Developments in the Built Environment 21 (2025) 100611

Contents lists available at ScienceDirect

Developments in the Built Environment

journal homepage: [Link]/journal/developments-in-the-built-environment

Enhancing quality inspection efficiency and reliability of unscreened

recycled coarse aggregates (RCA) streams using innovative mobile
sensor-based technology
Cheng Chang a,* , Francesco Di Maio a, Rajeev Bheemireddy b, Perry Posthoorn b,
Abraham T. Gebremariam a , Peter Rem a
a
Resource & Recycling, Department of Engineering Structures, Faculty of Civil Engineering and Geosciences, Delft University of Technology, Stevinweg 1, 2628 CN, Delft,
the Netherlands
b
DEMO, Delft University of Technology, Gebouw 36, Mekelweg 4, 2628 CD, Delft, the Netherlands

A R T I C L E I N F O A B S T R A C T

Keywords: Recycled coarse aggregates (RCA) from End-of-Life (EoL) concrete face resistance due to inconsistent quality. To
Recycled coarse aggregates (RCA) address this, a mobile, containerized sensor-based inspection system is developed, capable of processing over 100
Quality inspection tons of RCA per hour. Using advanced 3D scanning and laser-induced breakdown spectroscopy (LIBS), the system
Concrete recycling
ensures reliable real-time analysis of particle size distribution (PSD) (Root Mean Square Error: <5.5%) and
Conveyor belt
contaminant detection (Accuracy: 0.94). Incremental learning techniques dynamically update chi-square dis­
3D scanner gocator
Laser-induced breakdown spectroscopy (LIBS) tribution parameters as new spectral data becomes available, refining models continuously without full
retraining and sustaining high classification performance. Monitoring data are recorded on radio frequency
identification (RFID) tags, enhancing traceability. This innovation improves efficiency compared to traditional
methods, supporting sustainable practices in the construction industry. Its applications also extend to related
fields such as mining, waste management, and resource recovery, contributing to the circular economy, reducing
reliance on natural aggregates, and promoting environmentally friendly infrastructure development.

1. Introduction produce recycled coarse aggregates (RCA), which serve as substitutes for
natural aggregates. The increasing demand for construction materials,
The growing global population and rapid urbanization have signifi­ alongside the urgency for sustainable waste management, has propelled
cantly increased the generation of Construction and Demolition Wastes this practice as a progressively feasible solution. This practice enables
(CDWs), making their management a critical challenge for the con­ resource conservation, reduces environmental impact, and supports the
struction industry (Aslam et al., 2020; Kabirifar et al., 2020). CDWs are circular economy (Gálvez-Martos et al., 2018; C. Zhang et al., 2022).
among the largest waste streams globally, comprising diverse materials This transition aligns with global circular economy initiatives and
such as concrete, wood, metal, and plastics (Galán et al., 2019; Soto-Paz highlights the construction industry’s shift towards more sustainable
et al., 2023). Inefficient handling and improper disposal of these wastes practices.
not only accelerate environmental degradation but also deplete natural Recognizing the potential environmental detriments and the asso­
resources, emphasizing the urgent need for sustainable waste manage­ ciated carbon footprint of unfettered natural aggregate extraction, many
ment solutions (Akhtar and Sarmah, 2018; Amesho et al., 2023; K. countries are promoting sustainable practices in the construction
Zhang et al., 2023). In response to these challenges, many countries have domain (Al Martini et al., 2023; Aslam et al., 2020; Kabirifar et al., 2020;
implemented recycling and reuse strategies to reduce the environmental Sai Trivedi et al., 2023; Soto-Paz et al., 2023). The European Union (EU)
footprint of CDWs and conserve valuable resources (Lederer et al., 2020; has distinctly emerged as a frontrunner, adeptly incorporating sustain­
Luciano et al., 2022; Su et al., 2024). One promising approach involves able recycling methodologies into mainstream construction practices
the recycling of concrete from End-of-Life (EoL) infrastructures to (Akhtar and Sarmah, 2018; Gálvez-Martos et al., 2018; Lederer et al.,

This article is part of a special issue entitled: Novel Materials Testing published in Developments in the Built Environment.
* Corresponding author.
E-mail addresses: chang-cheng@[Link], [Link]-1@[Link] (C. Chang).

[Link]
Received 22 August 2024; Received in revised form 4 January 2025; Accepted 20 January 2025
Available online 24 January 2025
2666-1659/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license ([Link]
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

2020; Marique and Rossi, 2018; C. Zhang et al., 2022). This shift is not significant period, their inherent limitations, such as potential impre­
merely a reaction to environmental urgency, it marks a strategic tran­ cisions and inefficiencies, have become increasingly evident as the
sition towards resource conservation and a reduced carbon footprint for construction sector has progressed. Specifically, traditional methods for
the entire construction sector. RCA quality control, such as manual sieving and inspection, face several
However, the transition to RCA is fraught with challenges, chiefly critical challenges: (1) low accuracy: manual methods are prone to
concerning the assurance of their quality and purity (J. Kim, 2022; H. human error, resulting in inconsistent assessments; (2) inefficiency:
Wu et al., 2023). These challenges are accentuated when the RCA is labor-intensive processes are time-consuming and unsuitable for
sourced from a diverse range of dismantled infrastructures, bringing to high-throughput applications; (3) lack of real-time feedback: traditional
the fore issues related to the presence and detection of contaminants methods do not provide immediate insights, delaying necessary ad­
(Alaejos et al., 2013). These contaminants, if not properly identified and justments in production lines; and (4) limited adaptability: these
managed, can significantly compromise the integrity and applicability of methods struggle to operate effectively in dynamic environments such as
RCA in new construction projects (Poon and Chan, 2007; L. Wu et al., demolition sites. As the industry’s aspirations continue to shift towards
2024). Consequently, addressing these quality-related concerns is crit­ achieving enhanced operational efficiencies without compromising
ical to maintaining the performance, durability, and reliability of quality, the imperative for a more innovative, efficient, and precise
RCA-infused construction (Marín-Cortés et al., 2024; Vegas et al., 2015). quality inspection mechanism has grown more pronounced.
It requires a focused approach towards standardizing the quality Amidst the prevailing challenges, the emergence of sensor technol­
assessment methods and developing stringent guidelines to ensure that ogy provides a promising solution (Barri et al., 2020; Cabral et al., 2023;
the recycled aggregates meet or surpass the performance metrics of their Chang, 2025; Chang et al., 2022, 2025; Lotfi et al., 2015; Trotta et al.,
natural counterparts. 2021; Xia and Bakker, 2014). The rising demand underscores the po­
Historically, the assessment of RCA primarily relied on traditional tential of recent advancements in sensor technology (Bonifazi et al.,
methodologies, characterized by their labor-intensive nature, prolonged 2018; Nalon et al., 2022; Vegas et al., 2015), which offer real-time,
time frames, and manual procedures (Marie and Mujalli, 2019; Tuan on-site characterization of RCA. Such state-of-the-art methodologies
et al., 2022). While these methods have served the industry for a not only enable prompt feedback, expediting the decision-making

Fig. 1. Concrete to Cement and Aggregate (C2CA) technology.

2
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

process and any necessary recalibrations in production but also augment (3) Dynamic Adaptability: Incremental learning ensures the model
the overall efficiency of the RCA production and usage cycle. Further­ dynamically updates with new spectral data, eliminating the
more, the integration of real-time sensors into the RCA assessment need for full retraining and maintaining robust performance.
process signifies more than just technological evolution, it represents a (4) Mobility and Versatility: The containerized design enables on-site
broader shift towards sustainable construction practices. By allowing for operation, reducing transportation costs, optimizing resource
instant feedback and adjustments, these systems can reduce wastage, allocation, and enhancing efficiency in diverse demolition
optimize resource use, and ensure that the resulting product meets the contexts.
necessary quality benchmarks.
This research delves into the sensor-based quality inspection system The innovative use of advanced sensors enables real-time assess­
integrated into the Concrete to Cement and Aggregate (C2CA) technol­ ments on-site, which is particularly valuable in the dynamic context of
ogy (Fig. 1). Fig. 2 presents a flowchart summarizing the methodological demolition sites. By examining operational aspects and evaluating
steps of the quality inspection. This quality inspection system employs a effectiveness in real-world scenarios, this research aims to highlight how
3D scanner Gocator and a laser-induced breakdown spectroscopy (LIBS) this technology could improve the quality assessment of RCA. Through
sensor to provide a comprehensive granulometric analysis of RCA and experimentation and analysis, the study emphasizes the advancements
enable potential contaminants. The accuracy and validity of the particle in RCA processing, presenting a pathway for industries to achieve sus­
size distribution (PSD) measurement method are verified using 3D tainable growth while maintaining high-quality standards.
modeling with X-ray tomography, comparable to traditional manual
sieving. Additionally, incremental learning techniques update existing 2. Materials and methods
models as new spectral data becomes available, dynamically adjusting
chi-square distribution parameters. This approach ensures continuous 2.1. Material samples
model refinement without the need for complete retraining, enhancing
computational efficiency and sustaining high classification This study used recycled concrete aggregates obtained from dis­
performance. carded concrete, collected during the dismantling of various in­
To overcome the challenges posed by traditional quality control frastructures throughout the Netherlands. To maintain the purity of the
methods, this mobile sensor-based system offers significant advantages: EoL concrete, selective demolition methodologies were employed.
Additionally, other demolition residuals such as brick, foam, glass,
(1) Real-Time Precision: LIBS achieves consistent contaminant gypsum, mineral fibers, plastics, and wood were manually separated
detection under dynamic and static conditions with an accuracy from the collection sites. This segregation process aimed to preserve the
of 94%, and PSD estimation shows a Root Mean Square Error of integrity of the samples. The primary components of the materials used
less than 5.5%. in this study are detailed in Table 1.
(2) High Throughput and Scalability: Capable of processing over 100
tons of RCA per hour, the system supports industrial-scale oper­
ations, meeting the demands of large-scale construction and de­
molition projects.

