Concept for Monitoring-Based Stress Prediction in Structures by Means of Artificial Neural Networks and Synthetic Noise

An approach for monitoring and simulation-based stress prediction employing feed-forward neural networks (FNNs) is presented, where artificial noise is used to improve the robustness of structural stress prediction. The approach is demonstrated on a beam structure where FNNs are trained using synthetic data to predict the maximum bending moment based on displacements. The analytical solution of the inverse problem is also derived and presented for different load cases. To validate the approach, FNNs trained on synthetic data are applied to displacement measurements of a full-scale beam test. It is shown that perfectly trained networks cannot in general be straightforwardly applied to real data because of correlations in the synthetic training data, but the introduction of controlled artificial noise to the synthetic training data significantly helps to improve prediction performance and robustness. Multiple noise ranges and load cases are investigated, and the predictive performances of the FNNs are analyzed. The concept is finally applied to a tunnel structure for an artificial benchmark scenario, where advantages and limitations of the method are expounded.