Combined model-based and machine learning approach for damage identification in bridge type structures

In this work, we propose a combined approach of model-based and machine learning techniques for damage identification in bridge structures. First, a finite element model is calibrated with real data from experimental vibration modes for the undamaged or baseline state. Second, generic synthetic damage scenarios based on modal parameters are automatically generated with the model to train machine learning algorithms for damage classification (Support Vector Machine, SVM) and damage location and quantification (Neural Network, NN). For an initial validation of the method we use a lab scale truss bridge model, proving that specific damage scenarios can be assessed by the Supervised Machine Learning algorithms trained with generic damage scenarios including a certain variability. The NN provides an assessment in terms of damage location and quantification, whereas the SVM provides a damage severity classification with graphical indication of the damage location and quantification through a reduced dimension plot.