Estimation of Maximum Ground Surface Settlement due to Tunneling with Artificial Neural Network and Wave-net Network

Regarding to increase in cities population, the amount of internal trips enlarged. To provide suitable facilities on these movements, construction of urban railways has been adopted. Deportation of massive materials containing stone and soil in tunnel excavation causes stress variation in region around the tunnel and consequently settlement in ground surface. Therefore, to prevent probable damages in adjacent structures, assessment of settlement amount has to be considered before tunnel construction. Settlement amount due to excavation depends on different factors such as, soil type, ground water level and parameters refer to shield operation. Considering a lot of effective factors in settlement and also great distance of excavation, a large amount of calculations and studies are necessary to settlement estimation. In this field, the black box models can be used as a powerful method, because of their ability, independent performance of physical parameters and their relation among them. In this paper, we used two types to networks, feed forward back-propagation neural network and wave-net. Feed forward neural network is the most popular neural network. To construct wave-nets, sigmoidal activation functions of hidden layers in feed forward neural network are substituted with appropriate wavelet functions. In this paper, studies on the measurement of subway project in Bangkok have been done. In the Bangkok MRTA project, data on ground deformation and shield operation were collected. The tunnel sizes are practically identical and the subsurface conditions over long distances are comparable, which allow one to establish relationships between ground characteristics and EPB-operation on the one hand, and surface deformations on the other hand. At first, the impact of every parameter on the settlement was observed individually and then effective input data was determined for training of the network. Thirteen parameters were determined as inputs and surface settlement was output parameter of networks. Networks were trained and optimal models of networks were determined and then analysis results of ANN and wave-net were compared together. Analysis results demonstrate both networks capability. Furthermore, reduction in errors for wave-net shows its high accuracy in comparison to neural network.