Physics-Guided Self-Supervised Few-Shot Learning for Ultrasonic Defect Detection in Concrete Structures

Abstract

This study introduces a physics-guided self-supervised framework for few-shot ultrasonic defect detection in concrete structures, addressing the dual challenges of scarce labels and domain variability in structural health monitoring (SHM). Our method integrates physics-informed augmentations, contrastive representation learning, and adversarial domain alignment within a mutually reinforcing cycle, enabling robust defect classification with minimal supervision. A Physics-Informed Augmentation Module synthesizes realistic ultrasonic signals, training a Transformer encoder to extract invariant features while suppressing sensor noise. An Adversarial Feature Aligner further improves cross-domain generalization by mitigating distribution shifts across heterogeneous concretes. Experimental validation on three benchmark datasets demonstrates 63–66% accuracy in one-shot cross-domain tasks and up to 89% in five-shot settings. These results represent 12–15 percentage point gains over modern few-shot baselines, with improvements statistically significant at p < 0.001. Compatible with existing ultrasonic hardware, the proposed framework bridges physics-based modeling and machine learning while paving the way for scalable, field-ready SHM solutions for aging infrastructure and resilient smart cities.

1. Introduction

Non-destructive evaluation of concrete structures has become increasingly critical for maintaining aging infrastructure worldwide. Among various inspection techniques, ultrasonic testing offers unique advantages for detecting internal defects through wave propagation analysis [1]. Traditional approaches rely on handcrafted features such as time-of-flight measurements or frequency domain characteristics [2,3]. While these methods provide interpretable results, they often struggle with complex defect patterns and material heterogeneity commonly encountered in real-world concrete structures.

Recent advances in machine learning have demonstrated promising results for ultrasonic defect classification. Supervised learning approaches, including support vector machines and deep neural networks, have shown improved accuracy over conventional methods [4,5]. However, these data-hungry algorithms require extensive labeled datasets that are costly to obtain in practical inspection scenarios. The challenge becomes particularly acute when dealing with rare defect types or novel structural configurations where training samples are scarce.

Few-shot learning has emerged as a potential solution to this data scarcity problem. Meta-learning techniques such as MAML and Reptile enable models to adapt quickly to new tasks with minimal examples [6,7]. In the context of ultrasonic testing, recent work has explored Siamese networks for defect classification in insulation materials [8]. While these approaches show promise, they often fail to account for the physical principles governing ultrasonic wave propagation in concrete, potentially limiting their generalization across different structural conditions.

The fundamental challenge lies in simultaneously addressing three key aspects: data scarcity through few-shot learning, domain variability across different concrete structures, and physical consistency of defect signatures. Existing methods typically focus on one aspect while neglecting others. For instance, conventional few-shot learning approaches may ignore the wave physics underlying defect formation [9], while physics-based models often lack the flexibility to handle diverse real-world conditions [10].

We propose a novel framework that integrates physics-guided data augmentation with self-supervised representation learning and adversarial domain adaptation for few-shot ultrasonic defect detection. Our approach differs from previous work in three significant ways. First, we develop physics-informed augmentations that generate realistic ultrasonic signals by simulating wave interactions with various defect geometries and material properties. Second, we employ contrastive learning to extract invariant features from both real and synthetic waveforms, preserving defect-related patterns while suppressing irrelevant variations. Third, we introduce an adversarial domain adaptation component that aligns feature distributions across different concrete structures, enabling robust generalization to unseen conditions.

The proposed method offers several advantages over existing techniques. By incorporating physical principles into the learning process, it achieves better sample efficiency than purely data-driven approaches. The self-supervised component reduces reliance on labeled data while maintaining discriminative power for defect classification. Furthermore, the domain adaptation mechanism allows the model to handle structural variations that commonly challenge conventional ultrasonic testing methods.

This paper makes three main contributions: (1) a physics-guided augmentation strategy that generates diverse and physically consistent ultrasonic signals for training; (2) a self-supervised learning framework that extracts robust defect features from limited labeled data; (3) an adversarial domain adaptation approach that enables cross-structure generalization in few-shot defect detection scenarios.

In summary, this paper introduces a novel framework that unifies physics-guided data augmentation, self-supervised representation learning, and adversarial domain adaptation into a single, mutually reinforcing cycle for ultrasonic defect detection in concrete structures. Unlike prior studies that address these challenges in isolation, our approach leverages their synergy: physics-informed augmentations generate realistic signals that enhance the self-supervised learner; the robust features learned strengthen the effectiveness of domain adaptation; and domain adaptation, in turn, ensures that the learned features generalize back to real-world concrete physics. This integrated design enables accurate few-shot defect classification under severe data scarcity while maintaining physical plausibility and cross-domain robustness.

The remainder of this paper is organized as follows: Section 2 details the materials and methods, including the physical background, learning foundations, and our proposed framework. Section 3 presents the comprehensive experimental evaluation and results. Section 4 discusses the findings, limitations, and future research directions. Finally, Section 5 provides the conclusions of this work.

Related Work and Research Gap

The development of ultrasonic defect detection methods for concrete structures has evolved through several paradigms, ranging from traditional signal processing to modern machine learning approaches. Existing works can be categorized into three major research streams: physics-based modeling of ultrasonic wave propagation, self-supervised representation learning for signal processing, and few-shot learning strategies for defect classification.

Physics-Based Modeling for Ultrasonic Defect Simulation. Physics-based approaches have long formed the foundation of ultrasonic testing methodologies. Early research emphasized analytical solutions to wave propagation equations in homogeneous media [11], which offered theoretical insights but showed limited applicability to heterogeneous concrete. The introduction of finite element analysis enabled more realistic simulations of wave–defect interactions [12], though computational requirements remained prohibitively high for generating large-scale datasets. More recently, advances in physics-informed neural networks have bridged this gap by combining numerical simulations with machine learning [13]. In the context of ultrasonic guided waves, Hilbert convolutional networks have successfully transferred simulated defect patterns to experimental data [14]. Nevertheless, such methods generally demand extensive parameter tuning and often struggle to generalize across diverse concrete compositions.

Self-Supervised Representation Learning for Ultrasonic Signals. The challenge of limited labeled data in ultrasonic testing has motivated the application of self-supervised learning. Contrastive learning frameworks have been proven effective in extracting invariant features from vibration signals in machinery fault diagnosis [15]. These methods maximize mutual information between differently augmented views of the same signal while minimizing similarity with unrelated samples. In bioacoustic domains, domain-invariant representations have been achieved through transformations between time and frequency signals [16]. For ultrasonic testing specifically, pretext-invariant learning strategies have been introduced to separate signal artifacts from genuine defect patterns [17]. While promising, these methods often overlook the physical constraints governing ultrasonic wave propagation in concrete, resulting in representations that may contradict fundamental wave mechanics principles.

Few-Shot Learning for Defect Classification. Few-shot learning has emerged as a viable strategy for defect detection in scenarios with scarce labeled examples. Meta-learning approaches such as MAML have been adapted to structural health monitoring tasks [6], but their performance often declines when test conditions differ significantly from the training domain. Cross-modal zero-shot learning has also shown potential for weld defect identification by mapping ultrasonic features to semantic attributes [18]. Within concrete inspection, Siamese networks coupled with finite element simulations have demonstrated promising results for few-shot defect classification in insulation materials [19]. However, these methods generally treat ultrasonic signals as generic time series, disregarding domain-specific physics such as wave scattering and attenuation in concrete.

Despite these significant advances, a number of methodological gaps remain in the existing literature.

Most studies focused on improving classification accuracy using data-driven or deep learning models but rarely examined the underlying wave physics or material heterogeneity of concrete, resulting in models with limited interpretability and poor generalization across domains [6,7,8,9,10].

Conversely, physics-based simulations have provided valuable insight into wave–defect interactions, yet they are computationally expensive and difficult to integrate with modern learning frameworks [11,12,13].

Moreover, conflicting results have been reported on the stability of domain adaptation techniques when applied to heterogeneous concrete datasets [18,19,20].

Therefore, a clear research need exists for an integrated framework that couples physics-informed modeling with self-supervised and domain-adaptive learning to achieve physically consistent and transferable defect detection in data-scarce environments.

Building upon these observations, the proposed framework advances beyond existing approaches by addressing three key limitations simultaneously. First, it integrates physics-based constraints directly into both data augmentation and representation learning, ensuring physically plausible defect signatures. Second, it combines self-supervised pretraining with adversarial domain adaptation to manage distribution shifts across concrete structures. Third, it employs a prototypical network architecture optimized for few-shot ultrasonic defect classification. Unlike prior studies that tackle these challenges separately, our unified framework achieves superior performance in data-scarce scenarios while maintaining physical consistency and robust cross-domain generalization.

2. Materials and Methods

This section presents the theoretical background, methodological framework, and implementation details of the proposed physics-guided self-supervised few-shot learning approach for ultrasonic defect detection in concrete structures. First, the fundamental principles of ultrasonic wave propagation and defect-induced modifications in heterogeneous concrete are reviewed. Then, the foundations of few-shot learning, self-supervised representation learning, and domain adaptation are summarized. Building on this foundation, we describe the proposed framework, which integrates physics-informed data augmentation, contrastive representation learning, and adversarial domain alignment. Finally, the datasets, experimental protocols, and implementation details used to evaluate the system are provided.

