Accurate time synchronization of acceleration responses from distributed sensors is a fundamental requirement for Structural Health Monitoring (SHM) of bridges. However, wireless accelerometers commonly suffer from desynchronization due to clock drift, packet delays, and asynchronous duty cycles, leading to phase errors that significantly degrade modal identification and data-driven damage detection. While classical synchronization methods such as cross-correlation and coherence analysis are effective under ideal conditions, their performance deteriorates under noise, non-stationary excitation, or structural damage. This work introduces a damage-insensitive, physics-informed autoencoder framework designed to robustly estimate and correct time delays between accelerometer channels, using only vibration data collected under ambient excitation.

The proposed method exploits the physical directionality of structural wave propagation in bridge members and the inherent tri-axial nature of accelerometers. Instead of relying on stiffness- or frequency-dependent constraints—which are altered by damage—the model incorporates two physics-based principles that remain valid in both healthy and damaged states. First, a direction-aware tri-axial weighting strategy is introduced based on transforming global accelerations into the local coordinate frame of each beam or girder element using geometric rotation matrices obtained from the finite element mesh. The axial component of motion, aligned with the element’s longitudinal axis, carries the strongest and most coherent signatures of low-frequency structural waves and is highly sensitive to time misalignment but largely insensitive to moderate stiffness loss. This component is therefore emphasized in the reconstruction and latent-space learning process, while the transverse components are down-weighted. This introduces geometric, damage-insensitive physical structure into the learned representation.

Second, the model includes a cross-correlation consistency loss, which enforces that the latent vectors of two signals attain their maximum similarity exactly at the correct time alignment. Instead of computing cross-correlation directly in raw signal space, the autoencoder learns a delay-invariant latent representation in which aligned windows map to nearby points, while misaligned windows map farther apart. This contrastive cross-correlation-inspired loss ensures that the encoder’s latent space captures timing relationships that are physically meaningful, excitation-independent, and robust to changes in the structural state.

Synthetic training and evaluation data are generated using a finite element model of a multi-span bridge, discretized with beam elements and excited through random ambient-like loading. Multiple damage scenarios (e.g., localized stiffness reductions) are simulated to assess robustness. Artificial delays are introduced to each sensor channel, giving exact ground-truth alignment for supervised evaluation. The proposed physics-informed autoencoder is compared against classical methods including cross-correlation, phase-based delay estimation, and coherence maximization.

Results demonstrate that the direction-aware, tri-axial geometry weighting and cross-correlation latent consistency enable highly accurate delay estimation even under damaged, noisy, and non-stationary conditions where classical methods fail. Because the physics constraints depend solely on element geometry and vibration directionality—rather than stiffness or modal properties—the method remains effective across a range of damage states.

Overall, this work proposes a damage-insensitive, physically grounded synchronization framework for tri-axial accelerometer networks on bridges, offering a robust alternative to conventional correlation-based methods and supporting more reliable SHM analytics under real-world operating conditions.

Extending the service life of existing civil structures while ensuring acceptable safety and serviceability levels has led to a growing deployment of Structural Health Monitoring (SHM) systems. Vibration-based SHM, and in particular Operational Modal Analysis (OMA), is widely adopted as it extracts modal parameters from ambient excitations without interrupting operation and with relatively low instrumentation costs. However, in steel structures, damage frequently concentrates in joints, where local stiffness losses may have a limited impact on global dynamic properties, making damage difficult to detect with acceleration-only layouts.

This work presents a novel methodology for damage identification of semi-rigid joints in steel structures using a heterogeneous multi-sensor system. The methodology is demonstrated on a laboratory-scale steel frame subjected to various damage scenarios affecting its joints. The methodology aims to identify, locate, and quantify joint damage based on ambient vibration data. For this purpose, the frame was instrumented with two types of sensors, namely accelerometers and strain gauges, used for modal identification via OMA. On this basis, the proposed methodology first updates a finite element (FE) model of the structure using the modal parameters from the undamaged scenario. This updated model is then used to train a supervised damage identification algorithm, exploring both the use of metamodels based on artificial intelligence (AI) and substructuring. The presented results discuss their effectiveness for quantifying the damage at the joints and their strengths and limitations in terms of computational efficiency, interpretability of the results, and potential for being used to generate structural digital twins.

The accurate understanding of real structural behavior throughout a bridge’s life cycle remains one of the major challenges in modern civil engineering. Conventional inspection methods and isolated sensor-based monitoring provide only limited insight, as they rely on periodic measurements, simplified analytical assumptions, and often subjective evaluation. This paper presents the field validation of a high-fidelity digital twin, developed for and commissioned by ASFINAG Austria, that has been in continuous operation for more than four years on a prestressed concrete highway bridge. The system has been active from the beginning of construction through to full operation, enabling the unique opportunity to capture the entire structural evolution process.

