Journal of Nondestructive Evaluation最新文献

Instance segmentation of compound objects in XXL-CT imagery poses a unique challenge in non-destructive testing. This complexity arises from the lack of known reference segmentation labels, limited applicable segmentation tools, as well as partially degraded image quality. To asses recent advancements in the field of machine learning-based image segmentation, the ‘Instance Segmentation XXL-CT Challenge of a Historic Airplane’ was conducted. The challenge aimed to explore automatic or interactive instance segmentation methods for an efficient delineation of the different aircraft components, such as screws, rivets, metal sheets or pressure tubes. We report the organization and outcome of this challenge and describe the capabilities and limitations of the submitted segmentation methods.

In industry, rolling bearing damage acoustic emission (AE) signals are interfered with by complex transmission paths and strong noise. The signal-to-noise ratio of the AE signal is low. The bearing periodic fault pulse is weak, and fault feature extraction is challenging. To address these issues, combined with the characteristics of impulsiveness and rapid attention of the AE signal, an enhancement of the bearing weak fault signal based on the autocorrelation function (ACF) and improved multipoint optimal minimum entropy deconvolution adjusted (MOMEDA) method is proposed in this contribution. Firstly, in low signal-to-noise ratio, the target vector of the MOMEDA method is not optimal, and the diagnostic accuracy is low. To address this problem, this paper improves MOMEDA by using the gradient descent method, called GMOMEDA. Rolling bearing fault AE pulse signals are enhanced. Then, a method combination of ACF and GMOMEDA highlights the periodic elastic wave in the signal. Finally, the enhanced AE signal is processed by envelope demodulation to extract the frequency of the bearing fault signal. The experimental results show that the performance of the ACF-GMOMEDA method is better than the other five methods. The frequency features of bearing fault AE signal can be accurately extracted.

In this research, based on the weak nonlinear elasticity theory and finite deformation theory, a dynamic equation of guided waves in pre-stressed plates is established. Legendre orthogonal polynomials expansion method is employed to analytically solve wave equations. The dispersion curves and particle trajectory are obtained without root-finding algorithm. The solution is validated by comparing with the superposition of partial bulk wave method, and its convergence properties are analyzed. The effects of the pre-stress states on particle trajectory and ellipticity dispersion curves are investigated. Results confirm that the behavior of particle trajectory and ellipticity dispersion depends not only on the pre-stress states but also on frequency and mode. Next, ellipticity dispersion curve of fundamental Lamb modes with various stress state and propagation directions are calculated. Finally, the sensitivity of ellipticity as an indicator of stress is also analyzed. These results provide useful reference for the development of innovative nondestructive testing method for pre-stress states based on Lamb waves.

The system presented in this paper enables automatization of the two-dimensional calibration process (determination of Barkhausen noise (BN) intensity dependence on in-plane components of strain). Then, using dedicated software created by the authors in LabVIEW environment, and with the help of two dimensional calibration data one can effectively determine strain and stress distribution i.e. magnitude and orientation of main strain/stress components relative to measurement direction. BN signal measurements are performed using an advanced, multidirectional Barkhausen noise (BN) measuring sensor and a measurement system dedicated for cooperation with it. The system uses a robust algorithm for the strain components determination based on calibration surfaces, instead of usually applied curves, thus taking the influence of normal strain component directly into account instead of treating it as a correction factor (if not completely neglecting). The originality of the system arises also from the fact that this is the first BN measurement system that is self-calibrating (i.e. automatically loads the calibration sample in a pre-programmed way, performs BN signal measurements and calculates calibration planes), provided that the user possesses enough of the investigated material for calibration sample preparation.

Local damage or stress concentration that forms during manufacturing and long-term use of ferromagnetic materials has a direct impact on the safety of engineering structures. Thus, accurately identifying damage and stress conditions in these materials is crucial. In this study, martensitic stainless steel, a type of ferromagnetic material, is chosen as the subject for investigation. A weak magnetic detection device is engineered specifically for this purpose, and tests are conducted on the material using this device. The stress value of the material is determined using X-ray diffraction, while magnetic induction intensity is simultaneously recorded with a weak magnetic detection device along the same path. The stress value and magnetic induction intensity are normalized, and the results are analyzed to establish a correlation between weak magnetic signals and stress. Then, a signal processing technique combining blind source separation and eigenvalue recognition is introduced to achieve stress concentration and microdefect location identification. This method is based on the correlation analysis results between weak magnetic signals and stress, as well as supporting evidence from prior studies. The experimental results demonstrate that the location of stress concentration can be accurately determined by identifying the peak or valley value of weak magnetic signals, with an error range of less than 30%. The algorithm of blind source separation and eigenvalue recognition can pinpoint the location of stress concentration and microdefects from the obtained signal. This study presents a novel nondestructive testing method for stress concentration and microdefect identification in ferromagnetic materials.

