Use of a population-based approach for structural health monitoring of underground structures

Underground structures are becoming increasingly important in modern society, promoting the development of efficient and reliable infrastructure networks, especially in metropolises and densely populated areas. Given the growing mileage of underground structures designed for public transportation, such as metro tunnels, the development of Structural Health Monitoring (SHM) strategies capable of ensuring structural integrity, efficient operational conditions and user safety is paramount. Several monitoring techniques have already been proposed, both traditional and innovative, as well as static and dynamic approaches. However, the vibration-based methods developed for this type of structure are relatively few and often require the knowledge of a large amount of data from the monitored structure. In many practical applications, these datasets are missing, making data-driven approaches impractical in terms of time, cost and efficiency. These drawbacks could be addressed through Population-based Structural Health Monitoring (PBSHM), already used in other engineering applications, which exploits the transfer learning of damage-sensitive features across similar structures to investigate their state of health and build a large-scale monitoring approach, more automated and efficient. However, this monitoring methodology has not yet been applied to underground structures. This thesis first investigates the feasibility and accuracy of the Transmissibility function (TF) and Cross-correlation function (CCF) as vibration-based damage identification methods for underground structures. Their effectiveness is analysed by performing numerical simulations on a tunnel-soil finite element model; two procedures of hammer impact tests are proposed to extract the structure’s acceleration responses from which the transmissibility damage indicator (TDI) and cross-correlation damage indicator (CDI) are computed to localise an artificially introduced damage. The same investigations are also conducted in the presence of noise, imitating real measurements. Subsequently, a second tunnel-soil model, similar to the first one, is developed to extend the use of TFs as sensitive features in a PBSHM framework for a heterogeneous population of underground structures. Thus, a Domain Adaptation (DA) algorithm, based on Statistic Alignment (SA) and Normal Condition Alignment (NCA), has been implemented to share knowledge between the two numerical metro tunnels. The results confirm that both strategies (TFs and CCFs) appear to be quite sensitive to damage localisation in most of the cases considered, but showing also a non-negligible dependence on the input position. However, neither the TDI nor the CDI provide thresholds to assess the presence of damage. Conversely, using the proposed PBSHM approach, damage detection can be performed on the second structure, leveraging the knowledge gained from the first tunnel. The results exhibit precision and robustness, showing that TFs could be a promising feature for the development of a PBSHM approach for underground structures monitoring.