Journal of Nondestructive Evaluation最新文献

A 3D-CNN-based acoustic pattern recognition model is developed for accurate detection of abrasion faults in wind turbine blades. The model first processes acoustic vibration signals through empirical mode decomposition and wavelet denoising to account for local signal characteristics. The denoised signals are then subjected to frame splitting, windowing, and discrete Fourier transform to construct two-dimensional energy spectrograms, which are subsequently downscaled using Mel-filter banks to extract distinctive acoustic features associated with blade abrasion faults. These features are input into an innovative three-dimensional convolutional neural network for fault identification. Experimental results demonstrate the model’s effectiveness, achieving a peak recognition accuracy of 98.8% and consistent performance with accuracy rates above 94% across tests. The model exhibits a reliable capability to distinguish between normal operational sounds and various severities of blade abrasion faults, while maintaining low misjudgment rates.

The acoustic signal feature extraction and intelligent diagnosis method for tile debonding are of great significance to ensure building safety. This paper presents a detection method based on wavelet transform and Convolutional Neural Network (CNN) integrated with Support Vector Machine (SVM) to solve the problem that the detection results of traditional defect detection methods based on frequency domain characteristics are unstable under environmental noise imbalance. In this study, the acoustic vibration signals polluted by real environmental noise were collected. The time–frequency diagrams were obtained using the complex Morlet wavelet transform, which captured the temporal and spectral variations of the acoustic signals. The image recognition method of CNN was enhanced through the integration of SVM, which replaces the softmax classifier with SVM. Four tile-wall specimens with different degrees of debonding were crafted, and the tiles were divided into nine different regions for tapping. Model training and prediction were conducted on the acoustic signals acquired from identical regions across the four specimens, which verified the reliable classification performance of this method. The average test accuracy of the nine regions reached over 98%, which provides a basis for the study of debonding quantification. Moreover, the traditional CNN was also employed for model analysis, and comparative result revealed that the proposed method demonstrates superiority in accuracy and efficiency. In the future, more groups of experiments on debonding area gradients could be conducted to research whether this method can accurately classify the degree of bonding defects when the obtained data set is large and sufficient.

Wind turbine blades are prone to various types of damage under long-term fatigue loading, making accurate damage type identification critical for structural health monitoring and operation and maintenance decision-making. This study proposes a hybrid algorithm framework that integrates SOM, Principal Component Analysis (PCA), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and Random Forest (RF) to identify damage in wind turbine blades based on acoustic emission (AE) signals collected during fatigue testing. Firstly, nonlinear and linear dimensionality reduction is performed on the raw AE features using SOM and PCA, respectively, resulting in six representative feature parameters. Then, DBSCAN is employed to cluster and label the reduced-dimension samples, enabling unsupervised signal classification without requiring prior knowledge. Based on the clustering results, a Random Forest model is trained and evaluated in a supervised manner, with classification accuracy, F1-score, and generalization performance quantitatively assessed. Experimental results show that the proposed method achieves over 90% accuracy in a four-class classification task, significantly outperforming traditional methods in both precision and stability. The clustering process exhibits strong robustness and is suitable for monitoring damage evolution at various stages of fatigue for the blade. This study provides an efficient and scalable signal processing approach for damage identification in composite wind turbine blades, laying a methodological foundation for intelligent and automated AE-based monitoring systems.

Electromagnetic tomography (EMT) is a promising nondestructive imaging technique for metallic defect detection. However, its inherent soft-field diffusion, rapid sensitivity decay, and nonlinear eddy-current interaction lead to severe boundary blurring and low-resolution reconstructions, particularly for small or multiple defects. To address these limitations, this study proposes Edge-DeepLabv3+, a physics-informed deep reconstruction network specifically designed for metallic EMT. The model integrates SE-Res2Net for multi-scale conductivity encoding, DenseASPP for dense receptive-field expansion under strong diffusion, ESA for capturing long-range electromagnetic correlations, and a Boundary-Refinement (BR) decoder to compensate for soft-field-induced edge loss. Furthermore, we introduce an Edge-Focused Hybrid Loss (EHL), which combines global MSE, imbalance-aware Dice loss, and a boundary-supervision BCE applied to morphology-derived defect contours, enabling precise recovery of high-frequency conductivity discontinuities. A physics-based dataset comprising 12,960 samples is generated using COMSOL, incorporating coil misalignment, temperature drift, non-white noise, and mutual-coupling perturbations through domain randomization, ensuring robustness against practical domain shifts. Extensive experiments on both simulation and real EMT systems demonstrate that Edge-DeepLabv3+ significantly improves reconstruction accuracy, boundary fidelity, and robustness to noise compared with LBP, Tikhonov, SE-Res2Net, and DeepLabv3+. The proposed model achieves accurate reconstruction of 3–6 mm single and multiple defects, even under low-SNR (10 dB) conditions, highlighting its strong potential for reliable online metallic defect monitoring in industrial environments.

