A wheel end defect enhanced detection method and system

Abstract

The application discloses a wheel axle defect enhanced detection method, comprising: data standardization and basic enhancement of wheel axle ultrasonic B scan image; generating enhanced image and reconstructing high noise image by using generative data enhancement module combined with ultrasonic mechanism; inputting the enhanced image into a multi-scale dynamic attention target detection network for defect identification. Through the generative data enhancement module, the application synthesizes samples covering rare defects and complex noise at the data level, effectively solving the problems of uneven data distribution and weak model generalization ability caused by the scarcity of real defect samples; at the same time, through the detection network designed for the physical characteristics of B scan image and fusing the multi-scale dynamic attention mechanism, the sensitivity to slender and weak echo defects is enhanced and the background interference is suppressed. The cooperative design of data layer and model layer realizes the high-precision and high-robustness automatic detection of wheel axle defects.

Description

Technical Field

[0001] This application relates to the fields of nondestructive testing and computer vision technology, and in particular to a method and system for identifying ultrasonic defects in the wheel axles of high-speed trains based on generative data augmentation and multi-scale dynamic attention target detection. Background Technology

[0002] Train wheels are critical load-bearing and guiding components of the train's running gear, and their safety directly affects train operation safety. Currently, non-destructive testing of train wheels mainly uses fully automated ultrasonic phased array equipment to acquire multi-channel B-scan images, which are then manually inspected by maintenance personnel. This method suffers from problems such as high workload, poor interpretation consistency, and susceptibility to interference from ambient noise, leading to missed detections and false alarms.

[0003] In recent years, intelligent auxiliary defect detection technology based on deep learning has become a trend. However, existing technologies face three major challenges: (1) the scarcity of defect samples and the uneven distribution of data lead to insufficient generalization ability of the model; (2) the general target detection network is not optimized for the physical characteristics of B-scan images (such as sound path-channel structure, slender defect morphology, and strong structural echo), resulting in insensitivity to small defects and high false alarm rate; (3) the lack of collaborative design between data augmentation strategy and detection network, and the mismatch between augmentation samples and network structure in terms of statistical characteristics, which restricts the overall performance.

[0004] Therefore, there is an urgent need for a new comprehensive solution for ultrasonic intelligent-assisted damage assessment that can systematically solve the above problems. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing technologies and provide a wheel and axle defect enhancement detection method and system. Through the collaborative design of generative data augmentation and a multi-scale dynamic attention target detection network, the robustness, accuracy, and adaptability of intelligent wheel and axle defect identification are systematically improved. To achieve the above objectives, this application adopts the following technical solution:

[0006] In a first aspect, embodiments of this application provide a method for enhanced detection of wheel and axle defects, including:

[0007] Data standardization and basic enhancement processing were performed on the multi-channel B-scan images of the wheel axle ultrasonic phased array to obtain standardized B-scan images;

[0008] The standardized B-scan images are input into a generative data augmentation module that incorporates ultrasound mechanisms to generate enhanced B-scan images, thereby expanding the training data and reconstructing high-noise images.

[0009] The enhanced B-scan image is input into a multi-scale dynamic attention target detection network for training and inference, and the results of wheel axle defect identification and localization are output.

[0010] The generative data augmentation module and the multi-scale dynamic attention target detection network are designed in tandem to jointly improve the accuracy and robustness of defect identification.

[0011] Furthermore, the construction logic of the generative data augmentation module includes:

[0012] A defect simulation model was constructed based on the wheel and axle geometry and ultrasonic propagation mechanism, and a B-scan simulation segment with defects was generated.

[0013] The joint distribution of real B-scan images is learned using a deep generative model, and B-scan samples with specified features are synthesized through conditional control.

[0014] A balanced training dataset is constructed by mixing simulated samples, deep learning samples, and real-world samples.

[0015] Furthermore, the multi-scale dynamic attention target detection network includes:

[0016] The backbone feature extraction network uses strip convolution and cross-channel correlation modeling to extract depth and channel features from B-scan images;

[0017] A multi-scale feature fusion structure integrates shallow details and deep semantic information, enhancing the ability to express slender defects;

[0018] The multi-scale dynamic attention detection head dynamically adjusts scale, spatial and channel weights to highlight defect areas and suppress background interference.

[0019] Furthermore, the multi-scale dynamic attention detection head includes:

[0020] Scale attention branch dynamically adjusts the weights of features at each scale in the prediction;

[0021] Spatial attention branch highlights suspected defect areas and suppresses background response;

[0022] Channel attention branch, recalibrate feature channels, and strengthen channels related to defect patterns.

[0023] Furthermore, the deep generation model is a denoising diffusion probability model or a generative adversarial network, which can reconstruct and denoise high-noise B-scan images while preserving defect echo features.

[0024] Furthermore, the data standardization includes:

[0025] The B-scan images are corrected for time axis, registered for channel axis, resampled, and normalized for amplitude, and then unified into a two-dimensional matrix with a fixed resolution.

[0026] Furthermore, the basic enhancements include:

[0027] Perform random translation, cropping, flipping, and / or channel discarding operations in the image domain;

[0028] Gain perturbation and noise injection are performed in the amplitude domain to simulate changes in on-site operating conditions.

[0029] Furthermore, the detection method further includes:

[0030] During training, smaller defect samples and samples with low contrast defects are given higher loss weights, and a hard example mining strategy is introduced.

[0031] Furthermore, the detection method is applicable to the detection of one or more defects in the rim, flange, spoke, and axial transition zone of the axle of a high-speed train.

