Method, system, and medium for detecting surface defects in wind turbine blades.

Abstract

The present invention provides a method, system, and storage medium for detecting defects on the surface of wind turbine blades. [Solution] The method includes the following: Replacing the convolutional block structure within the CBL in the YOLOv5 backbone structure network with a deep separation convolutional structure, and introducing a channel attention mechanism to the CSP module in the backbone structure network to obtain an updated backbone structure network. Introducing a channel attention mechanism to the CSP module in the YOLOv5 neck structure network to obtain an updated neck structure network. Replacing the trained Kmeans algorithm with a trained Kmeans++ algorithm in the YOLOv5 detection structure network to obtain an updated detection structure network. Obtaining the final prediction box using the updated YOLOv5.

Description

Technical Field

[0001] (Cross - reference to Related Applications) This invention claims priority based on Chinese Patent Application No. 202411514600.X, "Method, System, and Medium for Detecting Defects on the Surface of Wind Turbine Blades", filed with the China National Intellectual Property Administration on October 29, 2024. The entire content thereof is incorporated by reference to form part of this invention and is incorporated herein in its entirety for all purposes.

[0002] This application relates to the technical field of defect detection on the blade surface, and particularly to a method, system, and medium for detecting defects on the surface of wind turbine blades.

Background Art

[0003] The blades of a wind turbine are important components for converting wind energy into mechanical energy, and at the same time, they are also the main load - bearing members of the wind turbine, playing an extremely important role in the safe operation of the entire wind turbine. However, the working environment of the blades is extremely harsh. Not only do aerodynamic loads, gravitational loads, and inertial loads act superimposedly, but they are also operated continuously in a harsh natural environment for a long time. External weather conditions have a great impact on the normal operation of the blades. In particular, bad weather such as typhoons, thunderstorms, snowfall, and sandstorms may damage the blades, and as a result, it may cause economic losses to wind power generation. Therefore, it is necessary to monitor the condition of the wind turbine blades in a timely manner, so as to quickly diagnose faults and prevent the occurrence of problems in the entire power system due to blade faults and cause economic losses.

[0004] Currently, commonly used one-stage object detection methods, such as SSD (Single Shot MultiBox Detector), demonstrate a significant advantage in detection speed, enabling rapid recognition and location of objects. However, this type of algorithm has certain limitations in the detection accuracy of the model, often failing to adequately meet the requirements of high-precision patrol inspection work. In contrast, two-stage object detection methods, such as Faster R-CNN (Faster Region-based Convolutional Network), achieve significantly higher accuracy than one-stage detection algorithms through a more complex network structure and sophisticated detection process. However, it suffers from the problem of relatively slower detection speed. [Overview of the project]

[0005] In view of the above-mentioned problems in the prior art, this application provides a system, method, and medium for detecting surface defects in wind turbine blades. The present invention aims to solve the problems of insufficient detection accuracy or reduced processing speed in conventional object detection techniques.