Reliable detection of aero-engine blade surface defects is hindered by weak defect saliency, long-tailed category imbalance, and strong geometric priors from curved surfaces. This paper proposes MSFF-YOLO, a multi-scale feature fusion framework built upon YOLOv11. The method integrates a Multi-scale Efficient Aggregation Module (MEAM) for enhancing subtle and edge-attached defects, a multi-scale FMSIoU loss for improving regression robustness under long-tailed distributions, and a Manhattan Self-Attention (MaSA) mechanism for modeling curvature-related spatial dependencies. Evaluated on the high-resolution AeBSDD dataset, MSFF-YOLO achieves an mAP₅₀ of 89.1%, surpassing YOLOv11 especially on nick, bent, and dent defects. Real-world illumination-disturbance tests and zero-shot evaluation on NEU-DET further verify its strong cross-scene and cross-domain generalization, demonstrating its robustness for industrial blade inspection.

Detection of defects in diverse domains requires a specialized object detection system capable of identifying various types of flaws of different sizes and shapes. This research addresses detection challenges across six critical domains: saline bottle level monitoring, screw defect detection, magnetic tile inspection, road crack analysis, fabric flaw identification, and potato leaf disease recognition. These applications exhibit unique visual characteristics, including variable defect morphologies, subtle texture variations, and domain-specific features, which conventional detection models often fail to adequately address. Standard Faster R-CNN implementations with ResNet-50 and VGG-16 backbones offer general feature extraction but lack domain-specific optimization for these specialized applications. We propose LightDefectNet-18, a custom CNN backbone for Faster R-CNN featuring dual residual blocks with skip connections, strategic kernel sizing, and progressive channel expansion, integrated with a Feature Pyramid Network (FPN) architecture. The FPN component creates a multi-scale feature hierarchy through top-down pathways and lateral connections, effectively detecting defects across scales. The architecture incorporates batch normalization layers, calibrated dropout, and proper weight initialization to enhance feature preservation and gradient flow. When integrated with Faster R-CNN, we implement refined anchor configurations optimized for multi-scale defect detection across our target applications, with tailored anchor sizes and aspect ratios for each pyramid level. The detection pipeline employs an adaptive optimization strategy with learning rate scheduling and early stopping mechanisms. The quantitative evaluation demonstrates superior detection performance across all target applications compared to standard backbones, with significant improvements in Average Precision using a relaxed IoU threshold specifically calibrated for industrial defect detection scenarios. The model's FPN-enhanced architecture effectively addresses the challenges of capturing fine-grained visual features essential for distinguishing subtle anomalies at multiple scales in specialized materials while maintaining computational efficiency suitable for deployment in real-world industrial and agricultural monitoring systems, even with limited training data.

The prediction of fracture and progressive damage in solid materials can be assessed with acoustic emission (AE) monitoring, yet most quantitative AE-stress/strain relationships remain largely empirical. In this work, we present a unified stochastic framework that derives and generalizes these relationships from a microfailure distribution. By modeling the temporal occurrence of microfailures as a stochastic Poisson process, we establish direct statistical connections between AE activity and macroscopic mechanical variables such as stress and strain. This perspective allows us to reinterpret proposed empirical AE laws as particular statistical regimes, while also advancing a generalized formulation capable of capturing more complex behaviors often observed under dynamic loading. Extensive stochastic simulations further reveal that simple assumptions about internal deterioration rates naturally lead to heuristic quantitative laws for the number of AE events, thereby grounding empirical observations in probabilistic reasoning. The framework is validated against experimental datasets from cortical bone and collagenous soft tissue, confirming its robustness and predictive capacity. Beyond providing a rigorous theoretical foundation for empirical AE laws, our results demonstrate how microlevel statistical assumptions can explain macroscopic fracture signatures, offering new tools for structural health monitoring and prediction of fractures.

Steel defects can significantly diminish the corrosion resistance, wear resistance, and load-bearing capacity of the steel, leading to substantial financial losses. To effectively identify and locate steel surface defects, we propose a cross-scale feature fusion network. The process begins with the pre-processing of the input image through gray transformation and histogram equalization, followed by feature extraction using an enhanced backbone feature extraction network. Subsequently, a feature fusion network incorporating a gather-and-distribute (GD) structure is introduced to merge multi-scale feature maps, improving the robustness of information fusion across different scales. In the final stage, three detection heads of varying sizes undergo processing by a convolution module with a coordinate attention mechanism. The efficacy of the proposed method is validated using the Northeastern University surface defect database (NEU-DET) dataset, with experimental results demonstrating that the network achieves an 84.7% mean average precision (mAP) at an intersection over union (IoU) threshold of 0.5. Noteworthy contributors to the mAP of the proposed network include the image pre-processing module, the improved feature extraction network, the gather-and-distribute feature fusion network, and the detection network, contributing 4.1%, 2.9%, 2.6%, and 0.2%, respectively. The comparative experiments based on the attention mechanisms illustrate that the Squeeze-and-Excitation (SE) mechanism is the most suitable mechanism for the model proposed in this paper compared to other mainstream attention mechanisms. In comparison with other deep learning networks, our network demonstrates a significant enhancement in detection capability, showcasing superior performance in the identification of steel surface defects.

We present a note on the implementation and efficacy of a box-constrained (L_1/L_2) regularization in numerical optimization-based approaches to performing tomographic reconstruction from a single projection view. The constrained (L_1/L_2) minimization problem is constructed and solved using the Alternating Direction Method of Multipliers (ADMM). We include brief discussions of parameter selection, as well as detailed numerical comparisons with relevant alternative methods. In particular, we benchmark against a box-constrained TVmin and an unconstrained Filtered Backprojection in both cone-beam and parallel-beam (Abel) forward models. We consider both a fully synthetic benchmark and reconstructions from X-ray radiographic image data.