[0032] Secondly, embodiments of this application provide a detection system capable of implementing any of the aforementioned detection methods, comprising:

[0033] The data preprocessing module is used to standardize and perform basic enhancement processing on the multi-channel B-scan images of the wheel axle ultrasonic phased array to obtain standardized B-scan images;

[0034] The generative data augmentation module is used to receive the standardized B-scan image, synthesize the enhanced B-scan image based on the ultrasound mechanism and depth generation model, and reconstruct and denoise the high-noise image;

[0035] The multi-scale dynamic attention detection module is used for feature extraction, multi-scale fusion, and defect identification and localization of the enhanced B-scan image;

[0036] The generative data augmentation module works in conjunction with the multi-scale dynamic attention detection module to improve the system's accuracy and robustness in identifying wheel and axle defects.

[0037] Thirdly, embodiments of this application provide an electronic device, including: one or more processors;

[0038] A memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors are able to implement the steps of the detection method described in any of the preceding claims.

[0039] Fourthly, embodiments of this application provide a computer-readable medium storing a computer program, which, when executed by a processor, can implement the steps of the detection method described in any of the preceding claims.

[0040] This application utilizes a generative data augmentation module to synthesize samples encompassing rare defects and complex noise at the data level, effectively addressing the issues of uneven data distribution and weak model generalization ability caused by the scarcity of real defect samples. Simultaneously, through a detection network specifically designed for the physical characteristics of B-scan images and incorporating a multi-scale dynamic attention mechanism, the sensitivity to slender, weak-echo defects is enhanced and background interference is suppressed at the model level. The collaborative design of the data and model layers jointly achieves high-precision, highly robust, and automated detection of wheel and axle defects. Attached Figure Description

[0041] Figure 1 This application provides a core flowchart of a wheel and axle defect enhancement detection method.

[0042] Figures 2(a)-2(c) are a set of comparison figures of the performance of the UT-DDPM algorithm provided in the embodiments of this application;

[0043] Among them, EPLL: Enhanced Patch Prior, is an image denoising method based on maximum a posteriori probability estimation; KSVD: K-Singular Value Decomposition, is a sparse representation algorithm based on dictionary learning, which achieves efficient signal compression and denoising by iteratively optimizing dictionary atoms and sparse coding; WNNM: Weighted NuclearNorm Minimization, is an image denoising algorithm based on low-rank matrix recovery.

[0044] Figure 3 A schematic diagram of the target detection model architecture combined with Dy-ut-head provided in the embodiments of this application;

[0045] Figure 4 Comparison of noise reduction effects of real vehicle data provided in the embodiments of this application;

[0046] Among them, Bscan stands for Brightness-scan, a type B ultrasound scan;

[0047] Figure 5 The data generation effect display diagram provided for the embodiments of this application;

[0048] Figure 6 Comparison images of actual vehicle wheel inspection results provided in this application embodiment;

[0049] Figure 7 Comparison images of actual vehicle axle inspection results provided in this application embodiment;

[0050] Figure 8 Example diagram of detected wheel surface defects provided in the embodiments of this application;

[0051] Figure 9 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0052] To enable those skilled in the art to better understand the technical solutions of this application, exemplary embodiments of this application are described below with reference to the accompanying drawings, including various details of the embodiments of this application to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description. Unless otherwise specified, the various embodiments of this application and the features within those embodiments can be combined with each other.

[0053] As used herein, the term “and / or” includes any and all combinations of one or more of the associated enumerated entries. The terminology used herein is for describing particular embodiments only and is not intended to limit the application. As used herein, the singular forms “a” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated features, integrals, steps, operations, elements, and / or components is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded. Terms such as “connected” or “linked” are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0054] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It should also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having the meaning consistent with their meaning in the context of the relevant art and this application, and will not be interpreted as having an idealized or overly formal meaning, unless expressly so defined herein.

[0055] Train wheels are critical load-bearing and guiding components of the train's running gear, and their safety directly affects the stability of wheel-rail interaction and the overall operational safety of the train. Under prolonged conditions of high-speed rolling, frequent braking, and complex environmental loads, wheels are prone to various damage forms, including rim fatigue cracks, flange chipping, and internal spoke defects. Failure to identify and address these damages promptly during maintenance can lead to serious failures such as rim fracture and flange breakage, even jeopardizing train operation safety. Therefore, developing a highly reliable, automated, and intelligent non-destructive testing technology system for train wheels is one of the core foundational tasks in the modern high-speed railway operation and maintenance system.

[0056] Currently, non-destructive testing of EMU wheels mainly employs fully automated ultrasonic phased array equipment. This equipment uses a pre-set sector scanning process to acquire multi-channel B-scan images of key areas such as the wheel rim, flange, and spokes. Maintenance personnel then manually assess defects based on waveform and image characteristics. This method is highly dependent on the experience and effort of operators, resulting in high workload and difficulty in ensuring consistent interpretation. Furthermore, it is susceptible to factors such as insufficient tread cleaning, unstable coupling, probe fluctuation, improper gain settings, and electrical noise interference on-site. These factors can lead to significant background noise and numerous false echoes in the B-scan images, potentially obscuring or mixing true defect indications with structural echoes, thus causing false alarms and missed detections. This makes it difficult to fully meet the requirements of intensive EMU maintenance and high safety redundancy.