[0006] In its first aspect, this application provides a method for detecting defects on the surface of wind turbine blades. The method includes: To acquire defect images from the surface of wind turbine blades and construct a dataset of wind turbine blade surface defect images. To obtain defect annotation data for blade surface defect images within a wind turbine blade surface defect image dataset. To expand the number of images included in the wind turbine blade surface defect image dataset to a predetermined number and construct it as the final dataset. The final dataset is input into the letterbox function in YOLOv5 to obtain the first image of a predetermined size. The objective is to replace the convolutional block structure within the CBL module of the YOLOv5 backbone structure network (used for feature extraction) with a deep separation convolutional structure, and to introduce a channel attention mechanism into the CSP module within the backbone structure network to obtain an updated backbone structure network. Here, the deep separation convolutional structure is configured as a combination of channel-level convolution and point-level convolution, and the output result of the channel-level convolution is set to be input to the point-level convolution. The process involves introducing a channel attention mechanism into the CSP module within the YOLOv5 neck structure network to obtain an updated neck structure network. In the YOLOv5 detection structure network, the trained Kmeans algorithm is replaced with a trained Kmeans++ algorithm to obtain an updated detection structure network. The first image is input into the updated backbone structure network to obtain backbone output data, and this backbone output data is input into the updated neck structure network to obtain three types of feature maps with predetermined output scales. The feature maps are used as input data to the updated detection structure network to obtain the final prediction box. Furthermore, the number of images in the wind turbine blade surface defect image dataset was expanded to a predetermined number to form the final dataset, which specifically includes the following: Data augmentation processing is performed on images in a wind turbine blade surface defect image dataset using data augmentation techniques such as translation, rotation, inversion, affine transformation, and noise addition, thereby expanding the number of images in the dataset to a predetermined number. Furthermore, the CBL convolutional block structure in the YOLOv5 backbone network is replaced with a deep separation convolutional structure, specifically including the following: In the CBL of the YOLOv5 backbone network, the output parameters connected to the convolutional block structure are used as input parameters for a deep separation convolutional structure, and the output parameters of the deep separation convolutional structure are used as input parameters connected to the convolutional block structure of the CBL in the YOLOv5 backbone network, thereby replacing the convolutional block structure within the CBL in the YOLOv5 backbone network with a deep separation convolutional structure. Furthermore, by introducing a channel attention mechanism to the CSP module in the backbone structure network, an updated backbone structure network is obtained, specifically including the following: The output parameters of the CSP module within the backbone structure network are used as input parameters for the channel attention mechanism. Here, the input parameters are input feature maps. Next, the channel attention mechanism is used to apply weighting to the input feature maps, and the weighted feature maps are used as the final output data for the CSP module. Furthermore, in the YOLOv5 detection structure network, before replacing the trained Kmeans algorithm with a trained Kmeans++ algorithm to obtain the updated detection structure network, the following steps are further included: S1: Obtain the specified training dataset. S2: Randomly select one sample x from the given training dataset and set it as the initial cluster center a1. S3: Calculate the distance between each unsampled sample x and an already determined cluster center, and let this distance be D(x). Randomly select a new sample using a weighted probability distribution and choose it as the new cluster center. At this time, the probability that sample x is selected as the cluster center is:

number

[0007] Secondly, this application provides a detection system for detecting defects on the surface of wind turbine blades. The system includes: Construction Module: A module for acquiring defect images of wind turbine blade surfaces and constructing a dataset of wind turbine blade surface defect images. Image Module: This module acquires defect annotation data for blade surface defect images within a wind turbine blade surface defect image dataset, expands the number of images in the wind turbine blade surface defect image dataset to a predetermined number to construct a final dataset, and inputs the final dataset into the letterbox function in YOLOv5 to obtain a first image of a predetermined size. Update Module: This module replaces the convolutional block structure within the CBL in the YOLOv5 backbone structure network with a deep separation convolutional structure, and introduces a channel attention mechanism to the CSP module within the backbone structure network to obtain the updated backbone structure network. Here, the deep separation convolutional structure is a combination of channel-unit convolution and point-unit convolution, with the output of the channel-unit convolution being input to the point-unit convolution. The updated neck structure network is obtained by introducing a channel attention mechanism to the CSP module in the YOLOv5 neck structure network. In the YOLOv5 detection structure network, the updated detection structure network is obtained by replacing the trained Kmeans algorithm with the trained Kmeans++ algorithm. Acquisition Module: This module inputs the first image into the updated backbone structure network to acquire backbone output data, then inputs the backbone output data as input data into the updated neck structure network to acquire feature maps of three predetermined output scales. These feature maps are used as input data for the updated detection structure network to acquire the final prediction box. Furthermore, the image module includes an expansion unit. This method uses data augmentation techniques such as translation, rotation, inversion, affine transformation, and noise addition to process images within a wind turbine blade surface defect image dataset, thereby expanding the number of images in the dataset to a predetermined number. Furthermore, the update module includes the first update unit, In the YOLOv5 backbone network, the output parameters connected to the convolutional block structure within the CBL are used as input parameters for the deep separation convolutional structure, and the output parameters of the deep separation convolutional structure are used as input parameters connected to the convolutional block structure within the CBL in the YOLOv5 backbone network, thereby replacing the convolutional block structure within the CBL with a deep separation convolutional structure. Furthermore, the update module includes a second update unit, The output parameters of the CSP module within the backbone structure network are used as input parameters for the channel attention mechanism. Here, the input parameters are input feature maps. Next, the input feature map is weighted using a channel attention mechanism, and the weighted feature map becomes the final output data for the CSP module.