Fig. 2. Schematic of quality inspection process.

3
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Table 1 only reduces the costs associated with transportation but also minimizes
Principal components for each material. the overall expenses related to the recycling process, as the material is
Material Type Principal Components reused or repurposed immediately without additional handling or pro­
cessing steps.
Recycled Coarse Silica (SiO₂), Calcium Carbonate (CaCO₃), Calcium Silicate
Aggregates (RCA) Hydrate (C-S-H), Aluminum Oxide (Al₂O₃) Additionally, the container acts as a protective shield against adverse
Brick Silica (SiO₂), Aluminum Oxide (Al₂O₃), Iron Oxide (Fe₂O₃) weather conditions, ensuring the system’s functionality despite external
Foam Expanded Polystyrene ((C₈H₈)n), Blowing Agent (C₅H₁₂) environmental challenges. This safeguard is crucial for maintaining the
Glass Silica (SiO₂), Sodium Oxide (Na₂O), Calcium Oxide (CaO) accuracy of the sensors’ measurements by protecting the sensitive
Gypsum Calcium Sulfate Dihydrate (CaSO₄⋅2H₂O)
Mineral Fibers Silica (SiO₂), Calcium Oxide (CaO), Magnesium Oxide
equipment from damage. Furthermore, beyond its role in adverse
(MgO), Aluminum Oxide (Al₂O₃), Iron(III) Oxide (Fe₂O₃) weather conditions, the container’s presence demonstrates the system’s
Plastics Polyvinyl Chloride ((C₂H₃Cl)n) adaptability in different field conditions. Such adaptability not only
Wood Cellulose ((C₆H₁₀O₅)n), Hemicellulose (C₅H₈O₄), Lignin ensures the reliability of data but also strengthens the overall robustness,
(C₉H₁₀O₃)
resilience, and effectiveness. Consequently, it expands the applicability
Recycled Cement Paste Calcium Silicate Hydrate (C-S-H), Calcium Hydroxide (Ca
(RCP) (OH)₂), Ettringite (Ca₆Al₂(SO₄)₃(OH)₁₂⋅26H₂O), Calcium of the system across various demolition and recycling contexts.
Carbonate (CaCO₃) In summary, the containerized setup is fundamental in preserving
Recycled Fine Silica (SiO₂), Calcium Carbonate (CaCO₃), Residual data integrity, enhancing operational flexibility, and ensuring the sys­
Aggregates (RFA) Cement Paste (C-S-H, Ca(OH)₂) tem’s adaptability. These attributes are key to the technology’s effec­
tiveness in a variety of demolition scenarios.
2.2. Sensor-based quality inspection system
2.2.2. Composition
2.2.1. Containerization The sensor-based quality inspection system (Fig. 4) comprises two
The sensor-based quality inspection system, depicted in Fig. 3, plays primary sensors: a 3D scanner Gocator and a LIBS sensor. Both sensors
a critical role in the C2CA technology framework. Housed within a are positioned directly above the conveyor belt. The Gocator specializes
specialized container, this system performs several vital functions. in the granulometric analysis of RCA, pinpointing/measuring their PSD.
The container is partitioned into two areas: an inspection room and a This device is adept at generating high-resolution, three-dimensional
control room. These sections are separated to prevent any potential point cloud data, capturing detailed information about the surface to­
harm to operators from the lasers used during the inspection process. pology and granular distribution.
The inspection room is equipped with sensors for material analysis, and The LIBS sensor (Fig. 5) plays a crucial role in identifying contami­
it includes a vacuum system designed to reduce dust levels inside the nant compositions embedded within the RCA. It achieves this by
inspection room. Reducing dust is essential, as it enhances the sensors’ focusing ultra-short pulse lasers on the sample’s surface to create
accuracy by lessening their interference with dust particles. The control plasma, subsequently analyzing the emitted light spectrum from the
room is set up to receive and process various data collected in real-time. plasma to determine the material composition and content of the sam­
It also uses a monitoring system to oversee activities within the in­ ple. Specific experimental parameters used in this study are detailed in
spection room to ensure operational safety and efficiency. Table 2.
One of the primary advantages of the container is its mobility, which To mitigate interference caused by continuous laser-induced plasma
allows easy transport to various demolition sites. This mobility increases radiation, the detection system employed a time-resolved spectral
operational flexibility and optimizes resource allocation. The container acquisition method using temporal gating. This technique isolates the
is designed to facilitate the on-site recycling and testing of concrete emission from the plasma at its peak intensity, typically occurring
directly at demolition sites. By processing the demolished concrete on- within the first 20 ns post-laser excitation. A gated detector synchro­
site, the need to transport it to a remote facility is eliminated. This not nized with the laser pulses was used to collect only the desired spectral
data while excluding background radiation.
To prevent the loss of spectral information during high-speed LIBS
operations, several strategies were implemented:

(1) Optimized Laser Focus and Alignment: The laser was precisely
focused on the sample surface using the lens with a focal length of
300 mm. This ensured consistent plasma generation, avoiding
irregularities in the spectral data.
(2) High-Resolution Spectrometer: The LIBS system employed a
spectrometer with a resolution of 0.1 nm, capable of capturing
subtle variations in plasma emission.
(3) Temporal Gating: A temporal gate was applied to collect only the
most intense part of the plasma emission, reducing noise from
continuous plasma radiation or ambient light.
(4) Dust and Debris Mitigation: A vacuum system in the inspection
container reduced the presence of dust and debris, which could
obscure the laser-sample interaction or interfere with spectral
readings.

A noteworthy aspect of the system’s design is the ingenious incor­

poration of multiple reflective mirrors, which facilitates the simulta­
neous monitoring of RCA on two separate conveyor belts with the use of
only one Nd:TAG laser. This innovative approach not only reduces the
associated costs but also amplifies the system’s overall operational
efficacy.
Fig. 3. Sensor-based quality inspection system.

4
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 4. Sensor-based quality inspection system.

Fig. 5. Schematic diagram of LIBS.

2.2.3. Inspection
Table 2
The inspection process begins with the introduction of RCA into the
Experimental parameters.
system via a feeder, followed by their deposition onto the conveyor belt,
Parameter Value Description resulting in the formation of a triangular-shaped pile of RCA. This
Laser Pulse Wavelength 1064 Nd:YAG laser, suitable for high-precision triangular configuration is designed to ensure a uniform distribution of
(nm) spectroscopic analysis. the RCA, extending from the innermost region to the outer edges of the
Laser Pulse Frequency 100 Balances high-resolution data acquisition with pile (Fig. 6). This deliberate arrangement is particularly advantageous as
(Hz) system throughput.
Laser Pulse Energy 170 Generates plasma effectively while minimizing
it optimally facilitates the surface inspection conducted by the Gocator,
(mJ) surface damage. allowing for precise assessment and estimation of RCA properties.
Laser Pulse Duration 6 Provides a concentrated energy burst for plasma As RCA piles move along the conveyor belt, they are sequentially
(ns) generation. inspected by both the Gocator and LIBS sensors. The data generated
Focal Length of Lens 300 Optimizes laser energy density for uniform
during these inspections are instantaneously recorded in a computer
(mm) plasma generation.
Conveyor Belt Speed 0.529 Ensures uniform material inspection at high system and subsequently uploaded to a secure cloud storage platform for
(m/s) throughput. long-term archiving and retrieval. Additionally, monitoring data are
also linked to radio frequency identification (RFID) tags attached to the
piles to enhance traceability. The conveyor belt operates at a constant
velocity of 0.529 m/s, enabling a single conveyor belt to transport more

5
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

properties are derived from the dimensions and orientation of the fitted
ellipsoids. The key aspect of this ellipsoidal model is the choice of the
second shortest axis as the main parameter for measuring graduation
information. This particular axis is selected because it effectively rep­
resents the particle size, and consequently, helps in determining the
PSD. The use of this axis is beneficial because it strikes a balance, being
more informative than the shortest axis, which might be too small to
provide useful data, and less variable than the longest axis, which could
be too sensitive to minor changes in particle shape. This makes the
second shortest axis a reliable and representative measure for assessing
the characteristics of different particles in a sample.

