High-speed rail wheel macroscopic defect detection method and system based on deep learning

Abstract

The invention discloses a high-speed rail wheel macroscopic defect detection method and system based on deep learning. The method comprises the following steps: constructing an improved YOLOv11 target detection model; the improved YOLOv11 target detection model is deployed on a wheel production line, real-time image acquisition and defect detection are carried out, and a detection result is obtained; performing real-time classification and positioning based on a detection result to obtain defect information; wherein the improved YOLOv11 target detection model is obtained after training of a training set, the training set is a wheel surface image macroscopic defect data set, and the improved YOLOv11 target detection model is formed by introducing a multi-scale convolution MSGC module and an improved downsampling LightSEDown module into a backbone network and adding an SE attention mechanism. According to the invention, the detection efficiency and safety in the rail transit wheel production process are improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent nondestructive testing and computer vision technology, and in particular to a method and system for detecting macroscopic defects in high-speed rail wheels based on deep learning. Background Technology

[0002] With the continuous development of rail transit technology, rail transit wheels, as one of the key load-bearing components, are directly related to the operational safety of trains due to their defects. During long-term operation, wheels are susceptible to various factors, resulting in defects such as cracks, rust, dents, and black scale. If these defects are not detected and addressed in a timely manner, they may lead to serious safety accidents. Therefore, defect detection of rail transit wheels has become an important technical means to improve train safety.

[0003] Traditional methods for detecting defects in rail transit wheels mainly include manual visual inspection and automated inspection methods based on traditional image processing technology. While manual visual inspection offers advantages in flexibility and low cost, its results are easily affected by human factors, resulting in poor accuracy and low efficiency. Automated inspection methods based on image processing, although improving efficiency, suffer from uncertainties in the industrial environment. Their accuracy is easily affected by complex backgrounds and changes in lighting, leading to false positives and false negatives, and they struggle to meet the real-time and high-precision requirements of industrial applications.

[0004] With the development of deep learning technology, deep learning-based object detection methods have made significant progress in the field of image detection. Although the existing YOLO model has achieved excellent performance in object detection, it still faces the challenge of balancing computational efficiency and detection accuracy when dealing with defects in rail transit wheels.

[0005] Therefore, there is an urgent need for an intelligent detection technology that is highly accurate, fast-responding, and adaptable to meet the requirements of real-time detection of defects in rail transit wheels. Summary of the Invention

[0006] To address the problems of low detection accuracy, poor real-time performance, and low automation in existing wheel defect detection methods, this invention provides a deep learning-based method and system for detecting macroscopic defects in high-speed rail wheels. Compared with existing methods, this method offers higher detection accuracy and better real-time performance.

[0007] On the one hand, to achieve the above objectives, this invention provides a method for detecting macroscopic defects in high-speed rail wheels based on deep learning, comprising:

[0008] Construct an improved YOLOv11 object detection model;

[0009] The improved YOLOv11 target detection model was deployed on the wheel production line to perform real-time image acquisition and defect detection, and the detection results were obtained.

[0010] Based on the detection results, real-time classification and localization are performed to obtain defect information;

[0011] The improved YOLOv11 target detection model is obtained by training on a training set, which is a macroscopic defect dataset of wheel surface images. The improved YOLOv11 target detection model is composed of a multi-scale convolutional MSGC module and an improved downsampling LightSEDown module introduced into the backbone network, and an SE attention mechanism added.

[0012] Preferably, constructing the macroscopic defect dataset of the wheel surface image includes:

[0013] Image data of the wheel surface is acquired and stitched together to obtain the original dataset;

[0014] The Cutout method is used to randomly occlude image regions in the original dataset, and the AutoAugment strategy is used to automatically select image transformation strategies for initial enhancement. Then, the Random Erasing method is used to randomly select some regions of the image for erasure and replacement to generate image samples, expand the original dataset, and obtain the enhanced dataset.

[0015] The enhanced dataset was annotated using LabelImg annotation software to mark the surface defect areas of the medium-thick plate, thus obtaining the macroscopic defect dataset of the wheel surface image.

