System Identification of a Soil Tunnel Based on a Hybrid Artificial Neural Network–Numerical Model Approach

Abstract

System identification of structures is the primary goal of this study. Numerical simulation using methods such as finite element modelling or finite difference modelling is the practical solution to model the structure based on some accurate parameter values that are essential to identify the behaviour law of the structures. An optimisation method integrated with the numerical model (NM) can solve an inverse problem to provide a calibrated parameter set to improve the modelling. For this purpose, an application of artificial neural networks (ANNs) integrated with an NM with the title “hybrid ANN–NM” approach is proposed in this study so that the parameters and resulted deformations of a developed NM of the structure would enter to a designed ANN for training the network. The developed hybrid ANN–NM method to identify the system of a tunnel excavated in the soil is applied and reported in this paper. A set of variable material parameters with a set of displacement and strain data (in $$x$$ and $$y$$ directions) of critical points of the tunnel which obtained from 2D FDM in FLAC 7.00 software is provided to train the ANN. This study is divided into two methods: (1) fitting model to the displacements and strains to show the deformation behaviour of the tunnel related to the parameter changing, and (2) fitting model to the parameters related to the displacements and strains to estimate the optimum parameters for the NM. For the first method, the parameter dataset as input and the displacements and strains as the ANN’s output in one approach separately and in another approach altogether are considered to train the ANN. In the second method, the displacements and strains as input and the physical parameters of the NM as output are submitted to train the ANN. The MSE convergence of implemented ANN shows the power of ANN to model the behaviour of structure based on the token data. Also, the RMSE of residuals refers to the success of the proposed method for the fitting model. Besides, this method was able to provide the optimum parameters for the numerical model in minimum computing time regarding the implementation of the second method.