[0057] In recent years, with the rapid development of deep learning and computer vision technologies, introducing data-driven methods such as target detection networks and anomaly detection networks into wheel ultrasonic B-scan image analysis has become an industry trend. Some research has constructed multi-channel wheel B-scan datasets and achieved automatic annotation of rim cracks, flange defects, etc., based on improved target detection networks, which can significantly improve the detection rate and interpretation efficiency under typical working conditions. In engineering practice, pilot applications of intelligent ultrasonic-assisted wheel defect assessment have been carried out in some high-speed train depots, initially verifying the feasibility and engineering value of deep learning methods in wheel defect detection scenarios.

[0058] However, existing technologies still face several challenges under complex working conditions: most methods still rely directly on the original B-scan images or simple filtering results for identification, and have limited adaptability to conditions such as high noise and gain bias. When the quality of the B-scan signal is unstable, the defect echo is easily distorted or masked. Wheel flaw detection data has significant characteristics of "multi-channel, multi-scale, and multi-morphology", while traditional single-structure target detection networks often struggle to achieve a good balance between identifying small defects, suppressing echoes from complex structures, and inference efficiency.

[0059] Therefore, it is necessary to propose a novel comprehensive solution for ultrasonic intelligent-assisted defect assessment for high-speed train wheel applications. This solution uses advanced generative algorithms to compensate for the lack of data and combines a small defect optimization algorithm with a multi-scale dynamic attention mechanism based on ultrasonic principles to achieve high-quality reconstruction and high-precision defect identification of wheel B-scan images.

[0060] This application mainly focuses on solving the following key technical problems in the process of defect identification driven by ultrasonic B-scan data of EMU wheel axles:

[0061] 1. The problem of scarce defective samples and unbalanced data distribution

[0062] In ultrasonic testing of wheel and axle, the number of actual defect samples collected is limited, especially for early-stage fine cracks, complex-shaped defects, and defects under extreme operating conditions. Data from different maintenance sites, with different equipment, and with different process parameters vary significantly in terms of noise levels and structural echo patterns, resulting in a clear imbalance in training data regarding defect types, noise patterns, and operating condition coverage. Traditional supervised learning methods that rely solely on a limited number of real samples struggle to learn stable and reliable wheel and axle defect representations, and the models lack the ability to generalize to uncommon operating conditions and boundary conditions.

[0063] 2. Insufficient compatibility between B-scan characteristics and general target detection networks.

[0064] Ultrasonic B-scan images of wheel axles simultaneously possess physical meanings of "vertical direction as the sound path / depth axis and horizontal direction as the channel / circumferential axis." Defect indications often exhibit patterns such as elongation, curvature, and low contrast, while structural echoes such as interface waves, bottom waves, and bolt hole waves have strong energy and wide coverage. Mainstream target detection networks are mostly designed for natural images, and their feature extraction and detection head structure are not optimized for the temporal-spatial coupling characteristics and multi-channel coherence characteristics of B-scan images. This results in the network being insensitive to small and weak defects, easily misidentifying strong structural echoes as defects, and making it difficult to achieve a balance between detection accuracy and inference efficiency.

[0065] 3. Lack of collaborative design between data augmentation and detection networks

[0066] Existing data augmentation methods mostly remain at the level of general image enhancement such as geometric transformation, random cropping, and noise addition. They are unable to accurately characterize the physical mechanisms of ultrasonic propagation and defect echoes, nor can they provide targeted distribution expansion for the "scarce defects + complex noise" characteristics of wheel and axle scenes. At the same time, data augmentation strategies and detection networks are often designed independently, lacking a collaborative optimization mechanism for the entire "generative augmentation-feature learning-object detection" chain. This results in a mismatch between the augmented samples and the network structure in terms of statistical properties and representation capabilities, which restricts the improvement of overall performance.

[0067] To address the three key issues mentioned above, this application proposes a generative data augmentation method for ultrasonic B-scan data of EMU wheel axles and a collaborative design scheme of a multi-scale dynamic attention target detection network. By simultaneously addressing the data and model layers, this approach systematically improves the robustness and accuracy of intelligent wheel axle defect identification.

[0068] refer to Figure 1 One embodiment of this application proposes an enhanced detection method for wheel and axle defects, which may specifically include the following steps.

[0069] Step 1: Perform data standardization and basic enhancement processing on the multi-channel B-scan images of the wheel axle using ultrasonic phased array to obtain standardized B-scan images. Step 1 first uses a phased array ultrasonic testing system to perform a full-circumference scan of the wheel axle to acquire raw multi-channel B-scan data, which is then standardized.

[0070] (1) Unified data representation

[0071] Specifically, a phased array ultrasonic testing system is used to perform a full-circumference scan of the train's wheel axles, acquiring multi-channel B-scan images covering the rim, flange, spokes, and axial transition zone. The raw B-scan data undergoes time-axis (sound path / depth) correction and channel-axis (circumferential / position) registration, uniformly mapping B-scan data from different devices and sampling parameters to a fixed depth range and number of channels. Through resampling and interpolation, the B-scan images are standardized into a uniform resolution two-dimensional matrix representation, and amplitude normalization, logarithmic compression, or pseudo-color encoding are performed to provide standardized data for subsequent network input. This eliminates equipment and process differences, improving the model's adaptability across different devices.

[0072] More specifically, the original data undergoes time-axis (sound path axis) correction to eliminate depth deviations caused by differences in sound velocity or delay settings; channel-axis (circumferential axis) registration is performed to ensure alignment of each channel. Through bilinear interpolation resampling, all B-scan images are normalized into a two-dimensional grayscale matrix of 256 pixels (depth) × 512 pixels (channels). Finally, the amplitude is normalized to ensure its distribution range is between [0, 1].

