Journal of Nondestructive Evaluation最新文献

Veneer constitutes a critical structural component in engineered wood products such as plywood, laminated veneer lumber (LVL), and cross-laminated timber (CLT), which are extensively utilized in modern civil engineering infrastructure. However, the inherent heterogeneity and rapid growth of raw timber frequently introduce structural defects, including knots, cracks, and voids, that severely compromise the mechanical integrity and load-bearing capacity of the final assembly. Traditional non-destructive evaluation (NDE) methods and standard convolutional neural networks (CNNs) often fail to distinguish these structural anomalies from complex natural grain patterns, leading to potential safety risks in construction applications. To address these challenges, this paper proposes a material-property-aware feature co-evolution network named Veneer-AEPNet (VAEPNet), specifically engineered for the high-precision non-destructive grading of structural veneers. The architecture explicitly integrates a dual-stage Transformer mechanism to establish a global texture baseline, effectively identifying macro-scale topological deformations. Simultaneously, we introduce the Adaptive Anisotropic Granular Pooling (AAGP) module to capture directional fiber perturbations and the Global-Contrastive Feature Refinement (GCFR) module to suppress natural background noise by modeling deviations from a global texture reference. The methodology was validated on a large-scale industrial dataset containing 25,000 high-resolution images. Experimental results demonstrate that VAEPNet achieves a test accuracy of 94.47% and an F1-score of 93.91%, achieved the best overall performance among the evaluated baseline models. Furthermore, visualization analysis confirms that the model effectively decouples structural defects from natural wood features, providing a robust solution for automated quality control in NDE 4.0 environments.

Railway axles, though designed to last over 20 years, are fatigue-stressed components that may suffer premature failures with severe consequences. Ultrasonic techniques are among the most effective inspection methods, but their limitations–such as the need for physical coupling with the axle and long inspection times–have encouraged the exploration of innovative, contactless solutions. This paper presents the setup and analysis of ultrasonic data collected using Gas-Coupled Laser Acoustic Detection (GCLAD), while employing a pulsed laser as the emitting source. GCLAD is a contactless technology that enables point detection of waves in transmission and reflection at once, offering multiple opportunities to identify defects and discontinuities. The present preliminary study demonstrates the feasibility of inspecting railway axles without disassembling wheelset components, through a fully optical system that acquires and detects ultrasonic waves without contact. This approach maximizes automation and reduces inspection-related times. The devised experimental layout successfully identified defects in the wheel seat area (down to 2.7 mm), brake seat area (7.5 mm), and axle body (0.95 mm), even with mounted wheels. An industrial application highlights the potential suitability of GCLAD-based instrumentation for integration into actual inspection apparatus, supporting efficient monitoring of railway axles while paving the way for improved automation and reduced inspection times.

The curvature characteristics of aero-engine turbine blades introduce significant interference in eddy current testing signals, readily causing waveform distortion and thereby hindering quantitative defect assessment. To address this issue, a physics-informed random forest compensation method is proposed, and the signal distortion mechanism induced by curvature variation is systematically investigated. Considering that non-differentiable models such as random forests cannot directly incorporate physical priors, a physics information embedding mechanism based on iterative target-value correction is developed. A curvature-domain Laplace operator is employed to construct the physical residual, which is subsequently fed back to the training labels to dynamically refine the prediction targets. To validate the effectiveness of the proposed approach, comparative scanning simulations on curved and flat surfaces as well as experimental verification using Inconel 718 specimens with artificial defects were conducted. Experimental results demonstrate that, compared with the standard random forest model, the improved method reduces the root-mean-square error (RMSE) for flat reference signal compensation under varying curvatures by an average of 72.4% and decreases the standard deviation (STD) by 18.9%. Under the extreme curvature condition of 5 mm, the RMSE improvement reaches 75.5%. In experimental validation, the RMSE values of the compensated signals for different curvatures are reduced by 18% and 96.7%, respectively, while the STD values are reduced by 14% and 18.9%. The compensated signals obtained by the proposed algorithm exhibit substantially higher fidelity to the flat reference signal, effectively mitigating distortion phenomena and facilitating quantitative defect signal analysis. The algorithm demonstrates pronounced physical robustness in curved-surface eddy current signal compensation, providing an effective technical pathway for the correction of eddy current signals acquired on curved geometries.