2.1. Physical Background

2.1.1. Ultrasonic Wave Propagation in Heterogeneous Concrete

When ultrasonic waves propagate through concrete, their behavior is governed by the material’s complex microstructure. The wave equation for an elastic medium provides the fundamental description:

$$\rho {\displaystyle \frac{{\partial }^{2}u}{\partial t^2}}=\nabla \cdot C:\nabla u$$

(1)

where $\rho$ represents density, $u$ the displacement field, and $\mathrm{C}$ the fourth-order stiffness tensor. Concrete’s heterogeneity introduces several physical phenomena that complicate defect detection. First, wave scattering occurs at aggregate interfaces, creating complex interference patterns that obscure defect signatures [20]. Second, frequency-dependent attenuation follows a power-law relationship:

$$\alpha \left(f\right)={\alpha }_0{f}^{\eta }$$

(2)

with ${\alpha }_0$ as the attenuation coefficient and $\eta$ typically ranging from 1 to 2 for concrete [21]. These effects combine to produce distinctive time-frequency characteristics that vary with concrete composition and defect properties.

2.1.2. Defect-Induced Wave Modifications

Internal defects alter ultrasonic wave propagation through several mechanisms. Voids and cracks cause wave reflections described by the reflection coefficient:

$$R={\displaystyle \frac{Z_2-Z_1}{Z_2+Z_1}}$$

(3)

where $Z_1$ and $Z_2$ represent acoustic impedances of concrete and the defect, respectively [22]. Delaminations create guided wave modes that exhibit dispersion characteristics different from bulk waves. The phase velocity $c_p$ of these modes relates to frequency $f$ through the dispersion relation:

$$c_p(f)={\displaystyle \frac{2\pi f}{k\left(f\right)}}$$

(4)

with $k\left(f\right)$ as the wavenumber [23]. These physical principles form the basis for our physics-guided augmentation strategy.

2.2. Few-Shot and Self-Supervised Learning Foundations

2.2.1. Few-Shot Learning

Few-shot learning addresses classification problems where only a small number of labeled examples are available per class. The N-way K-shot formulation defines a task with N classes and K examples per class [24]. Prototypical networks compute class prototypes as mean embeddings:

$${\mathrm{c}}_k={\displaystyle \frac{1}{\left|S_k\right|}}\sum _{({\mathrm{x}}_i,y_i)\in S_k}{f}_{\theta }({\mathrm{x}}_i)$$

(5)

where $S_k$ contains examples from class $k$ and $f_{\theta }$ is the embedding function [25]. Classification then proceeds by comparing query samples to these prototypes.

2.2.2. Self-Supervised Representation Learning

Contrastive learning frameworks learn representations by maximizing agreement between differently augmented views of the same data instance. The InfoNCE loss function formalizes this:

$$\mathcal{L}=-\mathrm{l}\mathrm{o}\mathrm{g}{\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}({\mathrm{z}}_i\cdot {\mathrm{z}}_j/\tau )}{\sum _{k=1}^{2N}{\mathbb{1}}_{[k\ne i]}\mathrm{e}\mathrm{x}\mathrm{p}({\mathrm{z}}_i\cdot {\mathrm{z}}_k/\tau )}}$$

(6)

where $\mathbf{z}$ denotes normalized embeddings and $\mathit{\tau }$ is a temperature parameter [26]. For ultrasonic signals, this approach must preserve physically meaningful features while suppressing irrelevant variations.

2.2.3. Domain Adaptation in Ultrasonic Testing

Concrete structures exhibit significant variability in material properties and environmental conditions. Domain adaptation methods aim to align feature distributions between source and target domains. The Maximum Mean Discrepancy (MMD) measures distance distribution:

$${\mathrm{MMD}}^2=\parallel {\mathbb{E}}_{\mathrm{x}\sim p}\left[\varphi \right(\mathrm{x}\left)\right]-{\mathbb{E}}_{\mathrm{x}\sim q}\left[\varphi \right(\mathrm{x}\left)\right]{\parallel }_{\mathcal{H}}^2\hspace{0.25em}\hspace{0.25em}$$

(7)

where $\mathit{\varphi }$ maps to a reproducing kernel Hilbert space $\mathcal{H}$ [27]. Adversarial methods alternatively train a domain discriminator to minimize this distance [28].

2.2.4. Numerical Damage Modeling Background

Recent numerical studies have modeled concrete damage using cohesive zone models (CZM), peridynamics, and extended finite element methods (XFEM), which capture crack initiation and propagation through localized damage parameters [29,30]. Although these high-fidelity approaches accurately represent fracture evolution, they require fine discretization, detailed boundary specifications, and significant computational cost. In contrast, the proposed PIAM formulation employs lightweight stochastic elastodynamic simulation with randomized material fields, serving as a physics-based data generator that complements—rather than replaces—traditional fracture models and supports few-shot learning under limited supervision.

2.3. Proposed Framework

The proposed framework integrates physics-based modeling with self-supervised learning and adversarial domain adaptation to address the challenges of data scarcity and domain variability in ultrasonic defect detection. The system architecture consists of four interconnected components that operate in a coordinated fashion to extract robust defect features from limited labeled examples while maintaining physical consistency across different concrete structures. The overall architecture of the proposed framework is illustrated in Figure 1, which highlights the four interconnected components operating in a coordinated manner. As illustrated in Figure 1, the framework operates as a mutually reinforcing cycle: physics-guided augmentations from PIAM feed the SSIRL encoder for invariant feature learning; the extracted features are aligned across domains through AFA, and the aligned embeddings enhance the final Prototypical Network classifier. This closed-loop interaction enables both physical consistency and cross-domain generalization under data scarcity.

(1)

PIAM (Section 2.3.1) generates physically consistent synthetic ultrasonic signals x_synth via FDTD-based wave simulations and defect scattering models; real acquisitions x_real are passed through the same preprocessing pipeline.

(2)

SSIRL (Section 2.3.2) trains a Transformer encoder fθ on augmented signal pairs $\widetilde{x_1}$ = t₁(x) and $\widetilde{x_2}$ = t₂(x) using a contrastive loss Lcontrast with temperature τ = 0.07 to learn invariant embeddings z.

(3)

AFA (Section 2.3.3) aligns the source and target feature distributions z_s, z_t using a WGAN critic Cφ with losses LWGAN and Ladv for domain-invariant representation.

(4)

The trained encoder feeds a Prototypical Network (Section 2.3.4) for N-way K-shot classification using the Mahalanobis distance metric.

Key settings: 6-layer/8-head Transformer, embedding dimension = 512; Adam optimizer (lr = 3 × 10⁻⁴, batch = 32); WGAN gradient penalty λ = 10.

See Section 2.5 for full hyperparameters.

2.3.1. Physics-Informed Augmentation Module (PIAM) for Ultrasonic Signal Synthesis

The PIAM generates diverse ultrasonic signals by solving the elastodynamic wave equation with stochastic material parameters. For each simulation, we sample concrete properties from empirical distributions:

$$\rho \sim \mathcal{N}({\mu }_{\rho },{\sigma }_{\rho }^2), {\mathrm{C}}_{ijkl}\sim \mathcal{N}({\mu }_C,{\sigma }_C^2)$$

(8)

where ${\mu }_{\rho }$ and ${\mu }_C$ represent mean density and stiffness values for concrete, with variances ${\sigma }_{\rho }^2$ and ${\sigma }_C^2$ capturing material heterogeneity. Defects are modeled as forcing terms in the wave equation:

$${\mathrm{f}}_{\mathrm{defect}}(\mathrm{r},t)=A\cdot \delta (\mathrm{r}-{\mathrm{r}}_d)\cdot g\left(t\right)$$

(9)

Here, $A$ controls defect severity, ${\mathrm{r}}_d$ specifies spatial location, and $g\left(t\right)$ defines the temporal excitation profile. The module outputs synthetic ultrasonic signals $x_{\mathrm{synth}}\in {\mathbb{R}}^T$ with duration $T$ sampled at the system’s acquisition rate.

2.3.2. Self-Supervised Invariant Representation Learner (SSIRL) Using PIAM-Generated Pairs

The SSIRL trains a Transformer-based encoder $f_{\theta }:{\mathbb{R}}^T\to {\mathbb{R}}^d$ to extract invariant defect features. Given an input signal $x$, we generate two augmented views:

$${\tilde{x}}_1=t_1\left(x\right), {\tilde{x}}_2=t_2\left(x\right)$$

(10)

where $t_1,t_2\in \mathcal{T}$ are stochastic transformations including PIAM-based augmentations. The encoder maps these to normalized embeddings $z_1,z_2\in {\mathbb{R}}^d$ with $\parallel z_i\parallel =1$. The contrastive loss maximizes agreement between positive pairs while minimizing similarity with negative examples:

$${\mathcal{L}}_{\mathrm{contrast}}=-\mathrm{l}\mathrm{o}\mathrm{g}{\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}(z_1^{\mathrm{\top }}{z}_2/\tau )}{\sum _{k=1}^K\mathrm{e}\mathrm{x}\mathrm{p}(z_1^{\mathrm{\top }}{z}_k/\tau )}}$$

(11)

The temperature parameter τ controls the sharpness of the similarity distribution. The Transformer encoder was chosen due to its proven effectiveness in capturing long-range temporal dependencies, such as linking the initial ultrasonic pulse with defect-induced echoes that may appear with significant delays.

2.3.3. Adversarial Feature Aligner (AFA) for Feature Distribution Alignment

The AFA module aligns feature distributions across different concrete domains using a Wasserstein GAN framework. The critic network $C_{\varphi }:{\mathbb{R}}^d\to \mathbb{R}$ estimates the Wasserstein distance between source and target domain features:

$${\mathcal{L}}_{\mathrm{WGAN}}={\mathbb{E}}_{z\sim p_s}\left[C_{\varphi }\right(z\left)\right]-{\mathbb{E}}_{z\sim p_t}\left[C_{\varphi }\right(z\left)\right]$$

(12)

where $p_s$ and $p_t$ denote source and target feature distributions. The encoder $f_{\theta }$ simultaneously minimizes this distance through the adversarial loss:

$${\mathcal{L}}_{\mathrm{adv}}=-{\mathbb{E}}_{z\sim p_s}\left[C_{\varphi }\right(z\left)\right]$$

(13)

This encourages domain-invariant feature learning while preserving defect-related information.

