Abstract
A random forest model is trained on dataset generated by finite element (FE) simulation to predict the damage location on a metallic plate. An optimization framework is integrated to identify minimal number of sensors required and optimal sensor placements. The accuracy of the FE model was validated against analytical solutions and experimental measurements, where satisfactory agreement was found with analytical solutions. The out-of-plane displacement and the normal strain at the bottom of the plate surface were extracted from the simulation at different virtual sensor locations to serve as inputs to the random forest model. It was found that satisfactory prediction accuracy can be achieved with just two sensors for this specific problem. Using a response surface optimization method, the optimal sensor locations that minimizes prediction error are identified, and achieved higher accuracy than the baseline case with equidistant sensor placement. This study lays a strong foundation for developing a machine-learning-based structural health monitoring model within a digital twin framework.
