The high level of damage suffered by historical structures due to the interaction between materials' aging and seismic events that occurred during the last decade has revealed the need for developing reliable and long-lasting prevention techniques. In this context, long-term structural health monitoring (SHM) is a practice that has been spreading in recent years. A monitoring system is composed by an optimized network of sensors placed in strategic positions within the building. The data recorded by the sensors enable to track the structural behavior over time. In this study an automated Bayesian-based procedure, i.e., an inverse problem able to detect in real-time the occurrence of damage, is proposed. In particular, hourly data are periodically divided into subgroups and used for the automatic Bayesian update. The methodology can be summarized as follows: i) preliminary calibration of a Finite Element (FE) model able to reproduce the structural dynamic behavior identified by the experimental vibration data; ii) identification of damage-sensitive portions of the structure by performing non-linear static analyses (NLSA) on the FE model; iii) calibration of a digital twin of the structure by surrogate modeling and definition of the sensitivity damage chart; iv) real time Bayesian model updating of the selected uncertain parameters in order to continuously identify possible changes in the structural behavior that can be associated with a certain level of damage.