Deep Learning-Based Methods for Inspecting Sand Quality for Ready Mixed Concrete

Sand is a vital component within a concrete admixture for variety of structures and is classified as one of the crucial bulk material used. Assessing the Fineness Modulus (FM) of sand is an essential part of concrete production process because FM significantly affects the workability, cost-effectiveness, porosity, and concrete strength. Traditional sand quality inspection methods, like Sieve Analysis Test, are known to be laborious, time-consuming, and cost ineffective. Previous studies had mainly focused on measuring the physical characteristics of individual sand particles rather than real-time quality assessment of sand, particularly its FM during concrete production. This study introduces an image-based method for detecting flawed sand through deep learning techniques to evaluate the quality of sand used in concrete. The method involves categorizing sand images into three groups (Unavailable, Stable, Dangerous) and seven types based on FM. To achieve a high level of generalization ability and computational efficiency, various deep learning architectures (VGG16, ResNet-101 and MobileNetV3 small), were evaluated and chosen; with the inclusion of transfer learning to ensure model accuracy. A dataset of labeled sand images was compiled. Furthermore, image augmentation techniques were employed to effectively enlarge this dataset. The models were trained using the prepared dataset that were categorized into three discrete groups. A comparative analysis of results was performed based on classification performance metrics which identified the VGG16 model as the most effective achieving an impressive 99.87% accuracy in identifying flawed sand. This finding underscores the potential of deep learning techniques for assessing sand quality in terms of FM; positioning this research as a preliminary investigation into this topic of study.