Within the context of Structural Health Monitoring (SHM) of civil engineering structures, increasing attention has been recently given to development and application of the Deep Learning (DL) framework. Compared to traditional techniques, DL procedure allows to: (a) reduce the dimensionality of the input data; (b) simplify the computational path, not requiring any system identification and (c) implicitly model the environmental and operational variability (EOV) inside the input data. Among the DL architectures, the paper focuses on the use of Sparse Auto-Encoder (SAE) networks, aimed at detecting and localizing structural anomalies or damages. Unlike the usual SAE applications − where a specific SAE network is trained for each channel of data − the signals simultaneously collected by all available channels are used to train a unique SAE network. After training, the resulting network should be capable of accurately reconstructing the newly collected data until the structural condition does not change. On the other hand, an increase of the reconstruction error (i.e., the difference between measured and reconstructed signals) is conceivably expected if the monitored system departs from healthy condition; moreover, the increase in the reconstruction error should be more significant when referring to positions close to damage, so that localization of critical regions is attained as well. The effectiveness of the SAE methodology of damage assessment is demonstrated using accelerations data collected in the monitoring of the historical San Michele iron bridge (1889).