[0016] Preferably, the improved YOLOv11 target detection model includes:

[0017] Input module: Used to receive and preprocess input images;

[0018] Improved backbone network: for extracting multi-scale features from images through the multi-scale convolutional MSGC module and the improved downsampling LightSEDown module;

[0019] Feature fusion module: used to fuse extracted multi-scale features;

[0020] Decoupling head: Used to classify and locate targets based on the fused features.

[0021] Preferably, the multi-scale convolution MSGC module uses multi-scale depthwise separable convolution to divide the original feature map into two parts. The first part is processed by convolution, and the second part retains the original features. The original features are fused with the features extracted by multi-scale convolution, and channel information fusion is achieved by performing pointwise convolution through 1×1 convolution.

[0022] Preferably, the improved downsampling LightSEDown module adopts a channel segmentation mechanism to divide the original feature map into a third part and a fourth part, and divides it into processing path one and processing path two in the channel dimension. The segmented feature maps are processed separately, and finally the output results of processing path one and processing path two are spliced ​​and fused.

[0023] The first processing path is used to first perform illumination normalization on the third part of the feature map, then use multi-scale convolution to extract features, and introduce the SE attention mechanism to enhance the weight of important feature channels.

[0024] The second processing path is used to perform convolution and max pooling operations on the fourth part of the feature map, preserving the original details and reducing the spatial dimension.

[0025] Preferably, the SE attention mechanism performs two steps: compression and activation. In the compression stage, the SE attention mechanism first compresses the spatial dimension of the input feature map through a global average pooling operation to obtain global information for each channel. In the activation stage, the SE attention mechanism recalibrates the channel features through a fully connected layer and outputs the weight coefficient of each channel through a Sigmoid activation function. Finally, the weight coefficient of each channel is multiplied channel by channel with the original input feature map to complete the recalibration of the feature map.

[0026] Preferably, the processing procedure of the compression stage is as follows:

[0027] ;

[0028] In the formula, This represents the pixel value in the i-th row, j-th column, and c-th channel of the input feature map; This is the global average pooling result for channel c; H and W are the height and width of the input feature map, respectively.

[0029] The processing procedure for the incentive phase is as follows:

[0030] ;

[0031] In the formula, z is a C-dimensional vector obtained through global average pooling; and These are the weight matrices of the two fully connected layers; It is the ReLU activation function; It is the Sigmoid activation function; Weights for each channel;

[0032] The weight coefficient of each channel is multiplied channel by channel by the original input feature map to obtain:

[0033] ;

[0034] In the formula, The feature map is processed by the SE attention mechanism. It is the weight of channel c.

[0035] On the other hand, to achieve the above objectives, the present invention also provides a deep learning-based macroscopic defect detection system for high-speed rail wheels, used to implement a deep learning-based method for detecting macroscopic defects in high-speed rail wheels, comprising:

[0036] The model building unit is used to build an improved YOLOv11 object detection model. The improved YOLOv11 object detection model is obtained by training on a training set, which is a macroscopic defect dataset of wheel surface images. The improved YOLOv11 object detection model is composed of a multi-scale convolutional MSGC module and an improved downsampling LightSEDown module introduced into the backbone network, and an SE attention mechanism added.

[0037] The deployment execution unit is used to deploy the improved YOLOv11 target detection model to the wheel production line to perform real-time image acquisition and defect detection, obtain detection results, perform real-time classification and localization based on the detection results, and obtain defect information.

[0038] Information aggregation unit: used to feed back the defect information to the host computer in real time, provide defect type and severity data, and record and store historical detection results.

[0039] Compared with the prior art, the present invention has the following advantages and technical effects:

[0040] (1) The present invention deploys the improved YOLOv11 target detection model on the wheel production line for intelligent detection. Compared with the existing manual and traditional vision equipment detection methods, this method can provide higher detection accuracy under complex lighting and environmental conditions and has real-time detection function, which can significantly improve the detection efficiency and safety in the production process of rail transit wheels.

[0041] (2) While improving the detection accuracy and response speed, the present invention significantly enhances the adaptability and intelligence of the system, has good industrial application prospects, and is of great significance to the quality control and safety assurance of rail transit wheel manufacturing and maintenance. Attached Figure Description

[0042] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0043] Figure 1 This is a flowchart of a deep learning-based macroscopic defect detection method for high-speed rail wheels according to an embodiment of the present invention.

