Method and system for detecting surface defects of wind driven generator blade and medium

By applying the depth-separable convolutional structure and channel attention mechanism in YOLOv5, combined with the Kmeans++ algorithm, the problems of low accuracy and slow speed in the detection of surface defects of wind turbine blades are solved, and more efficient feature extraction and prediction box generation are achieved .

CN119048493AActive Publication Date: 2024-11-29CRRC WIND POWER(SHANDONG) CO LTD +2
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202411514600.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-29
Publication Date
2024-11-29
Estimated Expiration
2044-10-29

AI Technical Summary

Technical Problem

The existing target detection methods have problems such as low accuracy and slow calculation speed in the detection of surface defects of wind turbine blades, and cannot meet the requirements of high-precision inspection tasks.

Method used

By replacing the convolutional block structure in the backbone feature extraction network into a deep separable convolutional structure in YOLOv5, and adding a channel attention mechanism to the CSP module. At the same time, replacing the Kmeans algorithm in the detection structure network as a Kmeans++ algorithm, enhancing feature extraction Capability and prediction accuracy.

Benefits of technology

It improves the accuracy and speed of surface defect detection of wind turbine blades, reduces calculation complexity, and enhances the ability to identify and position the surface defects of fan blades.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048493A_ABST
    Figure CN119048493A_ABST
Patent Text Reader

Abstract

The invention discloses a method and a system for detecting surface defects of blades of a wind driven generator and a medium, mainly relates to the technical field of blade surface defect detection, and is used for solving the problem that an existing target detection method is relatively low in precision or relatively slow in calculation speed. Comprising the following steps: replacing a convolutional block structure in a CBL in a trunk feature extraction network in YOLOv5 with a deep separable convolutional structure, and adding a channel attention mechanism in a CSP module in the trunk feature extraction network to obtain an updated trunk feature extraction network; adding a channel attention mechanism into a CSP module in the neck structure network in YOLOv5 to obtain an updated neck structure network; after the trained Kmeans algorithm in the detection structure network in the YOLOv5 is replaced by the trained Kmeans + + algorithm, an updated detection structure network is obtained; and obtaining a final prediction frame through the updated YOLOv5.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of blade surface defect detection, and in particular to a method, system and medium for detecting surface defects of wind turbine blades. Background Art

[0002] The wind turbine blades are important components for wind turbine energy conversion, converting wind energy into mechanical energy. They are also the main load-bearing components of the wind turbine and play a key role in the safe operation of the entire wind turbine. However, the working environment of the blades is very harsh. Not only are they subject to the superposition of aerodynamic loads, gravity loads and inertial loads, but they also operate continuously in a harsh natural environment for a long time. The external climate will have a great impact on the normal operation of the blades, especially typhoons, thunderstorms, ice and snow, sandstorms and other severe weather conditions may cause damage to the blades, thereby causing economic losses to wind power generation. Therefore, it is necessary to monitor the status of the wind turbine blades in a timely manner to ensure that faults can be diagnosed in a timely manner to prevent the entire power system from failing due to blade failures, causing economic losses.

[0003] Currently popular one-stage target detection methods, such as SSD (Single Shot MultiBox Detector), have shown significant advantages in detection speed and can achieve rapid identification and positioning of targets. However, such algorithms have certain limitations in model detection accuracy and often cannot meet the requirements of high-precision inspection tasks. In contrast, two-stage target detection methods, such as Faster R-CNN (Faster Region-based Convolutional Network), achieve much higher accuracy than one-stage detection algorithms through more complex network structures and sophisticated detection processes. But at the same time, its detection speed is relatively slow. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present application provides a system, method and medium for detecting surface defects of wind turbine blades, so as to solve the problem that the existing target detection methods have low accuracy or relatively slow calculation speed.

