Fan blade defect detection method and system based on target detection model

Abstract

The invention relates to the technical field of wind power generation, in particular to a fan blade defect detection method and system based on a target detection model, and the method specifically comprises the steps: obtaining image data of a to-be-detected blade, and generating a fan blade defect image through employing an improved HP-VAE-GAN, so as to achieve the extension of a training set; in combination with a multi-exposure image fusion method based on a spatial frequency integration framework (MEF-SFI), the data quality is optimized; performing lightweight improvement on a YOLOV11 target detection model, and inputting the marked data set into the trained model to obtain a final fan blade state detection result; according to the method, the edge computing equipment is combined, the problems of large data acquisition amount, low transmission speed and the like are solved, and the accuracy, convenience and real-time performance of blade detection are improved on the whole.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and in particular to a method and system for detecting defects in wind turbine blades based on a target detection model. Background Art

[0002] Current wind turbine blade defect detection methods face multiple challenges, among which the most prominent problems are the limited coverage and uneven data distribution. In addition, wind turbine blade images are acquired in complex natural environments and are usually significantly affected by factors such as uneven lighting, low color saturation, and noise interference. In current wind turbine blade defect detection technologies, traditional detection models are inefficient when processing large-scale, high-resolution wind turbine blade images and cannot meet the needs of rapid detection and real-time monitoring. At the same time, detection accuracy is limited by the model's ability to recognize subtle defects in complex backgrounds, especially in the detection scenarios of unevenly distributed small targets (such as microcracks, pitting corrosion, etc.).

[0003] Existing models often exhibit problems such as low recognition rate and high false detection rate. The reason is that the models find it difficult to effectively extract and amplify the key features of these targets; and they also face problems such as slow data transmission speed and bandwidth bottlenecks during large-scale data collection.

[0004] Therefore, a method and system for wind turbine blade defect detection based on target detection model are proposed. Summary of the Invention

[0005] In order to overcome the problems of low recognition rate and high false detection rate often exhibited by existing models, the reason is that the model finds it difficult to effectively extract and amplify the key features of these targets; and the problems faced in the process of large-scale data acquisition, such as slow data transmission speed and bandwidth bottleneck, the present invention provides a method and system for wind turbine blade defect detection based on a target detection model, which uses an improved HP-VAE-GAN to generate wind turbine blade defect images to achieve the expansion of the training set; combines a multiple exposure image fusion method based on a spatial frequency integration framework (MEF-SFI) to optimize data quality; makes lightweight improvements to the YOLOV11 target detection model, and inputs the labeled data set into the trained model to obtain the final wind turbine blade status detection results; combines the present invention with edge computing equipment to solve the problems of large data collection volume and slow transmission speed, and improves the accuracy, convenience and real-time performance of blade detection as a whole.

[0006] The technical solution is: a method and system for wind turbine blade defect detection based on a target detection model, including the following modules:

[0007] First, a hexacopter drone was used to capture images and videos of wind turbine blades at the wind farm. The drone's stable flight capabilities and high payload capacity allow for flexible, high-precision photography of wind turbine blades, ensuring the quality of the captured images and videos meets subsequent analysis requirements.

[0008] Secondly, the data in the video is decomposed into image data at a pace of 2 frames / second, and the total amount of data after expansion;

[0009] After data collection, the images are first manually annotated;

[0010] Manually mark the defect areas and types in the image, select the defect areas, and mark the defect types (such as cracks, corrosion, deformation, etc.);

[0011] The improved HP-VAE-GAN is used to generate defect images of wind turbine blades to expand the training set.

[0012] The improved HP-VAE-GAN algorithm is specifically as follows:

[0013] (1) A single-sample generation model consisting of Patch-VAE and Patch-GAN;

[0014] (2) The encoder of HP-VAE-GAN contains three convolution blocks, and a convolution block is added to the encoder;

[0015] (3) By adding CBAM (Convolutional Block Attention Module) to the Patch-VAE encoder, the network’s feature extraction and representation capabilities are further improved;

[0016] This module can significantly improve the network's feature extraction and representation capabilities, and enhance the generative model's ability to render details when generating defect images of wind turbine blades.

[0017] A multiple exposure image fusion method based on the spatial frequency integration framework (MEF-SFI) is used for the images.

[0018] The multi-exposure image fusion method based on the spatial frequency integration framework (MEF-SFI) involves three basic elements:

[0019] Spatial fusion branch for local information fusion;

[0020] Frequency fusion branch for global illumination information fusion;

[0021] The attention-based fusion block is designed for measuring and integrating complementary information.

