Bearing surface defect detection system, equipment and medium
Through the combination of MEC-PA, PCAP and MUFF modules, the problem of high-precision detection of bearing surface defects under complex working conditions is solved, and efficient and accurate defect detection and quantitative evaluation are achieved.
Patent Information
- Application Number
- CN202510617944.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-14
- Publication Date
- 2025-09-19
Smart Images
Figure CN120673116A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of bearing detection, and more particularly to a bearing surface defect detection system, equipment and medium. Background Art
[0002] Bearings are one of the most common basic components in the machinery industry. With industrial development, the demand for bearings, as well as their quality and performance, continues to increase. However, bearings are prone to defects such as wear and scratches during use or production. These defects can significantly affect the accuracy, performance, and service life of equipment, and can even lead to serious consequences.
[0003] For bearing defect detection, traditional detection methods mainly rely on manual visual inspection or analysis technology based on vibration signals. Since the operating environment of bearings is usually very complex and easily interfered by environmental noise, there are problems such as low efficiency, strong subjectivity, and difficulty in capturing tiny defects.
[0004] With the development of image processing technology, automatic detection technology based on machine vision has gradually become mainstream. However, although deep learning technology has made significant progress in the field of defect detection, the complexity of bearing surface defect types leads to significant image background noise, making it difficult for traditional image processing algorithms to accurately segment defect areas. The defect types are diverse and the scales vary greatly. The existing deep learning models perform poorly in feature extraction and multi-scale fusion. Bearing surface defect detection based on machine vision faces great challenges in terms of accuracy and precision.
[0005] Therefore, there is an urgent need for a high-precision bearing surface defect detection technology suitable for complex working conditions. Summary of the Invention
[0006] In view of the above problems, the purpose of the present invention is to provide a bearing surface defect detection system, equipment and medium. First, the multi-scale feature image is extracted through the MEC-PA module, and the channel weights are dynamically adjusted to adaptively enhance the key channel features and suppress redundant information; secondly, the position attention and channel attention are integrated through the PCAP module to respectively enhance the shape positioning ability and semantic distinction ability of the defect area, and the dimensionality reduction operation is used to efficiently integrate multi-dimensional features to improve the target focusing ability under complex backgrounds; then, the deep features are up-sampled layer by layer to the original resolution through the MUFF module to ensure the continuity and integrity of the segmentation results; finally, the defect contour is extracted based on morphological operations through the output module, and parameters such as area and length are calculated, and a defect grading heat map is generated based on the preset threshold; thus, high-precision detection and industrial-grade quantitative evaluation of bearing surface defects are achieved.
[0007] A first aspect of the present invention provides a bearing surface defect detection system, the system comprising:
[0008] The MEC-PA module is used to extract a multi-scale feature image based on the input and dynamically adjust the channel weights of the multi-scale feature image to obtain a first multi-scale feature fusion image;
[0009] A PCAP module is configured to fuse images based on the first multi-scale feature-based attention augmentation mechanism,
[0010] Obtaining a first feature fusion image;
[0011] The MUFF module is used to fuse the images according to the first feature, starting from the deepest feature layer and upsampling layer by layer until the original resolution of the input image is restored, generating a multi-level output image containing features of different resolutions;
[0012] The output module performs 1×1 convolution dimensionality reduction based on the multi-level output images of features with different resolutions to obtain the quantitative evaluation results of defects in the input image.
[0013] In this solution, the MEC-PA module specifically includes:
[0014] A feature preprocessing unit extracts the input image through a 3×3 convolution to obtain a first feature image;
[0015] The parallel multi-scale convolution unit processes the input image through 3×3 convolution and 5×5 convolution in parallel to obtain the second feature image;
[0016] a multi-scale residual fusion unit, extracting a first multi-scale feature image by stacking at least two first feature images or second feature images;
[0017] An attention enhancement unit is used to dynamically adjust the channel-level weights of the first multi-scale feature image and determine a first weight parameter of each channel to obtain a first multi-scale feature fusion image.
