The identification of damage in the connection of steel–concrete composite beams is pursued by means of Convolutional Neural Networks. The data used for training the networks are gray level images obtained by converting a set of transmissibility functions of the beam. It is shown how simple Convolutional Neural Networks can be trained for the identification of the damage by evaluating its position and intensity. Data availability is thoroughly discussed highlighting the effects of various factors, among them the number and richness of the images used in the training phase, on the predictive capabilities of the network. In this regard, the results obtained are satisfactory as confirmed also by comparisons with data extracted from an experimental campaign. The problem of the unavoidable presence of modelling errors is also discussed and counteracted.
