Modular Neural Networks for Predicting Settlements during Tunneling

Abstract

This paper discusses back-propagation neural networks (NN) for predicting the settlement during tunneling. Three settlement parameters and 11 major affecting factors have been identified from analyzing the general tunneling operations. A general neural network model is trained and tested using the actual collected data from the 6.5 km Brasilia Tunnel in Brazil. The general model generates an average error of 70 mm for the predicted settlements compared with the actual values. To improve the prediction accuracy, modular NN models are studied based on the concept of integrating multiple NN modules in one system with each module being constrained to operate at one specific situation of a complicated real world problem. The modular concept can make better use of neural computation algorithms to improve the convergence in the training process. It has been studied on modeling multiple output variables and discrete input variables. After applying modular models to the same Brasilia Tunnel, the average prediction error is reduced to 33.4 mm, which shows a significant improvement over the general NN model.