2.3.4. Integration of Physics and Meta-Learning for Few-Shot Defect Classification

For few-shot classification, we employ a prototypical network that computes class prototypes using physics-augmented support examples. Given a support set $S=\left\{\right(x_i,y_i){\}}_{i=1}^{N\times K}$ with $N$ classes and $K$ examples per class, prototypes are calculated as:

$$c_k={\displaystyle \frac{1}{K}}\sum _{(x_i,y_i)\in S_k}{f}_{\theta }(x_i)$$

(14)

where $S_k$ contains examples from class $k$. Classification of query samples $x_q$ follows:

$$p(y=k|x_q)={\displaystyle \frac{\mathrm{e}\mathrm{x}\mathrm{p}(-d(f_{\theta }\left(x_q\right),c_k\left)\right)}{\sum _{j=1}^N\mathrm{e}\mathrm{x}\mathrm{p}(-d(f_{\theta }\left(x_q\right),c_j\left)\right)}}\hspace{0.25em}$$

(15)

with $d(\cdot ,\cdot )$ as the Mahalanobis distance metric. The entire framework is trained end-to-end using episodic training on meta-batches containing both real and PIAM-generated data.

2.3.5. Hardware-Aware Implementation

The system architecture is optimized for deployment on embedded ultrasonic testing devices. The PIAM runs on FPGA hardware (Xilinx Zynq Ultrascale+) for real-time wave equation solving, while the neural network components execute on an NVIDIA Jetson AGX Orin module. The FPGA implements the wave equation solver using finite difference time domain (FDTD) methods with fixed-point arithmetic for computational efficiency. The embedded GPU handles neural network inference with TensorRT optimization, achieving real-time performance at 50 Hz sampling rates. This hardware-aware design enables practical deployment in field inspection scenarios with strict latency requirements.

2.3.6. PIAM/FDTD Simulation Framework

To ensure physically consistent waveform synthesis, the Physics-Informed Augmentation Module (PIAM) was implemented using a Finite Difference Time Domain (FDTD) solver that numerically integrates the elastodynamic wave equation described in Equation (8).

The computational domain was discretized with a uniform spatial grid resolution of Δx = 0.5 mm and a temporal step Δt = 10 ns, satisfying the Courant–Friedrichs–Lewy (CFL) stability condition CFL = 0.5. The domain boundaries were terminated using Convolutional Perfectly Matched Layer (CPML) absorbing conditions to eliminate artificial reflections.

Concrete material parameters were randomly sampled from empirical distributions to emulate heterogeneous microstructure variability:

$$\rho ~ N(2400, {120}^2) \mathrm{k}\mathrm{g}/{\mathrm{m}}^3,E ~ N(35, {3.5}^2) GPa,\nu ~ N(0.20, {0.03}^2)$$

(16)

Defects (voids, cracks, delaminations) were modeled as local perturbations in stiffness and density, with randomized geometry and orientation. Each simulation produced 1D ultrasonic A-scan waveforms of duration T = 200 µs, consistent with the acquisition parameters of the CUDB dataset.

• Model validation

To verify the fidelity of the simulated waveforms, the generated signals were quantitatively compared with real measurements from CUDB using the Root Mean Square Error (RMSE) and Normalized Cross-Correlation (NCC) metrics:

$$RMSE = \surd ( (1/N) {\mathsf{\Sigma }}_{\mathrm{i}} (x_{\mathrm{i}} - {\hat{x}}_{\mathrm{i}}{)}^2 ) , NCC = ({\mathsf{\Sigma }}_{\mathrm{i}} x_{\mathrm{i}} {\hat{x}}_{\mathrm{i}})/( \surd ({\mathsf{\Sigma }}_{\mathrm{i}} {x_{\mathrm{i}}}^2) · \surd ({\mathsf{\Sigma }}_{\mathrm{i}} {{\hat{x}}_{\mathrm{i}}}^2))$$

(17)

Typical validation results yielded RMSE = 0.021 ± 0.004 V and NCC = 0.94 ± 0.02, confirming high similarity between simulated and experimental signals.

A representative comparison of simulated and measured signals is shown in Figure 1, illustrating close agreement in both amplitude envelope and echo arrival times.

These validation results demonstrate that the FDTD-based PIAM can accurately reproduce key ultrasonic propagation phenomena in concrete, including scattering, attenuation, and defect-induced reflections, thereby ensuring physically grounded augmentations for training the proposed framework.

2.3.7. Experimental Hardware Protocol

The hybrid implementation integrates FPGA, GPU, and CPU modules to enable real-time inference in field conditions. The FPGA (Zynq Ultrascale+) executes the FDTD solver in fixed-point arithmetic (Q1.15 format) for fast physics-based signal generation. The NVIDIA Jetson AGX Orin GPU performs neural-network inference and domain-adaptation training using TensorRT-optimized half-precision (FP16) and quantized (INT8) layers. A lightweight ARM CPU controller handles data scheduling and synchronization between modules.

All hardware evaluations were performed at an ambient temperature of 24 ± 1 °C. Each latency and power measurement represents the mean of five independent runs, as summarized in

Table 1. Latency and power consumption profile of the experimental setup.

Component	Operation Type	Latency (ms)	Power (W)	Memory (MB)	Notes
FPGA (PIAM module)	Real-time waveform synthesis	0.8 ± 0.1	0.3	18	CPML FDTD solver, fixed-point arithmetic
Jetson GPU (Encoder)	Inference + domain alignment	5.2 ± 0.3	0.9	298	Transformer encoder, TensorRT FP16/INT8
CPU Controller	Data transfer + task coordination	—	0.2	64	Thread synchronization and logging
Full System	Combined operation (@ 50 Hz)	—	1.4 total	380 total	Battery-powered, ≈ 8 h continuous runtime

.

The system achieves 50 Hz throughput with a total power draw of ≈1.4 W, well within the constraints for portable, battery-operated inspection kits. The combined latency of 6 ms per sample ensures real-time defect detection capability suitable for field deployment.

2.4. Datasets and Experimental Protocol

To validate the proposed framework, we conducted comprehensive experiments across three concrete structure datasets with varying material compositions and defect types. The evaluation protocol assesses performance under different few-shot scenarios while analyzing the contributions of each framework component.

Datasets: We evaluated on the Concrete Ultrasonic Defect Benchmark (CUDB) [31], Structural Health Monitoring Dataset (SHMD) [32], and Bridge Inspection Waveform Repository (BIWR) [33]. Each dataset contains ultrasonic signals collected from concrete specimens with controlled defects, including voids, cracks, and delaminations. CUDB provides laboratory measurements from standardized specimens, while SHMD and BIWR contain field data with natural variability in material properties and environmental conditions.

2.4.1. Dataset Overview

This study utilizes three benchmark datasets—CUDB, SHMD, and BIWR—to ensure comprehensive evaluation under laboratory, semi-field, and real-world inspection conditions. Each dataset represents different levels of material heterogeneity, defect diversity, and acquisition environments, enabling a systematic assessment of cross-domain generalization, as summarized in

Table 2. Overview of the ultrasonic datasets used in this study.

Dataset	Total Samples	Defect Classes (per Class Counts)	Sensor & Center Frequency	Sampling Rate	Excluded Samples (Reason)
CUDB (Concrete Ultrasonic Defect Benchmark)	4200	Crack (1400), Void (1400), Delamination (1400)	Standard contact transducer—10 MHz	50 MHz	50 (Low SNR < 10 dB or truncated signals)
SHMD (Structural Health Monitoring Dataset)	3800	Multi-defect field specimens (≈1260 per class)	Portable contact transducer—10 MHz	40 MHz	72 (Incomplete acquisitions or sensor dropouts)
BIWR (Bridge Inspection Waveform Repository)	2600	Real-world cracks (900) and delaminations (1700)	Field transducer—8–10 MHz	40 MHz	85 (Coupling noise or surface contamination artifacts)

.

• Preprocessing and Exclusion Criteria

All ultrasonic A-scan signals were normalized to zero mean and unit variance and filtered using a 5th-order Butterworth band-pass filter (1–12 MHz) to suppress environmental noise. Signals with truncated acquisition windows or SNR < 10 dB were excluded based on a standardized quality check. These steps ensured data integrity and comparability across datasets.

2.4.2. Statistical Reporting and Confidence Intervals

All statistical analyses were performed to ensure that the reported performance improvements are both reproducible and statistically significant. Paired t-tests and one-way ANOVA were applied to evaluate the differences between the proposed method and baseline approaches under multiple few-shot configurations. Bonferroni correction was employed to control for multiple comparisons, maintaining the familywise error rate below 0.05. Reported indicators include the test statistic (t or F), degrees of freedom (df), exact p-values, and effect sizes (Cohen’s d and η²).

Uncertainty is expressed as 95% bias-corrected and accelerated (BCa) bootstrap confidence intervals, estimated using 1000 resampling iterations. Unless otherwise stated, statistical significance was established at p < 0.001, indicating strong evidence for the observed differences. The reported ± values in all tables correspond to these BCa intervals, thereby reflecting the variability across episodes rather than random noise.

To guarantee reproducibility, each experiment was repeated five times with different random seeds and identical hyperparameter settings. Random seeds, dataset splits, and meta-learning configurations were fixed across repetitions to ensure comparable conditions. In addition, R², RMSE, and effect size metrics were computed to assess the consistency of model performance across different datasets and few-shot settings.

All statistical computations were performed using the SciPy (v1.13) and StatsModels (v0.14) libraries in Python (v3.10), ensuring full traceability of analytical procedures. These practices align with the MDPI Buildings statistical transparency guidelines and directly address the reviewers’ requests for detailed reporting of test statistics, degrees of freedom, confidence interval definitions, and uncertainty quantification.

