Organizations: German Aerospace Center (DLR), Institute of Maintenance, Repair and Overhaul, Aachen-Würselen, Germany · German Aerospace Center (DLR), Institute of Maintenance, Repair and Overhaul, Hamburg, Germany · RWTH Aachen University, Chair for Computational Analysis of Technical Systems, Aachen, Germany
Ultrasonic guided wave (UGW) inspection has become an effective method in structural health monitoring (SHM). While machine learning has enhanced damage detection capabilities, purely data-driven models remain constrained by extensive labeled datasets and limited physical interpretability. Standard multi-layer perceptrons (MLPs) further suffer from spectral bias (F-Principle) in high-frequency regimes, prioritizing low-frequency components while failing to resolve the high-frequency oscillations critical for UGW reconstruction. Physics-Informed Sinusoidal Representation Networks (PI-SIRENs) address these limitations by embedding governing partial differential equations directly into the training process. Building upon SIRENs, which utilize periodic activation functions to inherently capture high-frequency spectral content, the proposed framework is designed to operate effectively in sparse ultrasonic measurement regimes. This work introduces a frequency-domain PI-SIREN framework tailored for the Mindlin-Reissner plate theory to reconstruct ultrasonic wavefields without temporal discretization. The model parameters are optimized using a hybrid strategy comprising three phases: first-order Adam, residual-adaptive refinement (RAR) and quasi-Newton L-BFGS. Numerical and experimental evaluations confirm that the model serves as a physics-consistent baseline derived from sparse data, functioning as a computationally efficient surrogate for damage identification in SHM and non-destructive testing (NDT). Integrating the surrogate into a damage identification pipeline detects and localizes a milled notch from sparse experimental measurements with a localization error of 1 % of the plate diagonal. In contrast, data-driven benchmarks and MLPs with hyperbolic tangent activation functions, with and without Fourier feature embedding, are unable to identify damage due to higher reconstruction errors.
Figures & tables
Figure 1: Schematic representation of the frequency-domain PI-SIREN architecture for solving the normalized MR plate theory equations.
Figure 2: Schematic representation of the experimental robotic measurement setup.
Figure 3: Sequential damage identification pipeline on experimental data: The pristine baseline model is trained once offline and reused as pre-trained model in every online inspection for inference. Online phases I–IV: coarse field acquisition, baseline subtraction and candidate detection, adaptive resampling and refined damage localization.
Tag
Configuration
L2(wˉ^)
LPDE
M1
Data-only ( λPDE=λBC=0 )
0.4932
–
M2
PINN, Adam-only (70,000 iter.)
0.2881
2.7×10−4
M3
PINN, Adam + L-BFGS
0.1351
5.2×10−6
M4
PINN, Adam + RAR + L-BFGS
0.1037
5.6×10−6
Table 1: Reconstruction ablation study on the 25×25 grid. Relative test error L2(wˉ^) on a fixed dense test set and the final PDE residual LPDE .
Hyperparameter
Value
Hidden layers / neurons
5×256
Base frequency ω0 (all layers)
30
Data supervision
transverse wˉ^ only
Loss weights (λPDE,λBC,λdata)
(1.0,10.0,10.0)
Adam learning rate η
2×10−5
Adam iterations (FP32)
20,000
Table 2: Training configuration held constant across all numerical and experimental tests unless stated otherwise.
Figure 4: Test L2 versus training iteration for all four ablation variants on the 252 grid.
Figure 5: Pristine wavefield reconstruction on the 252 grid: FEM reference (left), M1 with L2=0.49 (center) and M4 PI-SIREN prediction with L2=0.10 (right). Top row: normalized transverse displacement Re(wˉ^)/Ws , bottom row: pointwise error ∣Δwˉ^∣/Ws .
Grid
ppw
L2(wˉ^)
15×15
1.16
0.262±0.0208
18×18
1.41
0.198±0.0834
20×20
1.58
0.265±0.0062
25×25
1.99
0.103±0.0032
30×30
2.41
0.021±0.0011
35×35
2.83
0.011±0.0001
Table 3: Sensor density study: ppw and pristine-field reconstruction error L2(wˉ^) (mean ± std over three seeds) for each uniform training grid.
Figure 6: Reconstruction L2(wˉ^) versus sensor density in ppw (mean ± std).
Figure 7: Median localization error versus sensor density in ppw for four simulated damage positions with 2mm acceptance threshold. Above the Nyquist limit (dashed line, 2 ppw) all positions converge to <1mm .
Figure 8: Localization error versus reconstruction L2(wˉ^) with damage at (130,50)mm . Successful localizations occur predominantly for L2≲0.10 .
Figure 9: Relative band error εB against training iteration: (a) A0 band, 0.85kA0≤∣k∣≤1.15kA0 , and (b) short-wavelength band above the A0 band, 1.15kA0<∣k∣≤2.0kA0 .
