Utilization of Prior Information in Neural Network Training for Improving 28-Day Concrete Strength Prediction

Abstract

A concrete mix design aims to obtain the optimal proportions of concrete ingredients, including cement, water, sand, coarse aggregates, and admixtures. Neural networks (NNs) have been widely applied in research to predict concrete strength, and the concrete ingredients and concrete strength metrics have been used as the input and output parameters, respectively. The objective of this study is to use the 3-day concrete strength as the prior information in the NN training to reduce overfitting and improve the 28-day concrete strength prediction capability. This study is unique because the 3-day concrete strength was not used as another input parameter in the NN training; instead, it was used as data for the initial weights and biases of the connection node in the hidden layer during the NN training. Accordingly, a prior information-based NN model (PI-NNM) was developed to obtain a 28-day concrete strength prediction model. According to the tests with data subsets, the PI-NNM showed a better prediction capability than the conventional NN model, which uses only input parameters for the concrete prediction. Moreover, an adjusted PI-NNM was applied to the actual concrete production; the results showed a high prediction capability for the 28-day concrete strength.