Automating concrete production control with computer vision-based aggregate characterisation

Concrete production is increasingly affected by fluctuations in the properties of natural and especially recycled aggregates. This paper investigates whether particle size distribution and material composition can be automatically determined from conveyor-belt image data during production. A backbone-agnostic deep-learning framework based on CNNs and Vision Transformers is applied to predict these properties and is extended with an additional branch that estimates aleatoric uncertainty directly from data via an uncertainty-aware loss formulation. The approach is evaluated on more than 80,000 real-world images collected using a camera-based sensor system installed on an operational concrete mixing plant. The results show accurate prediction of both grading curves and recycled material composition, providing a reliable basis for improved quality control for concrete producers and aggregate suppliers. The publicly available dataset enables further research and supports future progress towards fully automated, real-time quality assessment in concrete production. • Camera-based sensor system on conveyor belts for aggregate monitoring in concrete plants. • Backbone-agnostic framework for estimating particle size distributions and recycled-aggregate composition. • Aleatoric-uncertainty estimation for image-based aggregate property predictions. • Real-world data set of 80k+ conveyor-belt aggregate images from an operating concrete plant.