High-speed railway wheelsets operate under heavy loads and harsh environments, leading to rolling-contact fatigue, thermal cracks, and near-surface defects. Field ultrasonic B-scans are degraded by surface roughness, coupling fluctuations, and electrical interference, producing low SNR. In addition, defect rarity and safety constraints yield scarce, imbalanced annotations and substantial cross-condition domain shifts. To address these challenges and the NDE 4.0 requirements for interpretability and traceability, we propose an unsupervised, diffusion-based detection framework. A physics-prior-constrained diffusion augmentation module learns noise and defect-morphology distributions from real B-scans to generate diverse, controllable samples that mitigate scarcity and imbalance. An unsupervised diffusion detector with consistency constraints provides robust recognition under varying conditions. An uncertainty-guided closed loop enables semi-automatic annotation and continual model updating. Experiments on real multi-condition wheelset datasets show significant improvements over conventional methods in detection rate, robustness, and interpretability, reducing both missed and false detections. The system has been engineered and deployed as a closed-loop workflow that integrates data acquisition, diffusion-based enhancement, unsupervised detection, expert review, and traceable reporting, operating stably across multiple EMU depots and indicating strong potential for large-scale adoption.

The structured light projection profilometry technique is widely applied in various fields. However, for smooth surfaces, the technology faces challenges caused by specular reflection during application. These reflections introduce distortions into the modulation information of fringe patterns, causing phase errors in the subsequent image processing and impeding the accurate demodulation of corresponding 3D information. To address this issue, we propose a single-shot 3D measurement algorithm for reflective surfaces based on deep learning. The algorithm takes a single reflective fringe image as input and produces an optimized, high-quality fringe image without reflection interferences, along with generating the corresponding depth map. Experiments demonstrate the algorithm's efficacy in achieving reflective fringe image optimization with minimal phase accuracy loss (0.0382 radians) and high accuracy in single-shot 3D measurement with a mean absolute error (MAE) of similar to 0.48 mm. Notably, the algorithm requires only a single reflective image as input without additional hardware assistance, making it well-suited for dynamic measurement applications. The processing time of the algorithm is similar to 0.12 s per image. Furthermore, the algorithm can serve as a preprocessing tool for reflective fringe patterns, seamlessly integrating with various 3D measurement algorithms, demonstrating strong flexibility and transferability.

Frequency-domain plane-wave imaging (FPWI) enables ultrafast frame rates for ultrasonic nondestructive evaluation. However, image quality can still be degraded by noise and artifacts. Most coherence-weighting schemes are implemented in the time domain using sliding-window statistics, which introduce substantial computation and memory-access overhead. In this work, we compute coherence weights directly in the frequency domain using the fast Fourier transform (FFT). Inspired by spatial coherence principles related to the van Cittert–Zernike theorem (VCZ), we propose a normalized amplitude coherence factor (NACF) as an amplitude-domain alternative to the normalized energy coherence factor (NECF). Experiments on representative defect datasets show that FPWI with NACF improves contrast ratio (CR), signal-to-noise ratio (SNR), and array performance indicator (API) relative to baseline FPWI, while further sharpening defect boundaries and enhancing lateral resolution. These results demonstrate a favorable balance between artifact suppression and resolution enhancement, making the proposed method a promising solution for ultrasonic NDE scenarios requiring both high frame rates and high imaging quality.

The Total Focusing Method (TFM) in ultrasonic phased array imaging offers high resolution and promises improved detection of near-surface flaws. However, the irregular structure of wheel rim treads presents challenges for detecting such small volumetric defects. In this study, an Adaptive Total Focusing Method (ATFM) is proposed under water immersion phased array testing conditions. The technique reconstructs the interface using a pulse-echo approach and applies band-pass filtering to Full Matrix Capture (FMC) data via a Kaiser window. Based on Delay-and-Sum (DAS) beamforming, the effectiveness of six different coherence factor weighting methods is compared. The results demonstrate that the ATFM approach incorporating the Circular Coherence Factor (CCF) offers significant advantages in near-surface noise suppression and defect resolution. It successfully detects flat-bottom hole (FBH) defects with a diameter of 0.5 mm located 3 mm beneath the rim surface, achieving a signal-to-noise ratio (SNR) of 28.93 dB and a resolution of 45.79 dB.

