Journal of Nondestructive Evaluation最新文献

Rapid and reliable automated identification of wood species can be a boon for applications across wood scientific context including forestry and biodiversity conservation, as well as in an industrial context via requirements for timber trade regulations. However, robust machine learning classifiers must be properly analyzed and immunized against domain shift effects. These can degrade the automated system performance for input data variations occurring in many real-world scenarios. This work methodologically analyses the domain shift generated by using two differing sub-micro-scale and micro-scale computed tomography setups in the focused context of deep learning based binary wood classification from volumetric image data. To counteract this, we examine several mitigation strategies and propose primary data-level and narrow model-level strategies to effectively and successfully minimize the performance domain gap. Core elements of the data-wise strategy include the combined usage of phase-correction methods, low-pass pyramid representation of the data and adjustments of model normalization and regularization. Vanishing domain performance differences led to the conclusion that the combined strategy ultimately prompted the model to learn robust features. These features are discriminative for the utilized wood species data from both sub-micro-system and micro-system domains, despite the substantial differences in data acquisition setup that propagate into fundamental image quality metrics like resolution, contrast and signal-to-noise ratio.

The insulation of aviation cables is critical to aircraft safety but is vulnerable to defects such as cracks, ruptures, slices, and swelling. Reliable nondestructive testing (NDT) of these defects is challenging due to environmental interference, noise, and the limitations of existing inspection techniques. This work presents a novel NDT approach integrating reflective digital in-line holography with a Combined Anisotropic Total Variation (CATV) reconstruction algorithm and an Xception-based deep transfer learning model. The CATV reconstruction suppresses twin-image artifacts and preserves structural detail, enabling the generation of a phase-map dataset of multiple defect types. Using this dataset, the Xception-based classifier achieved 98% accuracy, surpassing state-of-the-art approaches. The contributions of this work are: (i) using CATV-based reconstruction for reflective holography of aviation cables, (ii) creating a phase-map dataset of insulation defects, and (iii) demonstrating the feasibility of a high-precision, non-contact inspection method for aviation safety applications.

A clinch point’s quality is usually assessed using ex situ destructive testing methods. These, however, are unable to detect phenomena immediately during the joining process. For instance, elastic deformations reverse and cracks close after unloading. In situ methods such as the force-displacement evaluation are used to investigate a clinching process, though deviations in the clinch point geometry cannot be derived with this method. To overcome these limitations, the clinching process can be investigated using in situ computed tomography (in situ CT). When investigating the clinching of aluminum parts in in situ CT, the sheet-sheet interface is hardly visible. Earlier investigations showed that radiopaque materials can be applied between the joining parts to enhance the detectability of the sheet-sheet interface. However, the layers cause strong artefacts, break during the clinching process or change the clinch joint’s properties significantly. In this paper, a minimally invasive method to enhance the interface detectability is presented. First, the aluminum oxide layer is removed by etching. Second, the specimen is electroplated with copper or gold, respectively. In some cases, a mask is applied to create a cross-shaped plating pattern. Then, the plated specimen is clinched with a non-plated counterpart and the interface detectability of the clinch points is assessed in CT scans. It is shown that a copper plating of 2.6–4 μm can visualize some parts of the interface, while 7–9 μm is suitable to enhance the detectability of the sheet-sheet interface almost continuously.