[0073] (2) Basic data augmentation

[0074] Specifically, on a uniformly represented B-scan image, geometric and structural enhancement operations such as random translation, cropping, scaling, flipping, partial occlusion, and channel discarding are employed to simulate probe position deviation, scanning step distance disturbances, and partial channel failures. Appropriate gain perturbations and noise injection are introduced in the amplitude domain to simulate observational differences under different gain settings and electrical noise levels, enhancing the model's robustness to changes in field conditions.

[0075] More specifically, the standardized image is horizontally flipped with a 50% probability to simulate reverse scanning; a 224×224 region is randomly cropped to simulate probe position deviation; and data from 1-2 channels is randomly discarded to simulate channel failure. In the amplitude domain, a gain perturbation of ±3 dB is randomly applied, and Gaussian white noise is injected to simulate different signal-to-noise ratios. This improves the model's adaptability to on-site disturbances.

[0076] Step 2: Input the standardized B-scan image into a generative data augmentation module that incorporates ultrasound mechanisms to generate an enhanced B-scan image, thereby expanding the training data and reconstructing the high-noise image. Step 2 aims to generate scarce defect samples and improve the quality of the input image.

[0077] (1) Generative data augmentation module combining ultrasound mechanism

[0078] Specifically, based on the wheel and axle geometry and ultrasonic propagation mechanism, simulation models are constructed for typical defect types (such as internal cracks in the wheel rim, defects at the wheel flange root, and discontinuities near the spoke holes). By adjusting parameters such as defect size, depth, orientation, and incident angle, a series of B-scan simulation segments with defects are generated. After being aligned with the real acquired data, these segments are added to the training set to cover the parameter space that is difficult to completely cover by real sampling.

[0079] More specifically, an ultrasonic propagation model of an internal crack in a wheel rim was constructed based on the finite element method. Parameters such as crack length (1mm-10mm), angle (0°-60°), and depth (5mm-50mm) were adjusted to generate 5000 sets of precisely labeled simulated B-scan image fragments.

[0080] (2) Deep Generative Model Enhancement

[0081] Specifically, using a large number of real wheel and axle B-scan images as training data, a deep generative model (such as a denoising diffusion probability model or a generative adversarial network) is constructed to learn the joint distribution between normal structure echoes, defect echoes, and noise. By controlling conditions (such as defect type, depth range, channel number, and noise level), B-scan samples with specified features are synthesized from the generative model, including rare defect samples and samples from complex noise conditions. Simulated generated samples, deep generated samples, and real-world acquired samples are mixed in a certain proportion to construct a more balanced wheel and axle ultrasonic training dataset in terms of defect type, noise morphology, and operating conditions.

[0082] More specifically, a conditional denoising diffusion probability model (DDPM) is trained using 10,000 real, defect-free, and high-quality defect B-scan images as the training set. The conditional vector includes the defect category, depth range, and noise level. Through this model, 2,000 B-scan images with specified features can be synthesized, such as crack samples under the conditions of "rim root, depth 20-30 mm, high noise".

[0083] (3) Generative reconstruction and denoising

[0084] Specifically, for high-noise B-scan images acquired on-site, a "reconstruction / denoising" operation can be performed using a trained generative model to map noisy observations into reconstructed images that approximate a "clean B-scan distribution." This suppresses random noise and some spurious echoes while preserving the morphology and energy characteristics of defect echoes, providing higher-quality input data for subsequent detection networks.

[0085] More specifically, the simulated images, depth-generated images, and 5000 real-world images are mixed in a 1:1:2 ratio to construct a balanced dataset covering a wide range of working conditions. For high-noise images acquired on-site, the trained DDPM algorithm can be input for reconstruction and denoising, outputting images with higher signal-to-noise ratios and clearer defect features for subsequent detection. A comparison of the UT-DDPM algorithm's performance can be seen in Figures 2(a)-2(c). This setup significantly expands the data distribution, covering extreme working conditions and rare defects, thereby improving the model's robustness.

[0086] (1)

[0087] in, : at time step The noisy data (i.e., the B-scan image after adding noise). : at time step The data (the B-scan image from the previous step). Noise scheduling parameters control the intensity of noise addition at each step; they are typically predefined hyperparameters. ). : Random noise sampled from a standard normal distribution (used to simulate electrical noise, environmental noise, etc. in ultrasound testing). Equation (1) describes the forward process of DDPM, that is, how to gradually add noise to the original B-scan image to simulate the data degradation process.

[0088] (2)

[0089] in, The cumulative noise scheduling parameter is defined as follows: , indicating from the initial step to step The overall noise level. Raw B-scan image (clean data without noise). Random noise sampled from a standard normal distribution. Equation (2) is a closed-form version of Equation (1), allowing direct sampling from the original data. Calculate noisy data at any time step t for efficient training.

[0090] (3)

[0091] in, : Noise predicted by the neural network (i.e., noise estimation of the output of the UT-DDPM model). The variance term is usually set as follows: Equation (3) describes the reverse process of DDPM, i.e., how to extract noise data x t Gradually reconstruct denoised data x （t－1） It achieves image reconstruction or generation by predicting and removing noise.

[0092] (4)

[0093] in, Noise actually added to the data (from the forward process). The neural network predicts the noise. Equation (4) defines the training loss function of DDPM, which optimizes the network parameters by minimizing the mean square error (MSE) between the predicted noise and the actual noise. This enables the model to accurately estimate noise.

