Keywords: artificial neural network, real-time prediction of displacement, mechanised tunnelling.

Although numerous empirical and analytical solutions and two- and three-dimensional numerical models have been developed to establish suitable relationships between the mechanised tunnelling process and the surface and subsurface deformations, it is still difficult to provide reliable predictions of surface settlements for shield tunnelling in soft soils. Recently, an advanced three-dimensional numerical model for shield supported tunnel excavation in partially and fully saturated soft soils has been developed, showing satisfactory agreement with measured settlements during the construction process [1]. However, the enhanced accuracy using sophisticated three-dimensional finite element models has to be paid for by a greater effort in the model generation as well as greater computational costs.

For the purpose of real-time predictions of surface settlements, a meta model is employed to substitute computationally demanding three-dimensional numerical simulations. Artificial neural networks (ANNs) are trained by means of a comprehensive, process-oriented simulation model for mechanised tunnelling and then used as a meta model to provide predictions for the tunnelling-induced settlements. Since the simulation model is characterised by the realistic consideration of all relevant components of the shield tunnelling construction process (tunnel boring machine (TBM), the hydraulic jacks, the frictional contact between the shield skin and the soil and the support measures at the face and the tail gap) with a three-phase model for soft soils, it provides a rich data base for the training of an ANN. In this paper relationships between soil characteristics and TBM operational parameters on the one hand, and surface settlements on the other hand are predicted using the model-based ANN. Providing a certain number of input parameters characterising the model geometry, the soil model and the shield tunnelling process [2], the ANN provides the evolution of the settlements through in space and time.

Different ANN models have been trained in order to establish an optimal training sample and network architecture. In order to improve the convergence rate, a local adaptation strategy using an independent learning rate for each connection weight is implemented in a backpropagation algorithm. The ANN model with best performance has been used to provide information about the relative and absolute degree of importance of input parameters on final settlements, using hierarchical analytical methods, such as the relative strength effect (RSE) [3] and the global potential RSE (GPRSE) [4] respectively.

1

F. Nagel, J. Stascheit, G. Meschke, "Process-oriented numerical simulation of shield tunneling in soft soils", Geomechanics and Tunnelling, 3, 268-282, 2010.

2

S. Suwansawat, H.H. Einstein, "Artificial neural networks for predicting the maximum surface settlement caused by EPB shield tunnelling", Tunnelling and Underground Space Technology, 21, 133-150, 2006. doi:10.1016/j.tust.2005.06.007

3

Y. Yang, Q. Zhang, "A heirarchical analysis for rock engineering using artificial neural networks", Rock Mechanics and Rock Engineering, 30(4), 207-222, 1997.

4

Y. Yang, Q. Zhang, "A new method for the application of artificial neural networks to rock engineering systems", International Journal of Rock Mechanics and Mining Science, 35(6), 727-745, 1998.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £70 +P&P)