Structural Utilization Prediction for the Health Monitoring of Tunnel Linings by Means of an Artificial Neural Network Ensemble

Abstract

The continuous expansion of underground structure networks and the necessity of guaranteeing the safety and functionality of the already built tunnels pose the necessity of developing techniques and methods to efficiently achieve these aims. Normally, measured data of different nature are available for tunnels due to sensors installed in the lining. However, adequate procedures have to be implemented to obtain valuable information out of the monitored data. A new method has been developed to achieve a real-time assessment of the stress state in segmental tunnel lining, given specific measured quantities as input. The method is based on the combination of finite element (FE) analyses and feedforward neural networks, which permits to exploit the advantages of both physics-based simulations, representative of the structure considered, and the predictive capabilities of machine learning tools. The FE model of the tunnel lining plays an important role for the reconstruction of the missing quantities, which are not available from monitoring campaigns, but which can be generated by numerical analyses. A Monte Carlo sampling procedure is performed for the definition of multiple sets of the input parameters used in the FE model for the generation of the training data of the metamodels. An ensemble of neural networks is created and assembled into a framework, which is validated against a full scale test where its predictive performances are investigated.