[0005] In a first aspect, the present application provides a method for detecting surface defects of a wind turbine blade, the method comprising: Collect surface defect images of wind turbine blades and construct a dataset of surface defect images of wind turbine blades; obtain defect annotation data of surface defect images of wind turbine blades in the dataset of surface defect images of wind turbine blades; expand the number of images in the dataset of surface defect images of wind turbine blades to a preset number as the final dataset; input the final dataset into the letterbox function in YOLOv5 to obtain the first image of a preset size; replace the convolution block structure in CBL in the backbone feature extraction network in YOLOv5 with a depthwise separable convolution structure, and add a channel attention mechanism to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network; wherein, the depthwise separable convolution structure is set It is a combination of channel-by-channel convolution and point-by-point convolution, and the output result of channel-by-channel convolution is input into point-by-point convolution; the channel attention mechanism is added to the CSP module in the neck structure network in YOLOv5 to obtain an updated neck structure network; the trained Kmeans algorithm in the detection structure network in YOLOv5 is replaced with the trained Kmeans++ algorithm to obtain an updated detection structure network; the first image is input into the updated backbone feature extraction network to obtain the backbone output data, and then the backbone output data is input as input data into the updated neck structure network to obtain 3 types of feature maps with preset output scales; the feature map is used as the input data of the updated detection structure network to obtain the final prediction box.

[0006] Furthermore, the number of images in the wind turbine blade surface defect image dataset is expanded to a preset number as the final dataset, specifically including: The images in the fan blade surface defect image dataset are augmented by translation, rotation, flipping, affine transformation and noise addition, so as to expand the number of images in the fan blade surface defect image dataset to a preset number.

[0007] Furthermore, the convolutional block structure in CBL in the backbone feature extraction network in YOLOv5 is replaced with a depth-separable convolutional structure, specifically including: The output parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 are used as the input parameters of the depthwise separable convolution structure, and the output parameters of the depthwise separable convolution structure are used as the input parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5, thereby completing the replacement of the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 with the depthwise separable convolution structure.

[0008] Furthermore, a channel attention mechanism is added to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network, which specifically includes: The output parameters of the CSP module in the backbone feature extraction network are used as the input parameters of the channel attention mechanism; wherein the input parameter is the input feature map; and then the channel attention mechanism is used to assign weights to the input feature map, and the feature map after weighting is used as the final output data of the CSP module.

[0009] Further, after replacing the trained Kmeans algorithm in the detection structure network in YOLOv5 with the trained Kmeans++ algorithm and before obtaining the updated detection structure network, the method further includes: S1. Obtain a preset training data set; S2. Randomly select a sample x from the preset training data set as the initial cluster center ; S3, calculate the distance between each sample x that has not been extracted and the cluster center of the determined sample, denoted as ; Use weighted probability distribution to randomly select a new sample as the new cluster center, where the probability of the sample being selected as the cluster center is: ; Where X represents the preset training data set; S4. Repeat steps S2 and S3 until K initial cluster centers are determined, thereby obtaining a trained Kmeans++ algorithm; wherein K is a preset constant.

[0010] In a second aspect, the present application provides a system for detecting surface defects of wind turbine blades, the system comprising: A construction module is used to collect surface defect images of wind turbine blades and construct a wind turbine blade surface defect image dataset; an image module is used to obtain defect annotation data of wind turbine blade surface defect images in the wind turbine blade surface defect image dataset; the number of images in the wind turbine blade surface defect image dataset is expanded to a preset number as the final dataset; the final dataset is input into the letterbox function in YOLOv5 to obtain a first image of a preset size; an update module is used to replace the convolution block structure in the CBL in the backbone feature extraction network in YOLOv5 with a depthwise separable convolution structure, and to add a channel attention mechanism to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network; wherein, the depthwise separable The convolution structure is set as a combination of channel-by-channel convolution and point-by-point convolution, and the output result of the channel-by-channel convolution is input into the point-by-point convolution; the channel attention mechanism is added to the CSP module in the neck structure network in YOLOv5 to obtain an updated neck structure network; the trained Kmeans algorithm in the detection structure network in YOLOv5 is replaced with the trained Kmeans++ algorithm to obtain an updated detection structure network; an acquisition module is obtained for inputting the first image into the updated backbone feature extraction network to obtain the backbone output data, and then the backbone output data is input as input data into the updated neck structure network to obtain three types of feature maps with preset output scales; the feature maps are used as input data of the updated detection structure network to obtain the final prediction box.

[0011] Furthermore, the image module includes an expansion unit, It is used to perform data augmentation on images in a wind blade surface defect image dataset by using translation, rotation, flipping, affine transformation, and noise addition, so as to expand the number of images in the wind blade surface defect image dataset to a preset number.

[0012] Furthermore, the update module includes a first update unit, It is used to use the output parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 as the input parameters of the depthwise separable convolution structure, and use the output parameters of the depthwise separable convolution structure as the input parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5, thereby completing the replacement of the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 with the depthwise separable convolution structure.

