Neural Network In The Building Materials Field Of System Identification

This paper is aimed at demonstrating the validity of adapting artificial neural networks (ANN) to model behavior of the building materials in macroscopic and microscopic aspects. A suitable network model will be selected to approximate the practical system.In macroscopic aspect, it is taken as an example about how to model the relationship between the mix proportion of raw materials and the compressive strength of concrete using radial basis function (RBF) neural networks, which can achieve a global minimum. Orthogonal least squares (OLS) method is introduced to model the strength of concrete. Compared with the popular BP network, the simulation shows that the learning speed of RBF network is substantially faster than that encountered in BP network and the generalization ability of this network is also better. So it is effective to use RBF network in the field of concrete.In microscopic aspect, taking no account of physical materials, how to build the relationship between raw materials (denoted in the form of molecular formulas) and the final compound components of building materials is discussed. Based on RBF neural network and perceptron neural network, a four-layer feed-forward neural network named radial basis perceptron (RBP) network is presented. This network can be summarized as follows: (1) About the networks' architecture, it is not fully connected but it uses selective connection between the units of two hidden layers. The number of these units is determined dynamically. (2) During the learning procedure, a new input-output clustering (IOC) method is adopted to select centers. The width parameter crof centers is self-adjustable according to the information included in the training samples. The validity of this network is illustrated using an example taken from component analysis of CaO-AL2O3- SiO2 system. Simulation shows that RBP network can predict the components of civil building materials successfully and gets good generalization ability.Altogether, as an advanced tool for system identification, neural networks will provide a new and effective way for studying building materials.