Prediction of surface settlement during shield TBM excavation using extreme learning machine

Tunneling-induced surface settlements in soft ground conditions threaten the stability of nearby structures, especially during subway tunnel excavation in urban areas. Therefore, controlling the surface settlement before excavation is the key to successful tunneling. There have been numerous means for predicting surface settlements during TBM tunneling such as empirical, analytical and numerical approaches. However, these techniques occasionally show poor predicting performance when being applied to actual excavation sites due to complex and unique surface settlement mechanisms. To circumvent limitations of the existing prediction methods, machine learning techniques such as the artificial neural network has been recently introduced. In this paper, the extreme learning machine (ELM), which is an improved version of the artificial neural network, is applied to verify its cost efficient neural network model for the prediction of surface settlements. 14 settlement-inducing features categorized as the tunnel geometry, TBM operating conditions and geological conditions are collected from the Hong Kong shield TBM tunneling site. The performance of ELM is compared with the well-known Levenberg Marquardt and the Bayesian Regularization algorithm for the same single-layered neural network. The obtained results show the significance of performance achieved by the ELM-based prediction of surface settlements.