[0022] This module can generate high dynamic range fused images with good visual effects and detail performance.

[0023] Perform a series of preprocessing operations on the collected images, such as image enhancement, image filtering, image segmentation, and feature extraction, to improve the effect of subsequent model training.

[0024] In terms of model selection, the present invention makes lightweight improvements to the YOLOv11 target detection model to improve the performance and efficiency of wind turbine blade defect detection.

[0025] In order to better embed YOLOV11 into mobile devices, it is necessary to make lightweight improvements to the YOLOV11 target detection model, specifically:

[0026] (1) The project replaces the backbone network of YOLOV11 with the EMO model, which can effectively improve the performance and efficiency when detecting defects in wind turbine blades;

[0027] (2) An EMA attention mechanism is added before the detection head to enhance the feature extraction of small target defects through an efficient multi-scale attention mechanism based on cross-space learning, thereby improving the target recognition ability in wind turbine blade defect detection;

[0028] (3) Combining C3K2 in the model with the lightweight attention network LANet effectively accelerates the defect recognition process;

[0029] (4) A new operator Partial_C_v1Detect is used in the detection head to improve the efficiency of spatial feature extraction in fan blade defect detection.

[0030] The improved model was trained on a self-built wind turbine blade damage dataset. By comparing it with the original algorithm, the impact of hyperparameter adjustment on defect detection performance was analyzed. Experimental verification and results analysis were then conducted.

[0031] Beneficial Effects: The data of this invention verifies that the improved model has excellent detection capabilities for the differentiated characteristics of different types of blade damage in complex industrial scenarios. This invention combines edge computing devices to solve problems such as large data collection volume and slow transmission speed, and improves the accuracy, convenience and real-time performance of blade detection overall. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is the overall framework diagram of YOLOV11 of the present invention.

[0033] Figure 2 This is an overall view of the improved HP-VAE-GAN of the present invention.

[0034] Figure 3This is the structure of the improved Encoder of the present invention.

[0035] Figure 4 This is the structure of the first spatial frequency fusion module of the present invention.

[0036] Figure 5 This is the structure of the second spatial frequency fusion module of the present invention.

[0037] Figure 6 It is a unified meta-movement block abstracted from the multi-head self-attention / feedforward network and the inverted residual block in the present invention.

[0038] Figure 7 This is the LANet network structure of the present invention.

[0039] Figure 8 This is a block diagram of the main research content of this invention.

[0040] Figure 9 This is the data collection flow chart and data analysis flow chart of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described below with reference to the embodiments shown in the accompanying drawings.

[0042] A method and system for detecting defects in wind turbine blades based on a target detection model include an improved image processing model framework that combines a variational autoencoder (VAE) and a generative adversarial network (GAN).

[0043] The framework consists of two parts: VAE for feature extraction and GAN for image generation. Figure 2 As shown in the figure, the improved HP-VAE-GAN model contains N+1 generators, denoted as G...GN. According to the generated results, from Rough to Fine; for each scale n=,...,N, the training sample x is downsampled to xn. Different from the original model, the improved HP-VAE-GAN adds CBAM to the encoder of Patch-VAE and adds a convolution block to the encoder;

[0044] The specific structure is as Figure 3 As shown in Figure 1, in Patch-VAE, a single sample x is input and cut into pieces of size r*r. The sample x is fed into the encoder (CE) containing the CBAM and outputs a feature map CE(x) of size H*W*C, where H, W, C represent the height, width, and channel of the feature map, respectively; and the number of image patches ρj (j=1...T) is T=H*W. From the perspective of the generative model, if T patches can be easily fitted, then a large number of new samples can be generated by directly sampling from this distribution.

[0045] This model draws on VAE theory, which achieves this by calculating the KL loss of the two. The formal expression of the KL loss function is as follows:

[0046]

[0047] The encoder of HP-VAE-GAN contains three convolutional blocks, each of which consists of a Conv2d convolutional layer, a BN layer, and a Leaky_Relu activation function;

[0048] The encoder structure in the modified Patch-VAE is as follows Figure 2 As shown, a convolution block is added to the original encoder;

[0049] And insert CBAM in the middle of the four convolutional blocks to enhance the extracted features. The functions and parameters used in the network are as follows: Figure 2 As shown;

[0050] GAN as a generator will generate images from the features learned by the VAE encoder through the decoding process. The GAN generator is responsible for generating new images;

