Conference Agenda

Standardised, exchangeable hydrogen tank modules potentially offer a flexible and scalable solution for mobile Gensets. To ensure operational safety across a large number of such modules under uncertain operational conditions, the structural integrity and operational loads of all hydrogen tanks must be monitored throughout their service life. Structural health monitoring (SHM) of hydrogen pressure vessels is particularly challenging due to manufacturing-related uncertainties and complex degradation mechanisms during operation. This study presents a fleet-level monitoring concept for modular exchangeable hydrogen tank units based on in-service surface-strain assessment using FBG sensors. The integration of fleet-wide data and a priori knowledge enables improved anomaly detection and early warning of systematic issues. In this study, a set of axisymmetric finite-element models of hydrogen pressure vessels serves as a priori input on expected surface strain during operation while accounting for manufacturing-related uncertainties. A second set of finite element models incorporates further sources of operational uncertainty and serves as simulated fleet data. This work explores the utilisation of Bayesian inference to update the a priori assumptions based on available fleet-data and assess the structural state of individual hydrogen pressure vessels. The framework is then evaluated based on its capability to detect damage in a third set of finite element models with local delamination within the dome region.

Ensuring the long-term integrity and health of bridge structures under diverse structural, environmental, and operational conditions remains a persistent challenge within the structural health monitoring (SHM) community. Although machine learning–aided unsupervised anomaly detection offers promising solutions for vibration-based SHM, its generalization and robustness across different bridge types and structural conditions are still limited. A model trained for one bridge often fails to perform effectively on another with distinct material properties, geometries, or boundary conditions. To overcome this limitation, this study introduces a novel hybridized unsupervised learning (HUL) framework for detecting structural changes in heterogeneous concrete bridge systems. The proposed framework employs daily measurement of modal frequencies as the primary dynamic features for vibration-based change detection, performing two essential tasks including data normalization by synergetic integration of Bidirectional Encoder Representation from Transformer (BERT) with statistical regularization of a Generative Adversarial Network (GAN), to suppress environmentally/operationally-induced outliers. At the last stage of HUL the change detection is conducted via applying a statistical anomaly detector on obtained residuals to identify genuine structural changes. Using a masked training strategy within the GAN structure, the generator utilizes a self-attention mechanism to reconstruct input sequences of modal frequencies where a specific portion of time steps are stochastically masked. By minimizing a dual-loss objective combining reconstruction error with adversarial loss, the model learns underlying temporal correlations of the healthy bridge state without requiring labelled damage data. Three concrete bridges, including one numerical and two experimental structures, with distinct configurations are used to validate the proposed method. Moreover, structural variations are categorized into four representative classes: detrimental (damage), innocuous (normal environmental or operational conditions), extreme (severe environmental and operational conditions), and mixed changes. Results indicate that the HUL framework can effectively detect true structural changes while mitigating false alarms arising from environmental and operational variability, thus providing a generalizable and robust approach for concrete bridge monitoring.

Deep learning-based damage identification often exhibits limited generalization when applied to structures that differ from those used for training. Although domain adaptation techniques can improve transfer performance under relatively mild structural variations, their effectiveness may degrade when the target structure exhibits changes in geometry, topology, or boundary conditions. To address this issue, this paper proposes a cross-structure transfer learning framework for graph-based damage identification using modal features. In the proposed framework, each structure is represented as a sensor graph, where nodes correspond to measurement points, node features are constructed from measured mode shapes, and edges are defined according to the physical adjacency between sensor locations. A graph sample and aggregate network is employed to learn topology-aware node embeddings by aggregating local neighborhood information. Damage localization is then achieved through edge-level anomaly detection based on the inner-product similarity between embeddings of adjacent nodes. A significant change in this similarity indicates a variation in the embedding relationship along the corresponding physical connection, thereby revealing the potential damage location. The proposed method learns transferable feature representations from a source structure to target structures with varying structural characteristics. Experimental validation on steel frame specimens demonstrates that the framework can achieve effective damage localization under different structural configurations. The results indicate that the proposed method provides a practical and transferable solution for structural damage identification across structures.

Population-Based Structural Health Monitoring (PBSHM) offers a new approach to asset management by leveraging data from entire populations of structures. The advancement of PBSHM methods, however, is constrained by the scarcity of comprehensive, multi-structure benchmark datasets from real-world assets. Synthetic datasets, such as the Population-based SHM Engineered Asset Resource (PEAR), have been developed to fill this gap by providing realistic structural models and data. While these datasets capture structural variability, they exist in isolation, lacking the real-world geospatial and operational context, such as location, environmental exposure, and network importance. These factors are what govern actual asset management decisions. This paper addresses the research question: How can a synthetic structural population from the PEAR dataset be grounded in a real-world environment using geospatial data? The paper presents a novel methodology that combines the synthetic PEAR bridge population with real-world geospatial data, successfully assigning structures to specific locations by matching structural parameters to specific geospatial constraints. The resulting dataset enables the validation of PBSHM methods at a network level for a spatially defined population and facilitates new research into location-dependent tasks. This work provides a robust template for enhancing synthetic data with real-world context, increasing the applicability for practical decision-making in infrastructure management.

Structural health monitoring of bridge infrastructures often faces the challenge of data scarcity, where reliable or labeled data is available for one portion of a structure but not for others. Transfer learning offers a solution by leveraging data from a source domain to improve diagnostic performance in a target domain. This study investigates a transfer learning approach for damage detection between two different spans of the same bridge: the riverbed and the floodplain span. To ensure physical consistency between the domains, four modal frequencies are selected for each span based on high modal assurance criterion values. Normal condition alignment is employed to minimize the distribution discrepancy between the source and target datasets under healthy conditions. To evaluate the effectiveness of the alignment, an artificial damage scenario is simulated in the target span by introducing a 0.1 Hz frequency shift in the final monitoring period. A Gaussian mixture model classifier is trained exclusively on the source data and then tested on the target data. Comparative analysis is performed by evaluating classification accuracy before and after statistic alignment. Normal condition alignment significantly enhances the detectability of damage in the target span by compensating for operational and structural variations between the two spans.