[0018] In this solution, the attention enhancement unit performs dynamic adjustment of channel weights, specifically including:
[0019] performing global pooling compression on feature information of the first multi-scale feature image to generate a channel-level statistical feature vector;
[0020] Adjust the convolution kernel size based on the preset number of channels, where the kernel size is an odd number;
[0021] Based on the convolution kernel size and the channel-level statistical feature vector, a channel weight matrix is generated;
[0022] The channel weight matrix is multiplied by the feature channel of the first multi-scale feature image to obtain a first multi-scale feature fusion image.
[0023] In this solution, the PCAP module specifically includes:
[0024] a position attention unit, which enhances the image of the feature points in the feature map by performing a batch dot product operation based on the position and shape of the first multi-scale feature fusion image to obtain a second multi-scale feature image;
[0025] a channel attention unit, which performs batch dot product operations based on the channel dimension based on the first multi-scale feature fusion image to enhance the correlation between defect categories and obtain a third multi-scale feature image;
[0026] The attention fusion unit superimposes the second multi-scale feature image and the third multi-scale feature image based on global average pooling and global maximum pooling to obtain a first feature fusion image.
[0027] In this solution, the MUFF module performs a layer-by-layer upsampling process, specifically including:
[0028] Starting from the deep feature map, bilinear interpolation upsampling is performed layer by layer to gradually restore the feature map resolution;
[0029] After completing one layer of feature sampling each time, the image information of the area with higher resolution than that in the restored feature map is retained;
[0030] The image information is superimposed to generate a multi-level output image containing features of different resolutions.
[0031] In this solution, the output module performs the evaluation steps, specifically including:
[0032] Perform morphological operations on the image after dimensionality reduction to extract the continuous contours of the defect area;
[0033] Calculate defect area, maximum length, minimum width and centroid position based on contour coordinates;
[0034] Defects are classified into minor, moderate, and severe levels based on preset thresholds, and a visual heat map is generated that shows defect type and location.
[0035] This solution also includes a DBLE module, which performs the optimization feature extraction step, specifically including:
[0036] According to the first feature image, a preset normalization algorithm is combined to accelerate convergence to obtain a third feature image;
[0037] The first characteristic image is superimposed on the third characteristic image to update the first characteristic image.
[0038] This solution also includes a CBAMT module, which performs a dynamic attention weight adjustment process, specifically including:
[0039] fusing the images according to the first feature to obtain a key spatial region;
[0040] For the image in the key spatial area, after convolution compression, the spatial attention output feature weight is obtained to generate a spatial attention weight map;
[0041] The spatial attention weight map is multiplied by elements of the first feature fusion image to update the first feature fusion image.
[0042] The second aspect of the present invention provides an electronic device, which includes a processor, a memory, a communication interface and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; the memory is used to store computer programs; and the processor is used to implement the operating steps of the bearing surface defect detection system described in any one of the above items when executing the program stored in the memory.
[0043] A third aspect of the present invention provides a computer-readable storage medium, which includes a bearing surface defect detection system program. When the bearing surface defect detection system program is executed by a processor, it implements the steps of the bearing surface defect detection system as described in any one of the above items.
[0044] The present invention provides a bearing surface defect detection system, equipment and medium. First, a multi-scale feature image is extracted through the MEC-PA module, and the channel weights are dynamically adjusted to adaptively enhance key channel features and suppress redundant information. Secondly, the position attention and channel attention are integrated through the PCAP module to respectively enhance the shape positioning ability and semantic distinction ability of the defect area, and a dimensionality reduction operation is used to efficiently integrate multi-dimensional features to improve the target focusing ability under complex backgrounds. Then, the MUFF module is used to upsample from deep features layer by layer to the original resolution to ensure the continuity and integrity of the segmentation results. Finally, the output module extracts the defect contour based on morphological operations, calculates parameters such as area and length, and generates a defect classification heat map based on preset thresholds. Therefore, high-precision detection and industrial-grade quantitative evaluation of bearing surface defects are achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope.