Reconstruction models with Transformer architectures have achieved state-of-the-art performance in anomaly detection, yet the issue of learning shortcuts persists. We address this problem in reconstruction tasks by eliminating residual connections in the Transformer structure. While traditional Transformers rely on residual connections to alleviate gradient vanishing, such designs in reconstruction tasks tend to allow models to bypass deep feature learning and directly copy shallow-layer features, resulting in performance degradation. By removing residual connections, we compel the model to learn deeper features more comprehensively, thereby improving reconstruction quality. Additionally, we design a novel feature processing mechanism based on features extracted by convolutional neural networks (CNNs): First, global contextual information is extracted through average pooling (AvgPool) for feature dimensionality reduction. Then, 1D convolution (Conv1D) is applied along the channel dimension to further fuse features. Finally, attention weights are generated via the Sigmoid activation function to highlight critical features. Under unified training and inference conditions, our method outperforms previous state-of-the-art approaches on the MVTec-AD dataset and demonstrates exceptional performance on a train component dataset.

Ultrasonic testing is a widely used nondestructive testing (NDT) method for detecting defects in critical industrial components. However, ultrasonic defect detection in high-speed rail (HSR) systems faces significant challenges due to limited sample availability and complex working conditions. These limitations often lead to subjective judgments by inspectors, increasing the risk of false positives and missed detections. To mitigate data scarcity, this study introduces a diffusion model for data augmentation, applied to real ultrasonic B-scan wheel defect data. By learning the probability and noise distribution through diffusion and reverse diffusion processes, the model generates synthetic data to improve detection accuracy. Experimental results show notable improvements in average precision and recall, increasing from 78.0 % to 66.0 %-93.3 % and 91.5 %, respectively. This method has been successfully deployed in practical applications, with plans for continuous updates as new data becomes available. The study addresses the challenge of limited defect data in industrial NDT and highlights the potential for broader applications in automated defect detection systems.

With the ease of acquiring RGB-D images from line-scan 3D cameras and the development of computer vision, anomaly detection is now widely applied to railway inspection. As 2D anomaly detection is susceptible to capturing condition, a combination of depth maps is now being explored in industrial inspection to reduce these interferences. In this case, this paper proposes a novel approach for RGB-D anomaly detection called Dual-Branch Cross-Fusion Normalizing Flow (DCNF). In this work, we aim to exploit the fusion strategy for dual-branch normalizing flow with multi-modal inputs to be applied in the field of track detection. On the one hand, we introduce the mutual perception module to acquire cross-complementary prior knowledge in the early stage. On the other hand, we exploit the effectiveness of the fusion flow to fuse the dual-branch of RGB-D inputs. We experiment on the real-world Track Anomaly (TA) dataset. The performance evaluation of DCNF on TA dataset achieves an impressive AUROC score of 98.49%, which is 3.74% higher than the second-best method.

This study presents an unsupervised deep learning framework aimed at improving defect detection in A-scan ultrasonic signals from high-speed train axles. By addressing the challenges associated with manual interpretation in traditional ultrasonic testing, we introduce a dual-encoder model with attention mechanisms that can analyze both global waveform patterns and localized anomalies without the need for labeled training data. The proposed architecture aims to isolate defect-induced signal variations from environmental noise through parallel feature processing. Evaluations on real-world industrial data indicate that the model performs well, with promising results (F1: 0.94, AUC: 0.98) in comparison to threshold-based methods, especially in low signal-to-noise scenarios. This approach has the potential to support more reliable large-scale axle inspection, reducing the reliance on annotated datasets and contributing to the improvement of rail safety through automated anomaly detection.

Fringe projection profilometry (FPP) technique is a widely used method in three-dimensional measurement and has broad applications in industrial measurement. This method typically requires projecting multiple phase-shifted fringe patterns to achieve high-precision imaging. However, in industrial contexts, the measured objects are often metallic components with high reflectivity, which would adversely affect the reconstruction quality of FPP. High Dynamic Range (HDR) technology can effectively address such issues; however, it requires an increased number of projected fringe patterns, which reduces measurement efficiency. This paper proposes a single-shot 3D measurement method that combines traditional HDR techniques with deep learning to measure objects with highly reflective surfaces. The method employs two network models: the Phase Recovery Network (PRNet) and the Coarse Phase Unwrapping Network (CPUNet). PRNet receives a single fringe image with saturated pixels and outputs the numerator and denominator terms for phase retrieval. CPUNet takes the wrapped phase data processed through an arctangent function from PRNet's output and predicts the coarse unwrapped phase to assist in phase unwrapping, yielding high-precision phase results. Experiments show that this approach requires only one shiny fringe image as input and effectively restores phase errors caused by high reflectivity, achieving a high reconstruction accuracy for the complete 3D shape of the object.