[0013] Furthermore, the update module includes a second update unit, It is used to use the output parameters of the CSP module in the backbone feature extraction network as the input parameters of the channel attention mechanism; wherein the input parameter is the input feature map; and then the channel attention mechanism is used to assign weights to the input feature map, and the feature map after the weighting is used as the final output data of the CSP module.

[0014] In a third aspect, the present application provides a non-volatile computer storage medium having computer instructions stored thereon, which when executed implement a method for detecting surface defects of a wind turbine blade as described in any one of the above items.

[0015] Those skilled in the art can understand that the present application has at least the following beneficial effects: This application replaces the convolution block structure in CBL in the backbone feature extraction network of YOLOv5 with a depth-wise separable convolution structure, wherein the depth-wise separable convolution structure is set as a combination of channel-by-channel convolution and point-by-point convolution, and the output result of the channel-by-channel convolution is input into the point-by-point convolution. Compared with the convolution block structure in the traditional CBL, the amount of parameters of the traditional 3-channel convolution is: number of filters * 3 * 3 * 3, and the way of inputting the output result of the channel-by-channel convolution into the point-by-point convolution is channel-by-channel convolution (3 * 3 * 3) + point-by-point convolution (1 * 1 * 3 * number of filters). It can be seen that the amount of parameters is significantly reduced, and the problem of relatively slow calculation speed is reduced.

[0016] In addition, this application adds a channel attention mechanism to the CSP module in the backbone feature extraction network. By introducing a channel attention mechanism in the CSP module (for example, the attention mechanism of the Squeeze-and-Excitation Network for deep convolutional neural networks), the network depth can be deepened and the network's feature extraction capability can be enhanced, thereby more accurately locating and identifying surface defects of wind turbine blades. In addition, by replacing the trained Kmeans algorithm in the detection structure network in YOLOv5 with the trained Kmeans++ algorithm, YOLOv5 can generate better anchor parameters, which can better fit the real target, thereby solving the problem of low accuracy of existing target detection methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Some embodiments of the present disclosure are described below with reference to the accompanying drawings, in which: Figure 1 It is a flow chart of a method for detecting surface defects of wind turbine blades provided in an embodiment of the present application.

[0018] Figure 2 It is a schematic diagram of the internal structure of a system for detecting surface defects of wind turbine blades provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] It should be understood by those skilled in the art that the embodiments described below are only preferred embodiments of the present disclosure, and do not mean that the present disclosure can only be implemented through the preferred embodiments. The preferred embodiments are only used to explain the technical principles of the present disclosure, and are not used to limit the protection scope of the present disclosure. Based on the preferred embodiments provided by the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work should still fall within the protection scope of the present disclosure.

[0020] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0021] The technical solution proposed in the embodiments of the present application is described in detail below with reference to the accompanying drawings.

[0022] The present application embodiment provides a method for detecting surface defects of a wind turbine blade, such as Figure 1 As shown, the method provided in the embodiment of the present application mainly includes the following steps: Step 110: Collect surface defect images of wind turbine blades to construct a wind turbine blade surface defect image dataset.

[0023] It should be noted that the wind turbine blade surface defect image dataset used can be provided by a wind farm in Northwest China, and the generator blade surface defect image can be taken by a camera mounted on the Amu P600 UAV gimbal. The dataset can be 800 blade defect images with an image size of 1280×852.

[0024] Step 120: Obtain defect annotation data of the fan blade surface defect image in the fan blade surface defect image data set.

[0025] It should be noted that this step can be: labelimg is used to label the wind turbine blade image. The specific operation process can be: (1) create a source file directory; (2) create a new label box; (3) enter the defect labeling data category name, which is erosion and gel coat shedding; (4) record the labeling file address. Finally, a labeling file with the suffix .xml will be generated, saved in VOC format, and stored in the target folder.

[0026] Step 130: Expand the number of images in the wind turbine blade surface defect image dataset to a preset number to serve as a final dataset.

[0027] It should be noted that the preset number can be 2400. The images in the fan blade surface defect image dataset can be amplified by translation, rotation, flipping, affine transformation, and noise addition to expand the number of images in the fan blade surface defect image dataset to a preset number. It can be understood by those skilled in the art that through these transformations, the data enhancement technology not only enriches the training dataset, but also simulates a variety of situations in real scenes and enhances the robustness of the model.

