Rolling bearings are key components of rotating machinery and their faults are a primary cause of industrial equipment shutdowns, with particularly significant impact in high-value fields such as wind power and rail transportation. Traditional methods relying on manual features lead to poor robustness, while deep learning methods exhibit weak generalisation under noise and complex operating conditions. Existing fusion approaches still suffer from feature redundancy and insufficient fusion due to inadequate modal feature concatenation and cross-modal interaction. This paper proposes a hybrid model integrating variational mode decomposition (VMD)-based multi-scale decomposition and cross-modal attention mechanisms (cross-attention). Using a Transformer-bidirectional long short-term memory (Transformer-BiLSTM) dual-channel architecture to extract spatiotemporal features, the cross-attention dynamically aligns and fuses modal and deep features, significantly improving the robustness and adaptability of bearing fault diagnosis under noisy and varying conditions. The experimental results show that, on the Case Western Reserve University (CWRU) public bearing dataset, this model demonstrates excellent diagnostic performance in both strong noise backgrounds and complex variable load conditions, with an average recognition accuracy of over 99%. Under extreme noise, it exhibits outstanding stability and generalisation; ablation experiments confirm that removing VMD or cross-attention leads to a 1.27% to 2.28% drop in performance. In multi-condition tests (0 HP to 3 HP loads), it maintains an accuracy of 99.02% to 99.15%, with precision, recall and F1 scores all stable above 0.99, confirming strong adaptability to complex conditions and providing a robustness solution for intelligent operation and maintenance of industrial equipment.
To address the signal saturation and poor quantification capability of conventional eddy current (EC) probes in deep crack detection, caused by the skin effect, this paper proposes a novel EC probe for inspection of deep cracks in thick-walled structures. The probe adopts a transmit-receive configuration with the transmitting coil (Tx) and the receiving coil (Rx) on opposing faces of the plate, enabling through-thickness magnetic field penetration, fundamentally altering the eddy current distribution and enhancing sensitivity to deep cracks. Theoretical analysis and finite element simulations demonstrate that this configuration effectively mitigates the dominance of surface currents, making the received signal primarily responsive to subsurface disturbances. Notably, the probe exhibits a highly linear relationship between normalised impedance variation and crack depth over a range of 6-32 mm (R 2 = 0.999), in stark contrast to the saturated non-linear response of conventional reflection-type probes. Leveraging this linear depth response, a direct inversion model is established for high-accuracy depth quantification, achieving a root mean square error (RMSE) of 0.23 mm. The results validate that the proposed transmit-receive probe overcomes the depth-sensing limitations of traditional EC testing and provides a reliable solution for quantitative evaluation of deeply buried cracks and surface-breaking deep cracks in thick-walled conductive structures.
In this paper, the non-destructive method of shearography is used to identify intentionally introduced defects in composite patches applied to repair fluid-carrying steel pipes. Composite repairs offer a practical and cost-effective alternative for restoring the integrity of fluid-carrying pipelines and their performance depends strongly on the manufacturing method. To evaluate this influence from a non-destructive testing (NDT) perspective, three common fabrication methods, namely hand lay-up with roller, filament winding and vacuum bag injection, are used to produce composite-reinforced steel pipe specimens. Two defect types, fibre-pipe debonding and interlayer delamination, are incorporated in controlled sizes and depths. The specimens are inspected using digital shearography under thermal loading. The results show that vacuum bag injection produces coatings with superior uniformity and allows all defects to be clearly identified. In contrast, the other two methods exhibit manufacturing-related artefacts that hinder defect detection. Overall, the findings demonstrate that the quality of the composite manufacturing process significantly affects the reliability of shearographic inspection for repaired pipes.
