Answers are grounded in this paper and cite their sources.
• Unsupervised FBG-based bondline damage detection in CFRP–steel joints. • Spatial Gradient Strain improves defect sensitivity vs absolute strain fields. • VAE and GANomaly achieve ∼ 0.99 AUC using healthy-state training only. • Feature representation more critical than deep model complexity for SHM. • C-GANomaly enables near real-time, high-accuracy structural monitoring. Bondline integrity between carbon fibre-reinforced polymer (CFRP) laminates and steel substrates is critical to the performance of composite-strengthened infrastructure, yet early interfacial defects such as kissing bonds remain difficult to detect because they produce only subtle strain perturbations under service-level loading. This study presents an unsupervised, data-driven framework for bondline damage detection using two parallel embedded Fibre Bragg Grating (FBG) arrays with eight sensing locations, requiring only healthy-state data for model development. Two CNN-based deep generative models, namely a Variational Autoencoder (VAE) and GANomaly, were trained using healthy measurements from an instrumented CFRP-strengthened steel coupon under multiple service-level loading conditions and evaluated through four-point bending tests on healthy and defective specimens. Manufacturing consistency was independently verified using three additional healthy specimens. A novel Spatial Gradient Strain (SGS) representation is proposed and benchmarked against the conventional Absolute Strain Field (ASF). By emphasising local strain discontinuities while suppressing global load-scaling effects, SGS provides a mechanics-informed, load-invariant representation of bondline behaviour. Both deep generative models achieved AUC-ROC values of approximately 0.99 using SGS, compared with only 0.70–0.81 using ASF, demonstrating that feature representation has a greater influence on detection performance than network architecture. The C-GANomaly (SGS) model achieved the best performance, with an F1-score and recall of 0.99 while maintaining near-real-time inference capability. Benchmarking against conventional one-class learning and multivariate statistical process control methods confirmed the superiority of the proposed framework, with PCA representing the strongest conventional baseline (F1-score = 0.89). Deep ensemble uncertainty analysis further showed well-calibrated confidence estimates (ECE = 0.011), supporting confidence-aware bondline monitoring. The proposed framework demonstrates proof-of-concept feasibility for asset-specific, physically interpretable, uncertainty-aware bondline monitoring under controlled laboratory conditions.
No takes yet. Share an insight, caveat, or question.
Mansourdehghan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: