Trustworthy AI for Structural Health Monitoring: Governance, Reliability, and Human Oversight in Sensor-Based Maintenance Systems
Main article
Abstract
The integration of artificial intelligence (AI) into structural health monitoring (SHM) systems has accelerated considerably in recent years, offering unprecedented capabilities in real-time damage detection, predictive maintenance, and anomaly identification from large-scale sensor networks deployed across civil and industrial infrastructure. However, the deployment of AI-driven SHM in safety-critical applications raises fundamental questions about trustworthiness: whether these systems can be relied upon to produce accurate, interpretable, and governed outputs under real-world operational conditions. This paper presents a comprehensive framework for trustworthy AI in sensor-based structural health monitoring, addressing four interlocking dimensions: model reliability and uncertainty quantification, explainability and interpretability of AI predictions, governance mechanisms for system accountability, and human oversight integration in the maintenance decision pipeline. Drawing on a multi-sensor testbed comprising 247 sensing nodes deployed across a prestressed concrete bridge and an industrial steel frame structure, we evaluate six deep learning architectures—LSTM, 1D-CNN, ResNet, XGBoost, autoencoder, and a proposed hybrid model—against standardized benchmarks for detection accuracy, false alarm rate, and anomaly localization precision. Our proposed hybrid model achieves a classification accuracy of 96.7%, an F1-score of 0.961, and an average detection delay of 1.4 seconds, outperforming baseline methods by 2.6 to 8.4 percentage points in accuracy. Explainability analysis via SHAP values reveals that sensor channels associated with mid-span deflection and modal frequency shift contribute most significantly to model decisions. A three-tier governance architecture encompassing data audit trails, model versioning, and human-in-the-loop validation protocols is proposed and evaluated through structured expert review. Our findings demonstrate that trustworthy AI frameworks can substantially improve both the technical performance and institutional accountability of AI-based SHM systems, making them viable for deployment in high-stakes infrastructure management contexts.