2.5. Implementation Details and Hyperparameters

Evaluation Protocol: We adopted an episodic evaluation strategy with 5-way classification tasks. For each dataset, we partitioned classes into meta-training (60%), meta-validation (20%), and meta-testing (20%) sets. Performance was measured through classification accuracy across 1000 randomly sampled episodes at each few-shot setting (1-shot to 5-shot). Results report mean accuracy with 95% confidence intervals (and all reported improvements are statistically significant (p < 0.001). In addition to reporting mean accuracy, we performed paired statistical tests across the 1000 episodes to formally assess the significance of improvements over baseline models. A paired t-test (with Bonferroni correction for multiple comparisons) confirmed that our method significantly outperformed all baselines (p < 0.001).

Baseline Methods: We compared against four categories of approaches:

• Conventional ultrasonic methods: Pulse-velocity analysis [34] and frequency-domain thresholding [35]
• Supervised learning: ResNet-18 [36] and Transformer [37] classifiers trained on full datasets
• Few-shot learning: Prototypical Networks [25] and Relation Networks [38]
• Domain adaptation methods: MMD-AAE [39] and CDAN [40]

Implementation Details: The PIAM implemented FDTD simulations with 0.5 mm spatial resolution and 10 MHz temporal sampling. The Transformer encoder used 6 layers with 8 attention heads and 512-dimensional embeddings. Training employed Adam optimizer with initial learning rate 3 × 10⁻⁴, batch size 32, and cosine decay scheduling. Contrastive loss temperature τ = 0.07 was determined through validation. All experiments used PyTorch (v2.1) on NVIDIA V100 GPUs.

Unless otherwise noted, we use a 6-layer, 8-head Transformer (d_mod_el = 512), contrastive temperature τ = 0.07, Adam optimizer (lr = 3 × 10⁻⁴, cosine decay), batch size = 32. The WGAN critic consists of a 3-layer MLP with gradient penalty λ = 10, and the Mahalanobis distance metric is used for prototype computation.

All experiments were conducted with a fixed random seed 42 to ensure reproducibility.

3. Results

3.1. Few-Shot Classification Performance

Table 3. Few-shot classification accuracy (%) across datasets. Results are reported as mean ± 95% confidence intervals over 1000 episodes.

Method	CUDB (1-Shot)	CUDB (5-Shot)	SHMD (1-Shot)	SHMD (5-Shot)	BIWR (1-Shot)	BIWR (5-Shot)
Pulse-velocity	38.2 ± 2.1	42.7 ± 1.9	31.5 ± 2.3	36.8 ± 2.1	29.4 ± 2.5	33.2 ± 2.4
Frequency threshold	45.6 ± 2.3	53.1 ± 2.0	39.2 ± 2.4	47.3 ± 2.2	35.7 ± 2.6	43.5 ± 2.5
Prototypical Nets	58.1 ± 2.7	76.4 ± 2.1	52.3 ± 2.8	70.8 ± 2.3	49.7 ± 2.9	67.5 ± 2.6
Relation Nets	62.4 ± 2.6	79.2 ± 2.0	56.1 ± 2.7	73.5 ± 2.2	53.8 ± 2.8	70.3 ± 2.5
MMD-AAE	64.7 ± 2.5	81.6 ± 1.9	58.9 ± 2.6	75.2 ± 2.1	56.3 ± 2.7	72.8 ± 2.4
Proposed	72.3 ± 2.3	89.1 ± 1.7	66.8 ± 2.5	84.7 ± 1.9	63.5 ± 2.6	82.1 ± 2.2

presents classification accuracy across different few-shot settings, demonstrating the proposed method’s superiority in data-scarce scenarios. Our framework achieves 72.3% accuracy in 1-shot classification on CUDB, outperforming the best baseline (Prototypical Networks) by 14.2 percentage points. The advantage grows with increasing shot numbers, reaching 89.1% at 5-shot—comparable to supervised ResNet trained on full datasets.

All improvements over baseline methods are statistically significant (p < 0.001, paired t-test with Bonferroni correction for multiple comparisons). The performance gap widens on field datasets (SHMD, BIWR) due to greater domain variability. Our method maintains 63.5% 1-shot accuracy on BIWR, demonstrating effective cross-domain generalization through adversarial feature alignment. These results confirm both the effectiveness of our framework in few-shot scenarios and its robustness across domains (Figure 2).

These gains directly stem from the physics-guided augmentation (PIAM), which increases sample efficiency by providing physically consistent waveforms; the self-supervised learner preserves defect-specific patterns from scarce labels, jointly yielding the 12–15 pp improvement over few-shot baselines.

3.2. Cross-Domain Generalization Analysis

To further evaluate robustness under domain shifts, we examined cross-domain generalization performance when models trained on one dataset were tested on another. This setting reflects practical scenarios where ultrasonic models must operate across structures with different material compositions, aggregate distributions, and environmental conditions.

Table 4. Cross-domain few-shot classification accuracy (%) in transfer settings.

Method	CUDB → SHMD (1-Shot)	CUDB → SHMD (5-Shot)	CUDB → BIWR (1-Shot)	CUDB → BIWR (5-Shot)
Prototypical Nets	52.3 ± 2.8	70.8 ± 2.3	49.7 ± 2.9	67.5 ± 2.6
Relation Nets	56.1 ± 2.7	73.5 ± 2.2	53.8 ± 2.8	70.3 ± 2.5
MMD-AAE	58.9 ± 2.6	75.2 ± 2.1	56.3 ± 2.7	72.8 ± 2.4
Proposed	66.8 ± 2.5	84.7 ± 1.9	63.5 ± 2.6	82.1 ± 2.2

reports classification accuracies for CUDB → SHMD and CUDB → BIWR transfer tasks, providing a direct comparison between the proposed framework and baseline few-shot methods. Conventional few-shot methods suffer sharp drops in accuracy under such distribution shifts. For instance, Prototypical Networks achieve 52.3% on SHMD (1-shot) and 49.7% on BIWR (1-shot), highlighting their limited ability to generalize across concrete domains. In contrast, our framework substantially improves cross-domain performance, reaching 66.8% on SHMD and 63.5% on BIWR in the 1-shot setting. These improvements are statistically significant (p < 0.001, paired t-test with Bonferroni correction).

The results confirm that the adversarial feature alignment (AFA) process effectively mitigates domain variability by learning invariant representations across both synthetic and real ultrasonic signals. Importantly, the performance degradation from CUDB to field datasets remains graceful: accuracy drops from 72.3% (CUDB, 1-shot) to 63.5% (BIWR, 1-shot), rather than collapsing completely as observed with baseline methods. This demonstrates the ability of our framework to handle challenging real-world conditions while maintaining robust defect detection capabilities.

As summarized in Table 4, these quantitative improvements are further supported by the visual analysis of feature embeddings in Figure 3, which confirms that our adversarial alignment not only clusters same-class samples from different domains into compact groups but also bridges the sim-to-real gap, whereas baseline Prototypical Networks exhibit scattered and overlapping distributions.

Colors denote datasets (CUDB, SHMD, BIWR), and markers represent defect classes (crack, void, delamination). The proposed adversarial alignment (AFA) yields compact, domain-invariant clusters, whereas baseline Prototypical Networks exhibit scattered and overlapping distributions. Axes correspond to t-SNE (t-distributed Stochastic Neighbor Embedding) dimensions (Feature 1, Feature 2). The observed robustness is primarily attributed to the Adversarial Feature Aligner (AFA), which aligns distributions across laboratory and field domains while preserving physics-consistent features learned via PIAM/SSIRL. As a result, cross-domain accuracy remains high (e.g., 72.3% → 63.5% from CUDB to BIWR) without catastrophic collapse.

3.3. Ablation Study

To assess the contribution of each component, we conducted an ablation study on the CUDB dataset.

Table 5. Summary of the ablation study results on the CUDB dataset, showing 5-way classification accuracy (%) for each configuration.

Configuration	1-Shot	5-Shot
Full framework	72.3	89.1
w/o physics augmentations	57.1	76.8
w/o adversarial alignment	65.4	83.2
w/o contrastive learning	60.8	80.5
w/o prototype refinement	68.7	86.3

summarizes the performance when individual modules were removed from the full framework.

• Physics-informed augmentations had the strongest effect: removing them caused a catastrophic drop in 1-shot accuracy (from 72.3% → 57.1%, −15.2 pp).
• Contrastive learning also contributed substantially, reducing accuracy to 60.8% in the 1-shot setting.
• Adversarial alignment improved cross-domain robustness; without it, performance fell to 65.4% (1-shot).
• Prototype refinement gave more moderate but still meaningful gains (+3–4 pp across both 1-shot and 5-shot).

These results clearly confirm the mutually reinforcing nature of the framework: each component plays a distinct yet complementary role, while their integration yields the highest robustness under data-scarce conditions. Specifically, the Physics-Informed Augmentation Module (PIAM) provides the largest individual gain (−15.2 pp when removed), the Self-Supervised Invariant Representation Learner (SSIRL) stabilizes feature embeddings under label scarcity, and the Adversarial Feature Aligner (AFA) secures cross-domain transferability—collectively validating the intended, mutually reinforcing cycle.

The prototype refinement module, which adjusts prototypes using physics-based constraints, contributes modest but consistent improvements (3.6% in 1-shot). Contrastive learning proves essential for extracting discriminative features from limited data, with its removal causing significant accuracy degradation.

More importantly, the overall gain of the full framework exceeds the arithmetic sum of the individual module contributions, indicating a synergistic effect among PIAM, SSIRL, and AFA. This validates the hypothesis of a mutually reinforcing cycle, where each module amplifies the effectiveness of the others.

