Preserving the world heritage: Post-earthquake monitoring based on structural break testing with deep temporal convolutional features

Built heritage faces nowadays increasing vulnerability due to the combined impact of climatic, seismic, and anthropogenic forcings. In this context, vibration-based monitoring has become a key non-invasive method for assessing the integrity of historical buildings. However, little attention has been given to the development of automatic tools, which are crucial for rapid and effective decision-making. This study examines San Cristobal Church, a 17th-century building located in the UNESCO World Heritage site of Cusco, Peru. The church has been continuously monitored during 17 months using a seismic sensor located on one of its walls. First, we develop machine learning models to predict the church’s natural frequencies based only on weather data. Then, we analyze deviations from the expected frequency variations to detect anomalies that may indicate structural changes in the building, especially following strong transient events such as earthquake-induced motions. We evaluate multiple machine learning approaches, including Ridge Regression, Feedforward Neural Networks, and Temporal Convolutional Networks, with the latter outperforming other models in capturing nonlinear temporal dependencies. To estimate the post-seismic recovery time of the natural frequencies following a Mw 4.2 earthquake occurred in August 13th, 2024, we employ the Bai-Perron test for structural break detection on the learned deep temporal convolutional features. As this recovery time is influenced by the damage state, changes in its duration can reflect alterations in masonry mechanical properties. By accurately assessing the post-seismic recovery time, our methodology offers a promising approach for developing early warning systems to identify damage in historical buildings.