[0094] Equations (1) and (2) together describe the forward noise addition process, and equation (2) is the cumulative form of equation (1). Equation (3) performs reverse denoising based on equations (1) and (2), and relies on the model trained by equation (4). Equation (4) is the core of the training process, ensuring that the model can learn the noise distribution, thereby supporting the reverse process of Equation (3).

[0095] Step 3: Input the enhanced B-scan image into a multi-scale dynamic attention target detection network for training and inference, and output the identification and localization results of the wheel axle defect. The generative data augmentation module and the multi-scale dynamic attention target detection network are designed collaboratively to improve the accuracy and robustness of defect identification. Step 3 uses the dataset enhanced in Step 2 to train an improved YOLOv11 detection network.

[0096] (1) Backbone feature extraction network design

[0097] Specifically, a backbone network combining multi-layer convolution and residual structures is used to perform bottom-up feature extraction on the uniformly expressed and enhanced B-scan images. One-dimensional or strip convolution / attention is introduced along the acoustic path direction to model the continuity and waveform evolution of defect echoes along the depth direction; cross-channel correlation modeling is introduced along the channel direction to capture the coherent echo features of the same defect in adjacent channels.

[0098] More specifically, in the backbone network of YOLOv11, strip convolution is introduced into the shallow network to specifically capture the continuous defect echo waveform features along the depth direction; at the same time, cross-channel attention is added along the channel direction to calculate the correlation between the feature maps of adjacent channels, so as to enhance the perception of the coherent echo patterns formed by the same defect in multiple channels.

[0099] (2) Multi-scale feature fusion structure

[0100] Specifically, by integrating features at different scales through top-down and lateral connections, high-resolution details at the shallow level are preserved while introducing deeper semantic capabilities. During the fusion process, considering that wheel and axle defects often present as elongated strips with low contrast, anisotropic convolutional kernels or attention mechanisms are used for features in different vertical and horizontal directions to enhance the expressive power for elongated targets.

[0101] More specifically, an FPN+PAN structure is used for top-down and bottom-up feature fusion. During fusion, anisotropic convolutions (such as a combination of 3×1 and 1×3 convolution kernels) are used to process the feature maps from shallow (high-resolution) layers to enhance the network's ability to express the elongated strip-like morphology commonly found in wheel and axle defects.

[0102] (3) Multi-scale dynamic attention detection head

[0103] Specifically, a detection head with a uniform structure is configured on the multi-scale feature map output by the feature pyramid. A multi-scale dynamic attention module, Dy-ut-head, incorporating ultrasonic mechanisms, is designed inside the detection head.

[0104] Scale attention branch: Based on the response of feature maps at different scales, dynamically adjust the weight of each scale in the final prediction, adaptively taking into account both large and small defects.

[0105] Spatial attention branch: By modeling the spatial distribution of feature maps, it highlights the response near suspected defect areas and suppresses the interference of large background areas on the detection results.

[0106] Channel attention branch: The feature channels are recalibrated through global statistics and learnable weights, giving higher weights to channels that are highly correlated with defect patterns and weakening channels that mainly reflect interface waves and bottom waves.

[0107] More specifically, a three-way attention module is embedded in front of the detection head at each scale:

[0108] Scale attention: Global average pooling is performed on feature maps at three scales, P3, P4, and P5. Weights for each scale are generated through a small MLP network, dynamically determining which scale contributes more to the current input image.

[0109] Spatial attention: Model the spatial location of the feature map to generate an attention heatmap, allowing the network to focus on suspected defect areas (such as areas with abrupt changes in echo energy) and suppress strong structural echo backgrounds such as spoke holes and interface waves.

[0110] Channel attention: A mechanism similar to the SE module is used to compress and excite feature channels, automatically learn and increase the weight of channels that are highly correlated with defect patterns such as "cracks" and "hole edge discontinuities".

[0111] This improves sensitivity to small-sized, low-contrast defects and reduces false alarm rates.

[0112] (4) Joint training and optimization strategy

[0113] Specifically, a hybrid dataset consisting of real-world data collection, simulation generation, and deep learning augmentation was used to train a multi-scale dynamic attention detection network end-to-end. In the loss function design, higher weights were assigned to small-sized and low-contrast defect samples, and a hard-example mining strategy was introduced to enhance the model's attention to difficult-to-identify defects. Through validation set performance monitoring and hyperparameter tuning, an optimal balance was achieved between detection accuracy, recall, false positive rate, and inference efficiency.

[0114] More specifically, end-to-end training is performed using a mixed dataset. In the loss function, a weight of 2.0 is assigned to the classification loss for defect boxes smaller than 32 pixels, and Focal Loss is used to balance positive and negative samples. The optimal learning rate and data augmentation combination is determined through hyperparameter search. A schematic diagram of the Dy-ut-head object detection model architecture can be found here. Figure 3 .

[0115] (5)

[0116] in, : The region of the predicted bounding box (the defect region output by the model). : The region of the true bounding box (the defect region marked manually). Equation (5) calculates the intersection over union (IoU), which measures the degree of overlap between the predicted box and the true box. The closer the value is to 1, the more accurate the detection.

[0117] (6)

[0118] in, CIoU loss value. Complete IoU, an improved IoU metric. Equation (6) converts CIoU into a loss function for optimizing the detection network.

