Bearing surface defect detection method and system based on improved YOLO-LMSE

Abstract

The application discloses a bearing surface defect detection method and system based on improved YOLO-LMSE, S1, collecting bearing defect images; S2, performing data enhancement and preprocessing on the original image set to obtain standardized image data; S3, dividing the standardized image data into a training set, a test set and a verification set; S4, constructing a YOLO-LMSE bearing defect detection model; S5, inputting the bearing defect data set into YOLO-LMSE for training; S6, performing comparative testing on the trained YOLO-LMSE and an original YOLOv11m model on the test set, and verifying the performance difference of the model in bearing defect position positioning, category identification and confidence output; S7, performing multi-dimensional comparison between YOLO-LMSE and mainstream target detection algorithms, comprehensively evaluating the detection performance, calculation efficiency and parameter scale of the model, and determining the industrial deployment applicability of the model. The application can effectively adapt to the deployment requirements of resource-limited industrial scenes, improve the efficiency and reliability of bearing defect detection, and provide a powerful guarantee for the stability of industrial production.

Description

Technical Field

[0001] This invention belongs to the field of bearing surface defect detection technology, and specifically relates to a bearing surface defect detection method and system based on an improved YOLO-LMSE. Background Technology

[0002] Bearings are core components in critical industrial sectors such as automotive and aerospace. Their surfaces are prone to defects such as grooves, wear, and scratches during production, assembly, and service, directly impacting equipment stability and safety. Therefore, efficient and accurate defect detection is crucial. Traditional manual visual inspection is affected by subjective factors, resulting in high false negative and false negative rates and low efficiency, making it unsuitable for automated production. With the advancement of industrial intelligence, machine vision and deep learning technologies offer new solutions for bearing defect detection, gradually replacing traditional manual inspection as the mainstream research direction.

[0003] Machine vision and deep learning technologies have become mainstream solutions. Among existing deep learning detection methods, two-stage algorithms have high accuracy but poor real-time performance, while one-stage YOLO series models balance accuracy and speed, but have shortcomings in detecting complex backgrounds, small targets and overlapping defects, and have high computational costs and are difficult to deploy.

[0004] Chinese invention patent CN118735743A discloses a bearing surface defect detection method based on an improved YOLOv8. This method improves the detection accuracy of bearing defects by introducing a CBAM attention mechanism and an improved feature fusion network. However, after introducing the attention mechanism and multi-scale fusion module, the number of model parameters increases significantly, resulting in insufficient lightweighting and difficulty in ensuring real-time inference at the edge. Furthermore, the method is not adequately optimized for small-sized bearing defects, still exhibiting a high false negative rate for dents and scratches with pixel sizes smaller than 8×8. Yao Jingli et al. replaced the Bottleneck structure in the original C2f module of YOLOv8 with a lightweight SCConv convolutional module. While this effectively reduced feature redundancy and resource consumption, it also significantly increased the detection time for bearing defects, offsetting the efficiency improvement brought by lightweighting with the decrease in detection speed.

[0005] In summary, while existing technologies have improved the performance of YOLO series models in bearing defect detection to some extent, they still suffer from insufficient detection accuracy, training efficiency, and ability to identify complex defects. Furthermore, they are characterized by redundant model parameters and high deployment difficulty. Developing a lightweight bearing surface defect detection model that combines high accuracy and real-time performance is the key to overcoming the current technological bottlenecks. Summary of the Invention

[0006] To overcome the above technical problems, the present invention aims to provide a bearing surface defect detection method and system based on improved YOLO-LMSE. Through multi-module collaborative optimization and network structure innovation, it significantly improves detection accuracy, complex defect identification capability and training efficiency, while achieving model lightweighting. This ensures excellent deployability and practicality in resource-constrained industrial environments, providing strong protection for the reliability, production efficiency and safety of industrial equipment.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The bearing surface defect detection method based on the improved YOLO-LMSE includes the following steps; S1. Collect bearing defect images, which cover three typical defect images: bearing grooves, wear, and scratches. At the same time, collect images of defect-free bearings as negative samples to construct the original image set. S2. Perform data augmentation and preprocessing on the original image set to obtain standardized image data; S3. Construct a complete bearing defect dataset from the standardized image data, divide it into training set, test set and validation set, and complete image annotation and data format conversion; S4. Based on the YOLOv11m model, improvements are made by introducing the MobileNetV4 structure to optimize the backbone, reducing the number of model parameters and computational cost; the EL-BiFPN module is used to enhance multi-scale feature fusion, and the C3K2-MSERes module is used to enhance feature capture and feature fusion; by introducing the MPDIoU loss function to replace the CIoU loss function, the model is further lightweighted while ensuring the accuracy remains unchanged, and the YOLO-LMSE bearing defect detection model is constructed. S5. Input the bearing defect dataset into YOLO-LMSE for training; S6. Compare and test the trained YOLO-LMSE with the original YOLOv11m model on the test set to verify the performance differences of the models in bearing defect location, category recognition and confidence output. S7. Compare YOLO-LMSE with mainstream target detection algorithms from multiple dimensions, comprehensively evaluate the model's detection performance, computational efficiency, and parameter scale, and determine its suitability for industrial deployment.

