One of the main motivations of the monitoring of concrete structures strengthened with fibre reinforced polymers (FRPs) in civil construction applications is to identify minor local damages occurring previously to a possible sudden and brittle failure mode, with the purpose of taking actions before a catastrophic failure can occur. Different approaches based on the electromechanical impedance (EMI) method, formulated from measurements obtained from PZT patches, have been proposed in the recent years for this purpose, given their ability for monitoring the performance and changes experienced by these strengthened beams at a local level, which is a key aspect considering their possible premature debonding failure modes. This work explores the design and application of Autoencoders based on raw impedance spectra for automatically detecting damage in FRP-strengthened RC beams. However, unlike the widely extended traditional procedure based on the reconstruction error, the suitability of reliable and effective damage indicators, directly defined on the latent space of autoencoders, is explored. This aproach would allow taking advantage of the hierarchical nature of deep models. The proposal has been verified on raw electromechanical impedance spectra obtained from PZT sensors bonded on an FRP-strengthened RC beam subjected to different loading stages and, therefore, damage levels. The results of the research demonstrate the capability and efficiency of the proposed approach.