Abstract:

A heterogeneous Mesh Graph Network (MGN)-based framework is presented, incorporating sparse Optical Fiber Sensors (OFS) to identify the existence, localization, severity, and extension of structural damage in flat composite laminates. The composite panel is discretised into regular Cartesian grids, with mesh nodes connected through 4-connected neighbors to represent the spatial continuity

of the composite structure. Sensor nodes are further integrated into the graph by connecting to their nearest mesh nodes and sensor nodes based on a K-nearest neighbor (KNN) strategy. The relative signal profile (εr), calcualted from the strain responses of damaged structure with respect to those of the healthy one (baseline), is generated through a finite element (FE) model performed on a flat composite plate and used as input for training the proposed MGN-based model. This model is trained independently by two separate datasets including either single large damage region or single small damage region. The proposed model demonstrated reliable damage prediction for large-damage scenarios, including cases with multiple damage regions despite being trained exclusively on single-damage configurations. Cross-material validation further demonstrated its ability to generalize to other composite systems. However, the performance of the model retrained using dataset with one small damage area descends substantially when employed to make predictions on two small damage scenarios, although it can provide reasonable results for previously unseen single-small damage cases