Predicting the life of concrete structures using neural networks

This paper outlines an investigation into the use of neural networks (NNs) to analyse data from natural exposure trials and durability surveys to assist in-service life prediction of concrete structures, The application area was restricted to the corrosion of steel in concrete, NNs for predicting chloride profiles, chloride binding and carbonation depth were developed in succession, each addressing specific issues concerning the application of NNs to service-life prediction; these NNs generally had average errors in the range 20-30%, individual studies necessarily only investigate subsets of the large number of variables involved. NNs can be used to combine and analyse data from separate studies to predict behaviour in situations not encountered before and to quantify the individual effects of the variables, Strategies are presented for difficult issues such as characterizing concretes and exposure environments, dealing with missing data, appropriate testing, predicting to times beyond the available data and the use of NNs in design.