The structural health monitoring of historical masonry structures presents unique challenges due to their complex behavior, limited accessibility, and the need for non-invasive assessment methods. Recent advances in deep learning, particularly transformer-based architectures, have opened new opportunities for computationally-efficient data-driven anomaly detection in time-series sensor data. This study investigates the application of transformer neural networks for detecting earthquake-induced anomalies in historical masonry buildings, focusing on the San Pietro bell tower in Perugia, Italy, a seven centuries-old structure subject to continuous monitoring and located in a seismically active region. The proposed method leverages the self-attention mechanism of transformers to learn temporal dependencies and latent representations from acceleration time series acquired by a network of sensors installed in the bell tower. The model is trained via unsupervised learning to reconstruct normal structural response sequences; deviations between the input and reconstructed signals, expressed in terms of carefully selected features, are used to identify anomalous behavior potentially associated with seismic events or structural degradation. Results from real monitoring data collected over an extended period, including during recent seismic activity occurred in 2016, demonstrate the model's effectiveness in identifying unusual dynamic patterns that are known to correspond to small structural changes. This approach offers a scalable and non-invasive tool for continuous condition assessment, contributing to the proactive preservation of vulnerable historic structures exposed to seismic hazards and enabling real-time alarms, standing as a viable monitoring alternative in comparison to more traditional approaches based on modal features identification.