[0046] Figure 1 A structural schematic diagram of a bearing surface defect detection system according to the present invention is shown;
[0047] Figure 2The figure shows a schematic structural diagram of the MEC-PA module provided in an embodiment of the present invention;
[0048] Figure 3 A flowchart of dynamically adjusting channel weights by an attention enhancement unit according to an embodiment of the present invention is shown;
[0049] Figure 4 A schematic diagram of the structure of a PCAP module provided by an embodiment of the present invention is shown;
[0050] Figure 5 A flowchart of performing layer-by-layer upsampling by a MUFF module provided in an embodiment of the present invention is shown;
[0051] Figure 6 A bearing surface defect detection device provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0053] Unless otherwise defined, all terms (including technical and scientific terms) used in the embodiments of the present invention have the same meaning as commonly understood by those skilled in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an idealized or extremely formal sense, unless explicitly defined in this manner in the embodiments of the present invention.
[0054] The words "first", "second" and similar terms used in the embodiments of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Words such as "one", "an" or "the" do not indicate a quantity limitation, but rather indicate the existence of at least one. Similarly, words such as "include" or "comprise" mean that the elements or objects preceding the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The steps before or after the method of the embodiment of the present invention do not necessarily have to be performed in exact order. On the contrary, the various steps may be processed in reverse order or simultaneously. At the same time, other operations may be added to these processes, or one or more steps may be removed from these processes.
[0055] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.
[0056] Figure 1 A structural schematic diagram of a bearing surface defect detection system according to the present invention is shown.
[0057] like Figure 1 As shown, the first aspect of the present invention discloses a bearing surface defect detection system 10, the system comprising:
[0058] The MEC-PA module 101 is configured to extract a multi-scale feature image based on the input, and dynamically adjust the channel weights of the multi-scale feature image to obtain a first multi-scale feature fusion image;
[0059] The PCAP module 102 is configured to obtain a first feature fusion image based on the first multi-scale feature fusion image based on an attention increase mechanism;
[0060] The MUFF module 103 is configured to fuse the image according to the first feature, starting from the deepest feature layer, and upsampling layer by layer until the original resolution of the input image is restored, thereby generating a multi-level output image containing features of different resolutions;
[0061] The output module 104 performs dimensionality reduction according to the multi-level output images of different resolution features using 1×1 convolution to obtain a quantitative evaluation result of defects in the input image.
[0062] It should be noted that the MEC-PA module is used to extract image features. First, it extracts initial features from the bearing surface image through 3×3 convolutions to generate the first feature map. Second, it performs 3×3 and 5×5 convolutions on the bearing surface image in parallel to capture local details and long-range contextual information, respectively, to generate the second feature map. Feature fusion is then achieved by stacking multiple first and second feature maps, significantly improving sensitivity to subtle defects. Finally, it dynamically adjusts channel weights to enhance key features and suppress noise, generating the first multi-scale feature fusion image. The PCAP module is used for feature enhancement, including position attention enhancement and channel attention enhancement. Position attention enhancement reshapes the first multi-scale feature fusion image into a spatial position matrix, calculates global correlations between pixels, and generates a spatial weight map to enhance defect shape and location information. Channel attention enhancement compresses and expands the channel dimension to generate a channel weight map, enhancing semantic channels related to the defect. The PCAP module then adds the spatial weight map and channel weight map together, and after dimensionality reduction through 1×1 convolution, generates the first feature fusion image, which accurately locates the defect area and suppresses background noise. The MUFF module is a module used for decoding and recognition. For the first feature fusion image, starting from the deepest feature layer, bilinear interpolation is performed layer by layer to upsample to the original resolution. After each upsampling, the shallow features of the corresponding coding layer are added through skip connections to retain high-resolution detail information. The preset model expression function is then applied to the fused features to enhance the model's expressiveness. The MUFF module uses recursive upsampling and skip connections to preserve deep semantics and shallow details, improving segmentation continuity. The output module is the module that outputs defect detection results. It compresses the multi-level feature map into a single-channel segmentation mask through 1×1 convolution to generate a defect probability map. Morphological operations are then performed on the mask to extract the defect contour, calculate parameters such as area and length, and generate a visual heat map and evaluation report, thereby replacing manual inspection, improving detection efficiency, and reducing false detection rates.