2.3.2. 3D modeling with X-ray tomography

Fig. 6. Layered formation of RCA piles.
To further verify the accuracy of algorithms simulating particle
morphology, medical imaging techniques are employed to scan samples
of RCA piles. It is important to note that this imaging process is not part
than 50 tons of RCA per hour.
of the container technology but is instead used for offline verification
and calibration purposes. This enabled the creation of 3D models of their
2.3. Analysis methods interior to capture the actual particle morphology for comparison.
Computed Tomography (CT) imaging, a commonly used modality (Basu
2.3.1. Particle morphology et al., 2011), employs X-rays to acquire multiple angular projections of
The primary objective of the particle morphology analysis is to an object, which are then used to reconstruct the object’s linear atten­
achieve a statistically representative and geometrically characterization uation coefficient distribution. The resulting images are typically
of RCA particles, ensuring that the morphological features such as size, assembled into a series of consecutive axial slices arranged in parallel
shape, and orientation are comprehensively and reliably captured. This (Pelc, 2014). CT scan information is digitally archived, frequently in a
characterization is crucial for understanding the behavior and perfor­ format referred to as Digital Imaging and Communications in Medicine
mance of RCA in various applications, particularly where particle shape (DICOM). This format arranges the information into an organized
and distribution significantly impact material properties. collection that includes both the imaging data and related metadata
The feeding method for the RCA piles is meticulously designed to (Fajar et al., 2022). Metadata parameters like slice thickness, instance
ensure symmetry, which justifies focusing the analysis on only one side number, pixel spacing, rescale slope, and rescale intercept found in
of the piles. This symmetric feeding method guarantees that the particle DICOM files are employed during the data preprocessing phase.
distribution within the heap is uniform, allowing the segmentation and The Hounsfield unit (HU) scale, employed in CT imaging, quantifies
analysis process to be both efficient and representative of the entire pile. the radiodensity of tissues and materials. It sets the baseline with water
Employing the Fastscape algorithm (Braun and Willett, 2013), at 0 HU and air at roughly − 1000 HU, where substances of greater
originally developed for terrain analytics, the system performs a density show higher HU values. The process of converting the linear
watershed segmentation process (Steer et al., 2022) of 3D point clouds attenuation coefficient of each material at a specified effective energy
to accurately delineate RCA particles. The segmented regions are then into HU uses the standard equation:
encapsulated within ellipsoidal envelopes, enabling a quantifiable ( )
μmaterial − μwater
analysis of particle morphology, structure, and orientation. This method HU = × 1000 (1)
μwater
addresses challenges such as over-segmentation and ensures accurate
geometrical representation through ellipsoidal fitting, providing a where μ represents the linear attenuation coefficient. In the process
comprehensive statistical and geometrical description of RCA particles. known as Hounsfield scaling or CT number scaling (Huda and Slone,
The process involves several steps: 1996), raw attenuation values derived from CT scans are converted into
Initial segmentation: The procedure begins by applying the water­ Hounsfield units using the equation:
shed algorithm, traditionally used for 2D digital elevation models, to
segment global 3D point cloud data. This adaptation allows the algo­ HU = (PV × RS) + RI (2)
rithm to effectively delineate individual particles by treating peaks in
the data as watershed ridges, thereby dividing the data into distinct where PV signifies the pixel value, representing the original value
segments. attributed to a pixel within a CT scan. RS, recognized as the rescale
Segmentation correction: Commonly, the initial segmentation can slope, is a scaling factor employed to modify the pixel values accord­
result in over-segmentation, where particles are divided into smaller, ingly. Meanwhile, RI, identified as the rescale intercept, denotes a
unnecessary parts. To correct this, the method merges particles that are particular offset value applied to alter the pixel values.
closely located. This merging is based on two criteria: the proximity of Retrieving pixel arrays from DICOM files, organized by the instance
the particles’ summits and the alignment of their surface normals. number found in the metadata, involves adjusting each pixel’s value in
Additionally, particles that are excessively flat or elongated are the CT image based on the rescale slope and intercept values. This
removed. These steps refine the segmentation and enhance the overall process standardizes the pixel values to accurately represent the actual
quality. attenuation coefficients. Following this standardization, the pixel values
3D ellipsoidal fitting: After the particles have been segmented and are then transformed into HU.
appropriately labeled, the next step is to characterize their geometrical To facilitate a comprehensive comparison with algorithms that
properties. This is done by fitting 3D ellipsoids to each particle. Ellip­ simulate particle morphology, adopting a distinct methodology
soidal fitting involves a complex optimization process where the best- involving the CT scanning of samples becomes imperative. Conventional
fitting ellipsoid is calculated to approximate the shape and size of CT scanning techniques typically involve making equidistant vertical
each particle. This step is crucial as it quantifies the particle structures, incisions through the sample to capture cross-sectional imagery. How­
which can be vital for further analysis and applications. ever, to enhance the comparability, it is crucial that slices are made at
Morphological Analysis: This method provides detailed geometrical uniform intervals along an oblique plane parallel to the surface of the
information like the size, shape, and orientation of each particle. These sample. This procedure demands the compilation of CT scan images into

6
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

a cohesive ensemble, which is then used to construct a 3D model of the (2) Parallel processing for enhanced principal component analysis
sample, employing the transformed HU values. Subsequently, this model computations
undergoes re-sectioning to align with the comparative analysis
requirements. Each laser pulse generates a spectrum denoted as S = (s1 , s2 , …, sN ),
where si (i = 1, 2, …, N) represents the intensity of plasma emission at
2.3.3. Dice similarity coefficient wavelengths λi . N is the total number of measured wavelengths. This
To assess the extent of overlap between results obtained from algo­ method positions each laser pulse into an N-dimensional space, creating
rithmic simulations and X-ray tomography, the Dice similarity coeffi­ distinct clusters for different materials based on their spectral signa­
cient (DSC) is employed. The Dice coefficient is a statistical metric used tures. New laser pulses are either assimilated into existing clusters or
to measure similarity, often applied in image processing to gauge spatial identified as outliers based on how much their spectra deviate from the
overlap. Its values range from 0 (indicating no overlap) to 1 (indicating norm.
perfect overlap). This measurement method provides a quantitative way The axes of the coordinate system of S can be scaled and rotated to
to compare the accuracy and alignment of the two sets of data, facili­ simplify the multi-dimensional normal distribution of points within a
tating an objective evaluation of how closely the simulation results cluster. This simplification is achieved by defining a new orthonormal
mimic the X-ray tomography findings. coordinate system, represented by N unit vectors α = (α1 , α2 , …, αN ). In
The Dice coefficient (Dice, 1945; Sørensen, 1948), also known as the this new coordinate system, the multi-dimensional normal distribution
Sørensen–Dice coefficient, is a statistical tool used to measure the sim­ decomposes to N independent one-dimensional normal distributions,
ilarity between two samples. In the context of image processing, the Dice each aligned with a new axis. Thus, the spectrum in the database
coefficient is often used to quantify the spatial overlap between two transforms into the new coordinate system as:
images. The Dice coefficient is defined as:
ξ = (ξ1 , ξ2 , …, ξN ) = (S⋅α1 , S⋅α2 , …, S⋅αN ) (4)
2 × |A ∩ B|
DSC = (3) Effective categorization does not require all N dimensions; a smaller
|A| + |B|
number n is sufficient. This reduction minimizes the influence of irrel­
evant or noisy parts of the emission spectra. The significant information
where:
is projected into the leading dimensions of the new coordinate system.
|A ∩ B| is the size of the intersection of two sets (in the context of
After the principal component analysis (PCA) process, the spectral
images, these would be pixel sets) – basically, the number of pixels that
database for a cluster consists of the number n, the principal component
are classified as the foreground (or as a particular object) in both images.
|A| is the number of pixels classified as the foreground in the first image. vectors (ξ1 , ξ2 , …, ξn ), and a center point ξ with variances σ2 , describing
|B| is the number of pixels classified as the foreground in the second the multi-dimensional normal distribution of the spectra.
image. By leveraging optimized Basic Linear Algebra Subprograms (BLAS)
The value of the Dice coefficient ranges from 0 to 1, where 0 in­ and Linear Algebra PACKage (LAPACK) libraries and implementing
dicates no overlap and 1 indicates perfect overlap. This coefficient is parallel processing (Abdelfattah et al., 2021; Psarras et al., 2022), we
particularly useful as it quantifies the similarity between two binary can significantly enhance the efficiency of PCA computations. These
images. In an ideal case, if the predicted segmentation matches the improvements are crucial for handling large datasets, particularly in
ground truth segmentation exactly, the Dice coefficient would equal 1. applications involving spectral data analysis, where computational de­
mands are high.
2.3.4. Contaminant detection The BLAS and LAPACK libraries are fundamental tools for perform­
Contaminant detection in RCA is achieved by employing spectral ing efficient linear algebra computations (Frison et al., 2018; Psarras
analysis. This study builds upon the cluster-based identification algo­ et al., 2022). BLAS provides low-level routines for common operations
rithm (Chang et al., 2022), which improved data representation. The such as vector addition, scalar multiplication, dot products, and matrix
current research focuses on refining this algorithm to further increase its multiplication. Optimized implementations, such as OpenBLAS and Intel
operational efficiency and adaptability. These refinements lead to more Math Kernel Library (MKL) (Frison et al., 2020; Yamazaki et al., 2018),
reliable contaminant identification by improving the accuracy and exploit modern CPU architectures to deliver significant performance
speed of the detection process. In the dynamic field of spectroscopic improvements through multi-threading and vectorization. LAPACK ex­
analysis, environmental factors influence LIBS spectra, making adapt­ tends BLAS functionalities by offering routines for more complex linear
ability to new data crucial for maintaining model accuracy and rele­ algebra problems, including solving linear systems, eigenvalue prob­
vance (T. Chen et al., 2020; Wang et al., 2021). Therefore, periodic lems, and singular value decomposition (SVD). By leveraging optimized
calibration is necessary. The incremental learning technique provides a BLAS libraries, LAPACK routines can achieve high performance across
practical approach for real-time spectral data analysis. This method various hardware architectures. Parallel computing libraries can paral­
significantly enhances performance and flexibility by eliminating the lelize independent tasks like computations on different data chunks.
need for complete model retraining, making it well-suited for applica­ Optimized BLAS libraries internally use multi-threading for operations
tions in environments where data is continuously generated. like matrix multiplication, which can be configured to utilize multiple
threads.
(1) Data preprocessing: To efficiently process spectral data, it is
necessary to preprocess the raw data. Analysis has shown that (3) Cluster-based identification
focusing on the wavelength range of 200–900 nm is sufficient to
achieve the desired outcomes, thereby improving computational Post-PCA, the database catalogs a material by assigning a finite
efficiency. This preprocessing includes standardizing the spectral number of principal components, n. This step defines a new coordinate
values to ensure uniform magnitude scales across different system through n orthonormal vectors, facilitating the depiction of
datasets, thereby highlighting unique data characteristics. Z- spectral data as multi-dimensional normal distributions around a
score standardization is used to maintain the data’s distribution centroid. Following normalization, principal component values are
while aligning its mean and standard deviation to zero and one, assumed to follow a chi-square distribution, which is crucial for classi­
respectively. This standardization captures essential data char­ fying materials based on their spectral data. The conformity to the ex­
acteristics, such as the distribution patterns of peaks and troughs. pected chi-square distribution is verified through P-value analysis,
determining the significance level for recognizing specific materials.

