A multivariate autoencoder-based anomaly detection for post-earthquake diagnosis of masonry structures using long-term accelerometer and ambient temperature recordings

This study introduces a multivariate convolutional autoencoder (1D-CAE)-based framework for unsupervised anomaly detection in historical masonry structures, vital for structural health monitoring (SHM) and proactive maintenance. Trained on healthy operational data, the 1D-CAE reconstructs dominant singular value features from spectral responses and related ambient temperature readings, flagging deviations as anomalies for unsupervised damage detection. To evaluate performance, six Temperature-Compensated Spectral Metrics are introduced, allowing for adjustments for temperature variability. Validation is conducted using data from the Consoli Palace in Gubbio, Italy, during a minor seismic event in May 2021, with assessments covering both short-term and long-term cross-seasonal analyses. The model demonstrates its adaptability to varying environmental conditions while maintaining consistent performance and reducing false positives over time. An Adaptive Dynamic Thresholding mechanism is integrated to dynamically adjust for reconstruction error shifts, making the 1D-CAE-based approach a scalable and efficient solution for SHM in heritage structures. • Autoencoder using dominant singular values and temperature data for anomaly detection. • Introduces temperature-compensated metrics for reliable damage detection. • Show the effectiveness of Adaptive Dynamic Thresholding in reducing false alarms. • Validated in a 14th-century Consoli palace during a May 2021 seismic event. • The proposed framework enables scalable and efficient monitoring for masonry buildings.