[0063] Figure 2 A schematic structural diagram of the MEC-PA module provided in an embodiment of the present invention is shown.
[0064] According to an embodiment of the present invention, Figure 2 As shown, the MEC-PA module specifically includes:
[0065] The feature preprocessing unit 201 extracts the input image through 3×3 convolution to obtain a first feature image;
[0066] The parallel multi-scale convolution unit 202 performs 3×3 convolution and 5×5 convolution on the input image in parallel to obtain a second feature image;
[0067] The multi-scale residual fusion unit 203 extracts a first multi-scale feature image by stacking at least two first feature images or second feature images;
[0068] The attention enhancement unit 204 is configured to dynamically adjust the channel-level weights of the first multi-scale feature image and determine a first weight parameter for each channel to obtain a first multi-scale feature fusion image.
[0069] It should be noted that the feature preprocessing unit extracts initial features from the bearing surface image through 3×3 convolution, and combines it with a preset normalization algorithm to accelerate convergence to obtain a first feature image. The parallel multi-scale convolution unit performs 3×3 and 5×5 convolutions on the bearing surface image in parallel, capturing contextual information of different scales and generating a second feature image. By collaboratively extracting multi-scale features, it adapts to the detection needs of defects of different sizes. The multi-scale residual fusion unit stacks at least two first feature images or second feature images; for example, by superimposing one first feature image and one second feature image, it achieves the fusion of local and global features, retains edge information, and then obtains the first multi-scale feature image. The attention enhancement unit performs global average pooling on the multi-scale features to generate a channel weight matrix. The channel weight matrix is then multiplied channel by channel with the original features to enhance the key channels, thereby improving the generalization ability for complex defects.
[0070] Figure 3 A flowchart of the attention enhancement unit provided by an embodiment of the present invention performing dynamic adjustment of channel weights is shown.
[0071] According to an embodiment of the present invention, Figure 3 As shown, the attention enhancement unit performs dynamic adjustment of channel weights, specifically including:
[0072] S302, performing global pooling compression on feature information of the first multi-scale feature image to generate a channel-level statistical feature vector;
[0073] S304, adjusting the convolution kernel size based on a preset number of channels, wherein the kernel size is an odd number;
[0074] S306, generating a channel weight matrix based on the convolution kernel size and the channel-level statistical feature vector;
[0075] S308 : Multiply the channel weight matrix by the feature channel of the first multi-scale feature image to obtain a first multi-scale feature fusion image.
[0076] It should be noted that the specific steps for dynamically adjusting channel weights performed by the attention enhancement unit include: First, global pooling-based compression: Global average pooling is performed on the first multi-scale input feature image to generate a channel-level statistical vector to compress the spatial dimension. Second, adaptive convolution kernel adjustment: The convolution kernel size is dynamically calculated based on the number of input channels, ensuring that the kernel size is odd to cover symmetric neighborhood information; for example, when the number of channels is 64, the kernel size is 3, and when the number of channels is 128, the kernel size is 5. Odd-valued convolution kernels optimize the interaction between local and global features and improve model robustness. Next, channel weights are generated by processing the statistical vector through lightweight 1×1 convolutions to generate a channel weight matrix. The sigmoid function is used to map the weights to the interval [0, 1] to indicate the importance of each channel. The weight adjustment mechanism effectively filters background interference and improves the signal-to-noise ratio of the defect area. Finally, the weight matrix is multiplied channel by channel with the original feature map to enhance the response value of key channels and suppress interference from redundant channels, such as background noise. The attention enhancement unit uses 1×1 convolution and global pooling to significantly reduce computational complexity and is suitable for real-time detection scenarios.