[0008] The preprocessing and data augmentation described in step S2 specifically include: Gaussian blur denoising is employed to smooth background noise interference and reduce the learning burden on the model. Subsequently, gamma transform is used to adjust image brightness and contrast, suppress background noise, and enhance the identification of defect areas. The transform formula is as follows:

[0009] in These are the transformed pixel values. It is a scaling constant. These are the original pixel values. This is the brightness adjustment coefficient; Gamma transform and Gaussian blur denoising can cancel each other out. Pre-Gaussian blur avoids the problem of gamma transform amplifying noise from the root, while subsequent gamma transform makes up for the shortcomings of Gaussian blur in weakening the details of defect edges and inability to correct uneven lighting and insufficient contrast. The two form a closed-loop complementarity in defect detection preprocessing.

[0010] Step S3, which involves constructing a complete bearing defect dataset, specifically includes: (3.1) Defect classification: The standardized images are classified into three categories according to the defect type: grooves, wear, and scratches. Images without defects are labeled as background. (3.2) Data set partitioning: The image data is divided into training set, test set and validation set in a ratio of 8:1:1 to ensure that the distribution of defect types in each set is balanced; (3.3) Image annotation: The Labelimg tool is used to perform multi-point contour annotation on the defect area, generate corresponding label files, and convert the label format into a txt format that the model can recognize; (3.4) Data reading: Based on the Python PyTorch library, the torch.utils.data function is used to read and load datasets in batches, thereby improving training efficiency.

[0011] The construction of YOLO-LMSE in step S4 specifically includes: (4.1) Module integration improvement: Based on the YOLOv11m model, four major modules were integrated in sequence to complete the improvement. The functions and implementation methods of each module are as follows: ① MobileNetV4 module integration: The MobileNetV4 module replaces the backbone network of YOLOv11m. With the Universal Inverted Bottleneck (UIB) as the core, it unifies multiple efficient structures and adds a deep convolution branch. While improving the network's computational efficiency, it strengthens feature extraction. Ultimately, while ensuring detection performance, it significantly reduces the number of model parameters and computational cost, effectively reducing the model size and complexity, and achieving model lightweighting. ② EL-BiFPN module integration: The original feature fusion structure of YOLOv11m is replaced by the EL-BiFPN module. Through the efficient weighted fusion mechanism, the multi-scale feature transfer is optimized, the feature adaptability of defects of different sizes is enhanced, and the feature fusion efficiency is improved. ③C3K2-MSERes module integration: The C3K2 module in the backbone network of the model is replaced with the C3K2-MSERes module. By optimizing the convolution kernel structure, multi-scale perception and edge feature enhancement are achieved. The residual structure is added to integrate feature information, reducing the number of model parameters and computational complexity, while improving the stability of feature extraction. ④MPDIoU Loss Function: The MPDIoU loss function is introduced to replace the CIoU loss function. While ensuring that the detection accuracy remains unchanged, the localization ability of the target and the robustness of the model are further improved, and the model is made lightweight. (4.2) Basic parameter settings: Set the optimal combination of learning rate, batch size, number of iterations, number of image channels, cropping size and learning rate momentum. The initial learning rate is set to 0.01, the batch size is set to 16, the number of iterations is set to 300, and the image cropping size is set to 512×512. The collaborative working logic of each module in step (4.1) is as follows: the multi-dimensional features extracted by the MobileNetV4 module are input into the EL-BiFPN module for cross-scale fusion. The fused features are then flexibly adapted to the C2K2-MSERes module structure to enhance the defect feature extraction capability, forming a "feature extraction-fusion enhancement" working link.

[0012] The model training process described in step S5 specifically includes: (5.1) Dataset input: Input the constructed bearing defect dataset into YOLO-LMSE in a preset batch and load the pre-trained weights to initialize the model parameters; (5.2) Training optimization: The XGBoost classifier is used to classify the bounding boxes, the Soft-NMS method is used to sort the confidence of the bounding boxes, the Shape-IoU loss function is introduced to calculate the loss between the predicted value and the true value, and the AdamW optimizer is combined for backpropagation and parameter update. (5.3) Iterative validation: After every 50 iterations, the model accuracy and recall are evaluated using the validation set, and the learning rate is dynamically adjusted until the preset number of iterations is reached or the performance of the validation set tends to stabilize, at which point training stops.