This study presents a novel vision-based methodology for damage detection in CFRP composite beams, combining optical flow analysis, statistical anomaly scoring, and deep learning (DL) models. Composite materials such as CFRP are widely used in structural applications due to their high strength-to-weight ratio, yet detecting internal damage remains a significant challenge. To address the limitations of traditional non-destructive evaluation methods, this study integrates non-contact optical flow techniques with a hybrid anomaly detection pipeline. The Lucas-Kanade optical flow method is used to extract displacement time series from video recordings of vibrating structures. These displacement signals are transformed into spectrograms using Short-Time Fourier Transform (STFT), and frequency-domain features are enhanced with added Gaussian noise to improve model robustness. Principal Component Analysis (PCA) is applied to reduce the dimensionality of the spectrogram features, and Mahalanobis Distance is computed to quantify deviations from the healthy state. The resulting Mahalanobis Distance time series is then used as input for three DL architectures—Autoencoder, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM)—which are trained to detect structural anomalies based on reconstruction error or pattern recognition. The proposed approach is experimentally validated on CFRP composite beams under multiple damage scenarios. Results show that leveraging Mahalanobis-based statistical features within DL models significantly improves anomaly detection accuracy, offering a robust and scalable framework for real-time structural health monitoring in civil, aerospace, and automotive domains.

Accurately picking acoustic emission (AE) arrival times remains a significant challenge, particularly for low signal-to-noise ratio (SNR) signals where manual picking is subjective and unreliable. This article introduces an improved manual picking method for AE arrival times, developed by integrating sensor acquisition principles with wave velocity attenuation laws. This method provides a derivation formula that enables the determination of “ground truth” arrival times for low SNR signals by leveraging characteristics from high SNR signals. These derived values serve as labels to train a two-dimensional convolutional neural network (2D CNN) for automated arrival time picking. A key innovation is converting the one-dimensional AE signal directly into a two-dimensional matrix using a transformation matrix as the CNN’s input, thereby significantly streamlining preprocessing by eliminating the need for additional feature extraction. The labeled 2D matrices are then fed into the 2D CNN for training to enhance its ability to recognize crucial temporal patterns. Finally, the AIC algorithm picks the arrival times picked from the CNN-processed signals. A major advantage of CNNs in this context is that it does not require additional feature extraction and can extract features from the original elements. In addition, it can identify high-order statistics and nonlinear correlations of images. The third convolutional neuron can process data in its receptive domain or restricted subregion, reducing the need for a large number of neurons with large input sizes and enabling the network to be trained more deeply with fewer parameters. Results demonstrate that the proposed method significantly outperforms mainstream detection methods, including AIC and Floating Threshold (FT), achieving high accuracy and stability, particularly in scenarios with limited data and low SNR.

X-Ray Computed tomography (XCT) has become an important non-destructive quality assurance technique in industry. Consequently, standards for quality insurance of XCT and on its performance are required to support industrial XCT users for reliable production. This performance is determined by analysis of the quality of the images produced and by the dimensional measurement accuracy achieved for a given XCT parameter setting. Until recently, standards assessed image quality solely in terms of contrast sensitivity and spatial resolution. Detection limits could not be predicted until now. A new term is introduced: The Detail Detection Sensitivity (DDS). It depends on the contrast sensitivity as a function of contrast and noise, and on the spatial resolution. The spatial frequency needs to be implemented into the analysis to consider sensitivity as a function of the size of an indication. The contrast sensitivity is quantified by the Contrast Discrimination Function (CDF) and the spatial resolution by the Modulation Transfer Function (MTF). The numerical DDS is determined for air flaws from the Contrast Detection Diagram (CDD) at 100% contrast. However, some XCT operators prefer visual determinations rather than numerical ones. To face this need, the SensMonCTII project proposes a new Image Quality Indicator (IQI), consisting of a disk with holes of different sizes for visual DDS determination. The project aims to produce a new ISO standard draft providing a practice to evaluate numerically the XCT image quality via MTF, CDF, CDD and DDS, as well as to evaluate visually the DDS from the hole visibility of a disk IQI. The paper does not address the performance of XCT in terms of dimensional measurement accuracy, but focuses on the performance of XCT in terms of image quality. It describes the methodology to evaluate the image quality, including DDS for the first time.