In the context of defect detection in high-speed railway train wheels, particularly in ultrasonic-testing B-scan images characterized by their small size and complexity, the need for a robust solution is paramount. The proposed algorithm, UT-YOLO, was meticulously designed to address the specific challenges presented by these images. UT-YOLO enhances its learning capacity, accuracy in detecting small targets, and overall processing speed by adopting optimized convolutional layers, a special layer design, and an attention mechanism. This algorithm exhibits superior performance on high-speed railway wheel UT datasets, indicating its potential. Crucially, UT-YOLO meets real-time processing requirements, positioning it as a practical solution for the dynamic and high-speed environment of railway inspections. In experimental evaluations, UT-YOLO exhibited good performance in best recall, mAP@0.5 and mAP@0.5:0.95 increased by 37%, 36%, and 43%, respectively; and its speed also met the needs of real-time performance. Moreover, an ultrasonic defect detection data set based on real wheels was created, and this research has been applied in actual scenarios and has helped to greatly improve manual detection efficiency.

As a critical operational component, railway wheels undergo ultrasonic testing, one of the most prevalent non-destructive testing methods. However, most ideal defects seldom manifest in real-world scenarios, leading to a scarcity of available cases. In light of this challenge, we propose a novel dual-network approach in this paper, designed for both data augmentation and the detection of extreme data. This approach aims to enhance defect detection performance. To validate the feasibility of our algorithm, we assembled a substantial dataset comprising simple ultrasonic B-scan defects. Our algorithm demonstrates high detection accuracy, surpassing 95% with augmented data, about a 15% improvement over the original dataset. Furthermore, employing augmented data for detecting real data yields commendable results. This paves the way for integrating deep learning with traditional nondestructive testing in future industrial inspections, ensuring efficient defect detection and bolstering railway safety and personnel training efforts.

The lack of real defect data samples has become a challenging problem for the effective application of deep learning networks in ultrasound target detection. This paper proposes a data augmented generative adversarial network (DCSGAN) aimed at overcoming the scarcity of welding ultrasonic defect data in training target detection networks. This network utilizes bilinear interpolation to expand the real data sample space, facilitating the extraction of high-dimensional defect spatial features through deeper networks. By obtaining a mixed dataset of generative data and real data, training and testing experiments are conducted on the object detection network. The experimental results demonstrate that the data augmentation method proposed in this paper effectively enhances the detection rate of ultrasonic welding defects in the target detection network, which has reference significance for similar application scenarios of ultrasonic defect detection.

Three-dimensional point cloud registration is a critical task in 3D perception for sensors that aims to determine the optimal alignment between two point clouds by finding the best transformation. Existing methods like RANSAC and its variants often face challenges, such as sensitivity to low overlap rates, high computational costs, and susceptibility to outliers, leading to inaccurate results, especially in complex or noisy environments. In this paper, we introduce a novel 3D registration method, CL-PCR, inspired by the concept of maximal cliques and built upon the SC2-PCR framework. Our approach allows for the flexible use of smaller sampling subsets to extract more local consensus information, thereby generating accurate pose hypotheses even in scenarios with low overlap between point clouds. This method enhances robustness against low overlap and reduces the influence of outliers, addressing the limitations of traditional techniques. First, we construct a graph matrix to represent the compatibility relationships among the initial correspondences. Next, we build clique-likes subsets of various sizes within the graph matrix, each representing a consensus set. Then, we compute the transformation hypotheses for the subsets using the SVD algorithm and select the best hypothesis for registration based on evaluation metrics. Extensive experiments demonstrate the effectiveness of CL-PCR. In comparison experiments on the 3DMatch/3DLoMatch datasets using both FPFH and FCGF descriptors, our Fast-CL-PCRv1 outperforms state-of-the-art algorithms, achieving superior registration performance. Additionally, we validate the practicality and robustness of our method with real-world data.

During the ultrasonic defect detection process of train wheel, electromagnetic interference and poor coupling of equipment can lead to disturbances or frame losses in the data. This results in some omissions and incorrect detections in subsequent defect target detection tasks, posing significant risks. The efficiency of classifying abnormal data through manual and traditional algorithms is not high. To address this issue, this study introduces for the first time the application of a deep learning classification network to the quality classification of ultrasonic B-scan data. Based on the size of our ultrasonic B-scan dataset, we selected ResNet34 as our classification network. Additionally, we augmented the ultrasonic B-scan dataset with enhanced data. The detection results demonstrate that the proposed network successfully differentiated between different types of samples in the test set, achieving classification accuracy of over 95%. This method effectively classifies abnormal data from normal data, thereby improving the efficiency of subsequent damage detection tasks. It holds significant reference value for image quality classification in similar applications.

