Simple Convolutional Neural Networks for the Damage Identification in Composite Steel-Concrete Beams

The identification of the damage in composite steel-concrete beams is addressed by implementing simple convolutional networks. By considering several damage scenarios, collections of images are generated by numerically evaluating a set of transmissibility functions relative to the generic damaged beam an by converting them into a gray level image suitably labeled. The images so generated are used to train simple convolutional networks capable to predict only the position or the position and the intensity of a single damage. The numerical experimentation carried out highlights the effectiveness of the proposed approach which does not require the adoption of predefined damage-related features.