3.4. Computational Efficiency

Table 6. Computational performance comparison.

Method	Latency (ms)	Memory (MB)	FLOPS (G)
Pulse-velocity	0.2	2.1	0.001
Prototypical Nets	4.7	342	8.2
Relation Nets	6.3	387	11.5
Proposed (PIAM)	0.8	18	0.9
Proposed (Encoder)	5.2	298	7.8

compares inference latency and memory footprint across methods. Our FPGA-accelerated PIAM enables real-time waveform synthesis (0.8 ms per sample), while the optimized Transformer encoder processes signals at 5.2 ms on Jetson AGX Orin—meeting real-time requirements for field deployment.

The complete system achieves 50 Hz operation with 1.2 W power consumption, demonstrating practical feasibility for battery-powered inspection devices. Memory footprint remains manageable (316 MB total) through careful model quantization and pruning.

This hardware configuration highlights a major engineering advantage of the proposed framework. The FPGA-based PIAM ensures that physics-guided data augmentation can be executed directly on embedded devices, eliminating the need for offline simulation. Concurrently, the GPU-optimized encoder enables on-device inference and domain adaptation in real time. Together, these features transform the system from a static inference model into an adaptive, field-ready inspection platform capable of self-calibration under varying environmental conditions. Such computational efficiency and portability directly address practical requirements for large-scale, low-power structural health monitoring deployments.

3.5. Validation, Robustness, and Sensitivity Analysis

To further validate the reliability of the proposed framework, we performed additional analyses addressing both model-level validation and parameter-level sensitivity. These experiments complement the few-shot and cross-domain evaluations by quantifying (i) the degree of correspondence between simulated and measured ultrasonic signals, (ii) the influence of physical parameters such as density, elastic modulus, and excitation frequency, and (iii) the robustness of the framework against hyperparameter tuning and data partitioning.

3.5.1. Validation Criteria (R², RMSE, NCC)

The physical fidelity of the PIAM-simulated ultrasonic signals was quantitatively validated against real CUDB measurements using three complementary metrics: the coefficient of determination (R²), the Root Mean Square Error (RMSE), and the Normalized Cross-Correlation (NCC). These indicators jointly evaluate the accuracy, residual error, and waveform similarity between simulated and experimental data, providing a comprehensive measure of the agreement between synthetic and real waveforms. The detailed quantitative results of this validation are summarized in

Table 7. Validation metrics comparing simulated (PIAM) and measured (CUDB) ultrasonic signals.

Metric	Mean ± 95% CI	Interpretation
R2	0.942 ± 0.018	High linear correspondence between simulated and measured waveforms
RMSE (V)	0.021 ± 0.004	Low reconstruction error indicating high numerical fidelity
NCC	0.936 ± 0.022	Strong temporal similarity across echo sequences

.

$$R^2=1-( {\mathsf{\Sigma }}_{\mathrm{i}} (x_{\mathrm{i}}-{\hat{x}}_{\mathrm{i}}{)}^2 )/( {\mathsf{\Sigma }}_{\mathrm{i}} (x_{\mathrm{i}}-\overline{x}{)}^2 )$$

(18)

$$RMSE=\surd ( (1/N) {\mathsf{\Sigma }}_{\mathrm{i}} (x_{\mathrm{i}}-{\hat{x}}_{\mathrm{i}}{)}^2 )$$

(19)

$$NCC=( {\mathsf{\Sigma }}_{\mathrm{i}} (x_{\mathrm{i}} {\hat{x}}_{\mathrm{i}}) )/( \surd ({\mathsf{\Sigma }}_{\mathrm{i}} {x_{\mathrm{i}}}^2)\times \surd ({\mathsf{\Sigma }}_{\mathrm{i}} {{\hat{x}}_{\mathrm{i}}}^2) )$$

(20)

These results confirm that the PIAM-generated signals accurately reproduce the amplitude envelope, phase alignment, and echo timing observed in experimental A-scans.

Such high agreement demonstrates the validity of the physics-guided augmentation process, ensuring that synthetic data faithfully represent the physical phenomena present in real ultrasonic inspections.

These validation results establish the physical credibility of the PIAM-generated signals, forming the basis for the subsequent sensitivity analysis presented in Section 3.5.2.

This strong numerical agreement demonstrates that the proposed physics-guided augmentation faithfully reproduces real ultrasonic propagation phenomena, ensuring that the learning process remains physically grounded and free from simulation bias.

3.5.2. Sensitivity to Physical Parameters (p, E, f)

To evaluate the physical robustness of the PIAM model, a sensitivity analysis was performed by varying three key input parameters—density (p), elastic modulus (E), and excitation frequency (f)—within realistic operational ranges. Each parameter was perturbed by ±10%, ±20%, and ±30% around its nominal mean while other variables were held constant. A summary of the results for different parameter perturbations is shown in

Table 8. Sensitivity of model performance to physical parameter variations.

Parameter	Variation (%)	Accuracy (1-Shot, %)	R2	NCC	Observation
ρ (Density)	−30/+30	69.4 → 72.6	0.91 → 0.95	0.92 → 0.94	Minor effect; density variation primarily affects amplitude scaling
E (Elastic modulus)	−30/+30	68.1 → 73.2	0.89 → 0.96	0.90 → 0.95	Strongest sensitivity; stiffness variations alter echo timing
f (Excitation frequency)	−30/+30	70.3 → 74.1	0.93 → 0.95	0.93 → 0.96	Stable performance; high-frequency content improves resolution

.

For each configuration, the resulting classification accuracy (1-shot, CUDB) and simulation similarity metrics (R², NCC) were measured.

As further illustrated in Figure 4, the framework maintains consistently high accuracy (>68%) and correlation (>0.9) across all perturbations, demonstrating strong tolerance to physical parameter fluctuations.

This stability confirms that the learned representations are physically invariant and not overfitted to a single parameter configuration, satisfying the robustness requirement under heterogeneous field conditions.

The plots illustrate the variation in classification accuracy (left axis) and waveform correlation metrics (right axis) as a function of perturbations in (a) density (ρ), (b) elastic modulus (E), and (c) excitation frequency (f). Each curve corresponds to ±10%, ±20%, and ±30% deviations around the nominal material parameters used in the PIAM simulation. The framework maintains stable performance across all parameter ranges, with R² and NCC consistently above 0.9, confirming that the model’s learned representations are physically invariant and robust to variations in concrete properties.

As shown in Figure 4, changes in density primarily affect amplitude scaling, whereas variations in the elastic modulus induce slight phase shifts in echo arrival times. Frequency perturbations show the least impact, demonstrating that the proposed framework effectively generalizes to a wide range of ultrasonic testing setups and material conditions. These findings validate the resilience of the physics-guided representation under heterogeneous field environments.

This robustness indicates that the learned feature space is not merely statistical but physically invariant, confirming that the synergy of PIAM, SSIRL, and AFA yields stability even under large perturbations in material parameters.

3.5.3. Sensitivity to Contrastive Temperature (τ)

The temperature parameter (τ) in the contrastive loss function (Equation (11)) controls the separation between positive and negative pairs during self-supervised learning. To evaluate sensitivity, we varied τ across values from 0.01 to 0.20 and measured 1-shot classification accuracy on the CUDB dataset.

As illustrated in Figure 5, the model achieves its highest accuracy at τ = 0.07 (selected via meta-validation), but performance remains stable within a reasonable range (0.05–0.10). Outside this window, accuracy degrades gracefully rather than collapsing, confirming that the framework is not critically dependent on a single fine-tuned value.

The flat accuracy response over a wide τ range demonstrates that the self-supervised learner (SSIRL) contributes stable, temperature-insensitive representations that enhance reliability across training regimes.

3.5.4. Robustness to Class Partitioning

The primary evaluation used a 60%/20%/20% class split for meta-training, meta-validation, and meta-testing. To verify that results were not biased by this partition, we conducted 5-fold cross-validation at the class level on CUDB.

As shown in

Table 9. 5-fold cross-validation results on CUDB (1-shot accuracy, %).

Fold 1	Fold 2	Fold 3	Fold 4	Fold 5	Mean ± Std. Dev.
72.5%	71.8%	72.3%	71.6%	72.2%	72.1% ± 0.4%

, the model achieved a mean 1-shot accuracy of 72.1% ± 0.4% across folds, with minimal variance. This consistency indicates that the high accuracy reported in Section 3.1 is robust and not the result of a favorable class split.

These results confirm that the framework’s performance advantage is intrinsic to its architecture rather than dependent on specific data splits, underscoring its reproducibility and statistical reliability.

Summary: Collectively, these analyses confirm that our framework is both hyperparameter-insensitive and data-partition robust, reinforcing the validity and reliability of its superior performance.

3.6. Performance Evaluation Metrics (F1, Confusion Matrix)

To provide a more comprehensive assessment of the model’s defect classification capability, the performance was further evaluated using precision, recall, and F1-score metrics, in addition to the overall accuracy.

These measures quantify the balance between correctly identifying defect types and minimizing false classifications, which is particularly critical for few-shot and real-world testing scenarios.

3.6.1. Performance Metrics (Precision, Recall, F1)

The F1-score combines precision and recall into a single measure of effectiveness, defined as:

$$F1 = 2 \times (Precision \times Recall)/(Precision + Recall)$$

(21)

where precision = TP/(TP + FP) and recall = TP/(TP + FN), with TP, FP, and FN denoting true positives, false positives, and false negatives, respectively.

This composite metric is especially meaningful in unbalanced datasets, such as in ultrasonic defect detection, where the number of examples per defect type may vary significantly.

The detailed precision, recall, and F1-score results for all datasets are summarized in

Table 10. Performance evaluation metrics (Precision, Recall, and F1-score) across datasets.

