The detection of defects using ultrasonic guided waves remains challenging because the defect size-dependent backscattered signals are typically low in amplitude compared to the incident waves, often leading to signal obscuration. While 2D-FFT directional decomposition effectively boosts the contrast of these backscattered signals, its main drawback is the requirement for a large array of sensors to achieve sufficient spectral resolution, resulting in high data and computational demands. To overcome these data-intensive and computational limitations, a novel deep learning framework based on a convolutional neural network (CNN) was implemented for efficient rail defect diagnosis.
A CNN was successfully trained on a robust hybrid dataset of 100 kHz array-signal B-scan images that integrated 3D COMSOL rail simulations with experimental EMAT array measurements. Customized noise augmentation was applied to enhance real-world reliability. This deep learning framework delivered outstanding performance, achieving a 98% defect-detection accuracy. Its defect-localization capability was equally strong, yielding a longitudinal error of just 1.765 mm over a 1-meter inspection range. Vertical localization errors for 15 mm transverse defects were similarly small: 2% in the web and 8.6% in the base. These superior results clearly demonstrate the practical viability and significant promise of deep learning for accurate guided-wave signal interpretation and defect localization in rail inspection.

Ultrasonic nondestructive evaluation can locate and characterize damage by analyzing a structure’s response to high-frequency ultrasonic waves. Local wavenumber estimation is a common damage-indicative metric that can locate hidden regions of damage, such as corrosion. However, depending on the thickness of the structure, multiple wave modes of differing wavenumber may be observed, resulting in an inconsistent wavenumber and damage estimation.

The proposed method implemented template matching between local wavenumber filtered data and material dependent dispersion curves to provide a consistent thickness estimate across the entire specimen surface. Acoustic steady-state excitation spatial spectroscopy (ASSESS) measurements mapped the structure’s response to a single sinusoidal excitation. The surface velocity data was rapidly collected with a steering laser Doppler vibrometer (LDV) and contextual geometry was collected with a LiDAR to allow surface perspective correction; the resulting data was a 2D complex-valued steady-state velocity wavefield.

The wavefield data and dispersion curves were processed through a series of wavenumber Gabor filters. Pearson’s correlation coefficient between the filtered data and the template was used to calculate the correlation between dominant wavenumbers and material thickness. The thickness estimates were calculated for different frequencies and thicknesses and evaluated for accuracy.

This study investigates how internal architecture affects guided-wave propagation and defect detectability in additively manufactured (AM) curved PLA shells. Twelve specimens with identical curvature and thickness but different infill topologies—solid, concentric, grid, and gyroid—and defect types (healthy, 5 mm hole, and embedded PTFE delamination) were fabricated via fused filament fabrication (FFF). SuRE signals were collected using surface-mounted PZTs across 50 tests per condition and analyzed using Short-Time FFT spectrograms. Results show that infill geometry influences ultrasonic energy transmission and waveform characteristics: solid and concentric patterns preserved amplitude and coherence, while grid and gyroid increased attenuation and scattering. Despite these differences in wave behavior, a ResNet-18 CNN trained to classify Healthy, Hole, and Delamination states achieved 100% accuracy across all infill types, demonstrating that the proposed method remains highly effective and robust regardless of internal architecture. These findings highlight both the need to consider infill-dependent impedance effects in NDT modeling and the strong potential of deep learning to provide reliable defect detection across heterogeneous AM geometries.

Jetty pipelines are essential for marine refinery operations, enabling the transfer of crude, feedstocks, and refined products between onshore facilities and offshore berths. Their coastal location and exposure to harsh marine environments increase the likelihood of both internal and external corrosion, creating a need for early detection to prevent leaks, unplanned shutdowns, and environmental impact. This case study presents the successful application of a permanently installed guided wave monitoring system on jetty pipelines to enable continuous condition assessment.

Using a cloud-based Monitoring Studio for automated data capture, analysis, and alarm management, the system provided ongoing measurement of wall-thickness combined with large area coverage and sensor health. After 1.5 years in service, the system detected a developing cross-sectional wall loss near a weld. Notification to the asset owner allowed follow-up visual and ultrasonic inspection which confirmed localized corrosion with approximately 1–3 mm of wall loss.

The results demonstrate that continuous guided wave monitoring improves the reliability of corrosion detection in remote, high-risk assets. Early indication of degradation enabled timely maintenance intervention and helped reduce inspection uncertainty. The case highlights the value of automated, data-driven NDT monitoring as a proactive integrity strategy for jetty pipelines and similar challenging-access assets.