Most industries dealing with pipelines face problems resulting from the buildup of deposits therein, such as reduced efficiency, downtime and increased maintenance costs. Although solutions to this issue have been sought for decades, no widely employed technique for monitoring growth of inorganic deposits (or ‘scaling’) in pipelines exists. In this research, a means of detecting the onset of scaling growth, by processing resonance testing data, was sought. For the resonance testing measurements the pipeline segment of interest is equipped with acceleration sensors which record signals generated by impacting the pipeline with a steel tip. The signals are Fourier transformed and analysed in the frequency domain, where a clear shift in frequency peak positions can be observed as the scaling thickness changes. How best to extract quantitative information from the generated frequency data is an open question. In this research, two data analysis approaches for scaling thickness prediction are compared: a supervised (binary classification) machine learning approach as well as a comparison-based change detection approach using cross-correlation. The supervised machine learning approach yields generalizable results for different acceleration sensors and impactor diameters whilst the change detection approach is sensitive from a scaling thickness of 0.5 mm. Whilst this research is specific to the pipe–scaling geometry—and type used in the experiments conducted, resonance testing can be applied to any pipe–scaling combination. The robustness of the data processing approaches presented in this work, when applied to other pipe–scaling materials and geometries, is the next point of research.

Ultrasonic testing is utilized to ensure weld quality during the fabrication of steel bridges by identifying discontinuities that are classified as either acceptable or rejectable. The classification of a discontinuity can be affected by differences in the acoustic properties of the material under test and the reference standard used for calibration. Differences in wave velocity affect the refracted angle and amplitude of refracted shear waves. As a result, indications can be missed or incorrectly classified, or incorrectly located in the material. The objective of this research study was to characterize the acoustic wave velocities in a sample of contemporary steels to better understand the range over which velocities may vary for common steels. To address this objective, a series of velocity measurements have been conducted for shear waves propagating through different directions in steel plates of different strengths and reported manufacturing processes. The study also examines the loss of signal amplitude that results from changes in the refracted angle of shear waves used for the inspection of welds. Beam splitting that may occur in anisotropic materials and the potential impact on signal amplitudes is also presented. It was shown in the research that relatively small differences in velocity between the material under test and the reference standard cause a loss of sensitivity of the test. Data presented in the paper documents wave velocity and anisotropic ratios for a population of contemporary bridge steels used for the fabrication of steel bridges and an assessment of how velocity differences affect the amplitude of reflected shear waves.

In the timber manufacturing sector, ensuring high-quality products is crucial, but conventional inspection methods often struggle to detect internal defects non-destructively. To tackle this challenge, an innovative approach has been proposed that integrates terahertz (THz) imaging with artificial intelligence (AI) algorithms. By harnessing the unique ability of THz radiation to penetrate timber wood and AI algorithms, defect classification, segmentation, and characterization can be made possible. Here, a custom-made convolutional neural network has been optimized for the classification of the defects in timber wood into four classes – no defect, knot, small knot, and decay, yielding a classification accuracy of over 96%. Further, the custom classification model has been extended for thicker wooden samples with internal hidden defects using transfer learning and has yielded a classification accuracy of over 93%. Following classification, a U-Net-based segmentation algorithm has been developed to delineate the defect boundaries in THz images accurately with a high dice coefficient of over 0.90. Further, a YOLO-based algorithm has been utilized to characterize the defects by localizing the position of the defect using bounding boxes with a high F1 score of over 0.97. An accurate prediction of the defect dimension has been demonstrated using this algorithm with a percentage error of less than 4% for all the types of defects in the timber wood. This advanced methodology, leveraging multiple AI algorithms on the THz images, significantly boosts the efficiency and accuracy of automatic defect identification and characterization, marking a transformative step forward in timber industry quality control processes.

Gearboxes, as critical components, often operate in demanding conditions, enduring constant exposure to variable loads and speeds. In the realm of condition monitoring, the dataset primarily comprises data from normal operating conditions, with significantly fewer instances of faulty conditions, resulting in imbalanced datasets. To address the challenges posed by this data disparity, researchers have proposed various solutions aimed at enhancing the performance of classification models. One such solution involves balancing the dataset before the training phase through oversampling techniques. In this study, we utilized the Sparse Autoencoder technique for data augmentation and employed Support Vector Machine (SVM) and Random Forest (RF) for classification. We conducted four experiments to evaluate the impact of data imbalance on classifier performance: (1) using the original dataset without data augmentation, (2) employing partial data augmentation, (3) applying full data augmentation, and (4) balancing the dataset while using Kernel Principal Component Analysis (KPCA) for dimensionality reduction. Our findings revealed that both algorithms achieved accuracies exceeding 90%, even when employing the original non-augmented data. When partial data augmentation was employed both algorithms were able to achieve accuracies beyond 98%. Full data augmentation yielded slightly better results compared to partial augmentation. After reducing dimensions from 18 to 11 using KPCA, both classifiers maintained robust performance. SVM achieved an overall accuracy of 98.72%, while RF achieved 96.06% accuracy.