[0013] The comparative test described in step S6 specifically includes: loading the weight file generated during model training, outputting the defect detection results of the two models on the test set, including defect location coordinates, category labels and confidence scores, calculating the false negative rate and false positive rate of the two models, and verifying the performance advantages of the improved model.

[0014] The multi-dimensional comparison described in step S7 uses mAP@0.5, mAP@0.5-0.95, FLOPs, and the number of parameters as evaluation metrics. It is compared with mainstream algorithms such as YOLOv10, YOLOv11, and Faster R-CNN to verify the comprehensive advantages of the improved model in terms of accuracy and efficiency.

[0015] A YOLO-LMSE bearing surface defect detection system, the system includes a defect image acquisition module, a data augmentation and preprocessing module, a dataset construction module, a YOLOv11m model improvement module, a model training module, and a performance testing and comparison module; Defect image acquisition module: used to collect images of bearing grooves, wear, scratches, and defect-free negative samples, ensuring the integrity and diversity of image data; Data augmentation and preprocessing module: used to perform gamma transformation, noise reduction and other processing on the acquired images, and output standardized image data; Dataset building module: used to implement defect classification, dataset partitioning, image annotation and data format conversion, and build a complete dataset suitable for model training; YOLOv11m model improvement module: used to integrate MobileNetV4, EL-BiFPN, C2K2-MSERes, and MPDIoU loss functions, optimize the YOLOv11m model structure, and build an improved YOLO-LMSE bearing defect detection model; Model training module: Used to load datasets to train improved models, and improve model detection performance by optimizing loss functions and adjusting parameters; Performance testing and comparison module: Used to compare the improved model with the original model and mainstream algorithms, output test results and performance evaluation reports, and provide a basis for industrial deployment.

[0016] The system can be adapted to resource-constrained environments in industrial settings, supports real-time bearing defect detection, and can effectively improve the reliability and production efficiency of industrial equipment.

[0017] The beneficial effects of the present invention.

[0018] By reducing model complexity with the MobileNetV4 module and optimizing multi-scale feature fusion with the EL-BiFPN module, the improved model significantly improves the mAP@0.5 metric on the bearing defect dataset compared to the original YOLOv11m model, effectively reducing false negatives and missed detections in complex backgrounds. The recognition accuracy for grooves, wear, and scratches is specifically optimized, resulting in a significant improvement in detection precision. Furthermore, by replacing the C3K2 module with the C3K2-MSERes module, the number of model parameters and computational complexity are reduced. Compared to the original YOLOv11m model, the number of parameters is significantly reduced, computational complexity is greatly decreased, and the model exhibits outstanding lightweight performance, making it suitable for resource-constrained environments such as industrial edge devices. The model meets the deployment requirements of various environments; the introduction of the MPDIoU loss function further enhances the target localization ability and model robustness; it exhibits excellent training efficiency and generalization ability. Through reasonable parameter configuration, optimized loss function, and iterative verification strategy, the model training convergence speed is significantly improved, demonstrating strong generalization ability in bearing image detection under different working conditions and with different defect degrees, adapting to complex detection scenarios in industrial sites; it has strong industrial applicability. The improved model ensures high detection accuracy while possessing low computational complexity and good real-time performance, meeting the real-time bearing defect detection needs of industrial production lines, providing reliable technical support for industrial equipment maintenance, thereby reducing equipment maintenance costs and improving production efficiency and safety. Attached Figure Description

[0019] Figure 1 This is a model improvement framework diagram of the YOLO-LMSE bearing surface defect detection method of the present invention.

[0020] Figure 2 This is a schematic diagram of the MobileNetV4 module of the YOLO-LMSE bearing surface defect detection method of the present invention.

[0021] Figure 3 A schematic diagram of the general inverted bottleneck UIB structure for the core structure of the MobileNetV4 module.

[0022] Figure 4 This is a schematic diagram of the EL-BiFPN structure of the YOLO-LMSE bearing surface defect detection method of the present invention.

[0023] Figure 5 This is a schematic diagram of the EUCB module structure of the YOLO-LMSE bearing surface defect detection method of the present invention.

[0024] Figure 6 This is a schematic diagram of the MSEResConv module structure of the YOLO-LMSE bearing surface defect detection method of the present invention.

[0025] Figure 7This is a schematic diagram of the C3K2-MSERes module structure of the YOLO-LMSE bearing surface defect detection method of the present invention.

[0026] Figure 8 The image shows a sample of a bearing defect provided in an embodiment of the present invention.