Train wheels are crucial components for ensuring the safety of trains. The accurate and fast identification of wheel tread defects is necessary for the timely maintenance of wheels, which is essential for achieving the premise of conditional repair. Image-based detection methods are commonly used for detecting tread defects, but they still have issues with the misdetection of water stains and the leaking of small defects. In this paper, we address the challenges posed by the detection of wheel tread defects by proposing improvements to the YOLOv8 model. Firstly, the impact of water stains on tread defect detection is avoided by optimising the structure of the detection layer. Secondly, an improved SPPCSPC module is introduced to enhance the detection of small targets. Finally, the SIoU loss function is used to accelerate the convergence speed of the network, which ensures defect recognition accuracy with high operational efficiency. Validation was performed on the constructed tread defect dataset. The results demonstrate that the enhanced YOLOv8 model in this paper outperforms the original network and significantly improves the tread defect detection indexes. The average precision, accuracy, and recall reached 96.95%, 96.30%, and 95.31%.

The weld inspection process under coating is complex and labour-dependent. Therefore, it is of great engineering significance to study weld defect inspection. Based on eddy current pulsed thermography (ECPT) method, this work theoretically analysed the mechanism of weld defect detection undercoating and proposed an inductor suitable for defect detection under weld coating to achieve non-contact, rapid and visual detection results. The simulation model for weld defect detection under coating was constructed, and the effect of coating on weld crack inspection was investigated. However, weld defect test blocks with different coating, thicknesses were verified by experiment, and the thermal characteristics of the circular hole area were analysed. The results demonstrate that the proposed method is effective in weld defect detection, which provides a promising way to extend the application of ECPT to weld defect inspection.

In recent years, a multitude of self-supervised anomaly detection algorithms have been proposed. Among them, PatchCore has emerged as one of the state-of-the-art methods on the widely used MVTec AD benchmark due to its efficient detection capabilities and cost-saving advantages in terms of labeled data. However, we have identified that the PatchCore similarity principal approach faces significant limitations in accurately locating anomalies when there are positional relationships between similar samples, such as rotation, flipping, or misaligned pixels. In real-world industrial scenarios, it is common for samples of the same class to be found in different positions. To address this challenge comprehensively, we introduce Feature-Level Registration PatchCore (FR-PatchCore), which serves as an extension of the PatchCore method. FR-PatchCore constructs a feature matrix that is extracted into the memory bank and continually updated using the optimal negative cosine similarity loss. Extensive evaluations conducted on the MVTec AD benchmark demonstrate that FR-PatchCore achieves an impressive image-level anomaly detection AUROC score of up to 98.81%. Additionally, we propose a novel method for computing the mask threshold that enables the model to scientifically determine the optimal threshold and accurately partition anomalous masks. Our results highlight not only the high generalizability but also substantial potential for industrial anomaly detection offered by FR-PatchCore.

Wheels are an essential part of railway trains;thus,defects on the wheel tread present serious risk regarding the safety of railway trains.Due to the limited samples of wheel tread defects in practice,the corresponding supervised detection model is insufficient.To solve this problem,an unsupervised knowledge distillation anomaly detection model is proposed to detect wheel tread anomalies.Accordingly,UNet is employed to segment the tread region and reduce the influence of non-tread regions on the anomaly detection model.An attention mechanism is then added after the multiscale feature fusion to improve the ability of the student network to reconstruct normal features in the reverse knowledge distillation structure,as well as enhance the reconstruction of normal features.From the experimental results,the improved model achieves the performance indexes of 93.8%area under receiver operating characteristic curve,82.3%precision,95.4%recall,and 87.0%accuracy considering the railway wheel tread dataset.Compared with the original model,the detection performance of the model is improved.

Partial point cloud registration is an important step in generating a full 3D model. Many deep learning-based methods show good performance for the registration of complete point clouds but cannot deal with the registration of partial point clouds effectively. Recent methods that seek correspondences over downsampled superpoints show great potential in partial point cloud registration. Therefore, this paper proposes a partial-to-partial point cloud registration network based on geometric attention (GAP-Net), which mainly includes a backbone network optimized by a spatial attention module and an overlapping attention module guided by geometric information. The former aggregates the feature information of superpoints, and the latter focuses on superpoint matching in overlapping regions. The experimental results show that the method achieves better registration performance on ModelNet and ModelLoNet with lower overlap. The rotation error is reduced by 14.49% and 17.12%, respectively, which is robust to the overlap rate.