This paper introduces a deep learning (DL)-enhanced X-ray computed tomography (CT) approach for detection of defects in composite structures. While X-ray CT offers high-fidelity defect detection, test specimen size limitations restrict its application to large aerospace components. Inclined CT (ICT) addresses these size constraints by keeping X-ray source and detector on the different sides of a stationary test specimen. This system geometry results in a limited angular data 3D reconstructions that produce significant artifacts that may represent defects incorrectly. This research demonstrates that DL techniques, particularly the fine-tuned Segment Anything Model (SAM), can improve defect detection from ICT data. Methodology employs fine-tuning of SAM with a dataset of 1,800 images across ten synthetic phantoms with varying defect sizes and locations. The fine-tuned model was validated on an as-built aluminum test specimen, achieving over 70% accuracy in defect detection and 98% accuracy in overall shape detection. Validation with carbon fiber reinforced polymer specimens containing Teflon inserts yielded improved results compared to ICT reconstruction methods, indicating practical applicability. The findings suggest that DL-enhanced ICT can offer detection capabilities comparable to full CT while preserving the large-structure compatibility of ICT, making it a viable non-destructive inspection method for aerospace industry applications.

Preserving aging bridges, which are vital to transportation networks, presents notable difficulties due to factors like intense usage, structural wear, and restricted maintenance resources. This research examines the deployment of Nondestructive Evaluation (NDE) techniques to optimize bridge maintenance strategies and maintain structural soundness. Over the course of infrastructure lifespans, vast amounts of NDE data are accumulated, yet processing and interpreting this information proves challenging due to intricate spatial and temporal interdependencies. In this study, we approach the problem as one of graph-based prediction, introducing two advanced methodologies to address it. The primary approach utilizes a Temporal Graph Convolution Network (TGCN), harnessing spatio-temporal patterns for predictive modeling. The secondary approach, a multi-modal TGCN, integrates data fusion techniques to combine diverse data sources for improved predictive accuracy. We evaluate the performance of these approaches using NDE data collected at Rutgers’ BEAST^® facility that includes five NDE modalities and 14 consecutive time intervals for assessing bridge deck conditions, comparing the results against a baseline Spatio-Temporal Autoregressive (STAR) model. While the STAR model established foundational forecasts, the TGCN method achieved superior results by managing nonlinearities. The multi-modal TGCN further enhanced performance, demonstrating the advantages of leveraging data fusion to incorporate multiple data types within TGCN frameworks.

Solid phase processing, such as friction stir processing, is an advanced manufacturing method that often results in ultrafine grain sizes and superior mechanical properties. The motivation of this study was to demonstrate ultrasonic testing as a nondestructive evaluation method to complement traditional destructive methods for characterizing material microstructure, with an emphasis on grain size determination using a method that may have future applications for real-time inline process monitoring and product validation. The method for measuring grain sizes of polycrystalline metals after solid phase processing was established using ultrasonic shear wave backscattering, building on prior studies on coarse-grained materials. The work involved measuring ultrasonic backscattering for a series of 316L stainless steel specimens with various grain sizes made by friction stir processing, calculating ultrasonic backscattering coefficients from experimental data based on a physical measurement model, measuring ground truth grain sizes of the specimens from electron backscatter diffraction grain boundary images, and building a correlation of ultrasonic backscattering coefficients versus the ground truth grain sizes. The grain sizes of a set of blind test specimens were successfully determined based on the correlation. This work successfully demonstrates the viability of an ultrasonic nondestructive evaluation method for microstructural characterization of material having ultrafine grain structure, as produced by an advanced manufacturing method.

In this study, the feasibility of a machine learning model for the automatic classification of impact-echo testing results was investigated. A machine learning model with features such as instantaneous frequency and spectral entropy extracted from time series data was compared with two different approaches, including conventional peak frequency and a deep learning model. To construct a robust and flexible model, an open-source database from two organizations performed by different testing operators and equipment was used to train and develop the universal classifier. The model was evaluated for its ability to classify the type of defects as well as their presence, and the results showed that shallow delamination can be detected more accurately than other types of defects. The proposed machine learning model showed reliable and promising results and has the potential to improve the efficiency of impact-echo testing in concrete structures.