[0028] Step 140: Input the final data set into the letterbox function in YOLOv5 to obtain a first image of a preset size. It should be noted that the preset size can be 640*640*3. It should be noted that 3 is the number of channels, and those skilled in the art can divide the first image into three parts for processing. The application here unifies the input image to effectively improve the speed of target detection and also reduce the amount of calculation of the model.

[0029] Step 150: Replace the convolution block structure in CBL in the backbone feature extraction network in YOLOv5 with a depthwise separable convolution structure, and add a channel attention mechanism to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network; add a channel attention mechanism to the CSP module in the neck structure network in YOLOv5 to obtain an updated neck structure network; replace the trained Kmeans algorithm in the detection structure network in YOLOv5 with the trained Kmeans++ algorithm to obtain an updated detection structure network.

[0030] It should be noted that the backbone feature extraction network involved in the present application includes a focus function, a CBL (Conv convolution layer, a Batch Normalization layer, and a LeakyReLU activation function layer) algorithm, and a CSP (Cross Stage Partial) module. The relationship between the focus function, the CBL algorithm, and the CSP module is that the output data of the focus function is input into the CBL algorithm, and the output data of the CBL algorithm is input into the CSP module. In addition, the backbone feature extraction network can add new modules according to the needs of technical personnel in this field. The depthwise separable convolution structure is set to a combination of channel-by-channel convolution and point-by-point convolution, and the output result of the channel-by-channel convolution is input into the point-by-point convolution.

[0031] Among them, the convolution block structure in CBL in the backbone feature extraction network in YOLOv5 is replaced with a depth-separable convolution structure, which can be specifically: The output parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 are used as the input parameters of the depthwise separable convolution structure, and the output parameters of the depthwise separable convolution structure are used as the input parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5, thereby completing the replacement of the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 with the depthwise separable convolution structure.

[0032] Among them, a channel attention mechanism is added to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network, which can be specifically: The output parameters of the CSP module in the backbone feature extraction network are used as the input parameters of the channel attention mechanism; wherein the input parameter is the input feature map; and then the channel attention mechanism is used to assign weights to the input feature map, and the feature map after weighting is used as the final output data of the CSP module.

[0033] In addition, after replacing the trained Kmeans algorithm in the detection structure network in YOLOv5 with the trained Kmeans++ algorithm, before obtaining the updated detection structure network, the method also includes a process of training the Kmeans++ algorithm: S1. Obtain a preset training data set.

[0034] S2. Randomly select a sample x from the preset training data set as the initial cluster center .

[0035] S3, calculate the distance between each sample x that has not been extracted and the cluster center of the determined sample, denoted as ; Use weighted probability distribution to randomly select a new sample as the new cluster center, where the probability of the sample being selected as the cluster center is: ; Where X represents the preset training data set.

[0036] S4. Repeat steps S2 and S3 until K initial cluster centers are determined, thereby obtaining a trained Kmeans++ algorithm; wherein K is a preset constant.

[0037] Step 160: Input the first image into the updated backbone feature extraction network to obtain backbone output data, and then input the backbone output data as input data into the updated neck structure network to obtain three types of feature maps with preset output scales; use the feature maps as input data of the updated detection structure network to obtain the final prediction box.

[0038] It should be noted that the neck structure network is the neck structure network in the existing YOLOv5 (wherein the channel attention mechanism is added to the CSP module). The detection structure network is the neck structure network in the existing YOLOv5 (wherein the trained Kmeans algorithm is replaced by the trained Kmeans++ algorithm).

[0039] besides, Figure 2 The present invention provides a system for detecting surface defects of wind turbine blades. Figure 2 As shown, the system provided in the embodiment of the present application mainly includes: The construction module 210 is used to collect surface defect images of wind turbine blades and construct a wind turbine blade surface defect image dataset.

[0040] The image module 220 is used to obtain defect annotation data of the fan blade surface defect image in the fan blade surface defect image dataset; expand the number of images in the fan blade surface defect image dataset to a preset number as the final dataset; input the final dataset into the letterbox function in YOLOv5 to obtain a first image of a preset size.

[0041] The image module 220 includes an expansion unit, It is used to perform data augmentation on images in a wind blade surface defect image dataset by using translation, rotation, flipping, affine transformation, and noise addition, so as to expand the number of images in the wind blade surface defect image dataset to a preset number.