High-strength X80 steel is extensively used in modern long-distance oil and gas pipelines. Prolonged exposure to harsh environments leads to material loss, primarily through corrosion under insulation or at coating defects. This wall thinning compromises structural integrity and can lead to catastrophic failure. Conventional piezoelectric ultrasonic testing (UT) is widely employed for in-service pipeline inspection; however, its performance can be limited under high-temperature, rough or coated surface conditions, where maintaining consistent acoustic coupling with liquid couplants is challenging. In contrast, electromagnetic acoustic transducers (EMATs) enable non-contact operation and eliminate the need for surface preparation. However, most existing bulk-wave EMATs are limited to exciting a single ultrasonic wave mode, either longitudinal waves or transverse waves, which restricts their defect characterisation ability. To address this issue, a bimodal EMAT is developed that integrates a U-shaped permanent magnetiser with spatially distributed coils within a closed magnetic circuit. The design employs a common magnetic field with tailored components to concurrently generate and receive both longitudinal and transverse waves. A finite element model (FEM) is developed to assess the influence of defects in X80 steel plates on EMAT signals. Experimental results demonstrate that the bimodal EMAT can simultaneously excite and receive the two wave types, enabling accurate defect-depth quantification. By cross-validating the measurements derived from each wave mode, potential misinterpretations of defect presence or absence are effectively mitigated.
Weak alternating current (AC) has advantages such as reducing interference, energy conservation and higher levels of safety. Firstly, a compressive stress-magnetomechanical coupling model was established based on the Jiles-Atherton (J-A) model. Subsequently, the established model was subjected to simulation analysis using MATLAB and COMSOL software. Finally, an experimental platform was built to conduct experimental research, thereby validating the theory and simulation results. The results indicate that as the compressive stress increases, the magnetic signal first increases and then decreases. Additionally, under a constant compressive stress, the magnetic signal increases with the increase in the magnetic field. This study provides a theoretical basis and experimental reference for weak AC compressive stress detection methods.
In modern industry, metal materials are widely applied but are highly susceptible to corrosion and existing detection methods have limitations. This paper proposes evaluating the diameter, depth and hidden depth of corrosion pits based on the interaction mechanism between the electromagnetic field and the corrosion pits, utilising the multi-dimensional characteristics of the transient magnetic field. The transient magnetic field is analysed from different perspectives and dimensions: the surface area is represented by two-dimensional planar image features, enabling intuitive identification of the size of surface corrosion pits but showing insensitivity to vertical characteristics. The analysis of the linear area reveals that although it is less intuitive than two-dimensional planar images, it can detect the radius, depth and hidden depth of corrosion pits with rich feature information by effectively extracting the one-dimensional magnetic field distribution. After improving the structure of the excitation coil, the coil area can comprehensively detect corrosion pit parameters; however, the curve changes are not intuitive and there is mutual coupling among various dimensional parameters. This study provides accurate and reliable technical support for non-destructive testing of corrosion pits, ensuring equipment safety. The conclusions indicate that this method, which combines the analysis of different areas, has its own advantages and disadvantages, providing an effective approach for the quantitative non-destructive testing of complete parameters of corrosion pits, enriching the theoretical system of electromagnetic testing and offering new ideas for practical engineering applications.
In order to accurately monitor the cross-sectional deformation of an underwater tunnel of a reservoir without damaging its structure, research has been conducted on the application of non-destructive testing technology in the measurement of underwater tunnels in reservoirs. A three-dimensional laser scanner is employed to acquire point cloud data from the underwater tunnel. Subsequently, this data, along with control point data, is imported into the Leica Cyclone software. After performing relative coordinate stitching and absolute coordinate stitching, statistical filtering is applied to remove large-scale noise points that are far from the main body of the point cloud. By extracting curvature geometric features, noise points near the main body of the point cloud are filtered out. After determining the central axis of the tunnel using an ellipse-fitting-based central axis extraction method, the cross-sections of the underwater tunnel are fitted and extracted with this central axis as the reference. Then, through the model subtraction method for overall deformation analysis, multi-period encapsulated surfaces of the tunnel are constructed. By comparing these surfaces, deformation measurement of the cross-sections of the underwater tunnel is achieved. The experimental results indicate that this technology can effectively filter out outlier noise and noise near the main body while preserving the features of the main point cloud. The extraction results for the central axis of the tunnel, as well as the major and minor semi-axes of the cross-sections, show small errors compared with actual measurements obtained using a total station. The maximum error for the major and minor semi-axes does not exceed 1 mm and the minimum error is only around 0.3 mm. This technology enables precise monitoring of tunnel deformation.