[0077] Figure 4 A schematic structural diagram of a PCAP module provided by an embodiment of the present invention is shown.
[0078] According to an embodiment of the present invention, Figure 4 As shown, the PCAP module specifically includes:
[0079] The position attention unit 401 enhances the image of the feature points in the feature map by performing a batch dot product operation based on the position and shape according to the first multi-scale feature fusion image to obtain a second multi-scale feature image;
[0080] A channel attention unit 402 performs a batch dot product operation based on the channel dimension based on the first multi-scale feature fusion image to enhance the correlation between defect categories and obtain a third multi-scale feature image;
[0081] The attention fusion unit 403 performs superposition based on global average pooling and global maximum pooling on the second multi-scale feature image and the third multi-scale feature image to obtain a first feature fusion image.
[0082] The first multi-scale feature fusion image
[0083] It should be noted that the position attention unit reshapes the first multi-scale feature fusion image into a three-dimensional tensor and calculates the inter-pixel correlation matrix. Through matrix multiplication, it generates a spatial weight map to strengthen the shape and position consistency of the defect area and enhance the continuity of the defect shape. The channel attention unit compresses the first multi-scale feature fusion image in the channel dimension and, through a preset normalization function, increases and enhances the channel weights related to the defect category, thereby improving the model's ability to distinguish defect categories. The attention fusion unit superimposes the outputs of the position attention unit and the channel attention unit, combines global average pooling and global maximum pooling to extract global features, and obtains the first feature fusion image.
[0084] Figure 5 A flowchart of the MUFF module provided by an embodiment of the present invention performing layer-by-layer upsampling is shown.
[0085] According to an embodiment of the present invention, Figure 5 As shown, the MUFF module performs a layer-by-layer upsampling process, specifically including:
[0086] S502, starting from the deep feature map, bilinear interpolation upsampling is performed layer by layer to gradually restore the feature map resolution;
[0087] S504, after completing one layer of feature sampling each time, retain the image information of the area with higher resolution than the restored feature map;
[0088] S506: Superimpose the image information to generate a multi-level output image containing features of different resolutions.
[0089] It should be noted that, first, sampling is performed through a recursive upward rule to obtain a feature map. Starting from the deepest feature layer (such as a resolution of 32×32), bilinear interpolation is performed layer by layer to upsample to the original resolution (256×256); wherein, the resolution doubles after each upsampling, for example, 32×32, 64×64, 128×128, and 256×256 increase upward in sequence. Then, after each upsampling, the current feature map is added to the shallow features of the corresponding layer in the encoding stage (such as the 64×64 layer) through a jump connection, thereby retaining the original high-resolution details and compensating for the information loss during the upsampling process. Finally, the sampled images are fused and superimposed to enhance the nonlinear expression ability, and the above steps are repeated until the original resolution is restored to generate a multi-level output map. The MUFF module gradually restores the resolution through recursive upsampling, thereby enhancing the model's ability to detect tiny defects.
[0090] According to an embodiment of the present invention, the output module performs an evaluation step, specifically including:
[0091] Perform morphological operations on the image after dimensionality reduction to extract the continuous contours of the defect area;
[0092] Calculate defect area, maximum length, minimum width and centroid position based on contour coordinates;
[0093] Defects are classified into minor, moderate, and severe levels based on preset thresholds, and a visual heat map is generated that shows defect type and location.
[0094] It should be noted that the output module performs morphological operations on the reduced-dimensional image, including but not limited to erosion and dilation, to eliminate isolated noise points and connect broken defect areas. Then, based on a contour algorithm, it extracts the coordinates of the defect boundary and generates a continuous closed contour. Based on this closed contour, the defect's area, maximum length, minimum width, and center of mass position are determined. The center of mass position refers to the coordinates of the center of mass, which are used to locate the specific location of the defect on the bearing surface.