[0027] Figure 9 The results of detecting bearing defects using the YOLO-LMSE algorithm are provided in the embodiments of the present invention.

[0028] Figure 10 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] like Figure 10 As shown, the YOLO-LMSE bearing surface defect detection method proposed in this invention includes the following steps: S1. Bearing Defect Image Acquisition: Collect bearing defect images, covering three typical defect images: bearing grooves, wear, and scratches. At the same time, collect images of defect-free bearings as negative samples to construct the original image set, ensuring that the image data covers different working conditions and different defect levels, providing rich data support for model training.

[0031] S2. Image Data Augmentation and Preprocessing: The original image set is augmented and preprocessed to obtain standardized image data. Gaussian blur denoising is used to smooth background noise interference, simplify redundant image information, reduce the learning burden on the model, and provide high-quality input for model training. Subsequently, gamma transform is used to adjust image brightness and contrast, suppress background noise, and enhance the identification of defect areas. The transform formula is:

[0032] in These are the transformed pixel values. It is a scaling constant. These are the original pixel values. This is the brightness adjustment coefficient; Gamma transform and Gaussian blur denoising can cancel each other out. Pre-Gaussian blur avoids the problem of gamma transform amplifying noise from the root, while subsequent gamma transform makes up for the shortcomings of Gaussian blur in weakening the details of defect edges and inability to correct uneven lighting and insufficient contrast. The two form a closed-loop complementarity in defect detection preprocessing.

[0033] S3. Construction of Bearing Defect Dataset: Construct a complete bearing defect dataset from standardized image data, specifically including: (3.1) Defect classification: The standardized images are classified into three categories according to the defect type: grooves, wear, and scratches. Images without defects are labeled as background, thus clarifying the distribution of each category in the dataset; (3.2) Data set partitioning: The image data is divided into training set, test set and validation set in a ratio of 8:1:1 to ensure that the defect type distribution in each set is balanced and to ensure the effectiveness of model training, testing and validation; (3.3) Image annotation: The labelimg image annotation tool is used to perform multi-point contour annotation on the defect area. The actual defect area is segmented by multi-point annotation on the contour of the defect area. For images containing multiple defects, multi-anchor box annotation is used to generate JSON format label files and convert them into txt format that can be recognized by neural networks. (3.4) Data reading: Use the torch.utils.data function of the PyTorch library in Python software to read and load datasets in batches, thereby improving the data input efficiency for model training.

[0034] S4. YOLOv11m Model Improvement and Construction: Based on the YOLOv11 model, multi-dimensional improvements are made to construct an improved YOLO-LMSE bearing defect detection model, achieving model lightweighting. Specifically, this includes: (4.1) Model Structure Improvement: Based on the YOLOv11m model, four core modules were integrated sequentially to complete the improvement. The functions and implementation methods of each module are as follows: ① MobileNetV4 module integration: The MobileNetV4 module replaces the backbone network of YOLOv11m. With the Universal Inverted Bottleneck (UIB) as the core, it unifies multiple efficient structures and adds a deep convolution branch. While improving the network's computational efficiency, it strengthens feature extraction. Ultimately, while ensuring detection performance, it significantly reduces the number of model parameters and computational cost, effectively reducing the model size and complexity, and achieving model lightweighting. ② EL-BiFPN module integration: The original feature fusion structure of YOLOv11m is replaced by the EL-BiFPN module. Through the efficient weighted fusion mechanism, the multi-scale feature transfer is optimized, the feature adaptability of defects of different sizes is enhanced, and the feature fusion efficiency is improved. ③C3K2-MSERes module integration: The C3K2 module in the backbone network of the model is replaced with the C3K2-MSERes module. By optimizing the convolution kernel structure, multi-scale perception and edge feature enhancement are achieved. The residual structure is added to integrate feature information, reducing the number of model parameters and computational complexity, while improving the stability of feature extraction. ④MPDIoU Loss Function: The MPDIoU loss function is introduced to replace the CIoU loss function. While ensuring that the detection accuracy remains unchanged, the localization ability of the target and the robustness of the model are further improved, and the model is made lightweight.

