CONCRETE FLOW TRANSFORMER: PREDICTING FRESH CONCRETE PROPERTIES FROM CONCRETE FLOW USING VISION TRANSFORMERS

Abstract

To improve sustainability, concretes are increasingly produced using recipes containing up to a dozen different raw materials. The increasing complexity of the composition leads to an increased sensitivity and decreased robustness of the concrete, making a reliable quality control of the concrete highly important. Despite that, current quality control is mainly conducted based on analogous and empirical tests. This paper presents a novel approach for an automatic quality assessment of fresh concrete on the construction site. Based on a camera sensor setup, delivering image sequences showing the concrete flow during the discharge process of a mixing truck, we propose the Concrete Flow Transformer, a deep learning approach based on Vision Transformers, for the prediction of fresh concrete properties. The performance of the proposed approach is evaluated on a challenging real-world data set, demonstrating highly convincing results for the prediction of both, the consistency and rheological parameters of the fresh concrete.