[0095] According to an embodiment of the present invention, a DBLE module is further included, and the DBLE module performs the optimization feature extraction step, specifically including:
[0096] According to the first feature image, a preset normalization algorithm is combined to accelerate convergence to obtain a third feature image;
[0097] The first characteristic image is superimposed on the third characteristic image to update the first characteristic image.
[0098] It's important to note that the DBLE module optimizes the feature extraction process. First, a 3×3 convolution is performed on the bearing surface image, combined with a preset normalization function to accelerate convergence. Then, the response of the defect area is enhanced by dynamically adjusting channel weights. For example, the defect channel weight is increased to 0.8, while the background channel weight is reduced to 0.2. This results in a third feature image. The first and third feature images are superimposed, preserving the original information while enhancing defect contrast, thereby improving the efficiency and accuracy of subsequent defect segmentation.
[0099] According to an embodiment of the present invention, a CBAMT module is further included. The CBAMT module performs a process of dynamically adjusting attention weights, specifically including:
[0100] fusing the images according to the first feature to obtain a key spatial region;
[0101] For the image in the key spatial area, after convolution compression, the spatial attention output feature weight is obtained to generate a spatial attention weight map;
[0102] The spatial attention weight map is multiplied by elements of the first feature fusion image to update the first feature fusion image.
[0103] It should be noted that the CBAMT module is used to enhance the spatial attention of the first feature fusion image. First, a key spatial region is obtained based on the first feature fusion image. This region is compressed to the same dimension as the attention output using a 1×1 convolution. The compressed conditional features are then added to the elements of the key spatial region to supplement the local detail information. A 3×3 convolution is performed on the fused features to generate a spatial attention weight map. Finally, the spatial attention weight map is multiplied by the elements of the first feature fusion image to update the first feature fusion image, effectively suppressing background areas and enhancing defect areas.
[0104] Figure 6 A bearing surface defect detection device provided by an embodiment of the present invention is shown.
[0105] like Figure 6 As shown, the second aspect of the present invention provides an electronic device, which includes a processor 601, a memory 602, a communication interface 603 and a communication bus 604, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; the memory is used to store computer programs; and the processor is used to implement the operating steps of the bearing surface defect detection system described in any one of the above items when executing the program stored in the memory.
[0106] A third aspect of the present invention provides a computer-readable storage medium, which includes a bearing surface defect detection system program. When the bearing surface defect detection system program is executed by a processor, it implements the steps of the bearing surface defect detection system as described in any one of the above items.
[0107] In summary, the present invention provides a bearing surface defect detection system, equipment and medium. First, a multi-scale feature image is extracted through the MEC-PA module, and the channel weights are dynamically adjusted to adaptively enhance the key channel features and suppress redundant information; secondly, the position attention and channel attention are integrated through the PCAP module to respectively enhance the shape positioning ability and semantic distinction ability of the defect area, and the dimensionality reduction operation is used to efficiently integrate multi-dimensional features to improve the target focusing ability under complex backgrounds; then, the deep features are up-sampled layer by layer to the original resolution through the MUFF module to ensure the continuity and integrity of the segmentation results; finally, the defect contour is extracted based on morphological operations through the output module, and parameters such as area and length are calculated, and a defect grading heat map is generated in combination with a preset threshold; high-precision detection and industrial-grade quantitative evaluation of bearing surface defects are achieved.
[0108] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0109] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.
Claims
1. A bearing surface defect detection system, characterized in that: The system comprises: The MEC-PA module is used to extract a multi-scale feature image based on the input and dynamically adjust the channel weights of the multi-scale feature image to obtain a first multi-scale feature fusion image; A PCAP module is configured to obtain a first feature fusion image based on the first multi-scale feature fusion image based on an attention increase mechanism; The MUFF module is used to fuse the images according to the first feature, starting from the deepest feature layer and upsampling layer by layer until the original resolution of the input image is restored, generating a multi-level output image containing features of different resolutions; The output module performs 1×1 convolution dimensionality reduction based on the multi-level output images of features with different resolutions to obtain the quantitative evaluation results of defects in the input image.
