ARTIFICIAL NEURAL NETWORK IN DAMAGE ASSESSMENT OF STEEL BRIDGES

Abstract

In civil engineering, bridge is a very important structure. Especially steel bridges, which are mostly used by railways in India, forms an important part of the countries infrastructure. Beside this, most of the bridges are situated in inaccessible area, where environmental impacts are very high, demands a thorough assessment of its serviceability in present condition. The most common methodology for damage identification includes either field testing or system identification technique. The main problem with these methods is either they are difficult to implement or they associates some time consuming and costly affair. In contrary to this neural network can be successfully used for damaged identification with very easy to implement way. Artificial neural network tries to mimic the information processing principles of biological neural networks. In recent past, it has attracted wide attention and found numerous no of applications in various areas of engineering. A multilayer neural network can be used to classify any data sample into desired no of space. For this study, neural network is used as a multi class discriminator. The objective of the present study is, to identify suitable neural network architecture and other related parameter for successful damage identification of a warren type truss bridge. For this purpose three different truss of various span and other geometrical property, is taken into consideration. Three different neural networks have been built for three trusses. Later on a generalize network considering data from all three trusses has been built. Each network is built with two different approaches using the same data set. Type I network is an ordinary feed-forward network and Type II network is modular network. Performance of both types of network is tested and Type II network found superior in certain identification task. During development of each network, first, training was carried out with a given training sample and then performance of ANN has been evaluated in terms of prediction accuracy on test data. Various neural network architectures have been evaluated to assess their performance on the prediction and generalization capability. From these, an optimal neural network model has been identified. The outcome of the present study shows that developed combined network has immense potential to make reasonably accurate prediction about damage condition of an entirely new truss.