In structural health monitoring (SHM), recent advancements in computer vision measurement techniques have superseded the constraints imposed by conventional sensors. This paper introduces a multi-point displacement measurement method that integrates the Hough transform with the optical flow method. The technique detects circular markers, extracts the centre of the circle using the Hough transform and utilises optical flow to track the centre of the circle, enabling multi-point displacement measurement. This approach is not only applicable to translational targets but also to rotational targets. To enhance marker recognition and tracking accuracy, super-resolution generative adversarial network (SRGAN) preprocessing is introduced to improve image details. The maximum vibration error during the shaker test was 0.09 mm. The experimental results demonstrate that the proposed method achieves high accuracy and stability under different vibration frequencies, rendering it suitable for non-contact dynamic monitoring of complex structures.
Given the complexities of friction stir welding (FSW), traditional monitoring methods often fail to provide comprehensive real-time insights. By leveraging advanced digital twin technology, this study proposes an innovative system tailored for FSW. This system is capable of real-time monitoring of the temperature and axial force, as well as predicting the tensile strength and microhardness of the joint. By establishing a prediction model for joint mechanical properties based on multi-source data fusion techniques, the system achieves real-time processing and analysis of force and thermal data, enabling online prediction ofjoint mechanical properties. The integration of the prediction model, MySQL database, 3D visualisation and virtual-real data interaction significantly enhances the capability of the system for dynamic evaluation of physical quantities and joint mechanical properties in real time. The validity of the proposed method and the developed system is verified by experiments. This research provides valuable insights for optimising FSW process control. This workadvances digital twin applications in solid-state welding by bridging the gap between multi-physics monitoring and real-time quality prediction.
In order to solve the problem that ultra-8-bit high-greyscaleX-ray film taken bya charge-coupled device (CCD) line scan camera cannot be directly visualised, an adaptive enhancement algorithm based on brightness gain compensation is proposed. Firstly, in order to solve the problem that a 12-bit image cannot be displayed, the prior eigenvalues of the image are extracted and a greyscale correction factor based on a feature prior coefficient is constructed. The greyscale distribution characteristics of the image are assessed by calculating its skewness to determine exposure levels, leading to the construction ofa non-linear global mapping function for visualising 12-bit high-greyscale images. To address the significant information loss in high-greyscale images processed by existing pseudo-colour algorithms, this paper establishes a relationship between the spatial function of high greyscale and factors such as blackbody radiation, light wave wavelength, temperature and greyscale value. A red, green and blue (RGB) colour space is also designed based on brightness gain compensation to achieve pseudo-colour transformation of high-greyscale images. To verify the effectiveness of the proposed method, it is applied to various images, including a 12-bitX-ray image of an oil pipeline, a 24-bit digital radiography (DR) image ofa steel pipe weld, a 14-bit infrared image and a precision casting image with defects. Quantitative experimental results indicate that the design method effectively highlights image texture details, demonstrating greater universality and adaptability for processing various types of greyscale image. Additionally, evaluation indicators reflecting human vision are improved.