2. A bearing surface defect detection system according to claim 1, characterized in that: The MEC-PA module specifically includes: A feature preprocessing unit extracts the input image through a 3×3 convolution to obtain a first feature image; The parallel multi-scale convolution unit processes the input image through 3×3 convolution and 5×5 convolution in parallel to obtain the second feature image; a multi-scale residual fusion unit, extracting a first multi-scale feature image by stacking at least two first feature images or second feature images; An attention enhancement unit is used to dynamically adjust the channel-level weights of the first multi-scale feature image and determine a first weight parameter of each channel to obtain a first multi-scale feature fusion image.
3. A bearing surface defect detection system according to claim 2, characterized in that: The attention enhancement unit performs dynamic adjustment of channel weights, specifically including: performing global pooling compression on feature information of the first multi-scale feature image to generate a channel-level statistical feature vector; Adjust the convolution kernel size based on the preset number of channels, where the kernel size is an odd number; Based on the convolution kernel size and the channel-level statistical feature vector, a channel weight matrix is generated; The channel weight matrix is multiplied by the feature channel of the first multi-scale feature image to obtain a first multi-scale feature fusion image.
4. A bearing surface defect detection system according to claim 1, characterized in that: The PCAP module specifically includes: a position attention unit, which enhances the image of the feature points in the feature map by performing a batch dot product operation based on the position and shape of the first multi-scale feature fusion image to obtain a second multi-scale feature image; a channel attention unit, which performs batch dot product operations based on the channel dimension based on the first multi-scale feature fusion image to enhance the correlation between defect categories and obtain a third multi-scale feature image; The attention fusion unit superimposes the second multi-scale feature image and the third multi-scale feature image based on global average pooling and global maximum pooling to obtain a first feature fusion image.
5. The bearing surface defect detection system according to claim 1, characterized in that: The MUFF module performs a layer-by-layer upsampling process, specifically including: Starting from the deep feature map, bilinear interpolation upsampling is performed layer by layer to gradually restore the feature map resolution; After completing one layer of feature sampling each time, the image information of the area with higher resolution than that in the restored feature map is retained; The image information is superimposed to generate a multi-level output image containing features of different resolutions.
6. A bearing surface defect detection system according to claim 1, characterized in that: The output module performs the evaluation step, which specifically includes: Perform morphological operations on the image after dimensionality reduction to extract the continuous contours of the defect area; Calculate defect area, maximum length, minimum width and centroid position based on contour coordinates; Defects are classified into minor, moderate, and severe levels based on preset thresholds, and a visual heat map is generated that shows defect type and location.
7. A bearing surface defect detection system according to claim 2, characterized in that: The DBLE module is further included, and the DBLE module performs the optimization feature extraction step, specifically including: According to the first feature image, a preset normalization algorithm is combined to accelerate convergence to obtain a third feature image; The first characteristic image is superimposed on the third characteristic image to update the first characteristic image.
8. The bearing surface defect detection system according to claim 1, characterized in that: The CBAMT module is further included, and the CBAMT module performs a process of dynamically adjusting attention weights, specifically including: fusing the images according to the first feature to obtain a key spatial region; For the image in the key spatial area, after convolution compression, the spatial attention output feature weight is obtained to generate a spatial attention weight map; The spatial attention weight map is multiplied by elements of the first feature fusion image to update the first feature fusion image.
9. An electronic device, characterized in that: The system comprises a processor, a memory, a communication interface and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; The processor is configured to implement the operating steps of the bearing surface defect detection system according to any one of claims 1 to 8 when executing the program stored in the memory.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer-readable storage medium includes a bearing surface defect detection system program. When the bearing surface defect detection system program is executed by a processor, the operation steps of the bearing surface defect detection system according to any one of claims 1 to 8 are implemented.