A physics-informed neural network (PINN) is developed for modeling time-harmonic Lamb-wave excitation in a two-dimensional elastic waveguide under surface loading. The displacement and stress fields are represented by the network, and its trainable weights are determined by enforcing the first-order elastodynamic system, the traction boundary conditions, and the Lamb-mode-based Dirichlet-to-Neumann (DtN) conditions in the loss function. This formulation avoids derivative boundary constraints and enables direct extraction of modal amplitudes and far-field responses. The method is validated against an analytical multimodal solution for uniform shear-stress excitation. For representative single-frequency cases, the PINN reproduces the near-field displacement patterns and projected modal coefficients with good accuracy. A frequency-parameterized model is further trained to predict broadband responses and captures the main trends of the propagating modes with moderate errors over the considered band. The proposed model provides an extensible framework for broadband guided-wave forward modeling, with potential applications in transducer design optimization.

Baseline subtraction is widely used to isolate damage-related information in guided wave (GW)-based structural health monitoring (SHM). In deployed sensor networks, however, reliable baselines are often unavailable; even when obtained from nominally intact states, environmental and operational variations can introduce significant uncertainty and misleading residuals. To address this issue, this work proposes a fully baseline-free damage localization framework based on compressive sensing, tailored to sparse measurements from embedded, resource-limited sensor networks. Damage-independent components, including direct and boundary-reflected waves, are explicitly modeled to eliminate the need for a healthy reference signal. In parallel, damage-related atoms are constructed by discretizing candidate defect locations and generating their corresponding scattering responses. Considering that GWs interacting with discontinuities undergo mode conversion, each candidate location is represented as a block containing all relevant mode-conversion outcomes. A block-structured Sparse Bayesian Learning (SBL) scheme is introduced to enforce physical consistency among scattering modes from the same location while exploiting cross-channel information within the sensor network. All subcomponents in a block share a common hyperparameter and are activated jointly, preventing spurious substitution by scattering components from other locations and promoting network-consistent sparse support across sensor pairs. Numerical studies on plate-like structures show that the proposed method achieves stable and uniquely identifiable damage localization using sparse distributed sensors, without requiring prior intact-state measurements. The results highlight that the proposed physics-informed block-sparse Bayesian framework provides a robust and interpretable path toward practical baseline-free guided-wave SHM in sparse sensor networks.

Ultrasonic Guided Waves (UGWs) are among the most effective tools for damage diagnosis and Structural Health Monitoring (SHM) of thin-walled structures. However, traditional SHM methods based on UGWs typically require baseline measurements and signal post-processing to extract damage indices, which can lead to the loss of important information contained in the raw data.
To face these challenges, recent studies have investigated machine learning approaches. However, most of these methods depend on large labeled datasets and use black-box models with limited interpretability.
In this work, a Physics-Informed Neural Network (PINN) framework is proposed to overcome these limitations. The PINN solves an inverse problem by simultaneously reconstructing the full ultrasonic wavefield and the spatial distribution of the wave velocity, expressed in terms of Young’s modulus, recognizing potential discontinuities caused by damage. This is achieved by minimizing a combined loss function that enforces both the agreement with sparse measurement data and the physical laws governing the propagation of UGWs.
The proposed method is numerically validated on a case study of an isotropic structure affected by damage.
The results demonstrate that the approach can accurately localize damage without the need for data post-processing. Moreover, the method is fully unsupervised and baseline-free, relying solely on sparse measurements and the underlying physical laws of UGW propagation.

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.