Figure 10: NTK eigenvalue spectrum λi/λ1 against the dominant wavenumber ki of each eigenvector at initialization and after training. The dashed line marks kA0=2π/λA0≈0.38rad/mm .
Figure 11: Experimental pristine reconstruction: Experimental measurement (left), M4 PI-SIREN prediction (center), pointwise error ∣Δwˉ^∣/Ws (right). The dashed box indicates the interior region ( >15mm ) used for the interior L2 calculation.
Figure 12: Demonstration of damage identification pipeline on experimental data. Left: pass-1 candidate regions ( ∘ , #1–10) and adaptive resampled distribution. The true damage is among the identified regions. Right: Damage indicator D recomputed on the adaptively refined sampling, with the true-defect region showing the strongest response ( 2.63 mm localization error).
Quantity
FEM
EXP
M4 trainable parameters
265,478
Training (one-time)
89.6 min
96.1 min
Adam / L-BFGS [min]
22.6/66.9
18.7/77.5
Inference ( 40,000 pts)
0.77 ms
0.76 ms
Table 6: Computational cost overview. Values for M4 trained on synthetic (FEM) and experimental (EXP) data.
Guided wave-based structural health monitoring (GWSHM) with onboard transducers offers significant potential for the early diagnosis of damage in engineering structures. However, the practical deployment of deep learning models is often hindered by the limited availability of labelled experimental data and the high computational cost of generating large-scale high-fidelity simulation datasets. This study presents a multifidelity transfer learning framework that integrates lightweight physics-based simulations, convolutional autoencoder (CAE)-based deep feature learning, a feed-forward neural network, and limited experimental measurements for accurate damage localisation and sizing in plate-like structures instrumented with piezoelectric transducers. A computationally efficient one-dimensional time-domain spectral element model is employed to generate a large synthetic dataset for pretraining, while transfer learning adapts the model to experimental domains using only a small amount of labelled data. The CAE-based transfer learning framework significantly outperforms its CNN-based counterpart in damage localisation accuracy. The model achieves excellent predictive performance with R2 scores exceeding 0.93 for damage localisation and 0.99 for damage sizing. Its generalisation capability is demonstrated on previously unseen data, showing high prediction accuracy for damage scenarios not represented during pretraining or fine-tuning. The results establish the proposed framework as an accurate, computationally efficient, and practically viable solution for real-world GWSHM applications.
Santosh Kapuria, Abhishek
Department of Applied Mechanics, Indian Institute of Technology Delhi, New Delhi, 110016, Delhi, India
Guided-wave structural health monitoring enables damage localization in composite plates using sparse networks of bonded piezoelectric transducers. However, supervised localization remains weakly constrained when measurements are available from only a limited set of damage locations. Because exhaustive spatial coverage is impractical, models trained at observed locations may generalize poorly to unseen regions. We propose WaveGraphNet, an inverse-forward graph-learning framework for guided-wave damage localization in carbon-fiber-reinforced polymer (CFRP) plates. The sensing network is represented as a graph, with transducers as nodes and measured pitch-catch paths as edges. The inverse branch uses order-invariant message passing and a geometry-constrained decoder to map path-wise energy deviations to a damage coordinate. An independently trained forward branch predicts the energy-deviation pattern associated with a candidate coordinate. At test time, gradients propagated through the forward branch update only the inverse-predicted coordinate to reduce the discrepancy between the measured and predicted response patterns. The framework is evaluated on the OGW-1 benchmark using three spatial hold-out splits in which complete damage regions are excluded from training. Comparisons include non-graph and graph-learning baselines together with the training-free Reconstruction Algorithm for Probabilistic Inspection of Damage (RAPID). Under the validation-selected settings, refined WaveGraphNet achieves the lowest reported test mean absolute error on all three splits. Refinement reduces the standalone localization error by 23.6 percent-67.9 percent, with an average relative reduction of 50.4 percent. These results demonstrate that learned forward-response consistency can improve guided-wave localization under limited spatial training coverage.
Vinay Sharma, Aditya Bharade, Olga Fink
EPFL, Intelligent Maintenance and Operations Systems, Lausanne, Switzerland
Neural surrogates provide computationally efficient auto-regressive modeling of wave propagation, enabling e.g. fast inverse solutions. Despite recent progress, the current surrogates are not yet sufficient for the needs of ultrasonic sensing, characterized by high frequency and heterogeneous propagation media. We show how methods building on Fourier Neural Operators (FNO) fail to capture essential signal features, such as reflections, whereas U-Net -based methods learn the wave fields well but lack the ability to transfer information across different spatial scales. We then propose a new surrogate that captures the best of the two worlds, by integrating FNOs and local-integral resampling into a U-Net architecture. We show it can be trained on limited data, and that the predictions are robust to input noise and work across spatial scales.
Burla Nur Korkmaz, Petro Moilanen, Ari Salmi +2
Department of Computer Science, University of Helsinki, Finland