The structural integrity of railway rails is essential for the safety and efficiency of modern transportation networks, where early detection of damage is crucial to preventing catastrophic failures and service disruptions. Guided wave-based structural health monitoring (SHM) offers long-range and high-sensitivity inspection capabilities for rail infrastructure. However, the complex propagation characteristics of ultrasonic guided waves and the presence of noise in operational environments pose significant challenges for traditional signal processing methods. In this study, a deep learning-based framework is proposed for rail damage detection utilizing guided wave SHM. The methodology involves denoising, normalizing, and transforming the acquired ultrasonic signals into time–frequency representations, which, together with raw waveforms, are used as inputs to a long short-term memory (LSTM) network. The LSTM model is designed to automatically learn temporal dependencies and extract discriminative features for accurate damage identification. The proposed approach achieves superior detection accuracy compared to conventional techniques and maintains robustness under elevated noise conditions. These findings underscore the potential of integrating deep learning with guided wave SHM for intelligent and automated rail defect detection, paving the way for scalable monitoring solutions and enhanced railway infrastructure reliability.

This work addresses a critical challenge for the implementation of Structural Health Monitoring (SHM) systems in the aerospace sector: the validation of Ultrasonic Guided Wave (UGW) models in curved geometries and under real conditions of fatigue crack growth. The geometric complexity of curved aeronautical composites generates dispersion, attenuation, and mode conversion that complicates UGW signal interpretation. While UGW monitoring is well-established for flat composites and fatigue damage is more commonly tracked via Acoustic Emission (AE), validated models for mixed-mode fatigue delamination in curved components remain scarce. The main objective is to validate a 2D Finite Element Model (FE) developed at Abaqus to simulate the precise interaction between guided waves (A0 and S0 modes) and a Mixed-Mode fatigue delamination in two curved specimens of different materials and thicknesses. A co-simulation model was implemented to replicate the Round-Robin tests with Macro Fiber Composite (MFC) transducers, in which crack propagation was experimentally monitored using UGW through two bonded MFC piezoelectric sensors acting as both actuators and receivers, while a 15 mm PTFE film was inserted during manufacturing to ensure crack growth in a single delamination plane. The delamination simulation was carried out by progressively releasing the nodes of the set of contact nodes between the upper and lower surfaces of the delamination, resulting in a free surface, replicating the Severity and Location of the Damage (SoD and LoD) observed experimentally. Numerical-experimental validation focused on comparing the trends of the same key signal features extracted from the experimental campaign and the FE simulations as the material was damaged. The results confirm a high correlation in the features, demonstrating that the FE model successfully captures the effect of delamination growth and wave-curvature interaction. Finally, this validated model is proposed as the ideal training environment for a Leave-One-Group-Out Cross Validation (LOGO-CV) scheme for material generalization, demonstrating robustness to manufacturing variations.

Precision in damage localization is critical for the safety and maintenance of engineering structures. While previous studies have utilized Convolutional Neural Networks (CNNs) paired with Continuous Wavelet Transform (CWT) scalograms of ultrasonic guided waves to regress damage coordinates, these deterministic approaches often fail to account for predictive noise and model uncertainty. This paper proposes a novel framework that integrates Bayesian Inference into the framework of Explainable AI (XAI) techniques to provide a transparent and reliability-aware diagnostic tool for structural damage localization. Within this framework, standard point estimates for model weights w are replaced by posterior probability distributions p(w|D), conditioned on the training dataset D, through the use of Variational Inference. Consequently, the model outputs a predictive mean for the spatial coordinates (x ̂,y ̂) and a corresponding predictive variance σ^2, which serves as a formal measure of localization uncertainty. To interpret these results, we adapt Gradient-weighted Class Activation Mapping (Grad-CAM) to explain not only the predicted location but also the source of the uncertainty. By analyzing the time-frequency representations of the acoustic signals through this lens, we can pinpoint which time-frequency features contribute to high-confidence detections versus those that induce predictive variance; in other words, this allows us to not only say: "sensor A” influenced the damage coordinate prediction, but also "sensor A” is the reason for the high uncertainty in this prediction.
The approach is validated using experimental ultrasonic signals from an aluminium plate subjected to variations in temperature at different localizations (hot-spots). Results demonstrate that the integration of Bayesian uncertainty effectively filters out spurious hot-spots caused by environmental fluctuations. Furthermore, the XAI component reveals a physical alignment between high-relevance pixels in the CWT domain and the theoretical Time-of-Flight (ToF) of structural Lamb waves.

Machine learning is widely applied in Structural Health Monitoring (SHM) due to its superior ability to recognize patterns and classify instances, it significantly improves damage assessment processes. Ultrasonic Guided Wave (UGW) systems collect data in the form of large-scale time series of pitch-catch signals. Convolutional Neural Networks (CNNs) are particularly well suited for processing such large datasets.