7
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Each component ξm (m = 1, 2, …, n) in the set (ξ1 , ξ2 , …, ξn ) follows a

1[ ]
normal distribution with a mean of ξm and a variance of σ2m . After σ2t = σ 2t− 1 + (ξt − ξt− 1 )2 − σ2t− 1 (10)
t
applying z-score standardization, each standardized value Zm is calcu­
lated as: The chi-square statistic for the new observation ξt is computed as:
n
( )2
ξ − ξ ∑ ξt,i − ξt,i
Zm = m√̅̅̅̅̅̅m (5) χ 2t = 2
(11)
σ2m i=1
σ t,i

These standardized values follow a normal distribution with a mean By continuously updating ξt and σ 2t , the model dynamically adjusts to
of 0 and a variance of 1. To test if the data comes from a χ n 2 -distribution, reflect the latest data distribution, thereby improving classification ac­
the chi-square value is calculated as: curacy and robustness.
This method demonstrated a substantial reduction in computational
n
∑ n
∑ (ξm − ξm )2 load compared to traditional batch PCA, while maintaining similar
χ2 = Zm 2 = (6)
m=1 m=1
σ2m levels of dimensionality reduction efficacy. Additionally, the dynamic
This chi-square value helps determine the likelihood that the data updating of chi-square parameters facilitated more accurate and
responsive classification, particularly as new data was introduced.
follows the expected χ n 2 -distribution. Each χ 2 can be converted into a
probability P-value for the χ n 2 -distribution. A higher χ 2 value corre­
3. Results and discussion
sponds to a lower P-value, indicating greater confidence that the data
does not fit the expected distribution. P-values below a predetermined
3.1. PSD calculation
significance level indicate statistical significance. When this significance
level is set as the acceptance criterion for a particular material, any
3.1.1. 3D point cloud data
spectra with P-value exceeding this threshold will be regarded as orig­
Given the observed uniformity and consistency in the distribution
inating from that material.
pattern of constituents within the RCA piles, it becomes feasible to
extrapolate the PSD characteristics of the entire pile through a focused
(4) Incremental Learning
analysis of the PSD associated exclusively with the outer surface layer of
the pile. This method presents an effective means of obtaining a repre­
Implement incremental learning techniques to update the model as
sentative understanding of the overall PSD of the RCA piles. Using the
new spectral data becomes available. This can be achieved through
Gocator scanner, a comprehensive scan of the RCA pile’s external sur­
online PCA methods and updating chi-square distribution parameters
face was conducted, yielding detailed 3D point cloud data (Fig. 7 (a)).
dynamically.
The spatial resolution of this 3D point cloud data is crucial as it af­
Given a new spectral observation St at time t (represents the index of
fects the accuracy and detail of the RCA particle measurements. In this
the data point being introduced, i.e., it is the time step or the sequential
context, it’s paramount to note that the resolution of the point clouds
position of the data point), the mean vector μt and covariance matrix Ct
under study demonstrated discernible variances across different spatial
of the spectral data in the original N-dimensional space are updated
orientations—a phenomenon that has broader implications for the
using the following equations:
precision of granulometric assessments.
1 In the direction parallel to the motion of the conveyor belt, the res­
μt = μt− 1 + (St − μt− 1 ) (7)
t olution of 3D point cloud data is influenced by two factors. The first
factor is the velocity of the conveyor belt, controlled by adjusting the
1[ ]
rotational speed of the motor drive. This rotational speed adjustment is
Ct = Ct− 1 + (St − μt− 1 )(St − μt )T − Ct− 1 (8)
t achieved through the manipulation of the motor’s output frequency and
Here, the covariance matrix Ct is updated prior to performing PCA, the number of poles. Specifically, the motor drive in this study operates
capturing the variability of the spectral data in the original N-dimen­ at 50 Hz with a 4-pole design, resulting in a rotational speed of 1500
sional space. This updated covariance matrix reflects the latest statistical rpm. Mechanical adjustments include a gearbox ratio of 19 and a wheel
characteristics of the data, which is essential for accurately calculating perimeter of 401.92 mm, resulting in a conveyor belt speed of 0.529 m/
the principal components. Subsequently, the principal components are s. The second factor affecting the resolution is the encoder resolution of
extracted from the updated covariance matrix Ct using eigenvalue the Gocator, which was calibrated to record 1024 ticks per revolution,
decomposition. In this approach, earlier spectral data are given the same translating to a point cloud resolution of 0.3925 mm along the conveyor
weight as later spectral data, ensuring equal consideration throughout belt’s path.
the time series. However, it is important to note that alternative methods Perpendicular to the conveyor belt, across its width, the resolution
exist, which prioritize more recent spectral data by assigning them varies between 0.375 mm and 1.1 mm. This range is inherently tied to
higher weights, thereby adapting more swiftly to recent changes in the the Gocator’s field of view at any given point in time. For accuracy, the
data. transverse resolution was standardized at 0.375 mm in this study.
Following PCA, the spectral data is transformed into a lower- Vertically, concerning the height or depth of the RCA piles, the
dimensional space where each principal component follows a normal Gocator’s advanced internal mechanisms come into play. The system
distribution. The chi-square distribution is then employed to evaluate provides a resolution gradient from a fine 0.092 mm to a coarser 0.488
the goodness-of-fit for new data points. To maintain the accuracy of this mm. This variability highlights the scanner’s versatility and adaptability
statistical test, it is essential to dynamically update the chi-square dis­ in handling different granulometric scenarios, establishing it as a crucial
tribution parameters as new data is integrated. tool in the comprehensive evaluation of RCA piles. In this study, the
Let ξt represent the vector of principal components at time t. The resolution was set at 0.092 mm in the vertical dimension.
mean ξt and variance σ 2t of the principal components are updated
3.1.2. PSD calculation
incrementally:
The PSD of RCA piles was determined using the method described
1 previously. This involved fitting the 3D point cloud data of the RCA piles
ξt = ξt− 1 + (ξt − ξt− 1 ) (9)
t into ellipsoidal models (Fig. 7 (b)) to capture morphological data for
each particle. The resulting PSD was effectively demonstrated through a

8
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 7. Point cloud processing.

Fig. 8. Cumulative percentage retained graphs.

cumulative percentage retained graph, as depicted in Fig. 8. This graph, trials. The results indicate a minimal difference between the two
supported by calculations of ellipsoidal volume and the RCA’s apparent methods, affirming a high degree of accuracy in the surface-based
density, provided a clear visualization of the PSD characteristics technique.
inherent to the RCA piles. The insights gained from this analysis are To further assess the precision and efficacy of this surface-based PSD
crucial for understanding the aggregate structure and its potential in­ measurement technique, the Root Mean Square Error was calculated
fluence on the performance of concrete. between the predicted and manually derived values, resulting in de­
Subsequently, a cross-referential analysis was conducted to validate viations of 4.93%, 5.38%, and 4.27% across three experiments. These
the accuracy and reliability of this non-intrusive, surface-based tech­ low Root Mean Square Error values confirm the robustness and precision
nique for PSD determination. This entailed comparing the PSD results of this surface-based technique in estimating the PSD of RCA piles on a
from the 3D point cloud data with those obtained through the tradi­ conveyor belt. Initial results demonstrate a high degree of concordance
tional, more invasive method - manual sieving. The comparison aimed between the two methods, suggesting that the 3D scanning approach
to ascertain the degree of correlation and consistency between the sur­ could effectively approximate the comprehensive PSD with a significant
face PSD measurements and the actual overall PSD of the entire RCA reduction in manual effort and time.
piles. Fig. 8 depicts this comparative study, showing the cumulative
percentage retained curves as predicted from 3D point cloud analytics 3.1.3. X-ray tomography validation
against those obtained from manual sieving, based on pilot scanning To further validate this surface-based PSD measurement technique,

Fig. 9. RCA sample.