[0035] (4.2) Experimental environment: The experimental platform in this embodiment is Windows 11 operating system, and the hardware is 16GB RAM + Nvidia GeForce RTX 4090 (16GB VRAM, training relies on GPU computing); the software environment is Python 3.10, with PyTorch-GPU 2.1.1 (CUDA 12.1 + cuDNN v8.8.01 acceleration); the constructed bearing surface defect dataset is input into the YOLOv11m network structure model, including obtaining bearing surface defect images, bearing surface defect image labels, YOLOv11m network structure model pre-training weights, and YOLOv11m network structure model configuration file; (4.3) Parameter Configuration: Set the optimal combination of parameters for learning rate, batch size, number of iterations, number of image channels, image cropping size, and learning rate momentum. Specifically: learning rate Batch sample size Number of iterations Image channel count Other parameter values ​​remain unchanged; S5. Model Training and Optimization: Input the constructed bearing defect dataset into YOLO-LMSE for training, specifically including: (5.1) Dataset input: Input the constructed bearing defect dataset into YOLO-LMSE with pre-set parameters according to the preset batch and train it. Load the pre-trained weights to initialize the model parameters. (5.2) Model training: The XGBoost classifier is used to classify the bounding boxes. The classification probability of each bounding box is sorted by the Soft Non-Maximum Suppression (Soft-NMS) method to obtain the predicted value of the bounding box with the highest confidence. An optimized loss function is introduced to calculate the loss value between the predicted value and the true value. Backpropagation is performed based on the loss value, and the AdamW optimizer is used to update the model parameters. (5.3) Iteration and adjustment: During the model training process, the model is validated on the validation set every 50 iterations to evaluate the model's accuracy and recall. The learning rate is dynamically adjusted based on the validation results, and a learning rate decay strategy is adopted to gradually reduce the learning rate and maintain the stability of the training process. The training process continues until the preset number of iterations is reached or the performance of the validation set no longer improves, and the network model training is completed.

[0036] S6. Model Testing and Performance Verification: The trained YOLO-LMSE model and the original YOLOv11m model are compared and tested on the test set. Specifically, the weight file generated after the model training is completed is loaded into the test file for model testing to obtain the test results of the improved model on the test set, including defect location, defect category, and confidence level. The test results are compared with those of the original YOLOv11m model to verify the difference in accuracy and efficiency between the two in detecting bearing defects and to clarify the performance advantages of the improved model.

[0037] S7. Multi-model comparison and comprehensive evaluation: YOLO-LMSE is compared with mainstream object detection algorithms (such as YOLOv10, YOLOv11, Faster R-CNN, etc.) in multiple dimensions to comprehensively evaluate the detection performance, computational efficiency and number of parameters of different models. mAP@0.5%, mAP@0.5-0.95%, precision (P), recall (R), FLOPs and number of parameters (Params) are selected as evaluation indicators to determine the industrial deployment applicability of the improved model.

[0038] This example focuses on improving model lightweighting and detection accuracy, using YOLOv11 as the baseline model and proposing a series of innovative improvements. The implementation example is described below: See Figure 1In the overall framework diagram of the YOLO-LMSE bearing surface defect detection method model improvement, MobileNetV4 is used to replace the original YOLOv11m backbone network, reducing the number of model parameters and computational cost, thus achieving model lightweighting. A BiFPN bidirectional feature pyramid network structure is introduced to replace the neck network, and the ordinary upsampling module and some convolutions in the BiFPN structure are replaced by the efficient upsampling module EUCB and the depthwise separable convolution DWConv, respectively, to design the EL-BiFPN structure. To address the network degradation problem of the MSEConv convolution, a residual structure is added, and a novel residual convolution MSEResConv is designed, consisting of parallel dilated convolutions of different scales. Multi-scale receptive fields capture spatial features at different levels, and high-frequency edge signals are dynamically enhanced through differential operations and sigmoid activation. The residual structure then integrates the feature information, outputting lightweight features that take into account multi-scale features and edge detail information, ensuring that the model does not become "weak" in depth. Meanwhile, the convolutional module was introduced into the C3K2 structure, and the C3K2-MSERes module was designed to improve the model's defect detection capability, while also achieving lightweight model and real-time model inference.