[0008] In a third aspect, this application provides a non-volatile computer storage medium. This medium stores computer instructions, and the execution of these instructions realizes the wind turbine blade surface defect detection method described in any of the above sections. As those skilled in the art will understand, this application has at least the following advantageous effects: This application replaces the CBL convolutional block structure within the YOLOv5 backbone network with a deep separation convolutional structure. Here, the deep separation convolutional structure is a combination of channel-level convolution and point-level convolution, with the output result of the channel-level convolution being input to the point-level convolution. Compared to the conventional CBL convolutional block structure, the number of parameters for one operation in a conventional 3-channel convolution is "number of filters * 3 * 3 * 3", whereas in the method where the output result of the channel-level convolution is input to the point-level convolution, the number of parameters is "channel-level convolution (3 * 3 * 3) + point-level convolution (1 * 1 * 3 * number of filters)". In this way, the number of parameters is significantly reduced, and the problem of relatively slow computation speed can be mitigated. Furthermore, this application incorporates a channel attention mechanism into the CSP module within the backbone structure network, thereby integrating the channel attention mechanism (for example, using a Squeeze-and-Excitation Network as the channel attention mechanism applied to deep convolutional neural networks) into the CSP module. This increases the network depth and enhances feature extraction capabilities, enabling more accurate localization and identification of defects on the wind turbine blade surface. Moreover, by replacing the trained Kmeans algorithm with a trained Kmeans++ algorithm in the YOLOv5 detection structure network, YOLOv5 can generate more optimal anchor parameters, enabling detection that is better suited to the actual target. This solves the problem of low detection accuracy in conventional object detection methods.

[0009] Some embodiments of this disclosure will be described below with reference to the attached drawings. [Brief explanation of the drawing]

[0010] [Figure 1] Figure 1 is a flowchart of a method for detecting surface defects of a wind turbine blade according to an embodiment of the present application. [Figure 2] Figure 2 is a schematic diagram showing the internal structure of a detection system for surface defects of a wind turbine blade according to an embodiment of the present application. [Figure 3] Figure 3 is a schematic diagram showing the network structure of YOLOv5 according to an embodiment of the present application.

Embodiments for Carrying Out the Invention

[0011] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of the present disclosure, and do not mean that the present disclosure is limited only to these preferred embodiments. These preferred embodiments are shown for explaining the technical principles of the present disclosure and do not limit the protection scope of the present disclosure. Based on the preferred embodiments of the present disclosure, all other embodiments that can be obtained by those skilled in the art without creative efforts are also intended to be included in the protection scope of the present disclosure.

[0012] Also, the terms "comprising", "including" or any variation thereof mean non-exclusive inclusion, and thus a process, method, product or apparatus containing a plurality of elements is not limited to only including these elements, but also intends to include other elements not explicitly described or elements inherent to this process, method, product or apparatus. Unless otherwise limited, the elements defined by the description of "comprising..." do not exclude the presence of the same other elements in the process, method, product or apparatus containing the element.

[0013] Hereinafter, the technical solutions according to the embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0014] In an embodiment of the present application, a method for detecting defects on the surface of a wind turbine blade is provided. As shown in Figure 1, the method of this embodiment mainly includes the following steps:

[0015] Step 110: Obtain defect images from the wind turbine blade surface and construct a wind turbine blade surface defect image dataset.

[0016] The wind turbine blade surface defect image dataset used may be provided by an actual wind power plant. Specifically, the defect images of the generator blade surface included in the dataset may be obtained by mounting a camera on the gimbal of a drone and photographing the defective surface of the generator blade. The dataset may consist of 800 blade defect images, with each image having a size of 1280*852.

[0017] Step 120: Obtain defect annotation data for blade surface defect images in the wind turbine blade surface defect image dataset.

[0018] Note that this step may also involve annotating wind turbine blade images using LabelImg. The specific procedure may be as follows: (1) Create a source file directory. (2) Create a new label box. (3) Enter the category names for the defect annotation data, namely "Erosion" and "Gelcoat Delamination". (4) Record the save address for the mark file. Finally, a label file with the extension .xml is generated, saved in VOC format, and stored in the specified target folder.

[0019] Step 130: Expand the number of images in the wind turbine blade surface defect image dataset to a predetermined number to create the final dataset.

[0020] The predetermined number of images may be 2400. Data augmentation processing can be performed on images in a wind turbine blade surface defect image dataset using data augmentation techniques such as translation, rotation, inversion, affine transformation, and noise addition, thereby expanding the number of images in the dataset to a predetermined number. As those skilled in the art will understand, through these transformations, data augmentation techniques not only enrich the training dataset but can also simulate a variety of situations in real environments, thereby increasing the robustness of the model.

[0021] Step 140: Input the final dataset into the letterbox function in YOLOv5 to obtain the first image of a predetermined size.

