PSI - Issue 78

Akshay Rai et al. / Procedia Structural Integrity 78 (2026) 891–898

892

the risk of catastrophic failures, particularly following events like earthquakes or significant environmental stresses [1], [2], [3]. However, masonry structures—such as historic towers, bridges, and walls—pose unique challenges for anomaly detection, due to material heterogeneity, complex dynamic Behaviour, age and Conservation, and En vironmental Sensitivity. Changes in temperature and humidity can a ff ect dynamic properties, making it hard to di ff erentiate between real anomalies and benign e ff ects [4], [5], [6]. Recent advancements in deep learning are enhancing anomaly detection for masonry structures, particularly after seismic events. Techniques like Temporal Fusion Transformers (TFTs) and Convolutional Neural Networks (CNNs) e ff ectively model normal vibrational behaviour and identify damage or abnormal events. For instance, TFTs analyse long-term data from heritage towers, learning their natural frequencies and detecting deviations from earthquakes or simulated damage [8] and [7]. Additionally, deep learning is being applied for crack detection on masonry fac¸ades, expanding the capabilities of anomaly detection. [10], [9] Multiple studies emphasise the considerable influence of ambient temperature and environmental factors on masonry structures. Key insights include: Frequency Shifts - The natural frequencies of masonry towers and bridges can vary significantly with temperature fluctuations, sometimes even more so than changes caused by structural damage. This can obscure or amplify indicators of genuine issues [4], [5], [6]. False Positives / Negatives - If these environmental e ff ects are not adequately considered, structural health monitoring (SHM) data may result in false alarms or overlooked detections. Data-driven integration is another realm where state-of-the-art deep learning models have the potential to incorporate environmental variables such as temperature, humidity, and wind. This might enable the models to di ff erentiate between normal environmental responses and actual damage more accurately [7], [8]. Some studies propose hybrid frameworks that combine clustering, anomaly detection, and environmental partitioning to mitigate the e ff ects of temperature-induced variability. [2], [11] Overall, it is essential to account for temperature e ff ects, as they significantly impact variations in structural response in masonry and other civil structures. Deep learning approaches can e ff ectively model both structural and environmental data, enhancing fault detection. Incorporating ambient temperature as a contextual variable is essential to avoid misdiagnosis. An accurate anomaly detection framework must explicitly consider ambient conditions to di ff erentiate between structural issues and routine variability, ensuring the e ff ective management of historically significant buildings. This study proposes a novel anomaly detection framework that combines structural response data with ambient temperature recordings, utilising a convolutional autoencoder (CAE) for unsupervised anomaly detection. The CAE captures the relationship between seismic responses and temperature variations, enabling the identification of genuine anomalies post-seismic activity. The framework is validated using the Consoli Palace, a 14th-century monument in Italy, serving as a relevant case study for monitoring masonry structures a ff ected by seismic events. The overarching purpose of the proposed anomaly detection framework is to facilitate early diagnosis of dam age in historically significant masonry structures. This early diagnosis aims to enable timely intervention and mitigate the risk of irreversible damage to these structures. The scope of this framework includes: • Integrating structural vibration data with ambient temperature sensor recordings. • Utilising a deep learning model, specifically a convolutional autoencoder (CAE), for unsupervised anomaly detection. The CAE is trained to capture the complex relationship between the structure’s seismic response and environmental temperature variations. This allows it to distinguish between normal fluctuations and deviations caused by damage. • Developing an error-based metric to identify novel anomalies that follow seismic activity systematically. • Monitoring masonry structures subjected to seismic events. • Providing a real-world application and case study to evaluate the robustness and e ff ectiveness of the frame work, as exemplified by its application to the Consoli Palace in Central Italy. 2. Proposed framework for anomaly detection

2.1. Monitoring data

The data acquisition process for the proposed anomaly detection framework is designed to monitor masonry buildings and capture their dynamic responses under various environmental and operational conditions. The present framework utilises the vibration signals x s ( t ) from S accelerometer sensors, as well as the structure’s external temperature ( T 1 ), and internal temperature ( T 0 ) readings. Therefore, each data Sample ( X ) acquired for analysis comprehensively includes: