Incorporation of cluster analysis into Arificial neural networks for estimating scour depth

Abstract

A local scour around a bridge pier is known one of important factor of bridge collapse. Two approaches are used in estimating a scour depth in practice. One is to use of empirical formulas, and the other is to use computational methods. Since the empirical formulas are not general to be applied, the use of empirical formulas is limited to predict scour depths under similar conditions to which the formulas were derived. Recent developments in computational fluid dynamics allow numerical simulations of local scour around a bridge pier. However, numerical simulations are currently too expensive to be applied to practical engineering problems. This study describes the application of artificial neural networks (ANN) to the prediction of scour depths around a bridge pier at an equilibrium state. The advantage of ANNs is that the exact relationship among variables does not necessarily need to be known. In addition, this study investigates various neural network algorithms for estimating scour depths, such as backpropagation networks, radial based function networks, and probabilistic neural network. Preliminary study shows that the ANN model results in very wide range of error in predicting scour depths. This problem seems to result from the difference of mechanism to make scour depth. So, this study incorporates cluster analysis into ANNs to classify a set of observations into two or more mutually exclusive unknown groups based on combinations of interval variables. This preprocess is expected to improve similarity between data sets which results in reducing the error range in the prediction of scour depth in practice.