Pulsed eddy current (PEC) testing is a key non-destructive evaluation technique, but conventional FFT-based analytical modeling is computationally expensive due to full-spectrum impedance change evaluations up to the Nyquist limit. This paper proposes a novel modeling method using the Morlet wavelet transform. The method establishes a direct mapping between the wavelet scale factor and angular frequency, enabling time-domain reconstruction from discrete scale-dependent frequencies without global FFT sampling. An adaptive scale-truncation strategy further reduces computational cost by omitting negligible high-frequency components. Numerical results demonstrate that, with the optimal parameters (wavenumber 25 and discretization parameter 0.1), the proposed method requires only 96 frequency points, as opposed to 1000 at a sampling frequency of 1 MHz. This translates to a computational saving of over 90%, while maintaining relative errors below 1% at signal peaks. The method's accuracy and robustness are validated across wide parameter ranges and against experimental data. This work offers an efficient, theoretically sound alternative to FFT-based modeling, with potential for real-time inverse applications.

Steel plate defect inspection based on electromagnetic tomography (EMT) has attracted increasing attention because of its ability to perform non-invasive imaging without direct contact with the inspected object. Nevertheless, accurate image reconstruction remains difficult owing to the highly ill-conditioned inverse mapping and the diffuse nature of electromagnetic fields, which often result in distorted defect boundaries and unstable reconstructions. Although deep neural networks have demonstrated considerable potential in EMT imaging, their predictions may deviate from underlying physical principles when trained solely from data. To improve both reconstruction accuracy and physical reliability, a physics-guided image reconstruction framework named PI-TransUNet is developed. The framework integrates a TransUNet-based inverse solver with a pretrained differentiable forward surrogate model, forming a closed-loop learning architecture. Spatial positional information is incorporated into the measurement sequence to strengthen topological awareness, while a physics-consistency objective is introduced to regularize the reconstruction process. Furthermore, a progressive weighting strategy is employed to stabilize optimization and facilitate the incorporation of physical constraints. Numerical simulations and laboratory experiments conducted on a nine-coil planar EMT platform demonstrate that the proposed approach achieves superior defect localization, enhanced boundary preservation, and improved robustness against measurement noise. Quantitative results indicate consistently high structural similarity and low reconstruction error, highlighting the effectiveness of PI-TransUNet for EMT-based steel plate defect imaging.

Monitoring the quality of frozen meat is crucial for ensuring food safety. This study aimed to evaluate multiple quality indicators of frozen-thawed beef using hyperspectral imaging (HSI), including pH, total volatile basic nitrogen (TVB-N), thiobarbituric acid reactive substances (TBARS) and carbonyl content. Conventional hyperspectral analysis methods rely on complex manual feature engineering and require separate model training for different attributes, which is cumbersome and inefficient. To address these limitations, this study proposed a hybrid deep learning framework integrating computer vision processing, multi-task learning and hyperparameter adaptive optimization, enabling simultaneous assessment of multiple quality indicators. Comparative experiments demonstrated that the proposed method achieved optimal predictive performance for pH, TVB-N, TBARS and carbonyl content, with all coefficient of determination for prediction (R²_P) values exceeding 0.91 and ratio of performance deviation (RPD) values greater than 3.3. By applying the established model to each pixel of the hyperspectral images, chemical distribution maps for the four indicators were generated, providing a comprehensive and visual assessment of beef quality. The results indicated the great potential of integrating HSI with deep learning for simultaneously evaluating multiple quality indicators in frozen-thawed beef.

This study investigates the evolution of fatigue-induced damage in AA5083-H112 aluminum alloy under low-cycle fatigue using linear and nonlinear ultrasonic techniques in conjunction with X-ray diffraction (XRD) and optical microscopy. Three independent tensile specimens and three independent fatigue specimens were tested to verify the repeatability of the experimental results. Residual stress and full width at half maximum (FWHM) gradually increased during the early fatigue stage because of dislocation accumulation, followed by a decrease as fatigue damage progressed. Acoustic velocity exhibited significant changes during the early fatigue stage, whereas the root mean square (RMS) amplitude and the time of flight (TOF) varied predominantly during the later stages of fatigue. Optical microscopy revealed potential microcrack initiation sites at approximately 2200 fatigue cycles, followed by clearly identifiable microcracks at 2700 cycles. During this transition, the nonlinear ultrasonic SPC-I exhibited a pronounced increase during the transition from potential microcrack initiation to visible microcrack formation, indicating high sensitivity to fatigue damage evolution prior to the development of visible surface cracks. The combined results demonstrate that the integration of SPC-I with acoustic velocity, TOF, residual stress, FWHM, and optical microscopy provides a reliable and complementary approach for monitoring fatigue damage evolution, particularly during the transition from microstructural damage accumulation to microcrack formation.