[0119] (7)

[0120] in, : Center point of the prediction box Center point of the real frame The Euclidean distance between them. : The diagonal length of the smallest closed rectangle that covers the predicted bounding box and the ground truth bounding box. Balance parameters, usually determined by control. : Aspect ratio consistency parameter. Equation (7) extends IoU by introducing center point distance and aspect ratio penalty terms, which can improve the accuracy of bounding box regression.

[0121] (8)

[0122] in, : The width and height of the actual bounding box. : The width and height of the predicted bounding box. Equation (8) quantifies the difference in aspect ratio between the predicted box and the ground truth box, and is used for the CIoU calculation in Equation (7).

[0123] (9)

[0124] in, The width and height of the smallest closed rectangle covering the predicted and ground truth bounding boxes. The EIoU loss further optimizes the difference in width and height, making it suitable for multi-scale defect detection.

[0125] (10)

[0126] in, True positives (the number of defects that were correctly detected). False positives (the number of defects that are falsely reported). Equation (10) calculates the accuracy, which measures the accuracy of the model in predicting positive examples.

[0127] (11)

[0128] in, False negatives (the number of defects missed). Equation (11) calculates the recall rate, which measures the model's ability to cover real defects.

[0129] (12)

[0130] in, Precision is a function of recall. Equation (12) calculates the average precision (AP), which is the area under the precision-recall curve and is used to evaluate the performance of single-class detection.

[0131] (13)

[0132] in, Number of defect categories (e.g., rim cracks, flange defects, etc.). : No. The AP values ​​for each category are calculated using Equation (13). The mean average accuracy (mAP) is then used to comprehensively evaluate the detection performance across multiple categories.

[0133] Equation (5) is the basic metric, and equations (6)-(9) are IoU-based loss function variants used for bounding box regression optimization during training. Equations (10)-(13) are evaluation metrics that depend on the model output after training, and equations (12) and (13) are calculated directly based on equations (10) and (11). These formulas together constitute the training (loss minimization) and evaluation (performance quantification) system of the object detection network.

[0134] Equations (1)-(4) of the Generative Data Augmentation Module (UT-DDPM) provide a mathematical description of the noise addition and denoising process, reflecting the integration of ultrasound mechanism in this application and enhancing the scientific nature of data augmentation. The UT-DDPM formula directly addresses the problem of "scarcity of defective samples and unbalanced data distribution" by expanding the data distribution through a generative model.

[0135] Equations (5)-(13) of the target detection network demonstrate the rigor of the loss function and evaluation metrics, proving the network's optimization capability in multi-scale defect detection. The target detection network formula addresses the problem of "insufficient adaptability between B-scan characteristics and general target detection networks" by improving the detection accuracy of slender, low-contrast defects through customized loss functions (such as CIoU / EIoU).

[0136] refer to Figure 4 By introducing a simulation sample construction and deep generation model that incorporates ultrasonic mechanisms, this application significantly expands the ultrasonic training data space for wheel axles across multiple dimensions, including defect type, defect size, location depth, and noise conditions. This substantially improves the previously severely imbalanced number and distribution of "normal / defective" samples. During training, the model is fully exposed to various typical and extreme operating condition samples, resulting in more stable performance in terms of versatility and generalization ability across equipment, processes, and sites.

[0137] refer to Figure 5 To verify the applicability and superiority of the algorithm in ultrasonic nondestructive testing, the applicant conducted a systematic comparative evaluation of the models before and after the improvement. The results show that the optimized algorithm performs better in terms of accuracy, robustness, and processing efficiency, especially demonstrating stronger adaptability in multi-scale defect identification and complex signal environments. While maintaining high detection performance, this algorithm also considers the requirements of lightweight model and rapid response, making it more suitable for widespread application in industrial scenarios such as ultrasonic testing.

[0138] The improved model effectively improves the detection rate of various probes in ultrasonic wheel flaw detection. It can not only solve defects in small targets, but also detect defects in large targets well. Furthermore, it avoids interference with the flaw judgment results caused by spoke holes, bolt holes, and press-fit interface waves in wheel axles.

[0139] Wheel performance comparison: YOLOv11 as the baseline detection network has missed detections. The target detection model based on ut-ddpm and dy-ut-head (improved YOLOv11) can identify more small target defects.

[0140] The multi-scale dynamic attention target detection network proposed in this application fully considers the physical structure and defect morphology features of wheel axle B-scan images: through longitudinal depth modeling and lateral channel correlation modeling, the network has higher sensitivity to small defect patterns that are continuous along the sound path and coherent across channels; multi-scale feature fusion and a scale-space-channel three-dimensional attention mechanism enable the network to accurately highlight low-contrast defect regions even against a background of strong structural echoes. Compared with baseline detection networks that do not adopt the structure of this application, under the same test conditions, this application significantly improves the detection rate of small-sized, weak-echo defects, while also greatly improving the performance in reducing false alarms.

[0141] Tests were conducted on a real dataset containing 6212 wheelsets, and the results are shown in Table 1. The auxiliary defect detection accuracy achieved by this application reached 95.33%, the system alarm rate was 4.67%, and the number of missed detections for out-of-limit defects was 0, demonstrating its high reliability and engineering application value. Figure 6 and Figure 7 As shown, compared with the baseline detection model YOLOv11, the method proposed in this application can effectively detect more small target defects and has better suppression ability for interfering structures such as spoke holes and bolt holes.