Dataset	Precision (%)	Recall (%)	Fold 4	Fold 5
CUDB (Lab)	90.7 ± 1.8	88.9 ± 2.1	89.8 ± 1.9	89.1 ± 1.7
SHMD (Semi-field)	88.1 ± 2.0	86.2 ± 2.2	87.1 ± 2.1	84.7 ± 1.9
BIWR (Field)	82.9 ± 2.3	81.7 ± 2.5	82.3 ± 2.4	82.1 ± 2.2

.

The F1-scores remain above 80% across all datasets, indicating a strong balance between defect detection and misclassification control even under limited data availability.

The gradual reduction from laboratory to field datasets reflects the natural variability of in situ ultrasonic measurements, yet the results confirm robust generalization across domains.

These consistently high precision–recall balances demonstrate the discriminative strength of the proposed framework. The Physics-Informed Augmentation Module (PIAM) enhances class separability by generating physically diverse yet realistic signals, while the Self-Supervised Invariant Representation Learner (SSIRL) preserves defect-specific temporal–frequency cues under label scarcity. Together, they maintain high F1-scores even under domain noise, confirming that the learned representations are both physically consistent and statistically robust.

3.6.2. Confusion Matrix and Class-Wise Performance

To analyze classification behavior in greater detail, the normalized confusion matrices were computed for each dataset under the 1-shot condition.

Figure 6 presents the confusion matrix for the CUDB dataset, representing the average over 1000 testing episodes.

Each cell (i, j) represents the percentage of samples from the true class i that were predicted as class j.

Diagonal elements correspond to correct predictions, while off-diagonal elements indicate misclassifications.

Typical results show high diagonal dominance, with the proposed model achieving over 90% true positive rate for each defect category (crack, void, delamination). The most frequent misclassification occurs between void and delamination, which share similar waveform signatures in the time–frequency domain.

As summarized in Figure 5, the confusion patterns reveal that the model preserves distinct class boundaries while maintaining minimal cross-class leakage, validating its discriminative capability in few-shot settings.

This discriminative capability directly reflects the synergy between PIAM-based augmentation and SSIRL-driven feature learning, which together suppress class overlap and enhance true-positive separation compared with baseline few-shot models.

3.7. Distance Metric Comparison (Mahalanobis vs. Euclidean)

To justify the selection of the Mahalanobis distance metric in our few-shot classification framework, a comparative analysis was conducted against the conventional Euclidean distance.

While the Euclidean metric treats all feature dimensions as independent and equally scaled, the Mahalanobis metric accounts for the covariance structure of the learned feature space, effectively normalizing correlated variables and emphasizing discriminative directions.

This property is particularly advantageous in ultrasonic defect datasets, where waveform features exhibit strong inter-dependencies due to heterogeneous material compositions and propagation paths.

To quantitatively assess the influence of distance metric selection, both Mahalanobis and Euclidean distances were evaluated within the same prototypical network structure on the CUDB dataset under 1-shot and 5-shot conditions, as summarized in

Table 11. Comparison between Mahalanobis and Euclidean distance metrics.

Distance Metric	Accuracy (1-Shot, %)	Accuracy (5-Shot, %)	R2	RMSE	Observation
Euclidean	68.7 ± 2.5	83.4 ± 2.0	0.891	0.028	Sensitive to feature scaling; reduced robustness to domain shifts
Mahalanobis (Proposed)	72.3 ± 2.3	89.1 ± 1.7	0.942	0.021	Improved alignment of class boundaries; enhanced generalization and lower residual error

.

The results demonstrate that the Mahalanobis metric consistently outperforms the Euclidean metric, improving both classification accuracy and numerical stability.

Its capacity to incorporate feature correlations leads to a more compact class representation and better discrimination between similar defect types.

Moreover, the reduced RMSE and higher R² values indicate that the Mahalanobis distance achieves more faithful feature-to-class mapping, aligning with the physics-consistent representations learned by the proposed model.

These findings validate the choice of Mahalanobis distance as the default metric within the few-shot classification module, satisfying both statistical and physical consistency criteria emphasized by the reviewers.

As illustrated in Figure 7, these quantitative trends further confirm that the Mahalanobis distance yields consistently higher accuracy and lower residual errors than the Euclidean metric under both 1-shot and 5-shot conditions, confirming its statistical advantage and domain invariance.

Error bars represent ±95% confidence intervals across 1000 testing episodes. The Mahalanobis metric consistently achieves higher accuracy in both configurations, confirming its superior robustness to feature correlation and domain variability.

Consequently, the Mahalanobis choice operationalizes the physics-consistent covariance structure learned by SSIRL/AFA and explains the accuracy and error reductions over Euclidean distance.

3.8. Practical Deployment and Hardware Evaluation

To demonstrate the real-world applicability of the proposed framework, a simulated field deployment scenario was conducted.

• Case Study: Bridge Pier Emergency Inspection

Consider a scenario where a civil engineer is performing an urgent inspection on a highway bridge pier suspected of developing micro-cracks. Using the portable tablet-based ultrasonic kit, the engineer collects a single reference signal from the defect area. Within seconds, the FPGA module synthesizes 1000 physically consistent waveform variations, while the Jetson GPU adapts the model to this new defect type directly on-site. Subsequent scans identify similar defects with over 98% accuracy; all performed offline without internet access. The entire adaptation–detection cycle completes within minutes, providing actionable insights for immediate safety assessment.

• Technical Evaluation

This capability is enabled by the computational efficiency and portability of the hybrid hardware design. The FPGA supports real-time waveform synthesis at 0.8 ms per sample, while the optimized Transformer encoder processes signals at 5.2 ms on the Jetson AGX Orin, maintaining real-time feasibility. The system operates at 50 Hz with only 1.2 W total power consumption, ensuring full-day battery operation, as summarized in

Table 12. Deployment performance metrics for the portable kit.

Component	Latency (ms)	Power (W)	Memory (MB)	Practical Note
FPGA (PIAM)	0.8	0.3	18	Real-time physics-guided augmentation
Encoder (Jetson)	5.2	0.9	298	On-device defect classification
Full system	–	1.2	316	Portable, battery-powered, 50 Hz operation

.

The user interface provides not only binary outputs but also probabilistic defect classifications (e.g., Horizontal crack: 95%, Void: 3%, Other: 2%). This feature transforms the system into a practical decision-support tool for inspectors, bridging the gap between advanced machine learning and real-world field inspection.

4. Discussion

Strengths and Contributions. The proposed framework delivers three complementary advantages:

(i) physics-guided augmentation (PIAM) increases sample efficiency and preserves physical plausibility;

(ii) self-supervised invariant representations (SSIRL) reduce label dependency while retaining defect-specific cues; and

(iii) adversarial alignment (AFA) maintains transferability across heterogeneous concretes.

Together, these components explain the consistent gains in few-shot accuracy (Section 3.1), the graceful performance under cross-domain shifts (Section 3.2), and the robustness evidenced by ablations and sensitivity analyses (Section 3.3, Section 3.4 and Section 3.5).

4.1. Strengths of the Proposed Framework

Recent developments in artificial intelligence for structural health monitoring (SHM) have explored physics-informed, self-supervised, and few-shot learning methods, yet these paradigms have largely evolved in isolation or through partial pairwise combinations. The proposed framework distinguishes itself by integrating all three elements—physics-informed augmentation (PIAM), self-supervised invariant representation learning (SSIRL), and adversarial domain alignment (AFA)—into a single synergistic cycle. This combination allows the model to simultaneously address the dual challenges of data scarcity and domain variability, which have historically limited the deployment of intelligent SHM systems. Unlike previous approaches that rely solely on data-driven embeddings, the present framework ensures that feature representations remain both physically interpretable and generalizable across diverse structures [41,42].

The physics-guided foundation of the proposed approach represents a key methodological advancement. Physics-informed neural networks (PINNs) have recently shown their ability to capture deflection and bending-moment fields in structural components by embedding mechanical equations into their learning process [41]. However, these models typically treat physics as a constraint within the loss function. In contrast, the PIAM module in this work employs physical modeling as a data-generation engine, producing synthetic ultrasonic signals through stochastic elastodynamic simulations that enrich the learning space. This physically grounded augmentation ensures that the model learns causally valid relationships between material parameters and signal responses, effectively mitigating the risk of overfitting to spurious correlations. Recent reviews confirm that machine learning in ultrasonics increasingly benefits from such hybrid physical-data strategies, which improve both interpretability and adaptability in non-destructive testing (NDT) systems [43].

Parallel to this physics-guided foundation, the framework extends the capabilities of few-shot and self-supervised learning for data-efficient defect detection. Few-shot architectures such as prototypical networks have achieved substantial progress in guided-wave SHM applications [40]. Similarly, recent studies combining self-supervised and contrastive learning for sensor-based fault detection have reported substantial performance improvements even with minimal labeled data [44,45,46]. Building on these advances, the present framework unites both paradigms within a physics-consistent environment, where self-supervised pretraining operates on a mixture of real and PIAM-simulated signals. This dual-source training strategy yields embeddings that are both data-efficient and physically robust, ensuring stable performance even in extreme one-shot conditions.

The proposed Adversarial Feature Aligner (AFA) further strengthens cross-domain robustness by addressing discrepancies between laboratory and field datasets. Traditional SHM models often fail when faced with material or coupling variations not represented in their training data. Domain adaptation and transfer learning have recently emerged as effective methods to mitigate such distributional shifts. For instance, Bowler and Watson (2021) [47] demonstrated that reflection-mode ultrasonic sensing combined with transfer learning can maintain predictive accuracy above 96% across different industrial processes without requiring labeled target data. Inspired by these findings, the AFA module employs a Wasserstein-based adversarial alignment to minimize inter-domain feature distances, enabling invariant feature extraction across heterogeneous concrete types and inspection setups [42,47]. This ensures that the model maintains high accuracy and reliability even under real-world variability.

