Uncertainty-aware Graph Learning Approach for Sparse-Sensing Structural Health Monitoring

Projektbeschreibung

Zuverlässige KI für die Zustandsüberwachung von Bauwerken

Aktuelle Systeme zur Zustandsüberwachung von Bauwerken stützen sich auf begrenzte Sensordaten und Black-Box-Modelle, was es schwierig macht, verborgene Schäden in großen Infrastrukturen zuverlässig zu erkennen. Mit Unterstützung der Marie-Skłodowska-Curie-Maßnahmen wird das Projekt U-GLASS ein physikalisch fundiertes, unsicherheitsbewusstes Rahmenwerk für das Graph-Lernen entwickeln, um die Zustandsüberwachung von Bauwerken anhand spärlicher Sensordaten zu verbessern. Durch die Verbindung von physikalischen Modellen reduzierter Ordnung mit graphischen neuronalen Netzen werden vollständige Strukturreaktionen rekonstruiert und Schäden in komplexen Infrastruktursystemen genauer identifiziert. Der Ansatz wird anhand von Simulationen, Laborstrukturen sowie realen Anwendungsfällen wie Eisenbahnbrücken getestet, mit dem Ziel, die Sicherheit und das langfristige Management kritischer Infrastrukturen zu verbessern.

Ziel

Structural health monitoring (SHM) emerged from the critical need to continuously assess the integrity and performance of aging infrastructure to ensure public safety. Current sparse-sensing schemes leave vast portions of structures unmonitored and vulnerable to undetected damage, which necessitates full-field response reconstruction to gain systematic structural insight. Classic machine learning (ML) approaches struggle with the ‘black-box’ issue and topological blindness in complex connectivity patterns, which limit their prediction accuracy and model applicability. The proposed project, ‘Uncertainty-aware Graph Learning Approach for Sparse-Sensing Structural Health Monitoring’ (U-GLASS), aims to develop a reliable physics-informed graph learning framework while accounting for uncertainty that addresses fundamental sparse-sensing limitations plaguing current SHM. To implement the project, four sub-objectives guided by the framework include: 1) Develop a reduced-order model (ROM) with parametric uncertainties and reduction errors quantification; 2) Construct a GNN embedded with ROM-derived physics and topological knowledge; 3) Develop a training protocol for GNN to reconstruct full-field responses from sparse measurements; and 4) Quantify structural damage severity/location using reconstructed responses on multi-scale systems. The framework will be first implemented on multi-scale simulation/laboratory cases and then applied to real-world systems such as Swiss railway bridges. By harmonizing uncertainty-aware ROM/topological physics with ML, the goal is to enable physics-enhanced full-field response reconstruction from sparse sensor data, transforming limited sensing data into actionable engineering insights. The magnitude of U-GLASS’s contribution is ultimately measured not just in technical innovation, but in its potential to fundamentally reshape how society manages critical infrastructure systems for future generations.