The use of artificial neural networks for modeling air void content in aggregate mixture

A database for various pavement mixtures which were tested at the IGMAT Building Materials Institute, Ljubljana, during the period from 1998 to 2009 was established. This database consists of 17,296 asphalt mixture analyses. Artificial neural networks were used in this work to estimate air void content in aggregate mixture of several stone fractions for 7 types of asphalt concrete mixtures (AC 32, AC 22, AC 16, AC 11, AC 11 PmB, AC 8, AC 8 PmB) produced according to EN 13108-1. The main aim of the paper is to model the relationship between different parameters and air void content in aggregate mixture with artificial neural networks and multiple linear regression. The proposed method uses feed-forward neural networks with error back-propagation algorithm. Two different programs for modeling with artificial neural networks, NTR2003 and WEKA toolkit, were used. Before modeling air void content in aggregate mixture outliers among data were determined. Then, the artificial neural network analysis and multiple linear regression were done for each asphalt mixture and also for all mixtures together. Modeling of air void content in aggregate mixtures in general showed that linear models work better than artificial neural network models in the cases of specific asphalt mixture. In the case of analysis of all asphalt mixtures together, neural networks detected real hidden relationships between data and are therefore more effective than the linear model. Feed-forward neural networks are entirely appropriate models for an effective preliminary estimate of air void content in various aggregate mixtures.