9
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

X-ray tomography was employed to capture the morphology of particles through layered scanning with X-ray tomography. This process involved
beneath the surface of RCA plies. The internal structure captured by X- compiling each scanned layer and isolating the particle components
ray tomography was compared to the ellipsoids simulated from 3D point using thresholding. Both models were incrementally sliced from top to
cloud data. A black box (as shown in Fig. 9 (a)) was used to simulate the bottom, parallel to the inclined surface, allowing observation of each
inclined surface on one side of the formed piles, with RCA spread inside. layer’s cross-section. Then the two obtained cross-sectional images of
The sample was scanned using both 3D scanning and X-ray tomography each layer were overlapped for comparative analysis. Six representative
techniques. The data processed from these methods are shown in Fig. 9 layers were selected in Fig. 10, following the top-down slicing depth. On
(b) and (c) respectively. the left are the cross-sectional details of the simulated ellipsoidal model.
The ellipsoidal fitting model illustrated in Fig. 9 (b) originated from In the middle are the cross-sectional details derived from the 3D model
3D point cloud data obtained through 3D scanning with the Gocator. constructed using X-ray tomography, reflecting the real particle distri­
This process involved segmenting the 3D point cloud data and fitting it bution. On the right are the overlapping cross-sections of the two cutting
to the ellipsoidal shape. The 3D model in Fig. 9 (c) was generated methods for comparison.

Fig. 10. Cross-sectional details.

10
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 10. (continued).

To assess the similarity between cross-sections obtained by the two presence of numerous smaller particles in that area. 3D scanning only
methods, the DSC was employed. Following the sequence depicted in captures information from the top layer, hence forming a single layer of
Fig. 10, from (a)–(f), the DSC values are 0.63, 0.71, 0.82, 0.85, 0.86, and smaller ellipsoids. In contrast, X-ray tomography does not encounter this
0.82. From the selected sections, an incremental increase in overlap limitation, as additional particles fill the same location, preventing the
between the two methods is evident, peaking at 0.86 before slightly appearance of blanks.
decreasing. Initially, lower overlap occurs due to irregularities on the This comparative analysis further validates the reliability of surface-
particle surfaces, challenging their representation in the simulated el­ based PSD measurement techniques. Consequently, the Gocator 3D
lipsoids. As the sections progress, the ellipsoids better emulate the in­ scanning can be employed for a quick and convenient estimation of PSD.
ternal particle distribution, maintaining an overlap of around 0.8.
However, in deeper sections of the ellipsoids, a gradual appearance of
blank spaces in the lower-left region is observed. This is attributed to the

11
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

3.2. Contaminant detection reflective mirrors. Spectral measurements from a wide range of mate­
rials were recorded under two conditions: while the conveyor belt was
This study aimed to evaluate the performance differences of the LIBS running and while it was stationary.
sensor in capturing spectra from objects under varying environmental Fig. 11 provides a visual representation of the findings. The results
conditions. The focus was on assessing its ability to adapt to new data to reveal minimal variation in the spectra from the materials, regardless of
maintain model accuracy and relevance in dynamic spectroscopic the conveyor belt’s motion. The minor differences in spectra observed
analysis. Additionally, the study aimed to understand how the move­ between the moving and stationary states of the conveyor belt do not
ment of the conveyor belt might affect the effectiveness of the system’s significantly impact the overall system performance. This uniformity

Fig. 11. Comparison of spectra of different materials in the operation and stationary states of the conveyor belt.

12
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 11. (continued).

demonstrates the reliability and effectiveness of the system’s design. It consistency is critical.
also supports the system’s ability to perform consistently under different Despite the general consistency, there were slight spectral differ­
operational scenarios. Its resilience under different operational condi­ ences noted between the moving and stationary states of the conveyor
tions underscores its potential as a reliable tool for real-time industrial belt. These differences, though minor, could lead to errors when clas­
applications, ensuring consistent and accurate data acquisition regard­ sifying materials with similar chemical compositions, such as recycled
less of conveyor belt activity. This is paramount for industries where cement paste powder (RCP), recycled fine aggregates (RFA), RCA, and
conveyor belt speeds might vary, and where maintaining measurement others. These subtle variations, although not significantly affecting

13
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 11. (continued).

overall system performance, must be considered in detailed analyses to robust performance metrics for this new system: achieving an accuracy
avoid misclassification and ensure accurate results. rate of 0.94, a weighted average precision of 0.95, a weighted average
After acquiring spectra from various materials, a systematic classi­ recall of 0.94, and an F1-score (weighted average) standing at 0.95 on
fication was conducted using a cluster-based identification algorithm the validation dataset.
enhanced by incremental learning techniques. This algorithm, recog­ These performance metrics confirm the effectiveness of LIBS as an
nized for its ability to group data by identifying inherent similarities, analytical tool, especially for analyzing recycled concrete aggregates
was applied to discern patterns within the updated spectral data. The within industrial contexts. These metrics highlight LIBS’s ability to
model’s performance underwent a thorough assessment to ensure its deliver both accuracy and precision quickly, without compromising the
reliability for practical applications. Remarkably, the model showcased quality of the outcomes. By adopting these advanced analytical

14
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 11. (continued).

techniques, industries can confidently make strides forward, optimizing experimental variables require closer examination to assess their influ­
the use of recycled materials and maintaining high-quality standards. ence on system stability and accuracy. This section delves into key
factors such as conveyor speed, sensor resolution, and environmental
conditions, providing insights into the system’s robustness and
3.3. Discussion on experimental conditions and limitations adaptability.

While the proposed sensor-based quality inspection system demon­

strates significant advancements in RCA quality assurance, several

15
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 11. (continued).

3.3.1. Conveyor speed data fidelity. Future studies should examine the trade-offs associated
The conveyor belt speed was maintained at a constant velocity of with different speeds, particularly in dynamic industrial settings where
0.529 m/s during all experiments. This speed was chosen based on an operational throughput requirements vary.
optimization process that balances throughput and inspection accuracy.
However, it is crucial to consider how variations in conveyor speed may 3.3.2. Sensor resolution
affect sensor performance. At higher speeds, potential challenges The spatial resolution of the 3D point cloud data varied depending on
include reduced time for laser-sample interactions in the LIBS sensor and the orientation relative to the conveyor belt. Horizontal resolution
lower resolution of 3D point cloud data from the Gocator scanner. ranged from 0.375 mm to 1.1 mm, while vertical resolution was set at
Conversely, slower speeds could increase inspection time but improve 0.092 mm. These settings ensured detailed granulometric analysis but

16
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 11. (continued).

may not account for scenarios involving smaller particles or high-density 3.3.3. Environmental conditions
piles. Additionally, the LIBS sensor’s temporal gating technique effec­ Environmental factors such as dust, lighting, and temperature fluc­
tively mitigated noise but required precise synchronization to achieve tuations were controlled during experiments through the use of a vac­
optimal spectral acquisition. Investigating how resolution changes uum system and an enclosed container. While these measures ensured
impact data accuracy under different particle size distributions and pile accurate sensor readings, real-world conditions at demolition sites may
configurations is essential for refining system performance. pose additional challenges. To provide a comprehensive understanding
of the sensor-based quality inspection system’s performance in practical

17
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Fig. 11. (continued).

scenarios, the following environmental factors and mitigation strategies alignment of optical components or the performance of electronic sys­
are discussed: tems, potentially leading to calibration drift.
Mitigation: Laboratory tests indicate that the system maintains stable
(1) Dust and debris interference performance within a broad operational temperature range. Incorpo­
rating temperature-compensating sensors or materials can enhance
Impact: High levels of dust and debris, common in demolition sites, resilience in extreme climates.
can obscure the laser-sample interaction, reducing the accuracy of both
3D scanning and LIBS measurements. (4) Material surface moisture
Mitigation: The system employs a vacuum-based dust suppression
mechanism, which enhances visibility and minimizes interference dur­ Impact: Surface moisture on RCA particles can alter the spectral
ing data acquisition. Experimental results confirm a significant response captured by LIBS, potentially leading to the misclassification of
improvement in measurement consistency when dust levels are reduced. contaminants.
Future designs could integrate advanced filtration systems to further Mitigation: Pre-drying mechanisms or spectral correction algorithms
mitigate this challenge. can be introduced to account for moisture-induced spectral variations,
ensuring more accurate material classification.
(2) Ambient lighting variations
(5) Vibrations and conveyor belt motion
Impact: Changes in ambient lighting, especially in outdoor applica­
tions, could potentially affect the optical components of the Gocator and Impact: Vibrations from on-site machinery or inconsistencies in
LIBS systems. conveyor belt motion may affect the stability of sensor readings, espe­
Mitigation: The use of enclosed, containerized setups effectively cially during LIBS operations.
shields the sensors from direct sunlight or variable lighting conditions. Mitigation: The use of time-resolved spectral acquisition and syn­
However, integrating adaptive calibration algorithms that adjust for chronization with belt movement ensures minimal signal distortion.
residual lighting variability can further enhance robustness. Additional mechanical stabilizers could further reduce the impact of
vibration-induced inaccuracies.
(3) Temperature fluctuations Understanding and mitigating these environmental factors are crit­
ical for deploying the system in diverse operational scenarios. While
Impact: Extreme temperature variations may influence the initial validations demonstrate the system’s robustness, field trials in