[0039] See Figure 2 To address the challenge of deploying bearing defect detection tasks in resource-constrained environments, and to further improve model detection performance while reducing computational and parameter requirements, this invention replaces the original YOLOv11m backbone network with MobileNetV4. This network is a lightweight and efficient architecture that significantly reduces model parameters and computational load while maintaining high accuracy, enabling efficient model operation and good environmental adaptability. The MobileNetV4 network follows a hierarchical design. The first three layers consist of Convbn modules, primarily responsible for extracting basic feature information, laying the foundation for subsequent deep feature processing. Following these three layers, the fourth and fifth layers comprise the core feature extraction structure of MobileNetV4—the Universal Inverted Bottleneck (UIB). These two core layers, through a lightweight design using deep convolution, effectively improve the overall computational efficiency of the network and enhance the model's ability to extract and capture multi-scale features. The UIB architecture's design highlight lies in the introduction of two optional deep convolutional layers. These two layers have different layouts: one is placed before the expansion layer, and the other is positioned between the expansion and projection layers. Furthermore, the UIB architecture achieves unified integration of various efficient network structures, covering mainstream structures such as the classic inverted bottleneck, ConvNext, and feedforward network (FFN). It also adds an ExtraDW variant. A detailed design diagram of the UIB architecture is shown below. Figure 3As shown, the two newly added deep convolutional layers play different roles in the feature extraction process, demonstrating differentiated characteristics: the deep convolutional layer before the expansion layer performs convolution operations independently on each channel of the input feature map, accurately extracting low-dimensional features through this single-channel convolution. Then, leveraging the channel transformation capability of 1×1 convolution, these extracted low-dimensional channel features are mapped and expanded into high-dimensional channel features. The deep convolutional layer between the expansion and compression layers continues to perform convolution operations on the already expanded high-dimensional feature map. Through further in-depth mining and processing of high-dimensional features, the feature extraction capability of the network backbone is continuously strengthened, allowing the model to capture richer and more refined feature information.

[0040] In summary, MobileNetV4, as a convolutional neural network designed for lightweight deployment, leverages the synergy of depthwise separable convolutions and pointwise convolutions to achieve efficient feature extraction while significantly reducing computational complexity and parameter count, ensuring the effectiveness of feature extraction within a lightweight architecture. This design drastically reduces the computational cost of convolution operations and the number of model parameters, allowing the network to fully capture both local spatial and global channel features of the data, achieving a balance between lightweight model design and high feature extraction performance.

[0041] See Figure 4 To achieve better contextual feature fusion and lightweight model design, this invention incorporates the BiFPN (Bidirectional Feature Pyramid Network) concept into the model to replace the original neck network. Simultaneously, it replaces the ordinary upsampling module in the BiFPN structure with an efficient upsampling EUCB module, and replaces some ordinary convolutions with depthwise separable convolutions (DWConv). Each channel of the input feature map is assigned a convolution kernel of a corresponding size, performing convolution operations only within a single channel, rather than the cross-channel fusion calculations of traditional convolutions, significantly reducing the model's computational complexity and parameter count.

[0042] Furthermore, the Block portion within the module utilizes the C3K2-MSERes module, ultimately resulting in the design of the efficient and lightweight bidirectional feature pyramid network EL-BiFPN (Efficient and Lightweight Bidirectional Feature Pyramid Network). The EUCB (Efficient Upsampling Convolutional Block) employs a design combining bilinear interpolation and depthwise separable convolution, which better preserves edge details of the feature map while achieving precise matching between feature map size, resolution, and the number of channels in the next stage. The structure diagram of this module is shown below. Figure 5 As shown.

[0043]

[0044] See Figure 6 MSEResConv (Multi-Scale Enhanced Residual Convolution) is a residual convolution structure that combines multi-scale perception with edge information enhancement. Its computational flow is as follows: First, a 1×1 convolution is performed to transform the dimension... The dimensions are Feature blocks are divided into Each feature patch is processed, and a complete output branch is saved as a residual path; subsequently, the size in each dimension is... The feature patches are processed by 1×1 convolutions with different dilation rates, 3×3 depthwise separable convolutions (3×3DWConv), and edge enhancement operations. Finally, all processed feature patches are concatenated, feature information is integrated through a residual structure, and the result is output after a final 1×1 convolution. The operation process of the MSEResConv module can be represented by the following formula:

[0045]

[0046]

[0047] in, The corresponding edge enhancement operation is as follows: high-frequency edge signals are highlighted by difference operation (subtracting the pooling result from the original feature image), the edge response weights are dynamically adjusted by combining the sigmoid activation function, and then output by 1×1 convolution. The operation is used to divide the segments. Each feature piece was reassembled.

[0048] At the same time, MSEResConv is introduced into the C3K2 structure: when the parameter "C3K2-MSERes=False", the C3K2-MSERes structure is as follows: Figure 7 As shown in (a), its Bottleneck directly uses the MSEResConv module; when "C3K2-MSERes=True", the C3K2-MSERes structure is as follows Figure 7 As shown in (b), its Bottleneck uses the C3K-MSERes module, the specific structure of which is as follows: Figure 7 As shown in (c).

[0049] Furthermore, this invention proposes MPDIoU as the bounding box regression loss, focusing on the core part of the bounding box to better measure the overlap area between the predicted and ground truth bounding boxes, thereby improving target localization and model robustness. The definition formula for MPDIoU is shown below:

[0050] In the formula: , These represent the coordinates of the top left and bottom right of the actual border, respectively. , represents the coordinates of the upper left and lower right of the predicted bounding box, respectively, and w and h represent the width and height of the input image, respectively.

