Bridges are critical infrastructure assets whose structural reliability is increasingly challenged by evolving traffic demands, operational policies, and stochastic loading conditions. Conventional assessment approaches based on standardized load models, such as Eurocode Load Model 1 (LM1), are not tailored to capture site-specific dynamic traffic characteristics, particularly under congestion and long-term growth. This study presents a probabilistic framework for evaluating traffic-induced load effects on a critical bridge section using high-resolution Weigh-In-Motion (WIM) data, Monte Carlo simulation, and extreme value theory. A stochastic traffic model is developed to simulate realistic vehicle interactions, incorporating variability in axle loads, vehicle spacing, and traffic states. Extreme load effects over a 20-year remaining service life are estimated using a Gumbel distribution. To enable efficient scenario analysis, a high fidelity power-law surrogate model is developed and validated, relating load effects to key operational parameters: load factor (L_f), gap factor (G_f ), and traffic growth rate (g). Sensitivity analysis reveals that extreme shear events pose a higher probability of exceeding limit states than bending extremes under vehicle clustering and legal weight increases within the 20-year service life horizon. Calibration against LM1 demonstrates that the Eurocode model is generally conservative under baseline conditions but may underestimate structural demand in extreme loading scenarios, where exceeds unity. Finally, a Digital Twin framework integrating WIM and Structural Health Monitoring (SHM) data is proposed to support real-time reliability assessment and proactive traffic management. The study highlights the importance of site-specific calibration and data-driven methodologies for adaptive bridge asset management.

Condition- and predictive-maintenance strategies rely on accurate, real-time diagnosis and prognosis, yet damage-sensitive features are often distorted by changing environmental and operating conditions. This challenge is addressed by combining feature normalisation via an autoencoder (AE) with a model-based particle filter for online state estimation and damage localisation. The problem faced is damage diagnosis on a vibrating n-degrees-of-freedom system, featuring a non-linear stiffness component characterised by a Bouc-Wen hysteretic behaviour and subject to varying temperature conditions; temperature-induced drifts are removed by an AE trained on healthy data to recover damage-related features. On top of a standard particle filter (PF), an Action-aided PF (A-PF) is proposed to augment classical likelihood weights with a physics-inspired action term, discouraging high-action (non-plausible) trajectories and promoting path-coherent posteriors. In physics, the principle of least action states that the actual path taken by a physical system between two configurations is the one for which the action is stationary, usually a minimum. Thus, an action-based weighting scheme can be proposed to encourage particle trajectories that follow low-action (more coherent) paths, such as smoother and more physically plausible evolutions. Using identical seeds, noise regimes, and resampling policies, PF and A-PF are compared across state estimation and detection/localisation performance. Results show improved performance of the A-PF compared with the conventional PF, while maintaining comparable computational cost. These findings indicate that a simple, domain-informed path regularisation in PF weights can materially improve SHM performance for damage diagnosis and localisation in realistic, non-stationary conditions.

This study investigates a novel technique to estimate new local vibration parameters to monitor existing buildings in the structural health monitoring field. The paper also gives a comprehensive discussion on their use for damage localization, as well as how these parameters are related to the variations of structural eigen-properties, with the aim of connecting local with global damage features. In order to estimate these new local vibrational parameters, the authors propose the implementation of a new modelling strategy based on peridynamics, which represents a novel continuum theory. This is performed under operational conditions assumptions, i.e., without the need to force the system with external inputs. The idea is to exploit the discrete formulation of peridynamics to build low-fidelity structural models, useful for rapid and scalable analyses. The focus of this study is to generate a network of points and bonds, where each point represents the spatial position of a significant location within a structure. The goal is to construct a bond force network that structurally connects these points, approximating the structural topology and connectivity of the system. Starting from the spatial coordinates of these points, the authors apply the peridynamic theory to determine the corresponding values of mass, stiffness, and damping parameters. This allows to build a simplified yet physically meaningful model of the structure, suitable for structural health monitoring. Within this framework, two parameters are obtained to localize damage: Bond Extremity Acceleration (BEA) and Bond Extremity Velocity (BEV). Both are derived from a micro-viscoelastic formulation of peridynamic bonds and can be indirectly estimated from acceleration response data using established techniques. The methodology is tested on a numerical reproduction of a three-story aluminum frame, modeled with 36 degrees of freedom and subjected to Gaussian noise. The conclusions of the study suggest that low-fidelity peridynamic models can be a useful basis for developing physical representations of structural networks, suitable for monitoring existing buildings and infrastructures.

Accurate prediction of crest displacement is a critical concern for structural health monitoring and safety assessment of concrete dams. Conventional predictive models typically approximate thermal effects using sinusoidal seasonal terms, ambient air temperature, or sparse concrete temperature sensors data. However, when dense temperature measurements are available across upstream, mid-body, and downstream in a cross-section at multiple elevations, selecting optimal thermal inputs poses a significant challenge.
This study presents a data-driven framework integrating Principal Component Analysis (PCA) to reduce the dimensionality of temperature datasets and extract dominant thermal patterns. These PCA-derived features are combined with hydrostatic load and time-dependent effects in a multiple linear regression (MLR) model. Applied to the Ridracoli arch-gravity concrete dam, a major dam in Italy, the PCA-MLR model outperforms conventional sinusoidal-term approaches in prediction accuracy. The results demonstrate that leveraging high-resolution sensor data through PCA enhances displacement modeling, offering a robust methodology for dam behavior analysis and informed decision-making.

Modal filtering transforms spatial vibration measurements into modal coordinates, simplifying tasks such as model correlation, force identification, and damage detection. Classical modal filters rely on a full modal model consisting of natural frequencies, damping ratios, and mode shapes, which limits applicability when modal parameters are difficult to obtain or when operational conditions vary. Recent results show that modal filters can instead be synthesized directly from measured frequency response functions using natural optimization, eliminating the need for modal analysis and demonstrating that the transformation to modal space is not tied to a single computational pathway.
In this work, we broaden this perspective and propose a framework of alternative data-driven pathways to modal coordinates. First, we summarize the optimization-based approach, where modal filters are obtained by maximizing filtration quality using evolutionary search applied to FRFs. Next, we introduce a physics-informed learning route in which a neural network is trained to map physical system parameters to their corresponding modal filters, showing that the transformation can be learned rather than derived. Building on this, we consider a more practical formulation where neural networks infer modal filters directly from FRFs, serving as fast surrogates for optimization-based synthesis. Finally, we outline an exploratory pathway in which structural representations, such as geometric or mesh-based descriptions, are used to estimate modal filters, implicitly learning dynamic behavior from structural information.

Together, these pathways illustrate that modal coordinates can be recovered through multiple computational strategies, enabling adaptive, generalizable, and efficient alternatives to classical modal analysis. We show proof-of-concepts for selected pathways based on simulated family of dynamic systems and discuss implications for SHM, model updating, and vibration-based diagnostics.