[0051] Given a grayscale image I∈R H*W , its discrete Fourier transform (DFT) I is as follows:

[0052]

[0053] Figure 4 The architecture of the proposed multi-exposure image fusion network;

[0054] The network consists of two feature extraction paths, each corresponding to an overexposed / underexposed image as input:

[0055] Each path has 4 convolution blocks and Spatial-FrequencyFusionModule;

[0056] After obtaining the fused representations from the four fusion blocks, a concatenation operator and a single convolution block are used to generate the final fused image;

[0057] The architecture of the multi-exposure image fusion technology is a dual-path fully convolutional neural network, such as Figure 4 Given a three-channel RGB image Io, Iu∈RHW3, we first convert the color space of the source image from RGB to YCbCr, and only apply the proposed fusion method on the Y channel;

[0058] Use the Y channel of the overexposed image and underexposed images The network takes as input, where H and W are the height and width of the image, and then applies two feature extraction paths to each input image. The feature extraction path uses multiple 3×3 convolutional layers with PReLU activation function as the basic feature extraction unit, and dense connections of DenseNet to strengthen the connection between different layers in each path;

[0059] The dual feature extraction path consists of four convolutional blocks, each of which corresponds to the spatial frequency fusion module (SFFM) integration function and Generate fused representation in spatial and frequency domains

[0060] The architecture of SFFM is as follows Figure 5 As shown, the fusion function F_f^i is generated by these fusion blocks and applies 1×1 convolution operation and tanh activation function to obtain the output I_f^Y∈RH×W×1 of the entire network. For Cb and Cr channels, we simply fuse them by weighted averaging as shown below:

[0061]

[0062] C1 and C2 are the Cb or Cr channel values of the input image pair, while τ is set to 128 in our work, and Cf is the fused Cb or Cr channel;

[0063] Concatenate the fused Y channel with the fused Cb and Cr channels and convert back to RGB color space to obtain the final high dynamic range image If∈RH×W×3;

[0064] The loss function consists of two parts: spatial loss and frequency loss. The function of spatial loss is to impose pixel-level constraints on the input image pair and the fusion result, that is, and

[0065] The frequency loss is to impose a global constraint between the input image pair and the fusion result in the Fourier domain, and its term is:

[0066] and

[0067] In the frequency domain, global features of the image are extracted;

[0068] In the spatial domain, the local features of the image are preserved;

[0069] Combining the features of the frequency domain and the spatial domain, the two images are effectively fused. The fusion process uses the channel attention mechanism to enhance the features, and the spatial attention mechanism is also used to improve the feature fusion effect:

[0070] Design a multi-exposure image fusion network, such as Figure 5 As shown:

[0071] The network consists of two paths, one for processing overexposed images and the other for processing underexposed images;

[0072] Each path consists of four convolutional blocks;

[0073] Each path also includes a spatial-frequency domain fusion module;

[0074] The feature information of different paths is merged through the concatenation operator;

[0075] The combined features are passed through a single convolution block to generate the final fused image;

[0076] The generated fused image has a high dynamic range;

[0077] The final image has good visual effects and detail expression, and is suitable for image enhancement.

[0078] The structure of the fusion network improves the expressiveness of the image through multiple exposure fusion technology.

[0079] Object detection model optimization:

[0080] Improvement point 1: Emo model

[0081] An improved Emo model is provided as a lightweight object detection model.

[0082] The Emo model is based on an improved inverted residual block design, such as Figure 6 As shown:

[0083] This model is dedicated to studying the efficient Inverted Residual Block (IRB) in MobileNetV2 and the effective Multi-Head Self-Attention (MHSA) / Feedforward Network (FFN) module in Transformer, hoping to integrate these two advantages at the infrastructure design level. Figure 6As shown in the paper, when a single residual IRB with inductive bias is introduced into the attention model, it is found that the two underlying submodules in the Transformer (i.e., FFN and MHSA) have similar structures to the IRB. Therefore, this model proposes to extend the CNN-based IRB to a dual-resident transformer and abstract a single-resident meta-mobile module (MMB) for lightweight model design. The MMB uses the parameter expansion rate λ and the efficiency operator F to instantiate different modules, namely IRB, MHSA, and FFN. The MMB can reveal the consistent essential expression of the above three modules and can be regarded as an improved lightweight condensed set of Transformers. This meta-paradigm can not only describe the current efficient modules, but also has guiding significance for the creation of new efficient modules.

