An innovative procedure for the structural damage assessment through the analysis by Artificial Intelligence (AI) of vibration data was implemented for the interpretation of shaking table tests. In particular, the proposed procedure is based on the application of Convolutional Variational Auto-Encoder (CVAE) as deep learning method. Firstly, CVAE is trained to reconstruct the vibration time history of the structure in undamaged conditions. Then, the Mean Square Error (MSE) and the Modified Original to Reconstructed Signal Ratio (MORSR) were considered as evaluation metrics of the analyzed vibration data. MSE and MORSR are used to assess the state of damage of the tested structure in terms of similarity between the vibration response in undamaged and damaged conditions. The structural damage was defined in terms of a Damage Index (DI) based on the decay of the first modal frequency of the tested prototype. The distance in the MSE-MORSR plane between a given DI cluster centroid from the undamaged cluster centroid (DI equal to 0) was considered. The experimental validation of the proposed procedure was carried out by application to shaking table tests of a reinforced concrete frame prototype. Seismic tests were based on the recording at the NRC station of the strongest shake of the central Italy sequence (2016), occurred on 30 October 2016. It was scaled and reproduced with incremental steps of 0.1 g of acceleration up to 0.8 g, alternated with white-noise dynamic characterization (WNDC) tests at 0.05 g. Vibration data at 27 measurement points of the prototype in the WNDCs were considered for the analysis. The results showed very promising potentialities for the damage assessment of existing structures.