Quantification of construction and demolition waste products that can be carbonated using a deep learning-based image analysis

Author(s): J-D. Lau Hiu Hoong, J. Lux, P-Y. Mahieux, Ph. Turcry, A. Aït-Mokhtar
Paper category: Proceedings
Book title: CO2STO2019 - International Workshop CO2 Storage in Concrete
Editor(s): Assia Djerbi, Othman Omikrine-Metalssi, Teddy Fen-Chong
ISBN:
e-ISBN: 978-2-35158-232-9
Publisher: RILEM Publications SARL
Publication year: 2019
Pages: 116-126
Total Pages: 11
Language : English

Abstract: The building sector is the largest consumer of natural aggregates and producer of waste.
Europe is taking a step to tackle this issue through the Horizon 2020 programme, which
promotes sustainable development by encouraging its member states to recycle at least 70%
of their non-hazardous and inert construction and demolition waste (CDW). The latter is a
mixture of aggregates of different natures (bricks and natural stones), particles of mixed
natures (usually natural stones coated with cement paste or bitumen) along with traces of other
materials (e.g. plastic). The composition of CDW is variable as it depends on the site from
which it originates. Usually, it is mostly made up of recycled concrete aggregates. All this
available cementitious material could uptake significant amounts of carbon dioxide thanks to
carbonation. In order to determine the precise composition of a batch of CDW, the
NF EN 933-11 standard recommends manual sorting. However, this method is
time-consuming and could become a bottleneck in the process of recycling larger quantities of
CDW. Our work focuses on the development of a novel method to determine the composition
of CDW. It makes use of deep learning and neural networks. To train our algorithm accurately,
we had to use more detailed classes of recycled aggregates than those given in the
NF EN 933-11 standard Provided that a convolutional neural network has been trained on a
representative and reliable database, it can identify the nature of every aggregate on a picture
of CDW to give its composition instantly. Our algorithm reaches an accuracy above 92 % for
this detailed classification task. Further pixel-by-pixel identification (semantic segmentation) of
individual aggregates can then be used in order to quantify the proportion of mortar available
on each aggregate. Finally, this procedure, applied to an image of a batch of CDW, could be
used to estimate the amount of CO2 that could be uptaken through carbonation of these
aggregates.

Online publication : 2019
Publication type : full_text
Public price (Euros) : 0.00