Presentation Information
[NpVBMPAo-09]Study on Aggregate Gradation Estimation Using Image Recognition AI and Accuracy Enhancement
*Mika Serita1, Maiko Hirano1, Saki Endo1 (1. MAEDA ROAD CONSTRUCTION Co.,Ltd (Japan))
Keywords:
image recognition AI,aggregate gradation estimation,real-time quality control,sieve analysis
In asphalt mixture production plants, sieve analysis of aggregates is conducted as part of quality control for asphalt mixtures. Although sieve analysis is a relatively simple testing method, it poses challenges such as requiring significant time to test all aggregate types and the inability to immediately reflect test results in the manufacturing process. To address these issues, this study aims to reduce the labor involved in conventional sieve analysis and enable real-time quality control by developing a gradation estimation technique using image recognition AI. The goal is to create a system capable of estimating aggregate gradation from images.
Data for constructing the estimation model required for system development were collected using aggregates from multiple plants. Aggregate images were obtained by spreading the aggregates in a common container and photographing them with a smartphone fixed at a constant height. Simultaneously, sieve analysis was performed on the aggregates used for imaging, and the measured gradation was used to train the estimation model. The target accuracy for the model was defined as a difference within ±5% between the actual and estimated passing mass percentage for each sieve size. A prototype system was developed using the constructed estimation model, and verification tests were conducted. However, the estimation accuracy for unseen data was initially low, prompting model improvements that successfully achieved high estimation accuracy.
Furthermore, to address the issue of lower estimation accuracy for fine aggregates compared to coarse aggregates, measures were implemented to improve the accuracy for fine sand. These included refining the data acquisition method and enhancing the estimation model, resulting in improved accuracy for fine aggregate gradation estimation. Based on these results, the developed system can be considered an effective method for immediately reflecting aggregate gradation in the quality control of manufactured asphalt mixtures.
Data for constructing the estimation model required for system development were collected using aggregates from multiple plants. Aggregate images were obtained by spreading the aggregates in a common container and photographing them with a smartphone fixed at a constant height. Simultaneously, sieve analysis was performed on the aggregates used for imaging, and the measured gradation was used to train the estimation model. The target accuracy for the model was defined as a difference within ±5% between the actual and estimated passing mass percentage for each sieve size. A prototype system was developed using the constructed estimation model, and verification tests were conducted. However, the estimation accuracy for unseen data was initially low, prompting model improvements that successfully achieved high estimation accuracy.
Furthermore, to address the issue of lower estimation accuracy for fine aggregates compared to coarse aggregates, measures were implemented to improve the accuracy for fine sand. These included refining the data acquisition method and enhancing the estimation model, resulting in improved accuracy for fine aggregate gradation estimation. Based on these results, the developed system can be considered an effective method for immediately reflecting aggregate gradation in the quality control of manufactured asphalt mixtures.