Aero engine gas path anomaly detection serves as a critical safeguard for the normal operation of aircraft. The aero engine gas path data comprises a substantial volume of normal data and a limited number of abnormal data and exhibits the characteristics of high dimensionality and strong coupling. Existing methods have difficulty in effectively extracting deep-level feature information from the original data, which leads to insufficient differentiation between normal and abnormal data, thereby seriously reducing the accuracy of anomaly detection. To address this issue, an unsupervised anomaly detection method based on the memory-adversarial autoencoder (Memory-AAE) is proposed. Firstly, the Memory-AAE model is constructed by integrating the memory network and adversarial training mechanism, which effectively enhances the capability to extract deep-level feature information. Then, a normal sample screening strategy is designed, which clusters the reconstruction errors of the original data by the K-means algorithm to obtain a refined normal sample dataset. Finally, the anomaly scoring mechanism based on Euclidean distance and mean absolute error is adopted to quantify the anomaly degree from both global and local dimensions, which effectively improves the detection capability of complex anomalies. In this paper, gas path anomaly detection experiments indicate that the precision, recall and F1 score of the proposed method reach 0.923, 0.915 and 0.919, respectively, and exhibit excellent robustness to noise, significantly outperforming other methods.
A novel sensor package for fibre Bragg grating (FBG) acoustic emission (AE) based on a lateral coupling cone structure is proposed in this paper. The package is designed to be sensitive and reusable. Numerical modelling of the sensor package structure is performed through the use of finite element analysis for determining the main parameters of the coupling cone structure. An AE source system was built for the experimental work and a pencil lead break (PLB) signal was employed as the artificial acoustic emission source signal. The PLB signal was detected using the new sensor, the narrowband PLB signal was extracted using the Shannon wavelet transform and the simulated annealing multiple signal classification (SA-MUSIC) algorithm was used to achieve the localisation of the direction of the wave. The experimental data demonstrated that the lateral coupling cone fibre grating AE sensor has a maximum error value of -1.3 degrees in the direction of the wave and a minimum error value of -0.2 degrees in the angular estimation. The minimum error of the distance from the source was 4.8 mm and the maximum error was -10.39 mm. The results show that the FBG AE sensor package with lateral coupling exhibits advantages of high sensitivity, flexibility and adjustable positioning; it has achieved positioning of the acoustic emission signal source with useful accuracy.
Bearings are key components in most rotating machinery, making their reliability crucial for machine performance. Bearing fault detection using vibration analysis has been extensively addressed in the literature. In recent years, advances in machine learning (ML) have significantly contributed to improving and automating the tasks of bearing fault detection. This paper presents a methodology using convolutional neural networks (CNNs) to automatically classify time-frequency representations of vibration signals associated with different bearing faults. These representations are obtained using the short-time Fourier transform (STFT), where several parameters affecting the generation of the images are evaluated. The paper also explores transfer learning to address the issue of limited failure data, using a CNN trained with one dataset to classify bearing faults in a second dataset through fine-tuning techniques. The optimal configurations identified include a fixed number of shaft revolutions instead of fixed time and a local colour normalisation for the STFT images. The proposed model achieves 100% test classification accuracy for both the first dataset and the second dataset after fine-tuning. The results also confirm that transfer learning leads to faster training and model convergence. This methodology significantly improves the reliability and performance of rotating machinery through advanced artificial intelligence (AI) techniques, offering a practical solution for industries facing data scarcity.
Quality control is of paramount importance in the modern industrial landscape. Infrared thermography (IRT) is a non-destructive testing (NDT) technique that in recent years has experienced notable advancements and found widespread applications in the domains of defect detection and material characterisation. Tone burst eddy current thermography (TBET) is an active IRT testing procedure that is based on the principle of electromagnetic induction. The method has already proven to be valuable due to its cost-effectiveness, fast inspection rates and non-invasive nature. These features make TBET an ideal tool for in-process evaluation of conductive materials and composites. With the emergence of object detection and segmentation algorithms, deep learning (DL) has been widely applied in the field of IRT. DL-empowered thermography can achieve automated real-time inspection of materials. Such systems align perfectly with the objective ofzero-defect manufacturing by enabling active inspection and detection of defects. However, the lack of sufficient training data is one of the major hurdles to realising an intelligent quality control system. This work presents a methodology for bridging the data gap using synthetic TBET measurements generated from finite element method (FEM)based simulations. The proposed approach utilises two object localisation deep neural networks: the faster region-based convolutional neural network (Faster R-CNN) with an Inception-v4 backbone and the You Only Look Once version 8 (YOLOv8) network, to enable interpretations from the synthetic thermal data for automated quality management. The results reveal the potential of adopting synthetically generated datasets for pre-training the DL algorithms and, henceforth, the effective deployment of an automated defect detection and quality monitoring system based on TBET.