[0044] Figure 2 This is a basic architecture diagram of the YOLOv11 network object detection model used in the embodiments of the present invention;

[0045] Figure 3 This is a structural diagram of the MSGC module according to an embodiment of the present invention;

[0046] Figure 4 This is a structural diagram of the LightSEDown module according to an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the layout of the inspection section of the wheel production line according to an embodiment of the present invention;

[0048] Figure 6 This is a schematic diagram of the on-site identification results according to an embodiment of the present invention;

[0049] Among them, 10 is a wheel bracket; 20 is a strip light source; 30 is an industrial camera; 40 is a ring light source; 50 is a wheel; 60 is a wheel fixing pulley; and 70 is a U-shaped groove. Detailed Implementation

[0050] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0052] This embodiment proposes a deep learning-based method for detecting macroscopic defects in high-speed train wheels, such as... Figure 1 ,include:

[0053] Construct an improved YOLOv11 object detection model;

[0054] The improved YOLOv11 target detection model was deployed on the wheel production line to perform real-time image acquisition and defect detection, and the detection results were obtained.

[0055] Based on the detection results, real-time classification and localization are performed to obtain defect information;

[0056] The improved YOLOv11 target detection model is obtained by training on a training set, which is a macroscopic defect dataset of wheel surface images. The improved YOLOv11 target detection model is composed of a multi-scale convolutional MSGC module and an improved downsampling LightSEDown module introduced into the backbone network, and an SE attention mechanism added.

[0057] Furthermore, constructing the macroscopic defect dataset of the wheel surface images includes:

[0058] Image data of the wheel surface is acquired and stitched together to obtain the original dataset;

[0059] The Cutout method is used to randomly occlude image regions in the original dataset, and the AutoAugment strategy is used to automatically select image transformation strategies for initial enhancement. Then, the Random Erasing method is used to randomly select some regions of the image for erasure and replacement to generate image samples, expand the original dataset, and obtain the enhanced dataset.

[0060] The enhanced dataset was annotated using LabelImg annotation software to mark the surface defect areas of the medium-thick plate, thus obtaining the macroscopic defect dataset of the wheel surface image.

[0061] Specifically, Hikrobot MV series cameras are used to acquire image data of defects on the wheel surface. Image acquisition cabinets are set up on the production line to acquire image data of the wheel surface and stitch them together to obtain the original dataset.

[0062] The acquired images are preprocessed, and the Cutout method is used to randomly occlude image regions to generate new samples, thereby achieving image diversification, enhancing the diversity of the dataset, and completing the initial stage of data augmentation.

[0063] The AutoAugment strategy automatically selects a set of optimal image transformation strategies (such as random rotation, color jitter, and shearing transformation) to perform a second-step augmentation on the dataset, thereby improving its robustness.

[0064] The Random Erasing method is used to randomly select a portion of the image for erasing and replacement, thereby generating more challenging image samples to further expand the dataset and complete the third step of image enhancement.

[0065] The LabelImg annotation software was used to accurately annotate the enhanced image dataset, marking the defect areas on the surface of medium and thick plates, and finally completing the construction of the macroscopic defect dataset of wheel surface images.

[0066] Two types of cameras are commonly used on wheel production lines: area scan cameras and line scan cameras. While line scan cameras can achieve higher image resolution and clarity in high-speed motion scenarios, their black-and-white images can cause wheel defects to blend into the background. Therefore, color area scan cameras are more suitable for macroscopic defect detection of rail wheels. In this embodiment, the initial image defect categories are mainly divided into: black skin, scars, scales, gum, and rust.

[0067] The Cutout image data augmentation method generates new training samples by randomly selecting rectangular regions in an image and occluding them (usually replacing them with zeros or constant values), thus completing the first step of image data augmentation. This method improves the model's robustness and generalization ability by reducing its dependence on local features, encouraging it to learn more generalized global features. The Cutout method enhances the model's adaptability to partially missing data by introducing local information loss, helping to prevent overfitting.