[0022] The specified dimensions may also be 608*608*3.

[0023] To further explain, 3 indicates the number of channels, and those skilled in the art can divide the first image into three parts for processing.

[0024] In this application, by unifying the input images, the speed of object detection can be effectively improved, and the computational complexity of the model can also be reduced.

[0025] Step 150: In the YOLOv5 backbone structure network, replace the convolutional block structure within the CBL with a deep separation convolutional structure, and introduce a channel attention mechanism into the CSP module within the backbone structure network to obtain the updated backbone structure network. In the YOLOv5 neck structure network, introduce a channel attention mechanism into the CSP module to obtain the updated neck structure network. In the YOLOv5 detection structure network, replace the trained Kmeans algorithm with a trained Kmeans++ algorithm to obtain the updated detection structure network.

[0026] As shown in Figure 3, the YOLOv5 backbone structure network used in this application includes a focus function, a CBL (Conv convolutional layer, Batch Normalization layer, and 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 may be such that the output data of the focus function is input to the CBL algorithm, and the output data of the CBL algorithm is input to the CSP module. Furthermore, the backbone structure network may have additional modules added according to the requirements of those skilled in the art. The deep separation convolutional structure is configured as a combination of channel-unit convolution and point-unit convolution, and the output result of the channel-unit convolution is input to the point-unit convolution.

[0027] Here, the convolutional block structure within the CBL in the YOLOv5 backbone network can be replaced with a deep separation convolutional structure, which may be configured as follows:

[0028] In the CBL of the YOLOv5 backbone network, the output parameters connected to the convolutional block structure are used as input parameters for a deep separation convolutional structure, and the output parameters of the deep separation convolutional structure are used as input parameters connected to the convolutional block structure of the CBL in the YOLOv5 backbone network, thereby replacing the convolutional block structure within the CBL in the YOLOv5 backbone network with a deep separation convolutional structure.

[0029] Here, a channel attention mechanism can be introduced into the CSP module within the backbone structure network to obtain the updated backbone structure network, which may be configured as follows.

[0030] The output parameters of the CSP module within the backbone structure network are used as input parameters for the channel attention mechanism. Here, the input parameters are input feature maps. Next, the channel attention mechanism is used to apply weighting to the input feature maps, and the weighted feature maps are used as the final output data for the CSP module.

[0031] Furthermore, in the YOLOv5 detection structure network, before replacing the trained Kmeans algorithm with a trained Kmeans++ algorithm to obtain the updated detection structure network, this method further includes the step of training the Kmeans++ algorithm: S1: Obtain the specified training dataset. S2: Randomly select one sample x from the given training dataset and set it as the initial cluster center a1. S3: Calculate the distance between each unsampled sample x and an already determined cluster center, and let this distance be D(x). Randomly select a new sample using a weighted probability distribution and choose it as the new cluster center. At this time, the probability that sample x is selected as the cluster center is:

number

[0032] Step 160: The first image is input into the updated backbone structure network to obtain backbone output data. This backbone output data is then input into the updated neck structure network as input data to obtain three types of feature maps with predetermined output scales. These feature maps are used as input data to the updated detection structure network to obtain the final prediction box.

[0033] As shown in Figure 3, the neck structure network is an existing YOLOv5 neck structure network (with a channel attention mechanism introduced in the CSP module). The detection structure network is an existing YOLOv5 neck structure network (with the trained Kmeans algorithm replaced by a trained Kmeans++ algorithm). The three predetermined output scales (in ascending order) are 20*20*255, 40*40*255, and 80*80*255, respectively.

[0034] In addition, Figure 2 shows a detection system for detecting defects on the surface of a wind turbine blade according to an embodiment of this application. As shown in Figure 2, the system according to an embodiment of this application mainly includes the following components.

[0035] Construction Module 210: A module for acquiring defect images of wind turbine blade surfaces and constructing a wind turbine blade surface defect image dataset.

[0036] Image Module 220: A module for obtaining defect annotation data for blade surface defect images within a wind turbine blade surface defect image dataset, expanding the number of images in the wind turbine blade surface defect image dataset to a predetermined number to construct a final dataset, and inputting this final dataset into the letterbox function in YOLOv5 to obtain a first image of a predetermined size.

[0037] The image module 220 includes an expansion unit. This method uses data augmentation techniques such as translation, rotation, inversion, affine transformation, and noise addition to process images within a wind turbine blade surface defect image dataset, thereby expanding the number of images in the dataset to a predetermined number.