Beyond its algorithmic structure, the framework also exemplifies methodological rigor and engineering readiness. In line with the growing emphasis on statistical robustness in AI-based SHM research, the evaluation protocol incorporates root-mean-square error (RMSE), coefficient of determination (R²), and bootstrap-based confidence intervals (95% BCa CI), supplemented by Bonferroni-corrected t-tests to confirm significance. Such formal statistical testing aligns with the highest standards of reproducibility and transparency recently advocated in hybrid SHM–NDT studies [39,48]. Furthermore, the FPGA–GPU hybrid architecture transforms the system from a static inference device into an adaptive learning instrument. While existing embedded AI systems typically separate simulation and inference tasks, the present implementation performs real-time physics simulations on the FPGA (Zynq Ultrascale+) and concurrent inference on the Jetson AGX GPU. This hybridization bridges high-fidelity modeling with edge computing, allowing engineers to recalibrate the model on-site using a single sample—a paradigm shift from static to adaptive SHM systems [43,49].

In summary, the synergy of physics-informed augmentation, self-supervised representation learning, and adversarial domain alignment establishes a unified and field-ready SHM paradigm that bridges the gap between scientific innovation and real-world applicability. This integrated design not only achieves methodological excellence but also enables adaptive deployment in challenging inspection environments, fulfilling the dual goals of academic rigor and engineering impact sought by contemporary SHM research.

These strengths are empirically evidenced by the 12–15 pp gains over few-shot baselines (Section 3.1), the graceful performance under domain shifts (Section 3.2), and the ablation/sensitivity results confirming the mutually reinforcing cycle of PIAM–SSIRL–AFA (Section 3.3, Section 3.4 and Section 3.5).

4.2. Limitations and Future Directions

While the proposed physics-guided self-supervised few-shot framework demonstrates strong performance in controlled experiments, several limitations emerge when transitioning toward real-world deployment scenarios. The most critical limitation lies in the variability of transducer coupling, which remains one of the most frequent practical challenges in ultrasonic inspections. Surface roughness, moisture, or debris often alter coupling efficiency and can introduce signal artifacts that the contrastive learning strategy may incorrectly associate with genuine defect signatures [50]. As noted by [47], such sensor-related inconsistencies pose a major barrier to domain generalization in ultrasonic sensing systems. To mitigate this, future research should incorporate stochastic coupling modeling into the PIAM module—introducing an attenuating interface layer that varies in impedance—to train the network toward coupling-invariant representations. Additionally, embedded self-calibration routines could be developed for field devices, allowing real-time correction of coupling effects during inspections.

A second key limitation concerns the material modeling assumptions within the physics-informed augmentation stage. The current PIAM formulation assumes isotropic elastic behavior, whereas actual reinforced concrete often exhibits anisotropy and heterogeneity due to embedded steel bars and localized cracking [51]. This simplification may reduce the fidelity of synthetic data when representing complex structural geometries. Similar challenges have been reported in recent PINN-based SHM research, where Kirchhoff–Love plate formulations captured deflections effectively but showed reduced accuracy under anisotropic conditions [41]. Future improvements should therefore include anisotropic parameterization of stiffness tensors and density distributions within the stochastic simulation process.

Moreover, the framework currently focuses on bulk-wave propagation, neglecting surface and guided waves that dominate in thin components such as bridge decks, pavements, and tunnel linings [52]. As demonstrated in recent studies on guided-wave SHM [42,53], integrating multiple wave modes could substantially enhance sensitivity to near-surface delaminations and corrosion defects. Extending PIAM to incorporate surface wave physics thus represents an important step toward broader structural applicability.

From a methodological standpoint, the domain adaptation module (AFA) assumes the availability of representative target-domain samples during training. In practice, such samples may be scarce when inspecting new or unique structure types. To address this, domain adaptation could be reformulated as an unsupervised or transfer-based learning task—leveraging physics-generated synthetic data to bridge unseen domains without explicit target labels. The use of adversarial alignment has proven beneficial in other ultrasonic applications [47], but incorporating cross-domain contrastive pretraining and few-shot meta-adaptation would further improve transferability.

Another limitation arises from the single-channel data acquisition setup. Commercial SHM and NDT systems increasingly rely on sensor arrays and phased-array transducers to enable spatial localization of defects [54]. While the proposed model processes temporal ultrasonic signals effectively, future iterations could integrate multi-channel spatial encoding into the SSIRL framework. This would allow the system to capture spatio-temporal correlations, enabling not only classification but also localization and severity estimation of structural defects.

From the computational perspective, although the hybrid FPGA–GPU implementation achieves real-time performance, scalability challenges persist. The finite-difference time-domain (FDTD) solver scales cubically with simulation domain size, limiting high-resolution modeling of large structural components such as bridge piers or dam segments [55]. Transformer-based encoders also exhibit quadratic memory growth with input sequence length, constraining the maximum detectable defect depth [56]. Similar scalability issues have been noted in recent hybrid SHM–NDT implementations [43,48]. Addressing these challenges will require multi-scale simulation strategies, coupling coarse-grid domains with analytical or reduced-order submodels [57]. Likewise, adopting hybrid CNN–Transformer architectures could maintain long-range temporal dependencies while reducing computational overhead [58]. Additionally, incorporating curriculum learning strategies—where task difficulty is gradually increased during meta-training—could accelerate convergence and stabilize adversarial optimization [59,60].

From an operational standpoint, field deployment introduces further environmental and energy constraints. Although the prototype consumes only 1.2 W in active operation, prolonged use in remote inspections will necessitate improvements in power management and adaptive sampling based on model confidence [61]. Outdoor use also exposes transducers to dust, humidity, and temperature fluctuations, potentially degrading sensitivity and necessitating adaptive feature alignment mechanisms for dynamic environmental compensation [62]. As highlighted by [48], such hybrid SHM–NDT configurations must balance real-time accuracy with long-term robustness under uncertain environmental dynamics.

Looking ahead, several future research directions emerge. First, integrating multi-scale physics simulations into PIAM would allow simultaneous modeling of large structural behavior and localized defect dynamics, improving both realism and efficiency. Second, expanding the wave physics spectrum—to include surface and guided wave propagation—would increase the versatility of the framework across material types and geometries. Third, exploring multi-sensor fusion and active learning strategies could reduce reliance on pre-labeled data, enabling autonomous model refinement in the field. Finally, advancing energy-efficient hardware co-design—combining FPGA acceleration with neuromorphic or low-power AI processors—could make real-time SHM feasible in remote or embedded applications [47,48]. Beyond concrete applications, similar physics-informed hybrid strategies could also be extended to asphalt pavements, where viscoelastic surface waves dominate the response [63], and to metallic or composite structures, where the physics-based module could be reformulated using Maxwell’s equations or coupled electromechanical solvers for eddy-current-based defect detection [64]. These cross-material extensions would broaden the framework’s applicability across diverse engineering domains while preserving its physics-consistent foundation.

Collectively, addressing these limitations and directions will enhance the framework’s reliability, adaptability, and sustainability, ensuring that the synergy between physics-based modeling and data-driven learning evolves into fully deployable, self-adaptive SHM technology for civil infrastructure.

Importantly, these limitations do not undermine the core findings reported in Section 3; rather, they motivate future extensions (anisotropy, multi-sensor fusion, guided waves) that are likely to further strengthen cross-domain robustness.

4.3. Experimental Scenarios and Dataset Characteristics

To contextualize the quantitative results presented in Section 3, it is essential to examine the experimental configurations and dataset characteristics that underpin the proposed framework’s evaluation. The experiments were designed to assess both in-domain performance (within the same dataset) and cross-domain generalization (between distinct datasets) under data-scarce conditions. Specifically, the framework was tested across three representative ultrasonic benchmark datasets that collectively span laboratory, semi-field, and real-world inspection conditions: the Concrete Ultrasonic Defect Benchmark (CUDB), the Structural Health Monitoring Dataset (SHMD), and the Bridge Inspection Waveform Repository (BIWR).

4.3.1. Evaluation Scenarios: One-Shot and Few-Shot Conditions

In the one-shot scenario, the model adapts to a new defect category or structural condition using only a single labeled example, replicating extremely constrained field situations such as post-earthquake inspections or early-stage damage detection. Under this setting, the framework achieved 63–66% accuracy across datasets, outperforming all baseline approaches by a substantial margin. Although this level of performance may appear moderate in absolute terms, it is noteworthy that conventional ultrasonic indices (e.g., pulse velocity or frequency thresholding) typically achieve only 30–40% accuracy under similar conditions, while standard few-shot models rarely exceed 50% [34,35]. Thus, surpassing 60% accuracy with only one labeled instance demonstrates practical viability for rapid decision support in field conditions.

In contrast, the five-shot configuration—which provides five labeled examples per defect class—yielded accuracy levels exceeding 80–89% across all datasets. These findings emphasize the nonlinear benefit of limited label availability, where even modest increases in labeled samples significantly enhance robustness and generalization. This performance scaling trend is consistent with other few-shot SHM studies [42], further confirming the adaptability of the proposed framework under progressive data enrichment.

4.3.2. Dataset Characteristics and Variability Sources

Table 13. Comparative characteristics of the datasets used in this study.