[0068] The AutoAugment method, an automated image data augmentation approach, enhances the dataset by searching for optimal image transformation strategies (such as random rotation, color jitter, and shearing) to complete the second step of image data augmentation. Reinforcement learning algorithms are used to automatically select the best data augmentation operation, thereby improving image diversity and model robustness. By selecting appropriate transformation strategies, AutoAugment effectively improves the model's generalization ability, reduces overfitting, and enhances the model's performance in different scenarios.

[0069] The Random Erasing image data augmentation method generates challenging image samples by randomly selecting partial regions of the image for erasure and replacing these regions with specified values ​​(such as zero or mean). This completes the third step of the image data augmentation method. This method simulates the loss of local information in the image, forcing the model to pay more attention to other parts of the image and improving its robustness to missing or incomplete information.

[0070] Furthermore, the improved YOLOv11 object detection model includes:

[0071] Input module: Used to receive and preprocess input images;

[0072] Improved backbone network: for extracting multi-scale features from images through the multi-scale convolutional MSGC module and the improved downsampling LightSEDown module;

[0073] Feature fusion module: used to fuse extracted multi-scale features;

[0074] Decoupling head: Used to classify and locate targets based on the fused features.

[0075] As attached Figure 2 As shown, the YOLOv11 basic model mainly includes: Input module, Improved Backbone network, Feature Fusion module, and Decoupling Head.

[0076] Input: A dataset consisting of RGB three-channel images with a size of 640*640;

[0077] Improved backbone network: Feature extraction network, based on the original Darknet-53 network with some improvements, is used to extract information from the input dataset images, as shown in the appendix. Figure 3 As shown, the MGSC module in this embodiment uses multi-scale depthwise separable convolution to divide the original feature map into two parts. One part is processed by convolution while the other part is not processed. Then, the module retains the original features and fuses them with the features extracted by multi-scale convolution. Finally, it performs pointwise convolution through 1×1 convolution to achieve channel information fusion and flexibly adjust the number of output channels.

[0078] The Conv convolution consists of three parts: a 2D convolution, a 2D batchnorm, and a SiLU activation function. The residual module uses an improved C3k2 module, which obtains more gradient flow information while ensuring lightweighting. Finally, an SPPF module is added, which consists of a Conv convolution module and three Maxpooling modules. The feature map without Maxpooling and the feature map obtained after each additional Maxpooling are concatenated to achieve feature fusion.

[0079] Feature Fusion Module: This module fuses the extracted image information. A PAFPN feature extraction architecture was designed, referencing the Feature Pyramid Network (FPN). The FPN operates top-down, passing down strong semantic features from higher layers. PAN adds a bottom-up pyramid structure after the FPN, supplementing it by passing up strong localization features from lower layers. Upsampling is performed first, followed by downsampling, with two cross-layer fusion connections between these two branches.

[0080] Decoupling head: as attached Figure 3 As shown, the current mainstream decoupled-head structure is adopted to separate the classification and detection heads.

[0081] Furthermore, such as Figure 4The improved downsampling LightSEDown module adopts a channel segmentation mechanism to divide the original feature map into a third part and a fourth part, and further divides it into processing path one and processing path two in the channel dimension. The segmented feature maps are processed separately, and finally the output results of processing path one and processing path two are spliced ​​and fused.

[0082] The first processing path is used to first perform illumination normalization on the third part of the feature map, then use multi-scale convolution to extract features, and introduce the SE attention mechanism to enhance the weight of important feature channels.

[0083] The second processing path is used to perform convolution and max pooling operations on the fourth part of the feature map, preserving the original details and reducing the spatial dimension.

[0084] Specifically, the improved downsampling LightSEDown module is used to further enhance the robustness and efficiency of the model in complex industrial inspection environments. This module integrates adaptive average pooling and illumination normalization strategies, effectively enhancing the model's resistance to external illumination changes and background noise (such as rust, oil stains, and other interference factors). While compressing the feature map space size and reducing computational costs, it ensures the stability and recognizability of wheel defect features under different illumination conditions.

