Neuro-Computing Techniques for Prediction of Compressive Strength of Concrete

Abstract

Neuro-computing techniques are being widely used these days and these techniques are considered good for forecasting applications. In recent years, these techniques have also been applied to many civil-engineering problems with reasonable success. These techniques are particularly useful in applications where the complexity of the data or task makes the design of such a function impractical. As the development of concrete strength is a complex non-linear process, depending upon many parameters, it is a problem well suited for applying neuro-computing techniques. In this work, we focused on the prediction of compressive strength of concrete with the help of neuro-computing techniques. Data for this work has been taken from the experiments conducted by Kumar (2003). For generating a reliable data bank on concrete compressive strength, he considered five parameters, namely, water-cementitious ratio, cementitious content, water content, workability and curing ages in his experiments. He performed all experiments in controlled laboratory conditions. A set of 15 cubes for each of the mixes so proportioned were cast and tested after 28, 56 and 91 days of curing. Thus, an extensive data bank for analyzing the compressive strength of concrete has been used in the present work. Factor analysis has been performed on the data in order to decide variables for predicting compressive strength of concrete with the help of SPSS and investigation, reveal that water-cementitious ratio is the leading predictor variable. Regression models have been also developed in this work. These have been developed: (i) To analyze the effect of workability on compressive strength; (ii) To analyze the effect of FA on compressive strength; and (iii) For predicting compressive strength of concrete with three different aggregate zones, i.e., Zone-A, Zone-B and Zone-C with and without FA. The regression models developed can predict compressive strength of various mixes very efficiently. However, a variation in data affects the regression coefficients to a large extent. Therefore, we have introduced the ridge parameter in regression equations. By introducing ridge parameter in regression equations, we could obtain more trustworthy and efficient predictive models for compressive strength of concrete that are not affected by the variations in dataset used for prediction. The development of neural network and GP models for the prediction of compressive strength of concrete has also been undertaken in this work. For the development of an efficient neural network model, a total of 1440 pilot experiments have been conducted before selecting the final architecture of neural network. As an outcome of this study, it has been inferred that the neural network approach has a great potential for prediction of compressive strength of concrete. The results obtained from neural network model are compared with regression model and GP model. It has been found that neural network models show a high degree of consistency with experimentally evaluated compressive strength of concrete specimens used. One can further refine the neural network models and GP models proposed in this work by considering a larger and reliable dataset. One can also explore other soft computing models for more efficient prediction of compressive strength of concrete.