The digital twin is designed to self-calibrate and update autonomously in real time by integrating continuous distributed sensor data with nonlinear finite element (FE) simulations, advanced inverse analysis, and artificial intelligence. The methodology combines physics-informed neural networks (PINNs), operational modal analysis (OMA), and reverse parameter identification to infer actual mechanical and material properties. More than 7 km of fiber-optic sensors are embedded within the deck, webs, and foundations, continuously recording strain and temperature profiles. The data are processed by the cloud-based WeStatiX SHM platform, which performs automated numerical recalibration and executes daily nonlinear simulations of the complete structure.

Over four years of continuous monitoring have provided an unprecedented dataset that reveals significant differences between the bridge’s actual structural response and the original design assumptions. Measured deformations and stress redistributions indicate that temperature gradients, long-term creep, and local prestressing effects dominate the behavior far more strongly than predicted by design codes. The continuously updated digital twin has captured these effects with high accuracy, reducing the discrepancy between measured and simulated strains to below1 %.

Validation was carried out through live load testing and operational verification campaigns, confirming the reliability of the self-calibrating approach. The system demonstrates stable performance under varying environmental conditions, with consistent prediction accuracy across multiple years. It provides objective, real-time decision support for operators, enabling proactive maintenance planning, early anomaly detection, and precise evaluation of residual capacity.

This work represents one of the first long-term, full-scale validations of an autonomous digital twin for civil infrastructure. By integrating sensing, simulation, and data-driven intelligence into a single self-learning framework, the system shifts structural health monitoring from reactive data interpretation to predictive knowledge generation. The results from this ongoing ASFINAG project highlight the potential of such high-fidelity digital twins to extend the service life of critical assets, optimize maintenance interventions, and build the foundation for a new generation of continuously self-updating, high-trust SHM systems.

Decision making in structural health monitoring (SHM) usually involves assessing the condition of civil engineering structures and prescribing repair or retrofit measures. The condition assessment is either data-driven, based on comparing SHM data from the current structural state to SHM data from a benchmark “undamaged” state, or physics-based, aided by numerical models for estimating the structural and material properties. Existing physics-based approaches suffer from discretization mismatches between typical finely meshed numerical models and sparse topologies of SHM systems, resulting frequently in ill-posed optimization problems. In this direction, this paper introduces a methodology for physics-based condition assessment of civil engineering structures leveraging reduced-order modeling to alleviate the discretization mismatch. The proposed methodology involves condensing finite element (FE) models of monitored structures to degrees of freedom measured by the SHM systems, ensuring that the structural dynamic characteristics of the full FE models are accounted for to the best possible extent. The condensation results in reduced-order models that are used together with SHM data to describe the structural behavior (Fig. 1). Condition assessment is achieved by creating a family of reduced-order models, corresponding to various damage scenarios, and by finding the model that best describes the structural behavior. Validation tests, conducted on a shear-frame structure, as well as on a real-world road bridge, showcase the capability of the proposed methodology to yield reduced-order models that best fit the SHM data, thus providing rich information on the structural condition of civil engineering structures. The work presented in this paper is expected to pave the way towards fully automating condition assessment (and, in turn, decision making) in state-of-the-art decentralized, autonomous SHM systems.

Ice accretion on aircraft wings remains one of the most critical challenges for flight safety, as even thin layers of ice can severely degrade aerodynamic performance and significantly reduce operational safety margins. Existing ice-detection systems used in aviation typically rely on environmental sensing, performance monitoring, or single-point detectors, which often provide indirect, delayed, or spatially limited information. To address these limitations, current research is increasingly exploring Structural Health Monitoring approaches capable of offering distributed, real-time, and physics-based detection of ice directly on the structure.

In this work, an ice-detection algorithm is proposed leveraging Lamb-wave interrogation with piezoelectric transducers. Guided ultrasonic waves are particularly suited for this purpose thanks to their ability to propagate over long distances while being highly sensitive to local variations in stiffness and mass introduced by ice accretion. When interacting with an ice layer, Lamb waves experience scattering, mode conversion, and local reflection patterns that encode information about the altered structural conditions.

The core of the proposed method relies on constructing a Gaussian Process Regression (GPR) model of the healthy, ice-free structure. A key aspect is in the choice of the training features. Instead of using raw sensor signals, the GPR is trained on the differences between autocorrelations of the reference Lamb-wave responses. Notably, this data-driven formulation requires no prior knowledge of the plate geometry, material properties, or expected reflection patterns, as the model learns the healthy wavefield directly from the reference correlations.

During inspection, each newly acquired signal is compared with the baseline response derived from healthy conditions, and a diagnostic feature quantifying their deviation is extracted. The GPR then outputs a probabilistic prediction quantifying the deviation from the healthy state, providing both a detection metric and an uncertainty estimate. This correlation-based representation naturally enhances reflection phenomena associated with ice, while suppressing global variations that do not carry diagnostic value, such as reflections from the plate boundaries or from other fixed structural features.

The method has been experimentally validated on an aluminum plate instrumented with piezoelectric actuators and sensors. The experiments, which include various ice thicknesses and spatial configurations, demonstrate that the algorithm can reliably detect the onset of ice even in the presence of complex multi-path reflections and temperature variations. The use of piezoelectric transducers for ice detection also paves the way for fully integrated systems capable of both identifying and removing ice in real time, offering a promising solution for future aeronautical applications.