[0085] The LightSEDown module employs a channel splitting mechanism, dividing the input feature map into two equal parts along the channel dimension to avoid parameter redundancy caused by channel expansion in traditional convolutional structures, and then processing each part using differentiated strategies:

[0086] Part 1 (x1): Introducing 3×3 and 5×5 heterogeneous convolutional kernels to construct a multi-scale receptive field enhances the detection capability for wheel defects of different sizes (such as micro-cracks and large-area damage). Features from multi-scale convolutions are concatenated and fused to effectively improve the richness and hierarchy of feature representation, while reducing model complexity and avoiding the parameter accumulation effect caused by channel expansion in a single-path structure. This path further integrates the Squeeze-and-Excitation (SE) attention mechanism, which dynamically enhances salient feature channels related to defects through global average pooling and channel weight adjustment, suppressing background interference signals, thereby significantly improving the model's recognition performance for complex targets such as blurred boundaries and small-scale defects.

[0087] The second part (x2) employs a 3×3 convolution combined with max pooling with a stride of 2 to downsample the input feature map, and then further adjusts the channel dimensions via a 1×1 convolution. Max pooling, as a zero-parameter, efficient operator, can preserve key local features without increasing computational cost, improving the model's responsiveness to abnormal regions. Finally, the feature maps from the first and second parts are concatenated and fused along the channel dimensions to output an enhanced feature map with multi-scale structural features, while also retaining semantic information and detail preservation capabilities. This significantly improves the overall model's feature representation ability, robustness, and real-time performance in object detection tasks.

[0088] Furthermore, the Squeeze-and-Excitation (SE) attention mechanism is a lightweight module designed to improve the performance of Convolutional Neural Networks (CNNs) by adaptively adjusting channel feature responses. The SE attention mechanism models the dependencies between feature channels by introducing an attention mechanism, thereby enhancing the network's focus on key information and suppressing unimportant features. The core idea of ​​the SE module is two steps: "Squeeze" and "Excitation." In the compression stage, the SE attention mechanism first performs spatial dimension compression on the input feature map using Global Average Pooling (GAP) to obtain the global information for each channel. Assuming the size of the input feature map is H×W×C (where H and W are the height and width, respectively, and C is the number of channels), the GAP operation compresses the spatial information of each channel into a single value, forming a 1×1×C vector. This vector represents the global feature information of each channel, expressed by the formula:

[0089] ;

[0090] In the formula, This represents the pixel value in the i-th row, j-th column, and c-th channel of the input feature map; It is the global average pooling result of channel c; H and W are the height and width of the input feature map, respectively.

[0091] During the activation phase, the SE attention mechanism recalibrates channel features through a fully connected (FC) layer. The FC layer maps the channel features to a smaller dimension (typically a reduction factor of the input dimension), followed by a non-linear transformation using the ReLU activation function. Next, another FC layer maps the features back to the original channel dimension, and a sigmoid activation function outputs the weight coefficient for each channel. This weight coefficient represents the importance of each channel, reflecting the degree of attention the network pays to different features at different layers.

[0092] The mathematical form of the stimulus module is as follows:

[0093] ;

[0094] In the formula, It is a C-dimensional vector obtained through global average pooling; and These are the weight matrices of the two fully connected layers; It is the ReLU activation function; It is the Sigmoid activation function; The weight for each channel.

[0095] Finally, the SE attention mechanism recalibrates the feature map by multiplying the weights s of each channel with the original input feature map channel by channel. This process adjusts the response of each channel, enhancing the expression of key features and suppressing unimportant features.

[0096] The formula is expressed as:

[0097] ;

[0098] In the formula, The feature map is processed by the SE attention mechanism. It is the weight of channel c.

[0099] By inserting the aforementioned modules into the YOLOv11 model's downsampling network architecture, the weights of each channel are adaptively adjusted, increasing the network's focus on key features and suppressing the influence of irrelevant features. Compared to convolutional operations, this method has lower computational overhead and fewer parameters, resulting in a significant performance improvement without a substantial increase in computational cost.

[0100] Furthermore, based on the parameter size of the processed wheel dataset, a specific target detection model of the YOLOv11 series algorithm was selected, and a training strategy was determined.

[0101] The optimizer is set to AdamW, the base learning rate to 0.001, the base weight decay to 0.01, the optimizer momentum to 0.9, the batch size to 64, the learning rate schedule to cosine annealing, the training epochs to 300, the warmup iterations to max(500, 2*iters_per_epochs), the input size to 640*640, and the exponential moving average decay (EMA) to 0.999. The pre-trained weights are then loaded and the dataset is trained.