[0038] Update Module 230: This module replaces the convolutional block structure within the CBL in the YOLOv5 backbone structure network with a deep separation convolutional structure, and introduces a channel attention mechanism to the CSP module within the backbone structure network to obtain the updated backbone structure network. Here, the deep separation convolutional structure is a combination of channel-unit convolution and point-unit convolution, with the output result of the channel-unit convolution being input to the point-unit convolution. The channel attention mechanism is introduced to the CSP module within the YOLOv5 neck structure network to obtain the updated neck structure network. In the YOLOv5 detection structure network, the trained Kmeans algorithm is replaced with the trained Kmeans++ algorithm to obtain the updated detection structure network.

[0039] The update module 230 includes the first update unit, In the YOLOv5 backbone network, the output parameters connected to the convolutional block structure within the CBL are used as input parameters for the deep separation convolutional structure, and the output parameters of the deep separation convolutional structure are used as input parameters connected to the convolutional block structure within the CBL in the YOLOv5 backbone network, thereby replacing the convolutional block structure within the CBL with a deep separation convolutional structure.

[0040] The update module 230 includes the second update unit, The output parameters of the CSP module within the backbone structure network are used as input parameters for the channel attention mechanism. Here, the input parameters are input feature maps. Next, the channel attention mechanism is used to apply weighting to the input feature maps, and the weighted feature maps are output as the final output data of the CSP module.

[0041] Acquisition Module 240: This module inputs the first image into the updated backbone structure network to acquire backbone output data, and then inputs this backbone output data as input data into the updated neck structure network to acquire three types of feature maps with predetermined output scales. These feature maps are used as input data for the updated detection structure network to acquire the final prediction box.

[0042] In addition, embodiments of this application also provide a non-volatile computer storage medium. This storage medium stores executable instructions, and the execution of these instructions enables the method for detecting defects on the wind turbine blade surface described above.

[0043] Up to this point, the technical solutions of this disclosure have been described with reference to the aforementioned embodiments, but as will be readily apparent to those skilled in the art, the scope of protection of this disclosure is not limited to these specific embodiments. Within the limits of not departing from the technical principles of this disclosure, those skilled in the art may separate or combine the technical means of each embodiment described above, and may also make equivalent modifications or substitutions to the relevant technical features. Any modifications, equivalent substitutions, improvements, etc., made within the scope of the technical ideas and / or technical principles of this disclosure shall all be included within the scope of protection of this disclosure.

Claims

1. Executed by a computer, To acquire defect images of wind turbine blade surfaces and construct a wind turbine blade surface defect image dataset, To obtain defect annotation data for blade surface defect images within a wind turbine blade surface defect image dataset, To expand the number of images included in the wind turbine blade surface defect image dataset to a predetermined number and construct it as the final dataset, The final dataset is input into the letterbox function in YOLOv5 to obtain a first image of a predetermined size, In the YOLOv5 backbone structure network, the convolutional block structure within the CBL is replaced with a deep separation convolutional structure, and a channel attention mechanism is introduced into the CSP module within the backbone structure network to obtain an updated backbone structure network, where the deep separation convolutional structure is a combination of channel-unit convolution and point-unit convolution, and the output result of the channel-unit convolution is input into the point-unit convolution, and a channel attention mechanism is introduced into the CSP module within the YOLOv5 neck structure network to obtain an updated neck structure network, and in the YOLOv5 detection structure network, the trained Kmeans algorithm is replaced with a trained Kmeans++ algorithm to obtain an updated detection structure network, This includes inputting the first image into the updated backbone structure network to obtain backbone output data, inputting the said backbone output data as input data into the updated neck structure network to obtain three types of feature maps with predetermined output scales, using these feature maps as input data to the updated detection structure network to obtain the final prediction box, and so on. A method for detecting defects on the surface of a wind turbine blade, characterized by the features described above.

2. The number of images in the blade surface defect image dataset is expanded to a predetermined number, and the final dataset is constructed. Specifically, it includes the following: Data augmentation processing is performed on images in a wind turbine blade surface defect image dataset using data augmentation techniques such as translation, rotation, inversion, affine transformation, and noise addition, thereby expanding the number of images in the dataset to a predetermined number. A method for detecting defects on the surface of a wind turbine blade according to feature 1.