[0042] The updating module 230 is used to replace the convolution block structure in the CBL in the backbone feature extraction network in YOLOv5 with a depthwise separable convolution structure, and to add a channel attention mechanism to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network; wherein the depthwise separable convolution structure is set as a combination of channel-by-channel convolution and point-by-point convolution, and the output result of the channel-by-channel convolution is input into the point-by-point convolution; a channel attention mechanism is added to the CSP module in the neck structure network in YOLOv5 to obtain an updated neck structure network; after replacing the trained Kmeans algorithm in the detection structure network in YOLOv5 with the trained Kmeans++ algorithm, an updated detection structure network is obtained.

[0043] The updating module 230 includes a first updating unit, It is used to use the output parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 as the input parameters of the depthwise separable convolution structure, and use the output parameters of the depthwise separable convolution structure as the input parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5, thereby completing the replacement of the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 with the depthwise separable convolution structure.

[0044] The updating module 230 includes a second updating unit, It is used to use the output parameters of the CSP module in the backbone feature extraction network as the input parameters of the channel attention mechanism; wherein the input parameter is the input feature map; and then the channel attention mechanism is used to assign weights to the input feature map, and the feature map after the weighting is used as the final output data of the CSP module.

[0045] The acquisition module 240 is used to input the first image into the updated backbone feature extraction network to obtain the backbone output data, and then input the backbone output data as input data into the updated neck structure network to obtain three types of feature maps of preset output scales; the feature maps are used as input data of the updated detection structure network to obtain the final prediction box.

[0046] In addition, an embodiment of the present application further provides a non-volatile computer storage medium on which executable instructions are stored. When the executable instructions are executed, a method for detecting surface defects of a wind turbine blade as described above is implemented.

[0047] So far, the technical solutions of the present disclosure have been described in combination with the above multiple embodiments, but it is easy for those skilled in the art to understand that the protection scope of the present disclosure is not limited to these specific embodiments. Without departing from the technical principles of the present disclosure, those skilled in the art can split and combine the technical solutions in the above-mentioned various embodiments, and can also make equivalent changes or replacements to the relevant technical features. Any changes, equivalent replacements, improvements, etc. made within the technical concept and / or technical principle of the present disclosure will fall within the protection scope of the present disclosure.

Claims

1. A method for detecting surface defects of wind turbine blades, It is characterized in that The method comprises: Collect surface defect images of wind turbine blades and construct a wind turbine blade surface defect image dataset; Obtain defect annotation data of a fan blade surface defect image in a fan blade surface defect image dataset; Expanding the number of images in the wind turbine blade surface defect image dataset to a preset number to serve as the final dataset; Input the final data set into the letterbox function in YOLOv5 to obtain the first image of the preset size; The convolution block structure in CBL in the backbone feature extraction network in YOLOv5 is replaced with a depthwise separable convolution structure, and a channel attention mechanism is added to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network; wherein the depthwise separable convolution structure is set as a combination of channel-by-channel convolution and point-by-point convolution, and the output result of the channel-by-channel convolution is input into the point-by-point convolution; the channel attention mechanism is added to the CSP module in the neck structure network in YOLOv5 to obtain an updated neck structure network; after replacing the trained Kmeans algorithm in the detection structure network in YOLOv5 with the trained Kmeans++ algorithm, an updated detection structure network is obtained; The first image is input into the updated backbone feature extraction network to obtain the backbone output data, and then the backbone output data is input into the updated neck structure network as input data to obtain three types of feature maps with preset output scales; the feature maps are used as input data of the updated detection structure network to obtain the final prediction box.

2. The method for detecting surface defects of wind turbine blades according to claim 1, It is characterized in that The number of images in the wind turbine blade surface defect image dataset is expanded to a preset number as the final dataset, including: The images in the fan blade surface defect image dataset are augmented by translation, rotation, flipping, affine transformation and noise addition, so as to expand the number of images in the fan blade surface defect image dataset to a preset number.

3. The method for detecting surface defects of wind turbine blades according to claim 1, It is characterized in that The convolutional block structure in CBL in the backbone feature extraction network of YOLOv5 is replaced with a depth-separable convolutional structure, including: The output parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 are used as the input parameters of the depthwise separable convolution structure, and the output parameters of the depthwise separable convolution structure are used as the input parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5, thereby completing the replacement of the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 with the depthwise separable convolution structure.