Additionally, various techniques exist for post-processing UGW time-series data to prepare the CNN input, either by extracting features or reducing the dimensionality of the data space. When using an image-based input, CNNs show great potential. There are multiple ways to convert time-series data into image representations. The aim of this paper is to evaluate different computer vision (CV) techniques for transforming time-series data into image-like inputs for CNN-based damage assessment systems.

In this work, several methods are implemented and compared: Short-Time Fourier Transform (STFT), Gramian Angular Field (GAF), and Shearlet Transform. STFT converts time-series data into spectrograms, representing the spectrum and amplitude (intensity) of signal frequencies over time. GAF encodes time-series data into images by computing GAF matrices, which convert data values into angles. The Shearlet Transform, a technique from geometric multiscale analysis of singularities, enables directionally sensitive sparse representations of images with anisotropic features.

The proposed techniques are applied to the Open Guided Wave (OGW) dataset to facilitate data preprocessing and enhance CNN-based frameworks for damage assessment. The OGW dataset was obtained from a carbon fiber reinforced plastic (CFRP) plate with artificial damage introduced by placing metal disks at various locations.

This paper demonstrates the effectiveness of the proposed approach and compares the aforementioned CV techniques as input enhancement methods for CNN-based damage assessment systems.

Structural Health Monitoring (SHM) refers to a set of techniques for assessing structural integrity and identifying potential damage, which aids in maintenance decision-making to enhance the performance levels of the structure. The ability to travel larger distances with very little attenuation and the sensitivity towards minor damages make the Guided Waves(GWs) the most suitable technique for SHM. The GWs are generated in the structure using Piezoelectric Wafer Transducers (PWTs) that are attached to the surface of the host structure using adhesives. These adhesives provide the proper mechanical coupling between the PWTs and the host structure, enabling the necessary strain transfer between them. However, structures in service often operate under varying Environmental and Operational Conditions (EOCs), leading to possible degradation of adhesive properties or partial debonding of PWTs over time. While complete debonding of a transducer can be easily detected from the acquired guided wave signals, partial debonding presents a major challenge, as the sensor continues to operate but exhibits subtle changes in signal characteristics. These changes can distort the wavefield and mimic damage signatures, potentially resulting in false damage detection. Therefore, ensuring the health of the sensor network, particularly the detection of partially debonded transducers, is essential prior to applying guided-wave damage detection algorithms.

This work proposes a data-driven framework based solely on GW measurements, employing a one-dimensional Convolutional Autoencoder (1D-CAE) to detect partially debonded piezoelectric sensors within a sensor network attached to the structure, thereby eliminating the need for additional sensing modalities such as EMI. Experiments are conducted on an aluminium plate equipped with a square PWT network, where sensors are intentionally installed with different degrees of debonding to mimic realistic degradation scenarios. The model is trained to learn intrinsic variations in GW signals caused by sensor debonding and EOCs, enabling robust separation of sensor-induced variations from actual structural damage signatures. The proposed approach achieves clear separation of bonding conditions in the latent feature space and accurate classification of sensor states, with 100% classification accuracy under experimental conditions. Robustness studies further demonstrate reliable performance at signal-to-noise ratios as low as 10 dB. A comparative assessment indicates that the CAE-based framework significantly outperforms conventional Maximum Amplitude Spectra (MAS)–based techniques, which are sensitive to sensor placement and excitation frequency. By explicitly incorporating sensor bonding assessment into the diagnostic pipeline, the proposed AI-driven approach reduces false alarms and improves confidence in real-time GW-based SHM systems using permanently bonded piezoelectric sensors.

Structural Health Monitoring (SHM) of aircraft components is of critical importance due to their operation under severe fatigue loads and strong noise interference. Guided-wave-based SHM offers advantages of long-range propagation and high sensitivity to small defects, however, the damage-induced responses are typically weak and easily masked by noise. To address this challenge, a meta-lens based local signal enhancement strategy is proposed to achieve reliable and sensitive damage detection in noisy environments. A dual-lens and dual-sensor configuration is designed for a two-dimensional monitoring region, enabling both damage quantification and localization using meta-lens enhanced signals. The signal characteristics are systematically compared under noise-free and three noise-level conditions, which are SNR=5, 10, 15 dB, and these features are further used to train a Gaussian Process Regression (GPR) model for probabilistic damage identification. The influence of practical manufacturing and assembly tolerances on the meta-lens performance is also investigated. Both numerical and experimental results confirm that the meta-lens significantly improves the robustness of feature extraction and the sensitivity of damage-related responses. The redistributed signal features facilitate more effective learning within the GPR model, achieving localization errors below 1 mm and quantification errors below 0.5 mm, with a remarkable reduction in estimation uncertainty. Moreover, the performance remains stable within machining and assembly deviations of ±0.02 mm and ±0.2 mm, respectively. These results provide a comprehensive reference for the design, signal processing, and noise robustness performance of meta-structure enhanced SHM systems, demonstrating the strong potential of meta-lens based physical signal amplification for practical damage identification.