Addressing the significant challenge of extracting fault features from rolling bearing vibration signals amid complex working conditions and noise interference, a fault diagnosis approach grounded in the combination of the secretary bird optimisation algorithm (SBOA), optimised variational mode decomposition (VMD) and the Transformer-bidirectional gated recurrent unit (Transformer-BiGRU) model is put forth. The Transformer-BiGRU model is a skilfully integrated fault diagnosis method that is predominantly intended to heighten the accuracy. The SBOA first optimises the VMD parameters to decompose the vibration signal into multiple modal components. To be specific, each modal component is adopted as an input for feature extraction and abstraction by Transformer. The features extracted by Transformer are subsequently subjected to time-series modelling and feature fusion by the bidirectional gated recurrent unit (BiGRU). Ultimately, the fused features are employed for the diagnosis and classification of bearing faults. As suggested by the research findings, the approach can achieve an average accuracy of 98.80% in the presence of noise interference. It can even reach 100% when the signal-to-noise ratio (SNR) is 6 dB and the model can also achieve a fault diagnosis rate of 97.90% when the load is changed. Hence, the fault diagnosis approach proposed features high accuracy and strong generalisation ability, providing a groundbreaking way of thinking for the fault diagnosis of rolling bearings.
Existing methods for detecting damage in conveyor belts are affected by multiple factors such as complex underground working environments, uneven illumination and high dust concentration. Furthermore, these methods exhibit low detection efficiency, insufficientaccuracyand both false and missed detections. Therefore, in this study, an improved You Only Look Once version 8 small (YOLOv8s) damage detection algorithm for mining conveyor belts was developed, called GCW-YOLO. The lightweight Ghost module was combined with the C2f module to form the C2f_Ghost structure, enabling a lightweight model by generating more feature maps to enhance network performance, which significantly reduces the number of parameters. Therefore, the model demonstrates high performance while reducing the number of parameters and computational complexity, while simultaneously accelerating detection speed. A coordinate attention (CA) mechanism was embedded into the backbone network to enhance feature extraction capability by adaptively assigning weights to features of different channels, highlighting important features while suppressing irrelevant ones, and improving feature representation in complex scenarios. The wise intersection over union version 3 (WIoUv3) was used as the bounding box loss function instead of the original loss function to improve model convergence and regression accuracy. The loss calculation was adjusted to help the model focus on important targets or regions by assigning different weights to various regions or target categories. Experimental results indicated that, compared with the original YOLOv8s algorithm, the improved algorithm achieved an average detection accuracy of 89.7% for the types of damage investigated, representing an improvement of4.9%. Furthermore, compared with the original algorithm model, the number of model parameters was reduced by 16%, the computational load decreased by3.7giga floatingpoint operations per second (GFLOPS) and the average detection speed reached 55 frames per second (FPS). The improved algorithm ensures accurate detection and improved detection speed of conveyor belt damage in coal mines, demonstrating significant practical value.
The catenary is a core structure of high-speed railway systems. It is responsible for providing a stable and reliable power supply to the train. In the process of train operations, catenary components are prone to failure. In order to ensure the safe operation of trains, it is essential to effectively detect and distinguish between different types of catenary component. In this paper, a catenary optical inspection method is proposed based on an attention-enhanced faster region-based convolutional neural network (Faster R-CNN). Firstly, the established Residual Network-50 (ResNet-50) is optimised by redesigning the bottleneck. In the bottleneck, a dual-path attentional enhancement mechanism is proposed by combining a parallel-convolutional blockattention module (P-CBAM) in different paths. Secondly, a new adaptive attention module (NAAM) is designed to improve the small-sized object detection accuracy. Then, a feature pyramid network (FPN) is redesigned using the NAAM to reduce information loss during the feature map generation process and enhance the feature representation capability for multi-sized objects. Moreover, an exponential linear unit (ELU) activation function is introduced to improve the performance of the algorithm. In this paper, images of the whole catenary system are used. A single catenary image contains various components and eight of them are selected for analysis. Compared to other models, the proposed method achieves the highest detection accuracy with 6.6% improvement in mAP@0.5 and 4.78% improvement in mAP@0.5:0.95 over the baseline Faster R-CNN. Furthermore, the average recall (AR) also shows a notable enhancement. The experimental results confirm that the proposed method has good detection performance for distinguishing between different types of catenary component in electrified railways.