[0102] To address the macroscopic defects on wheel surfaces, this paper improves the model by replacing the convolutional module in the Bottleneck module of the C3K2 architecture with a Multi-Scale Group Convolution (MSGC) module and introducing depthwise separable convolution to reduce the number of parameters. It also improves the backbone network by designing a novel downsampling module, Lightweight Squeeze-and-ExcitationDownsampling (LightSEDown), referencing YOLOv9 and Ghostnet. Furthermore, it introduces SE attention mechanism, adaptive average pooling, illumination normalization techniques, and channel segmentation strategies to improve the detection accuracy of YOLOv11, reduce latency, and increase detection efficiency.

[0103] The YOLOv11 series algorithm network models are divided into yolov11-n, yolov11-s, yolov11-m, yolov11-l and yolov11-x according to the number of parameters from smallest to largest. The initial weights are also arranged sequentially according to the detection task. Due to the large number of parameters for detecting multiple types of surface defects on wheels, the yolov11-m model is adopted, and the corresponding yolov11-m initial weights are loaded to speed up the training.

[0104] This embodiment also provides a deep learning-based macroscopic defect detection system for high-speed rail wheels, used to implement a deep learning-based method for detecting macroscopic defects in high-speed rail wheels, including:

[0105] The model building unit is used to build an improved YOLOv11 object detection model. The improved YOLOv11 object detection model is obtained by training on a training set, which is a macroscopic defect dataset of wheel surface images. The improved YOLOv11 object detection model is composed of a multi-scale convolutional MSGC module and an improved downsampling LightSEDown module introduced into the backbone network, and an SE attention mechanism added.

[0106] The deployment execution unit is used to deploy the improved YOLOv11 target detection model to the wheel production line to perform real-time image acquisition and defect detection, obtain detection results, perform real-time classification and localization based on the detection results, and obtain defect information.

[0107] Information aggregation unit: used to feed back the defect information to the host computer in real time, provide defect type and severity data, and record and store historical detection results.

[0108] Specifically, such as Figure 5 An improved YOLOv11 target detection model is deployed on a wheel production line. Wheels 50 are placed on wheel supports 10, and their movement and stopping are controlled by motors. A U-shaped groove 70 is installed at the bottom of the wheel support 10. An industrial camera 30, a strip light source 20, and a ring light source 40 are fixed to the U-shaped groove 70. A robotic arm moves along a bridge beam to connect different workstations on the production line, loading and unloading wheels. An air-cooled unit reduces the impact of the surrounding environment on detection. A server cabinet houses workstations to meet the computing power requirements of real-time detection. An electrical control cabinet provides unified switching for the production line's electrical systems and enables emergency shutdowns to ensure the safety of personnel and equipment. Wheels are first loaded by the robotic arm, then automatically identified by the host computer control system, and finally unloaded by the robotic arm. This process repeats, achieving real-time, high-precision classification and location of wheel surface defects on the industrial production line. The area array camera uses Hikrobot MV series cameras. The number of line array cameras is determined according to the production line width and the optimal field of view of the area array camera. The strip light source 20 and the ring light source 40 are installed on the concave groove. All camera light sources are equipped with dustproof lenses to ensure the quality of the acquired wheel surface defect images.

[0109] The defects in the medium and heavy plate images acquired by the industrial camera 30 are mainly classified as: black skin, rust, cracks, scratches, glue, and scale. When black skin, rust, cracks, scratches, glue, or scale are detected in the defect image, the host computer classifies and locates the defects.

[0110] The information aggregation unit feeds back the detected defect information to the host computer system in real time, providing detailed data such as defect type and severity, and records historical detection results through the storage module to provide data support for subsequent analysis and decision-making.

[0111] To verify the effectiveness and advancement of the improved YOLOv11 object detection model (hereinafter referred to as ML-YOLO) proposed in this embodiment, a set of comparative experiments was designed to systematically evaluate the performance of this model against current mainstream object detection algorithms. The comparative models include: Faster R-CNN, RT-DETR, SSD, YOLOv5n, YOLOv8n, YOLOv10n, the YOLOv11 series (n / base, m, s), YOLOv12n, etc., covering classic detection frameworks of different sizes and representativeness.

