Automated material flow characterization of WEEE in sorting plants using deep learning and regression models on RGB data
Abstract
Waste from electrical and electronic equipment (WEEE) is a rapidly growing waste stream. Notably, electronic equipment contains valuable and critical raw materials. State of the art in WEEE recycling uses a combination of automated comminution and separation processes. To optimize these processes, analyzing material flow composition is essential, which today is performed by labor- and cost-intensive manual sampling and sorting. Automated analysis can be achieved through sensor-based material flow characterization (SBMC). However, this method has not yet successfully been applied for shredded WEEE. In a pilot-scale sorting plant, we developed a three-step SBMC method for shredded WEEE, based on cheap and widespread RGB cameras. Novel features of this approach are the combination of deep learning for material type identification, regression models for predicting individual particle masses, and aggregating the masses towards a material flow composition. First, YOLO v11 object detection performed best in identifying ferrous-metals, non-ferrous metals, printed circuit boards and plastics, reaching an [email protected] of 0.990. Next, geometry and color features were extracted for a total of 70 particle features. These data were used to train 11 types of regression models for particle mass prediction. K-nearest neighbors regression achieved a mean relative error of below 5 %, calculating the material shares in the waste stream from predicted particle masses. Finally, the combined approach of YOLO and k-NN regression was used on a validation dataset, achieving 4.94%. Our method can be applied in WEEE sorting plants to monitor and control processes or to analyze experiments.
Related
- Machine learning and spatial regression approaches to estimating willingness to pay for ecosystem services in Jakarta and Taoyuan
- Effect of Learning Motivation on Science Literacy of 11th grade studentsof MAN on Climate Change material
- Improving subsurface characterization in tropical granitic terrains using machine learning and deep learning-enhanced resistivity–chargeability modeling
- Does spatial sorting occur in the invasive Asian toad in Madagascar? Insights into the invasion unveiled by morphological analyses
- Computational and experimental study of multi-grain-size sediment transport and sorting process
BibTeX
@article{vogelgesang2026automated,
title = {{Automated material flow characterization of WEEE in sorting plants using deep learning and regression models on RGB data}},
author = {Malte Vogelgesang and Victor Kaczmarek and Alice Do Carmo Precci Lopes and Chanchan Li and Emanuel Ionescu and Liselotte Schebek},
journal = {Waste Management},
year = {2026},
month = oct,
doi = {10.1016/j.wasman.2025.114904},
url = {https://hal.science/hal-05789804v1},
note = {CC BY},
}
RIS
TY - JOUR
TI - Automated material flow characterization of WEEE in sorting plants using deep learning and regression models on RGB data
AU - Malte Vogelgesang
AU - Victor Kaczmarek
AU - Alice Do Carmo Precci Lopes
AU - Chanchan Li
AU - Emanuel Ionescu
AU - Liselotte Schebek
PY - 2026
DA - 2026/10/09
JO - Waste Management
DO - 10.1016/j.wasman.2025.114904
UR - https://hal.science/hal-05789804v1
AB - Waste from electrical and electronic equipment (WEEE) is a rapidly growing waste stream. Notably, electronic equipment contains valuable and critical raw materials. State of the art in WEEE recycling uses a combination of automated comminution and separation processes. To optimize these processes, analyzing material flow composition is essential, which today is performed by labor- and cost-intensive manual sampling and sorting. Automated analysis can be achieved through sensor-based material flow characterization (SBMC). However, this method has not yet successfully been applied for shredded WEEE. In a pilot-scale sorting plant, we developed a three-step SBMC method for shredded WEEE, based on cheap and widespread RGB cameras. Novel features of this approach are the combination of deep learning for material type identification, regression models for predicting individual particle masses, and aggregating the masses towards a material flow composition. First, YOLO v11 object detection performed best in identifying ferrous-metals, non-ferrous metals, printed circuit boards and plastics, reaching an [email protected] of 0.990. Next, geometry and color features were extracted for a total of 70 particle features. These data were used to train 11 types of regression models for particle mass prediction. K-nearest neighbors regression achieved a mean relative error of below 5 %, calculating the material shares in the waste stream from predicted particle masses. Finally, the combined approach of YOLO and k-NN regression was used on a validation dataset, achieving 4.94%. Our method can be applied in WEEE sorting plants to monitor and control processes or to analyze experiments.
SN - 0956-053X
ER -