[0051] The loss function based on MPDIoU is defined as follows:

[0052] If the predicted bounding box and the ground truth bounding box have the same aspect ratio, the LMPDIoU value of the predicted bounding box being inside the ground truth bounding box is lower than that of the predicted bounding box being outside the ground truth bounding box. This ensures the accuracy of the bounding box regression and tends to provide predicted bounding boxes with less redundancy.

[0053] See Figure 8 The images of three types of bearing surface defects (grooves, wear, and scratches) collected by the above method are randomly input into the YOLOv11 model to obtain a comparison image of the detection results with defect location markings and displaying defect categories and confidence levels.

[0054] See Figure 9 The comparison images show the original image, the inference result of the YOLOv11 model, and the inference result of the improved YOLO-LMSE model of this invention. Among the inference results for the three categories of defects, the confidence level inferred by the improved model of this invention is improved compared to the original model, achieving both lightweight design and high accuracy.

[0055] Comparative experiments were conducted on the YOLOv8m, YOLOv11m, RT-DETR, and YOLO-LMSE models on the bearing surface defect dataset. The experimental data are shown in Table 1.

[0056] Table 1 Comparison Experiment of Bearing Surface Defect Datasets

[0057] In bearing surface defect detection, the YOLO-LMSE model proposed in this invention outperforms mainstream comparative models such as YOLOv8m, YOLOv11m, and RT-DETR in all core performance indicators, achieving a systematic improvement in the identification and location capabilities of typical defects such as bearing grooves, wear, and scratches: its core comprehensive accuracy mAP@0.5 reaches 95.9%, ranking first among comparative models, and is improved by 0.5, 1.8, and 1.7 compared to the benchmark models YOLOv11m, YOLOv8m, and the Transformer architecture's RT-DETR, respectively. The model achieves a significant advantage in overall detection accuracy, with a precision (P) of 95.6% and a recall (R) of 91.8%, both of which are the highest values ​​among all models. Compared with YOLOv11m, these figures are 0.4 and 0.7 percentage points higher, respectively; compared with YOLOv8m, they are 4.3 and 4.5 percentage points higher, respectively; and compared with RT-DETR, they are 1.2 and 1.0 percentage points higher, respectively. This effectively reduces the false detection rate of defect identification, ensures the reliability of detection results, and greatly enhances the ability to capture various defects in bearings, significantly reducing the risk of missed detections in industrial inspection scenarios.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A bearing surface defect detection method based on improved YOLO-LMSE, characterized in that, Includes the following steps; S1. Collect bearing defect images, which cover three typical defect images: bearing grooves, wear, and scratches. At the same time, collect images of defect-free bearings as negative samples to construct the original image set. S2. Perform data augmentation and preprocessing on the original image set to obtain standardized image data; S3. Construct a complete bearing defect dataset from the standardized image data, divide it into training set, test set and validation set, and complete image annotation and data format conversion; S4. Based on the YOLOv11m model, improvements were made by introducing the MobileNetV4 structure to optimize the backbone, using the EL-BiFPN module to enhance multi-scale feature fusion, and using the C3K2-MSERes module to enhance feature capture and feature fusion; a YOLO-LMSE bearing defect detection model was constructed. S5. Input the bearing defect dataset into YOLO-LMSE for training; S6. Compare and test the trained YOLO-LMSE with the original YOLOv11m model on the test set to verify the performance differences of the models in bearing defect location, category recognition and confidence output. S7. Compare YOLO-LMSE with mainstream target detection algorithms from multiple dimensions, comprehensively evaluate the model's detection performance, computational efficiency, and parameter scale, and determine its suitability for industrial deployment.

2. The bearing surface defect detection method based on improved YOLO-LMSE according to claim 1, characterized in that, The data augmentation and preprocessing described in step S2 specifically include: Gaussian blur denoising is used to smooth background noise interference and reduce the learning burden on the model; subsequently, gamma transform is used to adjust image brightness and contrast, suppress background noise, and enhance the identification of defect areas. The transform formula is as follows: in These are the transformed pixel values. It is a scaling constant. These are the original pixel values. This is the brightness adjustment coefficient.

3. The bearing surface defect detection method based on improved YOLO-LMSE according to claim 1, characterized in that, Step S3, which involves constructing a complete bearing defect dataset, specifically includes: (3.1) Defect classification: The standardized images are classified into three categories according to the defect type: grooves, wear, and scratches. Images without defects are labeled as background. (3.2) Data set partitioning: Divide the image data into training set, test set and validation set to ensure that the distribution of defect types in each set is balanced; (3.3) Image annotation: The Labelimg tool is used to perform multi-point contour annotation on the defect area, generate corresponding label files, and convert the label format into a txt format that the model can recognize; (3.4) Data reading: Based on the Python PyTorch library, the torch.utils.data function is used to read and load datasets in batches, thereby improving training efficiency.