[0112] Each model was trained and evaluated on a unified dataset. The main comparison metrics included: precision (P), recall (R), mAP@50, mAP@50–95, number of parameters (in millions), and computational complexity (GFLOPs). The test results are shown in Table 1.

[0113] Table 1

[0114]

[0115] As can be seen from Table 1, the ML-YOLO model proposed in this embodiment achieves the best results in several key performance indicators, as detailed below:

[0116] In terms of detection accuracy, ML-YOLO achieved the highest values ​​in P (precision), R (recall), mAP@50, and mAP@50–95. Specifically, it improved upon the baseline model YOLOv11n by 7.00%, 2.76%, 5.28%, and 2.24%, respectively, fully demonstrating the accuracy advantage of the proposed improved structure while maintaining a lightweight design.

[0117] In terms of model complexity and computational efficiency: ML-YOLO has 1.88M parameters, which is 27.13% less than YOLOv11n; its GFLOPs are 5.1, which is 19.05% less than YOLOv11n, significantly optimizing the model's inference efficiency and effectively supporting edge deployment and real-time detection tasks.

[0118] Compared to YOLOv5n, which is known for its lightweight real-time performance, ML-YOLO achieves a 4.7% improvement in mAP@50 while having lower computational complexity and an inference speed improvement of approximately 46.4%, further reducing the hardware load while maintaining high performance.

[0119] Furthermore, compared with mainstream detectors such as RT-DETR, Faster R-CNN, and SSD, ML-YOLO demonstrates superior performance across all metrics, exhibiting a good balance between accuracy and efficiency, as well as excellent model generalization performance. It has broad application prospects in scenarios with high real-time and high reliability requirements, such as high-speed rail wheel defects.

[0120] like Figure 6 As shown, the defects from left to right are: black skin, scratches, glue, and rust. Figure 6 The first row shows the original samples of different categories, the second row shows the manually labeled samples, the third row shows the detection performance of the YOLOv11 algorithm, and the last row shows the test results of ML-YOLO. Clearly, both YOLOv11 and ML-YOLO algorithms have good performance and can basically complete the type identification and localization of defects. However, the confidence level of ML-YOLO detection is generally higher than that of YOLOv11. Furthermore, ML-YOLO can compensate for some missed detections by YOLOv11. For example, YOLOv11 cannot accurately detect the images in the second and fourth columns, but ML-YOLO can (red boxes in the figure).

[0121] In summary, this embodiment provides a method for detecting surface defects in rail transit wheels. It applies the YOLOv11 target detection model to the industrial wheel manufacturing process, improves the backbone network, and addresses the impact of complex backgrounds, lighting variations, and boundary blurring on detection performance through multi-scale feature extraction, illumination normalization, and attention mechanisms. This significantly reduces computational complexity while maintaining high accuracy. The system features real-time detection, automatic alarm, and data storage capabilities, making it suitable for defect detection in industrial wheels and widely applicable to wheel quality monitoring in the rail transit industry. This invention effectively improves the automation and intelligence levels of wheel inspection, providing an efficient and reliable solution for the rail transit industry with significant social and economic value.

[0122] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting macroscopic defects in high-speed train wheels based on deep learning, characterized in that, include: Construct an improved YOLOv11 object detection model; The improved YOLOv11 target detection model was deployed on the wheel production line to perform real-time image acquisition and defect detection, and the detection results were obtained. Based on the detection results, real-time classification and localization are performed to obtain defect information; The improved YOLOv11 target detection model is obtained by training on a training set, which is a macroscopic defect dataset of wheel surface images. The improved YOLOv11 target detection model is composed of a multi-scale convolutional MSGC module and an improved downsampling LightSEDown module introduced into the backbone network, and an SE attention mechanism added.

