Graph neural networks for SHM: exploiting spatial interdependencies of strain data for diagnostics and prognostics

Abstract

Structural health monitoring using strain data faces a critical challenge: decoupling subtle structural degradation signatures from the dominant influence of operational loads. This paper introduces a novel methodology to address this by synergistically combining a custom health indicator (HI) with graph neural networks (GNNs). The proposed HI, derived from the cumulative absolute first derivative of strain over time, effectively isolates load-independent features indicative of damage progression. These features serve as input to our proposed GENConv with Edge Attributes (GENEA) model, a GNN that explicitly models the spatially distributed sensors as an interconnected network, leveraging spatial interdependencies and edge attribute information within the strain field to enhance damage assessment. This integrated approach enables accurate structural stiffness reduction estimation (diagnostics) and remaining useful life (RUL) prediction (prognostics). Applied to strain data from fatigue tests on representative aeronautical composite panels, the methodology is rigorously evaluated using Leave-One-Panel-Out cross-validation. The framework shows promising performance on unseen test data, although challenges in generalizing to out-of-distribution specimens were also identified, highlighting the importance of a diverse training set for real-world applicability. Experimental results confirm the framework’s superiority. The proposed GENEA model significantly outperforms both a fundamental multi-layer perceptron and a spatially aware convolutional neural network baseline, and successfully generalizes to an unseen panel with a different sensor count. This validates the benefits of using a tailored GNN framework to learn robust, geometrically invariant patterns from load-decoupled spatial strain data.