[0142] Table 1. Statistical table based on wheelsets

[0143]

[0144] Through actual wheel track flaw detection experiments and data analysis, the algorithm's ability to analyze defect data was tested. The results of manual analysis were compared with those from image deep learning algorithms to evaluate the application technology's flaw detection efficiency and accuracy. Example images of detected wheel surface defects can be found for reference. Figure 8 .

[0145] This application effectively overcomes key technical challenges in ultrasonic defect identification of EMU wheel axles, such as "scarcity of defect samples, complex data distribution, and mismatch between B-scan characteristics and general networks," through the collaborative design of generative data augmentation and multi-scale dynamic attention target detection networks. While ensuring engineering feasibility, it achieves a higher detection rate, a lower false alarm rate, and stronger adaptability to working conditions for wheel axle defects, and has high engineering application value and promotion significance.

[0146] Another embodiment of this application proposes a detection system capable of implementing the detection method described in any of the foregoing claims, comprising:

[0147] The data preprocessing module is used to standardize and perform basic enhancement processing on the multi-channel B-scan images of the wheel axle ultrasonic phased array to obtain standardized B-scan images.

[0148] The generative data augmentation module is used to receive the standardized B-scan image, synthesize the enhanced B-scan image based on the ultrasound mechanism and depth generation model, and reconstruct and denoise the high-noise image.

[0149] The multi-scale dynamic attention detection module is used for feature extraction, multi-scale fusion, and defect identification and localization of enhanced B-scan images.

[0150] The generative data augmentation module works in conjunction with the multi-scale dynamic attention detection module to improve the system's accuracy and robustness in identifying wheel and axle defects.

[0151] Specifically, the detection system can be integrated into an industrial computer or server, including:

[0152] Data preprocessing module: Receives raw B-scan data stream from ultrasound phased array equipment, calls normalization and basic enhancement procedures, and outputs processed images.

[0153] Generative Data Augmentation Module: Includes a pre-trained UT-DDPM model and a physics simulator. This module can work offline to expand the training dataset, or online to denoise and reconstruct real-time images from the data preprocessing module.

[0154] Multi-scale dynamic attention detection module: Loads the trained improved YOLOv11 model, receives images from the first two modules, outputs the defect category and bounding box information in real time, and displays and alarms on the human-computer interaction interface.

[0155] Compared with the prior art, this application has the following significant advantages:

[0156] 1. Fundamentally alleviates data bottleneck: By combining physical simulation with data-driven generation, the coverage of training data in terms of defect type, size, depth and noise conditions is systematically expanded, allowing the model to be fully exposed to multiple samples and multiple conditions during training, which greatly improves the model's generalization ability and cross-device adaptability.

[0157] 2. Significantly improves the detection rate and robustness of small defects: The detection network, designed specifically for the characteristics of B-scan images, especially the multi-scale dynamic attention mechanism, can effectively capture the features of slender, weak-echo defects and suppress interference in the background of strong structural echoes. Thus, while maintaining high sensitivity to small-sized, low-contrast defects, it significantly reduces the false alarm rate.

[0158] 3. End-to-end collaborative optimization is achieved: This application breaks the traditional pattern of isolated design of data augmentation and model training. Through the overall collaborative optimization of "generative augmentation - feature learning - object detection", the augmented data and network structure are highly matched in terms of statistical characteristics and representation capabilities, thereby achieving a leap in overall performance.

[0159] The aforementioned embodiments of a wheel and axle defect enhancement detection method and a wheel and axle defect enhancement detection system are technically related and can be referred to each other in terms of technical details and technical effectiveness, which will not be repeated here.

[0160] Based on the same inventive concept, embodiments of this application also provide an electronic device. Figure 9 This is a structural block diagram of an electronic device provided in an embodiment of this application. Figure 9 As shown in the embodiments of this application, an electronic device includes: one or more processors 101, a memory 102, and one or more I / O interfaces 103. The memory 102 stores one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement any of the wheel axle defect enhancement detection methods described in the above embodiments; the one or more I / O interfaces 103 are connected between the processor and the memory, configured to enable information interaction between the processor and the memory.

[0161] The processor 101 is a device with data processing capabilities, including but not limited to a central processing unit (CPU); the memory 102 is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read / write interface) 103 is connected between the processor 101 and the memory 102, and can realize information interaction between the processor 101 and the memory 102, including but not limited to a data bus (Bus).

[0162] In some embodiments, the processor 101, memory 102, and I / O interface 103 are interconnected via bus 104, and thus connected to other components of the computing device.

[0163] In some embodiments, the one or more processors 101 include a field-programmable gate array.

[0164] This application also provides a computer-readable medium. The computer-readable medium stores a computer program, which, when executed by a processor, implements the steps of any of the wheel axle defect enhancement detection methods described in the above embodiments. The computer-readable storage medium can be volatile or non-volatile.

[0165] This application also provides a computer program product, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code. When the computer-readable code is run in the processor of an electronic device, the processor in the electronic device executes the above-described wheel axle defect enhancement detection method.

[0166] Those skilled in the art will understand that all or some of the steps, systems, and apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software can be distributed on a computer-readable storage medium, which may include computer storage media (or non-transitory media) and communication media (or transient media).

[0167] As is known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable program instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), flash memory or other memory technologies, portable compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is known to those skilled in the art that communication media typically contain computer-readable program instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0168] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0169] The computer program instructions used to perform the operations of this application may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), are personalized by utilizing the status information of the computer-readable program instructions. These electronic circuits can execute the computer-readable program instructions to implement various aspects of this application.

