University)
Hao Cheng (School of Civil and Environmental Engineering)
Francesco Di Maio (TU Delft - Civil Engineering & Geosciences)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Ensuring consistent quality of recycled aggregates is essential for their wider use in circular construction. Laser-induced breakdown spectroscopy (LIBS) enables rapid elemental inspection, yet its performance in automated sorting systems is strongly shaped by how raw spectra are represented. This study adopts a representation-first benchmarking perspective and evaluates four representative feature families, namely variance-driven Principal Component Analysis (PCA), manifold learning-based Isomap, label-driven Partial Least Squares Discriminant Analysis (PLS-DA), and cepstral envelope-line separation, together with a raw-spectrum logistic-regression baseline and a histogram gradient boosting (HGB) reference model. The benchmark uses 24,000 single-shot spectra collected from ten material classes under conveyor-belt conditions, and repeated stratified random-split evaluation is used to assess the robustness of the comparative results across data splits. Across the evaluated models, cepstral features deliver the strongest overall performance, while PCA and the raw-spectrum logistic-regression baseline remain closely competitive. Isomap broadens the comparison toward non-linear manifold-based embeddings but does not improve performance in the present dataset, and PLS-DA shows the weakest stability under strong channel collinearity and class overlap. The results indicate that explicit feature extraction is not uniformly beneficial across all methods, but that spectral representation remains a major source of performance variation under controlled conveyor-like LIBS acquisition. In particular, cepstral features provide the most favourable balance among classification performance, robustness to baseline variation, and compactness, whereas PCA remains attractive when interpretability is prioritised. These findings provide a controlled benchmark and practical guidance for designing reliable and explainable LIBS-based quality-assurance pipelines for recycled-aggregate processing.
help_outline