Interpretable Dual-Network Feature-Selection Framework for Guided Wave Damage Diagnosis on Composite Panels
Rafael Junges, Abderrahim Abbassi, L. Lomazzi, Marco Giglio, F. Cadini, R. Rolfes
Ultrasonic guided waves are widely used for monitoring composite panels due to their high sensitivity to structural changes. While numerous damage indices (DIs) have been proposed to detect and localize damage, they are typically applied either independently, ignoring the potential synergy between specific indices and excitation frequencies, or collectively, which can lead to overfitting, reduced generalizability, and unnecessary costs. Consequently, it remains unclear which features (DI and frequency combinations) are most informative for specific diagnostic tasks. This work introduces an automated and interpretable framework that explores the synergy between different features. To achieve this, we employed a Dual-Network Feature Selection (DNFS), a method that trains two neural networks jointly: a selection network (SN) that produces a quasi-binary relevance mask over the features, and a task network (TN) that performs detection or localization using only the selected features. A composite loss function combines the task-specific objective with a sparsity regularization term, allowing the SN to suppress uninformative or redundant inputs while the TN maintains high performance. Because the relevance mask is global, it can be directly interpreted as a ranking of the physical and performance significance of each feature. Relevance scores expressed as SHAP values were then applied to the selected features after damage diagnosis, providing a cross-check of physical consistency and a visual cue of feature importance. The workflow was experimentally benchmarked on the Open Guided Waves public repository, involving a composite panel with various pseudo-damage configurations. For damage detection, the DNFS achieved accuracies above 95% with as few as 3 input features, which is approximately 0.05% of the total input features available (roughly 250 features were sufficient to guarantee perfect accuracy with 95% statistical confidence). As for localization, it successfully localized damage to within 4% of the plate’s surface area with greater than 90% accuracy, using as few as 80 features (approximately 1.26% of the total input features available). The mean accuracy when using all features was 96.5%. Moreover, by revealing which indices matter and why, the study contributes to the development of interpretable and computationally efficient SHM pipelines.