Bearing fault diagnosis plays a critical role in predictive maintenance of rotating machinery, where early and accurate detection can prevent catastrophic failures and reduce operational downtime. This study presents a comparative analysis of four supervised machine learning (ML) algorithms: decision tree, k-nearest neighbours (k-NN), logistic regression and Gaussian naive Bayes, for classifying bearing faults based on features extracted from vibration signals. The input features are standardised and class labels are encoded to ensure compatibility with ML workflows. A stratified train-test split is employed to maintain balanced class distribution across subsets. Each model is evaluated using key performance metrics: accuracy, precision, recall, F1 score and multi-class area under the curve (AUC). The results show that the decision tree classifier achieves the highest classification accuracy (93.57%), while logistic regression and k-NN record the highest AUC scores (99.26% and 98.60%, respectively), reflecting strong generalisation and discriminatory capabilities. The findings indicate that while tree-based models offer superior raw classification accuracy, probabilistic and distance-based methods demonstrate excellent potential in handling multi-class bearing fault prediction tasks. This research underscores the importance of model selection and performance trade-offs in developing robust condition monitoring systems for industrial applications.
This study addresses the issues of high signal interpretation complexity and low efficiency in manual defect identification associated with alternating current field measurement (ACFM) technology in practical industrial inspections. A defect identification method that integrates a bidirectional long short-term memory (BiLSTM) network with an attention mechanism is proposed. By constructing a hybrid model with the capability for temporal feature selection, the method enhances the extraction of subtle defect features from ACFM signals. The experimental results show that on a test set containing 500 multi-condition signals (covering defects in 20# steel plates with depths ranging from 0.5 mm to 3 mm and lengths from 10 mm to 50 mm), the model achieves an accuracy rate of 92.10% +/- 1.2%. Compared to traditional manual inspection, the average single determination time is reduced to 17.14 ms (an efficiency improvement of approximately 68%), with the smallest reliably identifiable defect size being 0.5 mm (depth) & times; 10 mm (length). Compared to the baseline long short-term memory (LSTM) model, the F1 score improves by 10.07 percentage points to 94.65% and the memory usage of the model is 2.28 MB, meeting the deployment requirements for industrial embedded devices. This method provides a new technical pathway for the automation ofACFM inspections.
To solve the problem ofa high false positive rate in leakage detection in branch pipelines, this paper proposes a pipeline detection method based on acousto-pressure signal information fusion. The transformer architecture is introduced as the basic mode of the acousto-pressure fusion diagnosis model to enhance the signal timing feature extraction ability. The cross-self-attention mechanism is used to make the sound signal and pressure signal interact and fuse in the process of feature extraction, so as to realise end-to-end branch pipeline fault diagnosis. Firstly, the noise reduction method of complementary ensemble empirical mode decomposition-least mean squares (CEEMD-LMS) is used to denoise the collected sound pressure signal. Secondly, an acousto-pressure cross-fusion model based on a cross-fusion transformer is proposed to fuse the information in two types ofdenoised signal and classify different leakage types. The experimental results show that the accuracy of leakage detection for a 2 mm leak size is 97.21%, while the accuracies of leakage detection for 5 mm and 8 mm sizes are 98.57% and 98.94%, respectively. The accuracy of the model for detecting the presence or absence of any leakage is 99.03%, which verifies the effectiveness of the proposed method.
