Model-based and data-driven methodologies toward predictive digital twins of structures

The digital twin concept has emerged as a paradigm to enable diagnostic and predictive capabilities that are not achievable with computational models alone. A digital twin is a personalized virtual representation of a physical asset that evolves over time. It relies on a set of computational models that are dynamically updated to persistently mirror the physical counterpart throughout its life cycle. Additionally, it features predictive capabilities that inform decisions tailored to realize value in the physical setting of interest. A bi-directional interaction between the physical and virtual domains, comprising either automated or human-in-the-loop feedback flows, is a critical enabler for digital twins. Enhanced computational efficiency is also required to handle the continuous assimilation of noisy and big data, as well as to quantify and propagate relevant uncertainties. This Thesis aims to advance digital twin technologies for monitoring the structural integrity of engineering systems by hybridizing physics-based modeling with artificial intelligence. Enabling a digital twin perspective for critical structural systems, whether for safety or operational reasons, can unlock the potential of condition-based and predictive maintenance practices, yielding numerous benefits throughout their life cycle. We begin by laying down a theoretical groundwork and integrating cutting-edge insights from data science and engineering. This encompasses: (i) model-based and data-driven approaches to parameter identification; (ii) full and reduced-order computational models in structural dynamics to formulate damage identification strategies within the simulation-based paradigm of structural health monitoring; and (iii) machine learning techniques to extract information from sensor data, forecast future system responses, and determine optimal courses of action. Subsequently, we propose four contributions that combine simulation-based supervision, essential for addressing the lack of experimental data related to damage onset and propagation, with learnable mappings aimed at assimilating vibration recordings, enhancing damage identification capabilities, and improving computational efficiency. Firstly, we introduce a data-driven diagnostic framework that leverages deep metric learning to map raw sensor recordings onto a damage-sensitive low-dimensional space. We investigate its ability to locate damage by exploring different training options within a Siamese architecture, demonstrating remarkable accuracy and robustness against variations in the extent and severity of damage. Secondly, we tackle the computational efficiency of surrogate models for structural health monitoring. We employ both full and reduced-order physics-based models to sequentially train surrogate components with increasing levels of approximation. The resulting multi-fidelity surrogates serve both model-updating purposes and the generation of large datasets, demonstrating superior accuracy compared to single-fidelity approximations without increasing computational costs. As a third contribution, we present a hybrid data/model approach for real-time damage identification, enhancing Bayesian model updating through learnable mappings. Damage-sensitive features are extracted from sensor recordings via deep metric learning, and then exploited within a Markov chain Monte Carlo sampler to assess damage with quantified uncertainty. The achieved estimation accuracy and computational efficiency overcome limitations linked to the slow convergence of sampling algorithms and the curse of dimensionality affecting the inference of low-sensitivity damage parameters. Finally, we propose a digital twin framework relying on a probabilistic graphical model to encode the asset-twin interaction via observational data and control inputs. The decision-making capabilities of health-aware digital twins are assessed through demonstrations of data-driven health monitoring, prediction, and planning. While experimental data are not used, all methodologies are assessed across various case studies using noisy synthetic measurements.

Il concetto di gemello digitale è un paradigma emerso recentemente per consentire capacità diagnostiche e predittive non raggiungibili con i soli modelli computazionali. Un gemello digitale è una rappresentazione virtuale di un asset o processo fisico che evolve nel tempo. Si basa su modelli computazionali che vengono aggiornati dinamicamente per rispecchiare la controparte fisica durante tutto il suo ciclo di vita. Inoltre, possiede capacità predittive in grado di informare decisioni mirate a realizzare valore. Un aspetto fondamentale di tale paradigma è l’interazione bidirezionale tra il dominio fisico e quello virtuale. Simultaneamente, è richiesta una forte efficienza computazionale per gestire la continua assimilazione di dati rumorosi e di grandi dimensioni, nonché per quantificare e propagare le relative incertezze. Questa tesi mira a far progredire l’utilizzo dei gemelli digitali per il monitoraggio dell'integrità strutturale di sistemi ingegneristici, attraverso l’uso combinato di modelli basati sulla fisica e intelligenza artificiale. Tale prospettiva permetterebbe infatti di sfruttare il potenziale delle pratiche di manutenzione predittiva, offrendo numerosi benefici estesi a tutto il ciclo di vita.