New approaches in Structural Health Monitoring: Model Updating using automated intelligent algorithms

The increasing necessity to ensure high levels of safety, efficiency, and dura-bility in civil infrastructure and existing structures is one of the central and current topics in structural engineering. Buildings, bridges, and other strategic works, often built according to design criteria that are now obsolete, are currently exposed to deg-radation processes and a significant increase in uncertainty regarding knowledge of material characteristics, constraints, and load conditions. The management of these uncertainties is not only a technical challenge, but also a social and economic one, as the risk of damage or malfunction can have a considerable impact on communi-ties and the environment. In this context, Structural Health Monitoring (SHM) has es-tablished itself as an essential tool for continuous and timely diagnosis of structural conditions. Modern monitoring campaigns, based on the systematic acquisition of dynamic responses and the application of Operational Modal Analysis (OMA), allow significant global dynamic parameters to be obtained, including frequencies and modal shapes. However, data acquisition alone is not sufficient to ensure a robust assessment of structural performance, especially in the presence of exceptional ac-tions, changes in use, or degradation phenomena. In these cases, the calibration phase of numerical models, known as model updating (MU), is of fundamental im-portance. It is necessary to obtain the digital twin of the structure, reduce the uncer-tainty of the analyses, and support more efficient predictive maintenance strategies. Traditional MU strategies, both deterministic and probabilistic, are widely used in scientific practice, but they have evident limitations: they require strong a priori as-sumptions, are computationally expensive, and are often difficult to apply to systems characterized by high dimensionality of uncertain parameters. Automation and intelli-gent efficiency of the updating process are now considered priority objectives by the international scientific community. The objective of this PhD research is to contribute to methodological advances in the field by integrating experimental data from real SHM campaigns with the latest generation of artificial intelligence techniques. The core of the proposal is to develop, validate, and compare an advanced framework for fully automatic model updating of complex structures, based on the combined use of physically based Reduced Order Modeling (ROM) and Deep Reinforcement Learning (DRL). The key idea is to manage calibration through an intelligent agent capable of interacting with the numerical model and real data, progressively learning (through reward mechanisms typical of DRL) the updating strategies that best reduce the dis-crepancy between the simulated response and the real response of the structure. The framework was designed and organised in five consecutive stages. Given the moni-toring campaign data, the definition of the target variables of interest and a numerical model of the structure in an externally accessible environment, the proposed ap-proach defines a DRL engine, in which a strategy for fine-tuning the structural control parameters was defined. At the same time, the training/testing phases of the intelli-gent agent were carried out to obtain the correct solution based on a statistical eval-uation of the results. The methodology, divided into DRUM-Av1.0 and DRUM-Av2.0 versions, is applied and validated both on theoretical case studies and on real infra-structures, such as bridges and reinforced concrete frames. In the advanced version of the framework (DRUM-Av2.0), the introduction of ROMs, constructed using de-composition techniques from Full Order Models (FOMs), allows for a significant re-duction in the calculation times required to train and validate the DRL agent, making the methodology applicable even to systems characterized by a high number of de-grees of freedom and unknown parameters. The quantitative results obtained in the various case studies show that the approach: (a) guarantees high accuracy in the es-timation of uncertain parameters; (b) provides higher computational efficiency than classical methods; (c) is robust with respect to interferences caused by incomplete information or experimental noise; (d) allows for almost complete automation of the calibration process. Furthermore, in-depth analysis highlights the possibility of extending the tar-get variables (e.g., by including modal shapes and data from in situ tests), as well as the future development of intelligent agents equipped with physics-based engineering knowledge, for an even more reliable and interpretable solution to structural updating problems. These prospects drive the methodology towards the achievement of an in-telligent and self-adaptive digital twin, able to learn and optimise structural manage-ment throughout the entire life cycle of structures. In conclusion, the work developed in this thesis defines and validates an in-novative, efficient and scalable approach for the automatic model updating of struc-tural models. The approach responds to the needs of modern research, proposing implementable tools for the predictive, resilient and adaptive management of civil en-gineering works, and provides a solid foundation for the evolution of structural digital twins in the field of structural engineering.