Conference Agenda

Fatigue crack initiation and propagation remain critical challenges in structural engineering due to their potential to cause unexpected structural failures with severe safety risks and economic consequences. This study proposes a digital-twin-enabled approach for in-service fatigue crack detection and quantification, where real-time experimental data from the physical entity are continuously integrated into the virtual model to maintain synchronization and adaptive model updating. Within this framework, online dynamic response measurements from piezoelectric (PZT) transducers are processed using a Dynamic Piecewise Linear (DPL) algorithm, which facilitates automated and real-time detection of crack initiation during service. Experimental validation under various loading frequencies, stress ratios, amplitudes, and multi-block loading scenarios demonstrates reliable detection of sub-millimetre cracks (<2 mm). For fatigue crack growth prediction, a Physics-Informed Long Short-Term Memory (PI-LSTM) model is developed by embedding Walker’s law as a physical constraint, while Bayesian optimization ensures optimal hyperparameter selection. The continuously updated digital twin shows superior prediction accuracy and generalization across different specimens under different experimental conditions, highlighting its capability for adaptive, real-time structural health assessment. This study advances the development of intelligent digital twins for fatigue life prediction and condition-based maintenance in complex service environments.

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.

Accurate time synchronization is essential for vibration-based structural health monitoring (SHM), particularly when acceleration responses are collected using distributed wireless sensors. Residual timing errors may remain after clock-level synchronization due to clock drift, triggering jitter, communication latency, and hardware-dependent acquisition delays, introducing phase distortion and reducing the reliability of modal identification and damage-sensitive analysis. This paper proposes a post-processing synchronization framework for multi-sensor tri-axial bridge acceleration signals under ambient vibration. The method combines learned waveform representations, graph-based sensor interaction, classical delay-estimation evidence, and physics-informed consistency constraints. A shared convolutional encoder extracts sensor-level latent features, while graph message passing models inter-sensor dependencies over a complete sensor graph. In parallel, correlation- and phase-based edge features provide physically meaningful delay cues for each sensor pair. Raw pairwise delay estimates are projected into cycle-consistent sensor offsets using a graph-based least-squares formulation, and the predicted offsets are used to compensate the measured signals. A two-stage inference strategy is further introduced, where the model first estimates delays from the original asynchronous signals, synchronizes the signals using the extracted offsets, and then performs a second inference step to refine the final estimates. The model is trained using healthy and mass-varied finite element simulations with artificial timing offsets, while damaged scenarios are reserved for testing to evaluate generalization to unseen structural states. The results show that the proposed two-stage framework provides more accurate and robust synchronization than direct inference and the classical GCC-PHAT baseline, supporting its use as a damage-insensitive signal-level synchronization approach for wireless bridge SHM.

The past decade has shown a trend to collect and mine large quantities of our data with the objective to assess the performance of materials and systems for various reasons. This rise of Artificial Intelligence has also largely affected the field of Structural Health Monitoring. A plethora of intelligent data science algorithms has been studied, developing the ability to recognise subtle changes in the (dynamic) behaviour of structures resulting from the presence of early stages of damage or deterioration. More recently, the direction has been altered again, towards inclusion of physics-based information. This is done on various levels and in different ways, but all with the shared objective to mitigate one of the most important drawbacks or limitations of pure data-centric methods: the limited performance in case of poor, incomplete, scares or non-representative data. This lack of data quality is not just solved by collecting more data and data with higher accuracy, but is also an inherent problem for monitoring systems of actual structures: maintenance interventions, securing the safe operation of the structure, imply hardly any data is collected of failed systems.

In addition, the blunt collection of data may work counterproductive if it comes to practical implementation of these methods. The collection and processing of this data is certainly not free of charge, while the additional information (hence the value) of more data is often limited. Hence, research is directed towards selection of relevant and informative features, reducing the data demand, while maximizing the information output.

Although these developments are doubtlessly of great value, a different perspective may provide new ideas on how to overcome some of the challenges of finding the right balance between knowledge and data to optimise the information output of a monitoring system. What is nature's perspective on how to extract information efficiently from large and diverse data sources?

Human sense can be considered as a sensor system with diverse set of sensors and a large variety and flexibility in how these are connected and are mutually communicating. The human sensory system collects a lot of data, but the processing of this data differs fundamentally from the way data is processed in data-centric methods. It is true that a neural networks mimics to some extent the way a brain works: general patterns of behaviour are trained at first, after which new observations are mapped on these patterns to identify the observed behaviour. In reality, human brains use a limited part of the data to guess what is happening. Despite clear evidence that human brains can be fooled, the general functionality of the approach is fairly good to say the least. This work shows the potential of applying this method, by introducing a layered approach and addressing how various sensor source can be combined and how decision can be made regarding the data be used and discarded.

Structural Health Monitoring (SHM) is becoming essential in civil engineering due to its ability to continuously assess the condition of infrastructures and detect potential damage. SHM techniques are generally categorized into data-driven (DD) and model-driven (MD) approaches. DD methods exploit long-term monitoring data to identify the standard structural behaviour and define relevant thresholds for anomaly detection. However, these methods typically lack labelled damage examples to train the algorithms. On the other hand, MD methods rely on updated finite element (FE) structural models to define the expected performance, paying the price of being computationally demanding, requiring substantial technical and economic resources. Recent research has therefore focused on transfer learning (TL) strategies that integrate the strengths of both paradigms. TL provides a powerful knowledge-generalization framework that allows information learned from an abundant, labelled source domain to be transferred to a related target domain where labelled data are scarce. Recent research has therefore focused on hybrid approaches (HAs) that integrate the strengths of both DD and MD paradigms. These strategies are designed to overcome the inherent limitations of each method, compensating for the lack of physical interpretability while mitigating the high computational demand. By merging physical consistency with the flexibility of statistical learning, HAs provide a more robust and reliable tool for SHM. In this work, a HA is applied to a span of a viaduct monitored with multi-year clinometer measurements. First, a FE model is developed and used to generate a large labelled dataset covering multiple damage scenarios. Subsequently, instrumental noise is stochastically injected into the synthetic data. This augmented dataset is then used to train and validate Physics-Aided Surrogate Models (SMs) based on neural networks, to ensure physical consistency while achieving high computational efficiency. Once trained, SMs are applied to real clinometer rotations to infer structural damage indicators. To further demonstrate the reliability of the system, the process is validated by integrating simulated damage scenarios with experimental monitoring data, thereby testing the framework’s robustness against complex, real-world conditions. Results, tested on a multi-set of real bridges, highlight the potential of this hybrid framework as a scalable and effective tool for next-generation SHM systems in bridges and large civil infrastructures.