2. The method for detecting macroscopic defects in high-speed train wheels based on deep learning according to claim 1, characterized in that, Constructing the macroscopic defect dataset of the wheel surface images includes: Image data of the wheel surface is acquired and stitched together to obtain the original dataset; The Cutout method is used to randomly occlude image regions in the original dataset, and the AutoAugment strategy is used to automatically select image transformation strategies for initial enhancement. Then, the Random Erasing method is used to randomly select some regions of the image for erasure and replacement to generate image samples, expand the original dataset, and obtain the enhanced dataset. The enhanced dataset was annotated using LabelImg annotation software to mark the surface defect areas of the medium-thick plate, thus obtaining the macroscopic defect dataset of the wheel surface image.

3. The method for detecting macroscopic defects in high-speed rail wheels based on deep learning according to claim 1, characterized in that, The improved YOLOv11 object detection model includes: Input module: Used to receive and preprocess input images; Improved backbone network: for extracting multi-scale features from images through the multi-scale convolutional MSGC module and the improved downsampling LightSEDown module; Feature fusion module: used to fuse extracted multi-scale features; Decoupling head: Used to classify and locate targets based on the fused features.

4. The method for detecting macroscopic defects in high-speed rail wheels based on deep learning according to claim 3, characterized in that, The multi-scale convolution MSGC module uses multi-scale depthwise separable convolution to divide the original feature map into two parts. The first part is processed by convolution, and the second part retains the original features. The original features are fused with the features extracted by multi-scale convolution. Channel information fusion is achieved by performing pointwise convolution through 1×1 convolution.

5. The method for detecting macroscopic defects in high-speed rail wheels based on deep learning according to claim 4, characterized in that, The improved downsampling LightSEDown module adopts a channel segmentation mechanism to divide the original feature map into a third part and a fourth part, and further divides it into processing path one and processing path two in the channel dimension. The segmented feature maps are processed separately, and finally the output results of processing path one and processing path two are spliced ​​and fused together. The first processing path is used to first perform illumination normalization on the third part of the feature map, then use multi-scale convolution to extract features, and introduce the SE attention mechanism to enhance the weight of important feature channels. The second processing path is used to perform convolution and max pooling operations on the fourth part of the feature map, preserving the original details and reducing the spatial dimension.

6. The method for detecting macroscopic defects in high-speed train wheels based on deep learning according to claim 5, characterized in that, The SE attention mechanism is used for two steps: compression and activation. In the compression stage, the SE attention mechanism first compresses the spatial dimension of the input feature map through a global average pooling operation to obtain global information for each channel. In the activation stage, the SE attention mechanism recalibrates the channel features through a fully connected layer and outputs the weight coefficient of each channel through a sigmoid activation function. Finally, the weight coefficient of each channel is multiplied with the original input feature map channel by channel to complete the recalibration of the feature map.

7. The method for detecting macroscopic defects in high-speed rail wheels based on deep learning according to claim 6, characterized in that, The processing procedure for the compression stage is as follows: ； In the formula, This represents the pixel value in the i-th row, j-th column, and c-th channel of the input feature map; This is the global average pooling result for channel c; H and W are the height and width of the input feature map, respectively. The processing procedure for the incentive phase is as follows: ； In the formula, z is a C-dimensional vector obtained through global average pooling; and These are the weight matrices of the two fully connected layers; It is the ReLU activation function; It is the Sigmoid activation function; Weights for each channel; The weight coefficient of each channel is multiplied channel by channel by the original input feature map to obtain: ； In the formula, The feature map is processed by the SE attention mechanism. It is the weight of channel c.

8. A deep learning-based macroscopic defect detection system for high-speed rail wheels, used to implement the deep learning-based macroscopic defect detection method for high-speed rail wheels as described in any one of claims 1-7, characterized in that, include: The model building unit is used to build an improved YOLOv11 object detection model. The improved YOLOv11 object detection model is obtained by training on a training set, which is a macroscopic defect dataset of wheel surface images. The improved YOLOv11 object detection model is composed of a multi-scale convolutional MSGC module and an improved downsampling LightSEDown module introduced into the backbone network, and an SE attention mechanism added. The deployment execution unit is used to deploy the improved YOLOv11 target detection model to the wheel production line to perform real-time image acquisition and defect detection, obtain detection results, perform real-time classification and localization based on the detection results, and obtain defect information. Information aggregation unit: used to feed back the defect information to the host computer in real time, provide defect type and severity data, and record and store historical detection results.

Images