3. In the YOLOv5 backbone network, the convolutional block structure within the CBL is replaced with a deep separation type convolutional structure. Specifically, it includes the following: In the CBL of the YOLOv5 backbone network, the output parameters connected to the convolutional block structure are used as input parameters for a deep separation convolutional structure, and the output parameters of the deep separation convolutional structure are used as input parameters connected to the convolutional block structure of the CBL in the YOLOv5 backbone network, thereby replacing the convolutional block structure within the CBL in the YOLOv5 backbone network with a deep separation convolutional structure. A method for detecting defects on the surface of a wind turbine blade according to feature 1.

4. A channel attention mechanism is introduced to the CSP module within the backbone structure network to obtain the updated backbone structure network. Specifically, it includes the following: This method uses the output parameters of the CSP module within the backbone structure network as input parameters for the channel attention mechanism, where the input parameters are input feature maps. Next, the input feature map is weighted using a channel attention mechanism, and the weighted feature map is used as the final output data for the CSP module. A method for detecting defects on the surface of a wind turbine blade according to feature 1.

5. Before obtaining the updated detection structure network by replacing the trained K-means algorithm with a trained K-means++ algorithm in the YOLOv5 detection structure network, the following is also included: S1: Obtain the specified training dataset, S2: Randomly extract one sample x from the given training dataset and set the initial cluster center a １ And S3: Calculate the distance between each unextracted sample x and the already determined cluster center, denote this distance as D(x), randomly extract a new sample using a weighted probability distribution, and select it as the new cluster center. The probability that sample x is selected as the cluster center is: [Math 3] Here, X represents a predetermined training dataset, S4: Steps S2 and S3 are repeated until K initial cluster centers are determined, thereby obtaining a trained Kmeans++ algorithm, where K is a predetermined constant, and including the following: A method for detecting defects on the surface of a wind turbine blade according to feature 1.

6. Including the following, A construction module for acquiring defect images of wind turbine blade surfaces and building a wind turbine blade surface defect image dataset, An image module for obtaining defect annotation data for blade surface defect images within a wind turbine blade surface defect image dataset, expanding the number of images in the wind turbine blade surface defect image dataset to a predetermined number to construct a final dataset, and inputting the final dataset into the letterbox function in YOLOv5 to obtain a first image of a predetermined size, This update module replaces the convolutional block structure within the CBL in the YOLOv5 backbone structure network with a deep separation convolutional structure, and introduces a channel attention mechanism to the CSP module in the backbone structure network to obtain an updated backbone structure network, wherein the deep separation convolutional structure is configured as a combination of channel-unit convolution and point-unit convolution, the output result of the channel-unit convolution is input to the point-unit convolution, the channel attention mechanism is introduced to the CSP module in the YOLOv5 neck structure network to obtain an updated neck structure network, and in the YOLOv5 detection structure network, the trained Kmeans algorithm is replaced with a trained Kmeans++ algorithm to obtain an updated detection structure network. The system includes an acquisition module that inputs the first image into the updated backbone structure network to obtain backbone output data, inputs the backbone output data as input data into the updated neck structure network to obtain three types of feature maps with predetermined output scales, uses these feature maps as input data to the updated detection structure network to obtain the final prediction box, and A system for detecting defects on the surface of wind turbine blades, characterized by the following features.

7. The image module includes an expansion unit, This method uses data augmentation techniques such as translation, rotation, inversion, affine transformation, and noise addition to process images within a wind turbine blade surface defect image dataset, thereby expanding the number of images in the dataset to a predetermined number. A system for detecting defects on the surface of a wind turbine blade according to feature 6.

8. The update module includes the first update unit, In the YOLOv5 backbone network, the output parameters connected to the convolutional block structure within the CBL are used as input parameters for the deep separation convolutional structure, and the output parameters of the deep separation convolutional structure are used as input parameters connected to the convolutional block structure within the CBL in the YOLOv5 backbone network, thereby replacing the convolutional block structure within the CBL in the YOLOv5 backbone network with a deep separation convolutional structure. A system for detecting defects on the surface of a wind turbine blade according to feature 6.

9. The update module includes a second update unit, The output parameters of the CSP module in the backbone structure network are used as input parameters for the channel attention mechanism, where the input parameters are input feature maps. Next, the input feature map is weighted using a channel attention mechanism, and the weighted feature map is used as the final output data for the CSP module. A method for detecting defects on the surface of a wind turbine blade according to feature 1.

10. Computer instructions are stored thereon, and when these computer instructions are executed, a method for detecting defects on the surface of a wind turbine blade according to any one of claims 1 to 5 is realized. A non-volatile computer storage medium characterized by the following features.

Images