18
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

harsher environments could provide further insights into its adapt­ 4.2. Contaminant detection
ability. By integrating advanced calibration techniques and environ­
mental compensation mechanisms, the system’s accuracy and reliability In the field of contaminant detection using sensors for industrial
can be sustained across varied conditions. applications, most studies have concentrated on identifying contami­
nants in images through visual sensors. However, few have specifically
3.3.4. Scenario variations and system adaptability addressed the challenges of detecting contaminants in RCA under real-
To evaluate the robustness of the system, further testing under world industrial conditions. For instance, Bobulski and Kubanek
diverse scenarios is recommended. For instance: (Bobulski and Kubanek, 2021) employed a conveyor belt system inte­
grated with a microcomputer for image processing to classify plastic
(1) Material Heterogeneity: The presence of varied contaminants, waste into four categories, achieving an accuracy of 74%. Lohumi et al.
such as metals or polymers, may influence LIBS classification (2021) developed an online inspection system using fluorescence and
accuracy, especially for materials with overlapping spectral color imaging to detect foreign materials unintentionally introduced
signatures. into fresh-cut vegetables. Nevertheless, their method failed to identify
(2) Pile Formation: The triangular configuration of RCA piles ensures contaminants such as transparent low-density polyethylene, Styrofoam,
uniform inspection but may not represent alternative deposition small stepper pins, and needles. Mewada et al. (2024) applied advanced
methods. Analyzing different pile geometries would help assess computer vision models for contamination detection in densely clut­
system scalability. tered waste environments. The YOLOv8-x model achieved a mean
(3) Operational Scale: Scaling up the system to higher throughput average precision of 0.463 with the aid of transfer learning.
levels or multiple conveyor belts might reveal bottlenecks in data Compared to other sensors, the LIBS sensor offers significant ad­
processing or sensor synchronization. vantages, including versatility, multi-element detection, and suitability
for field applications (Brunnbauer et al., 2023). The LIBS sensor is
The current study does not extensively quantify the effect of these commonly used for real-time detection of rapid changes in material
variables across all possible operational conditions. While preliminary composition, especially in the metalworking industry, where it aids in
results suggest robustness in controlled environments, real-world ap­ the classification of steel, various alloys, slag, and metal scrap (X. Chen
plications may reveal additional constraints. Future work should et al., 2024; H. Kim et al., 2021; Park et al., 2021). While LIBS sensor
incorporate advanced modeling and simulation techniques to predict excels in identifying components within a single material type, its
performance under varying conditions. Additionally, integrating adap­ application for real-time detecting diverse substances remains limited.
tive algorithms to adjust sensor parameters dynamically could further Furthermore, its use in industrial settings for rapid contaminant detec­
enhance system reliability in fluctuating environments. tion in RCA on conveyor belts is rare. For example, Huber et al. (2014)
demonstrated the use of the LIBS sensor for quasi-real-time identifica­
4. Comparison to contemporary work tion of chlorine-containing waste polymers in an industrial sorting plant.
This system enabled sorting based on chlorine concentration, particu­
4.1. PSD analysis larly for polyvinyl chloride. Merk et al. (2015) developed a LIBS-based
system for metal scrap identification directly on a conveyor belt. Simi­
Several recent studies (Bai et al., 2021; Fu and Aldrich, 2023; larly, Werheit et al. (2011)described the use of the LIBS sensor to
Hamzeloo et al., 2014; Z. Zhang et al., 2020) have primarily focused on identify aluminum alloys in post-consumer scrap during conveyor
using 2D image analysis for particle size detection in materials such as motion.
copper ore, coal, and rock. However, there is limited research specif­ Despite these advances, implementing effective algorithms to pro­
ically targeting particle size detection for RCA. In these studies, mate­ cess the spectra captured by the LIBS sensor is critical to ensure robust
rials are typically spread flat on a conveyor belt, and in some cases, the performance in dynamic industrial environments. Arbitrary sample
particles do not overlap. This arrangement results from the inherent shapes, surface contamination, and movement increase signal fluctua­
limitation of conveyor belt systems, which expose only one side of the tions compared to static conditions. Additionally, uncontrollable factors,
material. Moreover, these methods are rarely scaled for on-site, high-­ such as environmental variability, can lead to data sparsity, adversely
throughput applications, leading to inefficiencies in industrial contexts. affecting classification model performance.
Additionally, most studies emphasize PSD of surface-level particles Our system achieves comparable accuracy to existing methods while
while neglecting the distribution of entire particle batches. Lu et al. excelling in adaptability. The LIBS sensor’s ability to analyze materials
(2024) proposed an improved lightweight rock object detection method in real time under dynamic conditions offers a significant advantage for
using sieve sizes ranging from 20 mm to 200 mm for particle size studies. industrial applications. Unlike many systems that require stationary
While this method improves upon approaches, it struggles with smaller samples for precise detection, the proposed system maintains consistent
particles. Particles near or below 20 mm are often unclear in the images, performance during conveyor belt motion. Spectral measurements
resulting in a relative measurement error of up to 14.13%. Moreover, the reveal negligible variations between stationary and moving states,
distribution of the finer fractions is estimated using fitting models rather confirming its robustness. The incorporation of an incremental learning
than directly measured. This limitation is particularly significant for algorithm enables continuous adaptation to new spectral data without
RCA, where particle sizes typically range from 2.0 mm to 22.4 mm—a the need for retraining from scratch. Compared to other spectroscopic
size range inadequately addressed by current techniques. systems, this feature is particularly advantageous in environments with
To address these limitations, our system collects representative data diverse contaminants and evolving material compositions.
to achieve integrating portability, high throughput, and precise PSD
evaluation, making it suitable for industrial deployment. The system 5. Limitations and future orientations
processes over 50 tons of RCA per hour on a single conveyor belt,
marking a substantial improvement over laboratory-based methods, 5.1. Limitations
which are inherently constrained in scalability. Additionally, the results
can be immediately uploaded to cloud systems for further processing While the proposed sensor-based quality inspection system demon­
and traceability, facilitating faster decision-making and improved strates substantial advancements in the processing and evaluation of
quality control. RCA, it is not without limitations. These include:

19
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

(1) Operational Constraints: The system’s performance was evalu­ (3) Sustainability and Efficiency: The containerized design ensures
ated under controlled demolition site conditions. Variability in operational flexibility and enables on-site recycling at demolition
environmental conditions such as extreme weather, dust, and sites. This reduces transportation costs, minimizes environmental
unexpected debris might impact system accuracy and reliability. impact, and aligns with circular economy principles, promoting
(2) Contaminant Spectrum: Although the LIBS sensor accurately sustainable construction practices.
identifies a wide range of contaminants, certain low-abundance (4) Validation and Reliability: Comprehensive validation using X-ray
or novel contaminants may remain undetected due to spectral tomography and manual sieving confirms the accuracy of PSD
overlap or insufficient training data. measurements. The Root Mean Square Error values—below
(3) Computational Demand: The real-time nature of the system ne­ 5.5%—highlight the system’s robustness in PSD estimation.
cessitates significant computational resources, which may limit Similarly, the LIBS sensor’s performance was validated under
deployment in resource-constrained environments. dynamic conditions, demonstrating minimal spectral variation
(4) Scalability Challenges: While the current configuration processes between moving and stationary states of the conveyor belt. The
over 100 tons of RCA per hour, further scaling to larger demoli­ system achieved an accuracy rate of 0.94, a weighted average
tion sites or simultaneous multi-site operations could pose precision of 0.95, and a weighted F1-score of 0.95 on the vali­
logistical and technical challenges. dation dataset, underscoring its reliability in real-time contami­
(5) Economic Viability: The cost-effectiveness of implementing the nant detection. These results collectively affirm the system’s high
system on a large scale, particularly in regions with limited re­ precision and reliability for both PSD and contaminant analysis.
sources, remains to be further assessed. (5) Applications and Impact: The system’s adaptability and scal­
ability extend beyond construction to other industries, such as
5.2. Future orientations mining and waste management. Its ability to deliver real-time
data supports informed decision-making, ensuring consistent
To address the limitations and build upon the current findings, the material quality and compliance with industry standards.
following directions are proposed: (6) Future Implications: The demonstrated precision and consistency
of LIBS in contaminant detection, even during conveyor motion,
(1) Environmental Adaptability: Developing robust calibration al­ underscore its potential for widespread industrial adoption. In­
gorithms and protective measures to enhance system perfor­ cremental learning capabilities provide a foundation for adapting
mance under diverse environmental conditions, including to evolving material compositions, ensuring long-term relevance
temperature extremes, high humidity, and airborne particulates. and applicability.
(2) Enhanced Contaminant Detection: Expanding the spectral data­ (7) Contributions to the Field: This research highlights the integra­
base of contaminants using advanced machine learning tech­ tion of cutting-edge technologies into RCA processing, bridging
niques to improve detection accuracy for rare or emerging the gap between innovation and practical application. It advances
contaminants. the sustainable agenda by harmonizing efficiency and precision,
(3) Distributed Computing Models: Exploring cloud-based distrib­ paving the way for eco-friendly infrastructure development
uted computing solutions to optimize the processing of high- without compromising quality standards.
dimensional spectral data, reducing on-site computational
burden. CRediT authorship contribution statement
(4) Modular Scalability: Engineering modular systems that can be
easily configured for larger-scale or multi-site operations, Cheng Chang: Writing – original draft, Visualization, Validation,
enabling seamless scalability. Software, Methodology, Investigation, Formal analysis, Data curation,
(5) Lifecycle Assessment: Conducting comprehensive cost-benefit Conceptualization. Francesco Di Maio: Writing – review & editing,
and lifecycle assessments to evaluate the economic and envi­ Supervision, Project administration, Methodology, Funding acquisition,
ronmental impacts of deploying the system at scale. Conceptualization. Rajeev Bheemireddy: Software, Data curation.
(6) Integration with Circular Economy Practices: Investigating the Perry Posthoorn: Software, Data curation. Abraham T. Gebremariam:
integration of real-time RCA data with broader circular economy Writing – review & editing, Resources. Peter Rem: Writing – review &
frameworks to optimize material recovery and reuse in con­ editing, Supervision, Project administration, Methodology, Funding
struction projects. acquisition, Conceptualization.
(7) Cross-Disciplinary Applications: Exploring applications of the
sensor-based system in other industries such as mining, waste Fundings
management, and resource recovery, to enhance its versatility
and impact. This research was supported by the European Union’s Horizon fun­
ded Project "Innovative Circular Economy Based solutions demon­
6. Conclusion strating the Efficient Recovery of valuable material resources from the
Generation of representative end-of-life building materials" (ICEBERG,
(1) Innovation in Quality Inspection: This study introduces an grant agreement No. 869336).
advanced, mobile sensor-based quality inspection system for
RCA, demonstrating its capability to process over 100 tons per Declaration of competing interest
hour. The integration of 3D scanning and LIBS provides a
comprehensive evaluation of PSD and contaminant detection, The authors declare that they have no known competing financial
setting a new benchmark for non-intrusive RCA quality interests or personal relationships that could have appeared to influence
assessment. the work reported in this paper.
(2) Technological Advancements: By utilizing incremental learning
techniques and cloud-based data management, the system en­ Acknowledgment
sures continuous model updates without requiring complete
retraining. This innovation significantly enhances computational The authors acknowledge P.M. Meijvogel-de Koning from the Lab
efficiency and sustains high classification performance, address­ Geoscience and Engineering at Delft University of Technology for her
ing the limitations of traditional RCA assessment methods. assistance with X-ray tomography.