[0170] The computer program product described herein can be implemented specifically through hardware, software, or a combination thereof. In one alternative embodiment, the computer program product is specifically embodied in a computer storage medium; in another alternative embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0171] Various aspects of this application are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0172] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0173] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0174] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0175] Exemplary embodiments have been disclosed herein, and while specific terminology has been used, it is used and should be interpreted only in a general illustrative sense and is not intended to be limiting. In some embodiments, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of this application as set forth by the appended claims.

Claims

1. A method for enhanced detection of wheel and axle defects, characterized in that, include: Data standardization and basic enhancement processing were performed on the multi-channel B-scan images of the wheel axle ultrasonic phased array to obtain standardized B-scan images; The standardized B-scan images are input into a generative data augmentation module that incorporates ultrasound mechanisms to generate enhanced B-scan images, thereby expanding the training data and reconstructing high-noise images. The generative data augmentation module synthesizes enhanced B-scan images based on ultrasound mechanisms and a depth generation model. Based on wheel and axle geometry and ultrasound propagation mechanisms, simulation models for typical defect types are constructed. By adjusting defect parameters, a series of B-scan simulation segments with defects are generated. The deep generation model is a denoising diffusion probability model, which learns the joint distribution between normal structure echoes, defect echoes and noise; through conditional control, B-scan samples with specified features are synthesized from the deep generation model, including rare defect samples and complex noise condition samples. The enhanced B-scan image is input into a multi-scale dynamic attention target detection network for training and inference, outputting the identification and localization results of wheel axle defects; the multi-scale dynamic attention target detection network includes: The backbone feature extraction network uses strip convolution and cross-channel correlation modeling to extract depth and channel features from B-scan images; A multi-scale feature fusion structure integrates shallow details and deep semantic information, enhancing the ability to express slender defects; A multi-scale dynamic attention detection head dynamically adjusts scale, spatial, and channel weights to highlight defect areas and suppress background interference; the multi-scale dynamic attention detection head specifically includes: Scale attention branch dynamically adjusts the weights of features at each scale in the prediction; Spatial attention branch highlights suspected defect areas and suppresses background response; Channel attention branch, recalibrate feature channels, and strengthen channels related to defect patterns; The generative data augmentation module and the multi-scale dynamic attention target detection network are designed in collaboration to improve the accuracy and robustness of defect identification. The detection method is applicable to the detection of one or more defects in the rim, flange, spoke, and axial transition zone of EMU wheel axles.

2. The detection method according to claim 1, characterized in that, The construction logic of the generative data augmentation module includes: A defect simulation model was constructed based on the wheel and axle geometry and ultrasonic propagation mechanism, and a B-scan simulation segment with defects was generated. The joint distribution of real B-scan images is learned using a deep generative model, and B-scan samples with specified features are synthesized through conditional control. A balanced training dataset is constructed by mixing simulated samples, deep learning samples, and real-world samples.

3. The detection method according to claim 1, characterized in that, The data standardization includes: The B-scan images are corrected for time axis, registered for channel axis, resampled, and normalized for amplitude, and then unified into a two-dimensional matrix with a fixed resolution.

4. The detection method according to claim 1, characterized in that, The basic enhancements include: Perform random translation, cropping, flipping, and / or channel discarding operations in the image domain; Gain perturbation and noise injection are performed in the amplitude domain to simulate changes in on-site operating conditions.

5. The detection method according to claim 1, characterized in that, The detection method further includes: During training, higher loss weights are assigned to pre-defined small-sized defects and pre-defined low-contrast defect samples, and a hard example mining strategy is introduced.

6. A detection system capable of implementing the detection method according to any one of claims 1-5, characterized in that, include: The data preprocessing module is used to standardize and perform basic enhancement processing on the multi-channel B-scan images of the wheel axle ultrasonic phased array to obtain standardized B-scan images; The generative data augmentation module is used to receive the standardized B-scan image, synthesize the enhanced B-scan image based on the ultrasound mechanism and depth generation model, and reconstruct and denoise the high-noise image; Based on the wheel and axle geometry and ultrasonic propagation mechanism, a simulation model for typical defect types is constructed; by adjusting the defect parameters, a series of B-scan simulation segments with defects are generated; the deep generation model is a denoising diffusion probability model, which learns the joint distribution between normal structure echo, defect echo and noise; through condition control, B-scan samples with specified characteristics are synthesized from the deep generation model, including rare defect samples and complex noise condition samples. The multi-scale dynamic attention detection module is used for feature extraction, multi-scale fusion, and defect identification and localization of the enhanced B-scan image; The multi-scale dynamic attention detection module includes: The backbone feature extraction network uses strip convolution and cross-channel correlation modeling to extract depth and channel features from B-scan images; A multi-scale feature fusion structure integrates shallow details and deep semantic information, enhancing the ability to express slender defects; A multi-scale dynamic attention detection head dynamically adjusts scale, spatial, and channel weights to highlight defect areas and suppress background interference; the multi-scale dynamic attention detection head specifically includes: Scale attention branch dynamically adjusts the weights of features at each scale in the prediction; Spatial attention branch highlights suspected defect areas and suppresses background response; Channel attention branch, recalibrate feature channels, and strengthen channels related to defect patterns; The generative data augmentation module and the multi-scale dynamic attention detection module work together to improve the system's accuracy and robustness in identifying wheel and axle defects. The detection system is applicable to the detection of one or more defects in the rim, flange, spoke, and axial transition zone of EMU wheel axles.

Images