Neural network based approach for dissemination of field measurement information

This paper presents a neural network based approach to disseminating information relating to experimental and field observations in engineering. Although the methodology is generic and can be applied to many areas of engineering science, attention is focussed here solely on geotechnical engineering applications. Field data relating to the settlement of foundations presented by Burland and Burbidge (1985) which led to their well known equation for calculation of settlement, now included in most text books, is re-visited. A part of the data, chosen randomly, is used to train an Artificial Neural Network (ANN), which relates foundation settlement to various causes as identified by the authors. Predictions are made for situations for which data were not used in training. These indicate sufficient accuracy when compared to the original field data. Accuracy of predictions is further improved when all the data are included in the training set. The finally trained ANN is shown to represent these data more accurately than the Burland and Burbidge equation. Based on the above heuristic example, an ANN is presented as an alternative to developing equations and design rules in geotechnical engineering practice. Significant advantages are shown to arise by using this methodology. Ease of updating the ANN, as and when additional data becomes available, being the most important one. Loss of transparency, however, seems to be the main disadvantage.