Physics-Guided CNN Detection of Crack-Associated Events from Embedded Fiber Bragg Grating Sensors

Crack detection in composite structures remains a central challenge in structural health monitoring, particularly when sensing must rely on a small number of embedded multiplexed fiber Bragg gratings (FBGs). Here, we present a physics-guided convolutional neural network (CNN) framework for crack-associated event detection from multiplexed FBG interrogator signals acquired during the three-point bending of glass-fiber-reinforced polymer (GFRP) beams. The dataset was constructed from raw interrogator recordings and synchronized force-displacement metadata while preserving the cracked and non-cracked loading stages present in the experiments. Each candidate response was encoded by 13 synchronized optical, loading, and mechanics-guided descriptors, including Euler-Bernoulli expected strain and residual terms, where the residual denotes the difference between the measured response and the elastic response predicted by beam theory. A compact one-dimensional CNN operating on 30-response sequences was evaluated on 64 experimental runs under strict leave-one-run-out validation. At the selected operating point, the model reached window-level precision of 0.900, recall of 0.910, F1 score of 0.905, and balanced accuracy of 0.942, while the corresponding run-level decision reached a precision of 0.833, a recall of 1.000, an F1 score of 0.909, and a balanced accuracy of 0.969. Bootstrap resampling over runs yielded 95% confidence intervals of 0.787-0.978 for window-level F1 and 0.769-1.000 for run-level F1. To probe generalization beyond the initial fabrication batch, the final frozen pipeline was also tested once on seven later-batch runs from two newly manufactured specimens, where it reached a window-level precision of 0.908, a recall of 1.000, an F1 score of 0.952, a balanced accuracy of 0.969, an ROC-AUC of 0.979, a PR-AUC of 0.955, and perfect run-level classification. These results show that a compact sequence CNN, enriched with mechanics-guided strain interpretation, can extract robust crack-event signatures from multiplexed FBG measurements while preserving a simple and reproducible modeling pipeline.