A study on coarse aggregate gradation detection based on shape factor estimation using GAT and Bayesian inference

Abstract

The mass distribution of particles of different sizes is characterized by aggregate gradation, which is regarded as a key parameter determining mixture performance. Existing machine vision-based gradation detection methods exhibit considerable errors, mainly due to their excessive reliance on two-dimensional particle size features while neglecting particle morphological information. To address this issue, a novel method based on Graph Attention Networks (GAT) was proposed. Based on extracted aggregate size information, inferred particle shape factors were integrated to provide a more accurate representation of aggregate morphology. Dimensional features such as the equivalent Feret minor axis and the equivalent ellipse minor axis, together with multidimensional shape features including circularity and rectangularity, were first extracted. Subsequently, multi-temporal and multi-angle shape paradigms were constructed, through which the shape factors of aggregate particles were predicted by GAT. On this basis, an equivalent-sphere distribution theory was proposed, in which an aggregate set with unknown mass was transformed into a set of equivalent spheres with known mass, enabling the fusion of dimensional and morphological features. Finally, the gradation distribution was efficiently inferred by a Markov Chain Monte Carlo (MCMC) algorithm within a Bayesian statistical framework. Experimental results show that the absolute error of gradation detection is controlled within ± 3.0%. This represented a significant improvement over traditional size-feature-based methods and satisfied the ± 5% error tolerance required in engineering practice.