CO₂ pipelines are critical for carbon capture, utilisation, and storage (CCUS) and industrial applications. Gas leaks from these pipelines can pose a significant safety hazard, impact the environment, and cause economic losses. When leakage occurs from the high-pressure pipeline, it creates stress waves due to localised energy release, which can be detected by using acoustic emission (AE) sensors. There have been considerable attempts made in literature to monitor leaks and perform source localisation of leaks in CO₂ pipelines using AE, as this technique is non-invasive and can be monitored in real-time. Most of these studies were performed using a straight CO₂ pipeline. However, most of the pipelines using CO₂ transportation will have bends, joints, and valves, which may impact the strength of the AE signal. Also, the economic viability of using the AE for commercial implementation for pipeline monitoring in industries is not well understood. Therefore, to address these knowledge gaps, this paper focuses on understanding the economic viability through an experimental study. Experiments were conducted on a six-meter pipeline with joints, bends and valves. Static test experiments were conducted under different flow coefficients. Also, attempts were made in the study to identify the source localisation of the leak by adjusting the sensor positions. From the experiments conducted, the minimum leak that could be identified from the valve was 0.5 L/min and from the pipeline was 42 L/min. It was also possible to identify the leak location with good accuracy. From the study conducted and based on international prices of CO₂ gas in December 2025, we observed that the industry could save at least $225,482 USD annually by monitoring leaks from a single valve and approximately 18.9 million USD by monitoring leaks from pipelines using acoustic emission. This study lays a strong basis for industries to further investigate the commercial implementation of AE sensors for CO₂ pipelines.

High-voltage cable defect detection is an important part of urban power system maintenance. Due to variations in cable thickness, large weight, and challenging operability, existing signal-processing and artificial-intelligence-based methods may face trade-offs among detection accuracy, cost, efficiency, and applicability. Moreover, some existing methods are invasive in nature. Therefore, this paper proposes an improved multiscale attention-based YOLOv8 model, designs a portable X-ray non-destructive testing (NDT) system, and creates a larger dataset based on the CycleGAN-based (Cycle Consistent Adversarial Networks-based) Pix2Pix (Pixel-to-Pixel Translation) model. First, the improved YOLOv8 model introduces EfficientNet to replace CSPNet (Cross Stage Partial Network) as the backbone network and incorporates a Squeeze-and-Excitation (SE) module after each convolutional block. Second, a bottom-up feature pyramid PAN (Path Aggregation Network) is integrated with the FPN (Feature Pyramid Network) to construct a hybrid multi-scale feature fusion structure, enabling multi-level and multi-directional information transfer. Moreover, a task-specific box refinement strategy combining K-means initialization with IoU-guided matching is employed to improve localization accuracy. The experimental results show that, compared with the other detection models, EMA-YOLOv8 improves the average precision and mAP50 by 4.68 and 3.9 percentage points, respectively, while reducing the average number of parameters and GFLOPs by 35.37% and 27.76%, respectively. Therefore, the proposed model has great potential.

This study addresses the dynamic identification of typical faults, such as rotor cracks and reversed installation of blades, in automotive power steering pumps by proposing a method for fault feature extraction and classification using multi-sensor acoustic emission (AE) signals. Initially, two AE sensors are positioned near the pump's inlet port and return port to synchronously capture AE signal time-domain waveforms under various working conditions, using a sampling frequency of 1 MHz. AE sub-signals for a single period are then extracted. After whitening, signals from the four sensors were arranged as column vectors to form a matrix. The singular value decomposition (SVD) was applied to extract four orthogonal mode vectors, which were then decomposed via three-layer wavelet packet decomposition (WPD). The frequency-domain fractal dimension (FFD) of the first four nodes is calculated. The random forest was used to assess feature importance and select the top six parameters. These features were fed into a support vector machine (SVM), and the performance of different kernel functions was compared. Experiments were conducted under pump body pressures of 4 MPa, 7.5 MPa, and 13 MPa, with a rotational speed of 1000 r/min. Results showed minimal differences among kernel functions, though radial basis function (RBF) performed best for the normal state. By incorporating SVD for multi-sensor information fusion and feature extraction, the classification accuracy of rotor states including normal, cracked, and reversed blade installation is improved from 87.77% to 97.55%, demonstrating the method's effectiveness and practicality.