Estimation of model parameters and ground movement in shallow NATM tunnel by means of neural network

摘要

Currently an increasing number of urban tunnel with small overburden are excavated according to the principle of the New Austrian Tunneling Method (NATM). However, it is still difficult under the present state of the art to deterministi-cally predict the behavior of the ground and tunnel system in the planning and design state, thereby resulting in an extensive difference between the predicted values and the actual behavior after excavation. This paper discusses back-propagation artificial neural network (ANN) for identification of unknown model parameters, Young's modulus, E, horizontal stress state, K_0, residual strength/original strength, η and increment of maximum strain during which strength drops from peak to residual value, Δγ. The back-propagation algorithm performs gradient descent analysis on this error surface by modifying each weighting factor in proportion to the gradient of the surface at its location. This paper trains to approximate the results of FE simulations. The trained ANN is able to reproduce the correlation between model parameters with the available measurement data. A FE simulations used here incorporate reduction of shear stiffness, as well as strain softening effects of given material strength parameters. Based on the identified material parameter, the ground movement is predicted. The proposed approach produced strain distribution, deforma-tional mechanism, surface settlement profile, which were in agreement with the results of the field measurement results by case study. Furthermore, the method offers a practical way for predicting final displacement of tunnel at earlier stages of construction, enabling rational safety management scheme to be employed. This makes a good starting point for optimizing ground support by reducing surface settlement in consideration of a particular nature of the deformational mechanism of shallow tunnels.