4. The bearing surface defect detection method based on improved YOLO-LMSE according to claim 1, characterized in that, The construction of YOLO-LMSE in step S4 specifically includes: (4.1) Module integration improvement: Based on the YOLOv11m model, four major modules were integrated in sequence to complete the improvement. The functions and implementation methods of each module are as follows: ① MobileNetV4 module integration: Select the MobileNetV4 module with the Universal Inverted Bottleneck (UIB), unify multiple efficient structures and add a deep convolution branch to achieve lightweight model; ② EL-BiFPN module integration: The EL-BiFPN module is selected, and multi-scale feature transfer is optimized through an efficient weighted fusion mechanism; ③C3K2-MSERes module integration: Select the C3K2-MSERes module and optimize the convolution kernel structure to achieve multi-scale perception and edge feature enhancement; ④MPDIoU loss function: The MPDIoU loss function is selected to further improve the target localization ability and model robustness while ensuring that the detection accuracy remains unchanged, thereby achieving model lightweighting; The multi-dimensional features extracted by the MobileNetV4 module are input into the EL-BiFPN module for cross-scale fusion. The fused features are then flexibly adapted to the C2K2-MSERes module structure to enhance the defect feature extraction capability, forming a "feature extraction-fusion enhancement" workflow. (4.2) Basic parameter settings: Set the optimal combination of learning rate, batch size, number of iterations, number of image channels, cropping size and learning rate momentum.

5. The bearing surface defect detection method based on improved YOLO-LMSE according to claim 1, characterized in that, The model training process described in step S5 specifically includes: (5.1) Dataset input: Input the constructed bearing defect dataset into YOLO-LMSE in a preset batch and load the pre-trained weights to initialize the model parameters; (5.2) Training optimization: The XGBoost classifier is used to classify the bounding boxes, the Soft-NMS method is used to sort the confidence of the bounding boxes, the Shape-IoU loss function is introduced to calculate the loss between the predicted value and the true value, and the AdamW optimizer is combined for backpropagation and parameter update. (5.3) Iterative validation: After iteration, the model accuracy and recall are evaluated using the validation set, and the learning rate is dynamically adjusted until the preset number of iterations is reached or the performance of the validation set tends to stabilize, at which point training stops.

6. The bearing surface defect detection method based on improved YOLO-LMSE according to claim 1, characterized in that, The comparative test described in step S6 specifically includes: loading the weight file generated during model training, outputting the defect detection results of the two models on the test set, including defect location coordinates, category labels and confidence scores, calculating the false negative rate and false positive rate of the two models, and verifying the performance advantages of the improved model.

7. The bearing surface defect detection method based on improved YOLO-LMSE according to claim 1, characterized in that, The multi-dimensional comparison described in step S7 uses mAP@0.5, mAP@0.5-0.95, FLOPs, and the number of parameters as evaluation metrics to compare with the algorithms of YOLOv10, YOLOv11, and Faster R-CNN, verifying the comprehensive advantages of the improved model in terms of accuracy and efficiency.

8. A YOLO-LMSE bearing surface defect detection system for implementing the method according to any one of claims 1-7, characterized in that, The system includes a defect image acquisition module, a data augmentation and preprocessing module, a dataset construction module, a YOLOv11m model improvement module, a model training module, and a performance testing and comparison module. Defect image acquisition module: used to collect images of bearing grooves, wear, scratches, and defect-free negative samples, ensuring the integrity and diversity of image data; Data augmentation and preprocessing module: used to perform gamma transformation and noise reduction on the acquired images and output standardized image data; Dataset building module: used to implement defect classification, dataset partitioning, image annotation and data format conversion, and build a complete dataset suitable for model training; YOLOv11m model improvement module: used to integrate MobileNetV4, EL-BiFPN, C2K2-MSERes, and MPDIoU loss functions, optimize the YOLOv11m model structure, and build an improved YOLO-LMSE bearing defect detection model; Model training module: Used to load datasets to train improved models, and improve model detection performance by optimizing loss functions and adjusting parameters; Performance testing and comparison module: Used to compare the improved model with the original model and mainstream algorithms, output test results and performance evaluation reports, and provide a basis for industrial deployment.

9. The application of the YOLO-LMSE bearing surface defect detection system according to claim 8, characterized in that, The system can be adapted to resource-constrained environments in industrial settings, supports real-time bearing defect detection, and can effectively improve the reliability and production efficiency of industrial equipment.

Images