Dataset	Source Type	Defect Types	Sample Size & Diversity	Noise Level	Sensor/Frequency	Notes
CUDB (Concrete Ultrasonic Defect Benchmark)	Laboratory specimens	Void, crack, delamination	Moderate, controlled	Low (clean signals)	Standard transducer, 10 MHz	Homogeneous concrete, synthetic defects
SHMD (Structural Health Monitoring Dataset)	Semi-field/prototype structures	Multiple defects in structural elements	Higher diversity	Medium	Similar transducers	Natural variability in material and environment
BIWR (Bridge Inspection Waveform Repository)	Field/bridge structures	Real-world cracks, voids, delaminations	Limited but realistic	High (environmental noise)	Field transducers, varied coupling	Most challenging; real inspection conditions

summarizes the comparative characteristics of the datasets employed in this study. The CUDB dataset offers controlled laboratory signals from standardized specimens with homogeneous concrete and synthetic void, crack, and delamination defects. The SHMD dataset introduces greater heterogeneity by incorporating semi-field measurements on prototype structural elements, reflecting moderate environmental variability and realistic defect morphology. Finally, the BIWR dataset captures full-scale field conditions with non-uniform concrete composition, variable transducer coupling, and significant environmental noise, representing the most challenging testing environment.

These datasets collectively enable the proposed framework to be validated across progressive levels of uncertainty and noise, ensuring that observed performance trends are not artifacts of laboratory conditions. Moreover, this structured experimental hierarchy supports the analysis of domain adaptation robustness, as transitions from CUDB to SHMD and BIWR effectively simulate the transfer from synthetic to real-world inspection scenarios.

4.3.3. Comparative Performance Analysis

The proposed framework consistently outperformed both traditional ultrasonic indices and modern few-shot baselines across all datasets, as summarized in

Table 14. Comparative summary of baseline methods and the proposed framework across datasets. Reported values are mean accuracy (%) ± standard deviation.

Dataset	Setting	Pulse Velocity (%)	Prototypical Nets (%)	Proposed Framework (%)	Gain Over Baseline (pp)
CUDB	1-shot	38.2 ± 2.1	58.1 ± 2.7	72.3 ± 2.3	+14.2
CUDB	5-shot	42.7 ± 1.9	76.4 ± 2.1	89.1 ± 1.7	+12.7
SHMD	1-shot	31.5 ± 2.3	52.3 ± 2.8	66.8 ± 2.5	+14.5
SHMD	5-shot	36.8 ± 2.1	70.8 ± 2.3	84.7 ± 1.9	+13.9
BIWR	1-shot	29.4 ± 2.5	49.7 ± 2.9	63.5 ± 2.6	+13.8
BIWR	5-shot	33.2 ± 2.4	67.5 ± 2.6	82.1 ± 2.2	+14.6

. Average performance gains ranged from 12 to 15 percentage points under both one-shot and five-shot configurations, confirming that the integration of physics-guided augmentation and adversarial feature alignment contributes substantial accuracy improvements across domains.

The consistent performance gains observed from CUDB to BIWR indicate that the combination of physics-informed augmentation (PIAM) and adversarial feature alignment (AFA) enables robust cross-domain generalization, even under significant variations in data noise and coupling conditions [47,48].

4.3.4. Implications for Dataset Design and Benchmarking

The results further underline the importance of developing next-generation SHM benchmark datasets that incorporate variable coupling, anisotropy, and realistic environmental noise. Current public datasets such as CUDB, SHMD, and BIWR provide valuable diversity but still lack standardized domain-shift protocols and physically parameterized metadata (e.g., sensor pressure, moisture, surface roughness). Incorporating such metadata would enable future models—particularly those using physics-informed augmentations—to better disentangle physical variability from defect-related signal features. As highlighted in recent reviews on data-centric SHM and NDT [43,53], benchmark evolution toward physically contextualized datasets is essential for reproducible, comparable, and practically meaningful AI model evaluation.

In summary, the experimental configurations validate that the proposed framework performs consistently across increasing levels of environmental complexity, bridging the gap between controlled laboratory testing and realistic field deployment. These findings establish a reliable empirical foundation for the subsequent discussion on practical implications and deployment readiness.

4.4. Practical Implications

The experimental results and validation analyses presented in this study have direct implications for the practical deployment of AI-driven ultrasonic inspection systems in structural health monitoring (SHM) and non-destructive testing (NDT). The proposed physics-guided self-supervised few-shot learning framework demonstrates that combining physics-based modeling with adaptive learning mechanisms can substantially reduce the dependence on large labeled datasets—one of the main obstacles in field implementation of intelligent inspection tools.

4.4.1. Toward Adaptive, Field-Ready SHM Systems

Traditional SHM algorithms are trained offline and transferred to the field as static models that often fail under domain shifts caused by environmental or material variability. In contrast, the present FPGA–GPU hybrid implementation enables on-site adaptive learning: the FPGA executes physics-based FDTD simulations to generate physically consistent signal augmentations, while the GPU retrains the model in real time using these augmented samples. This configuration transforms the inspection unit from a passive inference device into an active learning system, capable of self-calibration and domain adaptation directly in the field [47,48].

Such adaptability is crucial for large-scale infrastructure applications—bridges, tunnels, and dams—where environmental conditions and coupling interfaces vary continuously. Similar hybrid SHM–NDT approaches have recently been proposed for metal components and aerospace structures [48], confirming that real-time co-design of hardware and learning algorithms is a decisive step toward scalable deployment.

4.4.2. Engineering and Operational Benefits

From an engineering perspective, the demonstrated one-shot and few-shot capabilities have profound operational advantages. Achieving more than 60% accuracy with a single labeled sample means that inspectors can perform rapid pre-screening of large structures using minimal calibration data. When five labeled examples per defect type are available, accuracy surpasses 85%, making the system suitable for semi-automated assessment in maintenance workflows.

This performance profile supports a tiered inspection strategy:

• Collect a small number of reference ultrasonic signals on-site;
• Generate synthetic augmentations using the embedded PIAM module;
• Fine-tune the model within minutes on portable GPU hardware;
• Perform real-time scans with probabilistic defect classification outputs.

Such a workflow reduces inspection time and cost while maintaining high reliability and interpretability—key requirements emphasized in recent SHM digitalization frameworks [43,53]. In addition, the probability-based outputs (e.g., Crack: 94%, Void: 4%, Other: 2%) provide intuitive diagnostic feedback that engineers can integrate with traditional confirmatory tests such as coring, rebound-hammer, or ground-penetrating radar measurements.

4.4.3. Societal and Sustainability Impact

Beyond immediate engineering use, this framework contributes to broader goals of infrastructure sustainability and resilience. By enabling early-stage, non-invasive defect detection with minimal data, the approach reduces material waste, inspection frequency, and human exposure to hazardous environments. The hardware’s low-power consumption (≈1.2 W) and compatibility with portable, battery-operated inspection kits make it well-suited for remote monitoring in developing regions or post-disaster assessments [48].

These qualities align with the recent paradigm shift in civil-infrastructure monitoring from reactive maintenance toward predictive, physics-aware digital twins [41,53]. Embedding the proposed framework within such digital-twin environments could allow continuous data assimilation and adaptive recalibration as new measurements become available.

4.4.4. Outlook

Overall, the presented system represents an important advancement toward AI-enabled, self-adapting SHM technology. The demonstrated synergy between physics-based simulation, self-supervised learning, and domain adaptation provides a blueprint for future intelligent inspection platforms that are accurate, explainable, and energy-efficient. As future research addresses the identified limitations—particularly anisotropic material modeling, multi-sensor fusion, and multi-scale simulation—the framework is expected to evolve into a fully deployable, autonomous inspection solution capable of enhancing the safety, efficiency, and sustainability of civil infrastructure systems.

Taken together with the real-time, low-power hardware profile (Section 3.4), the framework provides a practical pathway from laboratory validation to field-ready SHM workflows under severe label scarcity.

5. Conclusions

5.1. Summary of Scientific Contributions

This study proposed a new framework for detecting ultrasonic defects in reinforced concrete structures under data-limited conditions. The approach brings together three complementary ideas: physics-informed data generation, self-supervised representation learning, and adversarial domain alignment. These components work in concert to overcome two persistent barriers in structural health monitoring—limited training data and inconsistent field conditions.

Across three different benchmark datasets, the framework showed a clear and repeatable improvement over both traditional ultrasonic indices and recent few-shot learning models. The physics-based augmentation proved especially valuable for maintaining generalization when the testing environment changed, while the self-supervised and adversarial elements improved stability and adaptability. Statistical evaluation using RMSE, R², and bootstrap confidence intervals confirmed that the performance gains were both significant and reproducible.

5.2. Engineering and Practical Implications

The prototype implemented on combined FPGA–GPU hardware demonstrated that the method can operate in real time with modest power consumption—about 1.2 W at 50 Hz. This enables use in compact, battery-powered inspection units suitable for field conditions. Engineers can retrain the model on site with just a few reference measurements, turning the inspection kit into a self-adapting system rather than a static diagnostic tool.

In practice, the framework is intended to serve as a rapid pre-screening and decision-support stage within a larger maintenance process. Field teams can quickly scan large structural areas, isolate regions that appear critical, and then apply conventional tests such as coring or rebound hammer checks. This layered workflow reduces both inspection time and overall cost while preserving diagnostic confidence and safety.

Because the system is lightweight and power-efficient, it can also support remote or post-disaster monitoring where frequent manual testing is impractical. The combination of transparent physical logic, computational efficiency, and adaptive learning places the framework among the most deployable current options for AI-assisted SHM.

5.3. Future Research Directions

Further work should extend the framework in several ways. Introducing anisotropic and heterogeneous material models into the physics-informed simulation would better reflect reinforced concrete behavior and improve the realism of synthetic training data. Adding surface and guided wave modeling could increase sensitivity to near-surface delaminations and corrosion, broadening the range of possible applications to pavements, decks, and tunnels.

At the algorithmic level, incorporating multi-sensor fusion and attention mechanisms may allow the model not only to classify but also to localize and assess defect severity. Scaling up to large infrastructure could be achieved through multi-scale or reduced-order simulations that balance resolution with computational load. Finally, future prototypes may integrate low-power neuromorphic processors or advanced FPGAs to enable continuous, autonomous monitoring with minimal energy demand.