Advanced Anomaly Detection in Structural Health Monitoring: A Case Study on Inclinometer Data from Italian Viaducts

Civil engineering structures, such as bridges and viaducts, are subjected to environmental stresses, material fatigue, and excessive loads, all contributing to gradual degradation. This deterioration poses significant risks to both safety and functionality. Structural Health Monitoring (SHM) has become essential for ensuring the integrity and longevity of these infrastructures through continuous assessment. By leveraging advanced sensors and data-driven techniques, SHM enables early anomaly detection, facilitating timely interventions and optimized maintenance strategies. This thesis explores the integration of Machine Learning (ML) and Deep Learning (DL) to further enhance SHM by improving the processing of large-scale data and boosting the accuracy of anomaly detection. In the first part of the thesis, data from 142 MEMS-based inclinometers and temperature sensors installed on 18 Italian Viaducts have been collected and organized in a unified structured dataset. Each record is hand-labelled to build a comprehensive dataset consisting of 87,161,948 records, with 9 anomaly types labelled. These anomalies are categorized into two broad groups: instantaneous rotation accumulation (e.g., sensor detachment, heavy load stationing) and progressive rotation accumulation (e.g., thermal deformation). This distinction is made to differentiate between anomalies characterized by a sudden dramatic variation in acceleration values between successive readings and those that exhibit a gradual mean shift in acceleration due to progressive rotation. Lastly, as inclinometer data are collected with non-constant sampling rates, we resample them to a fixed time interval, and training, validation, and test sets are generated using sensor-based and time-based splitting strategies. In the second part, two distinct anomaly detection pipelines have been developed to leverage this newly labelled dataset for ML- and DL-based SHM systems. The first pipeline focuses on the anomaly prediction capabilities of a Random Forest model, while the second pipeline adapts the Temporal Convolutional Network (TEMPONet), originally designed for EMG classification, to our anomaly detection use case. Both pipelines are trained for two classification tasks: a 10-class task, which includes the 9 distinct anomalies and the regular signal, and a 3-class task, where the anomalies were grouped into the two broader groups alongside the regular signal. In the latter task, the Random Forest model achieves a macro-average accuracy of 55% for instantaneous detection, while TEMPONet demonstrates macro-average accuracy of 88% for the same task. These results highlight the effectiveness of ML and DL techniques in enhancing SHM processes, providing valuable insights into the early detection of structural issues.