20
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Data availability Frison, G., Sartor, T., Zanelli, A., Diehl, M., 2020. The BLAS API of BLASFEO: optimizing
performance for small matrices. ACM Trans. Math Software 46 (2). [Link]
10.1145/3378671.
Data will be made available on request. Fu, Y., Aldrich, C., 2023. Online particle size analysis on conveyor belts with dense
convolutional neural networks. Miner. Eng. 193, 108019. [Link]
References [Link].2023.108019.
Galán, B., Viguri, J.R., Cifrian, E., Dosal, E., Andres, A., 2019. Influence of input streams
on the construction and demolition waste (CDW) recycling performance of basic and
Abdelfattah, A., Costa, T., Dongarra, J., Gates, M., Haidar, A., Hammarling, S., advanced treatment plants. J. Clean. Prod. 236, 117523. [Link]
Higham, N.J., Kurzak, J., Luszczek, P., Tomov, S., Zounon, M., 2021. A set of batched jclepro.2019.06.354.
basic linear algebra Subprograms and LAPACK routines. ACM Trans. Math Software Gálvez-Martos, J.L., Styles, D., Schoenberger, H., Zeschmar-Lahl, B., 2018. Construction
47 (3). [Link] and demolition waste best management practice in Europe. Resour. Conserv. Recycl.
Akhtar, A., Sarmah, A.K., 2018. Construction and demolition waste generation and 136, 166–178. [Link]
properties of recycled aggregate concrete: a global perspective. J. Clean. Prod. 186, Hamzeloo, E., Massinaei, M., Mehrshad, N., 2014. Estimation of particle size distribution
262–281. [Link] on an industrial conveyor belt using image analysis and neural networks. Powder
Al Martini, S., Sabouni, R., Khartabil, A., Wakjira, T.G., Shahria Alam, M., 2023. Technol. 261, 185–190. [Link]
Development and strength prediction of sustainable concrete having binary and Huber, N., Eschlböck-Fuchs, S., Scherndl, H., Freimund, A., Heitz, J., Pedarnig, J.D.,
ternary cementitious blends and incorporating recycled aggregates from demolished 2014. In-line measurements of chlorine containing polymers in an industrial waste
UAE buildings: experimental and machine learning-based studies. Construct. Build. sorting plant by laser-induced breakdown spectroscopy. Appl. Surf. Sci. 302,
Mater. 380, 131278. [Link] 280–285. [Link]
Alaejos, Pilar, de Juan, M.S., Rueda, J., Drummond, R., Valero, I., 2013. Quality Huda, W., Slone, R., 1996. Review of radiologic physics. Phys. Med. Biol. 41 (12), 2807.
assurance of recycled aggregates. In: Vázquez, E. (Ed.), Progress of Recycling in the Kabirifar, K., Mojtahedi, M., Wang, C., Tam, V.W.Y., 2020. Construction and demolition
Built Environment: Final Report of the RILEM Technical Committee 217-PRE. waste management contributing factors coupled with reduce, reuse, and recycle
Springer, Netherlands, pp. 229–273. [Link] strategies for effective waste management: a review. J. Clean. Prod. 263, 121265.
Amesho, K.T.T., Chinglenthoiba, C., Samsudin, M.S.A.B., Lani, M.N., Pandey, A., [Link]
Desa, M.N.M., Suresh, V., 2023. Microplastics in the environment: an urgent need for Kim, H., Lee, J., Srivastava, E., Shin, S., Jeong, S., Hwang, E., 2021. Front-end signal
coordinated waste management policies and strategies. J. Environ. Manag. 344, processing for metal scrap classification using online measurements based on laser-
118713. [Link] induced breakdown spectroscopy. Spectrochim. Acta B Atom Spectrosc. 184,
Aslam, M.S., Huang, B., Cui, L., 2020. Review of construction and demolition waste 106282. [Link]
management in China and USA. J. Environ. Manag. 264, 110445. [Link] Kim, J., 2022. Influence of quality of recycled aggregates on the mechanical properties of
10.1016/[Link].2020.110445. recycled aggregate concretes: an overview. Construct. Build. Mater. 328, 127071.
Bai, F., Fan, M., Yang, H., Dong, L., 2021. Image segmentation method for coal particle [Link]
size distribution analysis. Particuology 56, 163–170. [Link] Lederer, J., Gassner, A., Kleemann, F., Fellner, J., 2020. Potentials for a circular economy
PARTIC.2020.10.002. of mineral construction materials and demolition waste in urban areas: a case study
Barri, K., Jahangiri, B., Davami, O., Buttlar, W.G., Alavi, A.H., 2020. Smartphone-based from Vienna. Resour. Conserv. Recycl. 161, 104942. [Link]
molecular sensing for advanced characterization of asphalt concrete materials. RESCONREC.2020.104942.
Measurement 151, 107212. [Link] Lohumi, S., Cho, B.K., Hong, S., 2021. LCTF-based multispectral fluorescence imaging:
MEASUREMENT.2019.107212. system development and potential for real-time foreign object detection in fresh-cut
Basu, S., Kwee, T.C., Surti, S., Akin, E.A., Yoo, D., Alavi, A., 2011. Fundamentals of PET vegetable processing. Comput. Electron. Agric. 180, 105912. [Link]
and PET/CT imaging. Ann. N. Y. Acad. Sci. 1228 (1), 1–18. [Link] 10.1016/[Link].2020.105912.
j.1749-6632.2011.06077.x. Lotfi, S., Di Maio, F., Xia, H., Serranti, S., Palmieri, R., Bonifazi, G., 2015. Assessment of
Bobulski, J., Kubanek, M., 2021. Deep learning for plastic waste classification system. the contaminants level in recycled aggregates and alternative new technologies for
Applied Computational Intelligence and Soft Computing 2021 (1), 6626948. https:// contaminants recognition and removal. EMABM 2015: Proceedings of the 15th
[Link]/10.1155/2021/6626948. Euroseminar on Microscopy Applied to Building Materials, pp. 17–19.
Bonifazi, G., Palmieri, R., Serranti, S., 2018. Evaluation of attached mortar on recycled Lu, B., Zhou, J., Zhang, Y., Liu, Y., Wang, Q., 2024. An alternative rotating object
concrete aggregates by hyperspectral imaging. Construct. Build. Mater. 169, detection method for rock particle size distribution analysis. Powder Technol. 444,
835–842. [Link] 120059. [Link]
Braun, J., Willett, S.D., 2013. A very efficient O(n), implicit and parallel method to solve Luciano, A., Cutaia, L., Altamura, P., Penalvo, E., 2022. Critical issues hindering a
the stream power equation governing fluvial incision and landscape evolution. widespread construction and demolition waste (CDW) recycling practice in EU
Geomorphology 180–181, 170–179. [Link] countries and actions to undertake: the stakeholder’s perspective. Sustainable
GEOMORPH.2012.10.008. Chemistry and Pharmacy 29, 100745.
Brunnbauer, L., Gajarska, Z., Lohninger, H., Limbeck, A., 2023. A critical review of Marie, I., Mujalli, R., 2019. Effect of design properties of parent concrete on the
recent trends in sample classification using Laser-Induced Breakdown Spectroscopy morphological properties of recycled concrete aggregates. Engineering Science and
(LIBS). TrAC, Trends Anal. Chem. 159, 116859. [Link] Technology, an International Journal 22 (1), 334–345. [Link]
TRAC.2022.116859. JESTCH.2018.08.014.
Cabral, J.S., Menegatti, C.R., Nicolodelli, G., 2023. Laser-induced breakdown Marín-Cortés, S., Fernández-Álvarez, M., Enríquez, E., Fernández, J.F., 2024.
spectroscopy in cementitious materials: a chronological review of cement and Experimental characterization data on aggregates from construction and demolition
concrete from the last 20 years. TrAC, Trends Anal. Chem. 160, 116948. [Link] wastes for the assistance in sorting and recycling practices. Construct. Build. Mater.
org/10.1016/[Link].2023.116948. 435, 136798. [Link]
Chang, C., 2025. In: Quality Assessment & Monitoring of Recycled Coarse Aggregates. Marique, A.F., Rossi, B., 2018. Cradle-to-grave life-cycle assessment within the built
Chang, C., Di Maio, F., Bheemireddy, R., Posthoorn, P., Gebremariam, A.T., Rem, P., environment: comparison between the refurbishment and the complete
2025. Rapid quality control for recycled coarse aggregates (RCA) streams: multi- reconstruction of an office building in Belgium. J. Environ. Manag. 224, 396–405.
sensor integration for advanced contaminant detection. Comput. Ind. 164, 104196. [Link]
[Link] Merk, S., Scholz, C., Florek, S., Mory, D., 2015. Increased identification rate of scrap
Chang, C., Maio, F. Di, Rem, P., Gebremariam, A.T., Mehari, F., Xia, H., 2022. Cluster- metal using Laser Induced Breakdown Spectroscopy Echelle spectra. Spectrochim.
based identification algorithm for in-line recycled concrete aggregates Acta B Atom Spectrosc. 112, 10–15. [Link]
characterization using Laser-Induced Breakdown Spectroscopy (LIBS). Resour. Mewada, D., Agnew, C., Grua, E.M., Eising, C., Denny, P., Heffernan, M., Tierney, K., Van
Conserv. Recycl. 185, 106507. [Link] De Ven, P., Scanlan, A., 2024. Contamination detection from highly cluttered waste
RESCONREC.2022.106507. scenes using computer vision. IEEE Access 12, 129434–129446. [Link]
Chen, T., Zhang, T., Li, H., 2020. Applications of laser-induced breakdown spectroscopy 10.1109/ACCESS.2024.3456469.
(LIBS) combined with machine learning in geochemical and environmental Nalon, G.H., Santos, R.F., Lima, G. E. S. de, Andrade, I.K.R., Pedroti, L.G., Ribeiro, J.C.L.,
resources exploration. TrAC, Trends Anal. Chem. 133, 116113. [Link] Franco de Carvalho, J.M., 2022. Recycling waste materials to produce self-sensing
10.1016/[Link].2020.116113. concretes for smart and sustainable structures: a review. Construct. Build. Mater.
Chen, X., Kroell, N., Feil, A., Greiff, K., 2024. Sensor-based sorting. Handbook of 325, 126658. [Link]
Recycling: State-of-the-Art for Practitioners, Analysts, and Scientists 145–159. Park, S., Lee, J., Kwon, E., Kim, D., Shin, S., Jeong, S., Park, K., 2021. 3D sensing system
[Link] for laser-induced breakdown spectroscopy-based metal scrap identification.
Dice, L.R., 1945. Measures of the amount of ecologic association between species. International Journal of Precision Engineering and Manufacturing-Green
Ecology 26 (3), 297–302. Technology. [Link]
Fajar, A., Sarno, R., Fatichah, C., Fahmi, A., 2022. Reconstructing and resizing 3D images Pelc, N.J., 2014. Recent and future directions in CT imaging. Ann. Biomed. Eng. 42,
from DICOM files. Journal of King Saud University-Computer and Information 260–268.
Sciences 34 (6), 3517–3526. Poon, C.S., Chan, D., 2007. Effects of contaminants on the properties of concrete paving
Frison, G., Kouzoupis, D., Sartor, T., Zanelli, A., Diehl, M., 2018. BLASFEO: basic linear blocks prepared with recycled concrete aggregates. Construct. Build. Mater. 21 (1),
algebra subroutines for embedded optimization. ACM Trans. Math Software 44 (4). 164–175. [Link]
[Link] Psarras, C., Barthels, H., Bientinesi, P., 2022. The linear algebra mapping problem.
Current state of linear algebra languages and libraries. ACM Trans. Math Software 48
(3). [Link]