4. The method for detecting surface defects of wind turbine blades according to claim 1, It is characterized in that A channel attention mechanism is added to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network, which specifically includes: The output parameters of the CSP module in the backbone feature extraction network are used as the input parameters of the channel attention mechanism; where the input parameters are the input feature maps; Then, the channel attention mechanism is used to assign weights to the input feature map, and the weighted feature map is used as the final output data of the CSP module.

5. The method for detecting surface defects of wind turbine blades according to claim 1, It is characterized in that After replacing the trained Kmeans algorithm in the detection structure network in YOLOv5 with the trained Kmeans++ algorithm and before obtaining the updated detection structure network, the method further includes: S1. Obtain a preset training data set; S2. Randomly select a sample x from the preset training data set as the initial cluster center ; S3, calculate the distance between each sample x that has not been extracted and the cluster center of the determined sample, denoted as ; Use weighted probability distribution to randomly select a new sample as the new cluster center, where the probability of the sample being selected as the cluster center is: ; Where X represents the preset training data set; S4. Repeat steps S2 and S3 until K initial cluster centers are determined, thereby obtaining a trained Kmeans++ algorithm; wherein K is a preset constant.

6. A system for detecting surface defects of wind turbine blades, It is characterized in that The system comprises: A construction module is used to collect surface defect images of wind turbine blades and construct a wind turbine blade surface defect image dataset; An image module is used to obtain defect annotation data of a fan blade surface defect image in a fan blade surface defect image dataset; expand the number of images in the fan blade surface defect image dataset to a preset number to serve as a final dataset; and input the final dataset into a letterbox function in YOLOv5 to obtain a first image of a preset size; An updating module is used to replace the convolution block structure in CBL in the backbone feature extraction network in YOLOv5 with a depthwise separable convolution structure, and to add a channel attention mechanism to the CSP module in the backbone feature extraction network to obtain an updated backbone feature extraction network; wherein the depthwise separable convolution structure is set as a combination of channel-by-channel convolution and point-by-point convolution, and the output result of the channel-by-channel convolution is input into the point-by-point convolution; a channel attention mechanism is added to the CSP module in the neck structure network in YOLOv5 to obtain an updated neck structure network; after replacing the trained Kmeans algorithm in the detection structure network in YOLOv5 with the trained Kmeans++ algorithm, an updated detection structure network is obtained; The acquisition module is used to input the first image into the updated backbone feature extraction network to obtain the backbone output data, and then input the backbone output data as input data into the updated neck structure network to obtain three types of feature maps with preset output scales; the feature map is used as input data of the updated detection structure network to obtain the final prediction box.

7. The system for detecting surface defects of wind turbine blades according to claim 6, It is characterized in that The image module includes an extension unit, It is used to perform data augmentation on images in a wind blade surface defect image dataset by using translation, rotation, flipping, affine transformation, and noise addition, so as to expand the number of images in the wind blade surface defect image dataset to a preset number.

8. The system for detecting surface defects of wind turbine blades according to claim 6, It is characterized in that The update module comprises a first update unit, It is used to use the output parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 as the input parameters of the depthwise separable convolution structure, and use the output parameters of the depthwise separable convolution structure as the input parameters of the docking with the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5, thereby completing the replacement of the convolution block structure in the CBL in the backbone feature extraction network of YOLOv5 with the depthwise separable convolution structure.

9. The system for detecting surface defects of wind turbine blades according to claim 6, It is characterized in that The update module comprises a second update unit, Used to use the output parameters of the CSP module in the backbone feature extraction network as the input parameters of the channel attention mechanism; where the input parameters are the input feature maps; Then, the channel attention mechanism is used to assign weights to the input feature map, and the weighted feature map is used as the final output data of the CSP module.

10. A non-volatile computer storage medium, It is characterized in that Computer instructions are stored thereon, and when the computer instructions are executed, a method for detecting surface defects of a wind turbine blade as described in any one of claims 1 to 5 is implemented.

Citation Information

Patent Citations

  • Fan blade defect detection method and system based on improved SSD model

    CN114663376A

  • Wind turbine surface defect detection method

    CN115719337A

  • Steel surface defect detection method based on improved YOLOv5s

    CN115829991A

  • PCB defect detection method based on improved YOLOv5 and storage medium

    CN116245817A

  • Edge end steel surface defect detection method based on YOLOv5s-CM

    CN117934363A