21
C. Chang et al. Developments in the Built Environment 21 (2025) 100611

Sai Trivedi, S., Snehal, K., Das, B.B., Barbhuiya, S., 2023. A comprehensive review fundamental understanding to data processing. TrAC, Trends Anal. Chem. 143,
towards sustainable approaches on the processing and treatment of construction and 116385. [Link]
demolition waste. Construct. Build. Mater. 393, 132125. [Link] Werheit, P., Fricke-Begemann, C., Gesing, M., Noll, R., 2011. Fast single piece
CONBUILDMAT.2023.132125. identification with a 3D scanning LIBS for aluminium cast and wrought alloys
Sørensen, T.J., 1948. A Method of Establishing Groups of Equal Amplitude in Plant Sociology recycling. J. Anal. Atomic Spectrom. 26 (11), 2166–2174. [Link]
Based on Similarity of Species Content and its Application to Analyses of the Vegetation on C1JA10096C.
Danish Commons. I Kommission Hos E. Munksgaard. Wu, H., Liang, C., Zhang, Z., Yao, P., Wang, C., Ma, Z., 2023. Utilizing heat treatment for
Soto-Paz, J., Arroyo, O., Torres-Guevara, L.E., Parra-Orobio, B.A., Casallas-Ojeda, M., making low-quality recycled aggregate into enhanced recycled aggregate, recycled
2023. The circular economy in the construction and demolition waste management: cement and their fully recycled concrete. Construct. Build. Mater. 394, 132126.
a comparative analysis in emerging and developed countries. J. Build. Eng. 78, [Link]
107724. [Link] Wu, L., Sun, Z., Cao, Y., 2024. Modification of recycled aggregate and conservation and
Steer, P., Guerit, L., Lague, D., Crave, A., Gourdon, A., 2022. Size, shape and orientation application of recycled aggregate concrete: a review. Construct. Build. Mater. 431,
matter: fast and semi-automatic measurement of grain geometries from 3D point 136567. [Link]
clouds. Earth Surf. Dyn. 10 (6), 1211–1232. [Link] Xia, H., Bakker, M.C.M., 2014. Reliable classification of moving waste materials with
2022. LIBS in concrete recycling. Talanta 120, 239–247. [Link]
Su, Y., Xu, Y., Bao, Z., Ng, S.T., Gao, Q., 2024. Stakeholder interactions of government talanta.2013.11.082.
intervention in construction and demolition waste recycling market: a game theory Yamazaki, I., Kurzak, J., Wu, P., Zounon, M., Dongarra, J., 2018. Symmetric indefinite
approach. Developments in the Built Environment 18, 100391. [Link] linear solver using OpenMP task on multicore architectures. IEEE Trans. Parallel
10.1016/[Link].2024.100391. Distr. Syst. 29 (8), 1879–1892. [Link]
Trotta, O., Bonifazi, G., Capobianco, G., Serranti, S., 2021. Recycling-oriented Zhang, C., Hu, M., Di Maio, F., Sprecher, B., Yang, X., Tukker, A., 2022. An overview of
characterization of post-earthquake building waste by different sensing techniques. the waste hierarchy framework for analyzing the circularity in construction and
Journal of Imaging 7 (9). [Link] demolition waste management in Europe. Sci. Total Environ. 803, 149892. https://
Tuan, N.M., Kim, Y., Lee, J.-Y., Chin, S., 2022. Automatic stereo vision-based inspection [Link]/10.1016/[Link].2021.149892.
system for particle shape analysis of coarse aggregates. J. Comput. Civ. Eng. 36 (2). Zhang, K., Qing, Y., Umer, Q., Asmi, F., 2023. How construction and demolition waste
[Link] management has addressed sustainable development goals: Exploring academic and
Vegas, I., Broos, K., Nielsen, P., Lambertz, O., Lisbona, A., 2015. Upgrading the quality of industrial trends. J. Environ. Manag. 345, 118823.
mixed recycled aggregates from construction and demolition waste by using near- Zhang, Z., Liu, Y., Hu, Q., Zhang, Z., Wang, L., Liu, X., Xia, X., 2020. Multi-information
infrared sorting technology. Construct. Build. Mater. 75, 121–128. [Link] online detection of coal quality based on machine vision. Powder Technol. 374,
10.1016/[Link].2014.09.109. 250–262. [Link]
Wang, Z., Afgan, M.S., Gu, W., Song, Y., Wang, Y., Hou, Z., Song, W., Li, Z., 2021. Recent
advances in laser-induced breakdown spectroscopy quantification: from

22