[0084] MMB only contains a derived iRMB that absorbs the advantages of lightweight CNN and Transformer. Taking the image input X(∈RC×H×W) as an example, MMB first uses an ExpansionMLPe with an output / input ratio equal to λ to expand the channel dimension:

[0085] X E =mlp e (X)(∈R λC×H×W )

[0086] The intermediate operator F can further enhance the image features, such as identity operator, static convolution, dynamic MHSA, etc. Considering that MMB is suitable for efficient network design, the concept of F as an efficient operator is proposed, which is expressed as

[0087] X_f=f(Xe)(∈rλC×H×W)

[0088] Build a shrinkage MLP model with an inverted input-output ratio equal to λ to reduce the channel size:

[0089] X_S=mlps(Xf)(∈rC×H×W)

[0090] Improve performance through element-by-element multiplication, element-by-element addition, and residual network structure;

[0091] The Emo model stacks multiple inverted residual blocks to avoid complex operations and diverse module designs;

[0092] The design of the Emo model improves the usability and uniformity of the model.

[0093] Improvement point 2: Add EMA attention mechanism

[0094] Use a new efficient multi-scale attention (EMA) module;

[0095] Multi-scale feature extraction:

[0096] Process input features using 1×1 and 3×3 convolution kernels in parallel

[0097] Generate multi-scale features: X_1 = Conv_{1×1}(X), X_3 = Conv_{3×3}(X)

[0098] Cross-dimensional interaction: Split the channel dimension into g groups, and reshape each group of channels into the batch dimension: X'_1 = Reshape(X_1, (B×g, C / g, H, W)), X'_3 = Reshape(X_3, (B×g, C / g, H, W)), and calculate the inter-channel dependency through cross attention:

[0099] Feature fusion: Use dynamic weights to fuse multi-scale features: Y = α⊙X_1+(1-α)⊙X_3, where α is generated by the spatial attention module, focusing on the coordinates of the crack area, Attn = softmax((W_qX'_1)(W_kX'_3)^T / √d)

[0100] The goal is to preserve the information on each channel and reduce computational overhead;

[0101] Reshape some channels into batch dimensions and group the channel dimensions into multiple sub-features so that the spatial semantic features are evenly distributed in each feature group;

[0102] This process improves the detection capability of small targets;

[0103] The enhancement of small target features improves the accuracy of wind turbine blade defect detection.

[0104] Improvement point 3: MobileViT network application

[0105] MobileViT is used as the backbone network for feature extraction.

[0106] MobileViT combines the local feature extraction of CNN and the global feature modeling capabilities of Transformer. Figure 7 Shown:

[0107] When the c3k parameter is set to False, the C3K2 module adopts a standard Bottleneck structure, which is consistent with the traditional C2F module. The standard Bottleneck is usually composed of convolutional layers, batch normalization (BatchNormalization) and activation functions (such as ReLU), aiming to reduce the amount of computation while maintaining good feature representation capabilities. The depth controllability of the Bottleneck structure enables the network to have good scalability and computational efficiency in wind blade defect detection, ensuring fast processing and accurate recognition. On the other hand, when the c3k parameter is set to True, the C3K2 module replaces the Bottleneck with the C3 module. The C3 module is an improved Bottleneck structure that enhances the expressive power of feature extraction by introducing more complex convolution operations and skip connections. The C3 module usually uses multiple convolution kernels (such as depthwise separable convolution or grouped convolution) to process input features, which can not only better capture multi-scale information, but also reduce information loss and improve gradient flow, thereby improving the feature expression capability in wind blade defect detection;

[0108] The network improves image classification and object detection performance while maintaining computational efficiency;

[0109] Multiple encoder modules of the input gradually extract features;

[0110] After the features are fused through multiple EFA (Enhanced Feature Fusion) modules, their performance is further improved;

[0111] The AFF decoder fuses the EFA-processed features with the sampled features in the decoder;

[0112] This fusion enhances the understanding and representation of input features.

[0113] Improvement point 4: Detection head improvement

[0114] Provides a new type of convolution operation that performs convolution only on a specific area:

[0115] A new Partial_C_v1Detect is proposed, which can more efficiently extract spatial features by simultaneously reducing redundant computation and memory access. Based on PConv, FasterNet is further proposed, a new family of neural networks that achieves significantly higher running speeds than other networks on a wide range of devices without compromising accuracy in various vision tasks. For example, on ImageNet-1k, a small FasterNet-T is 31 times, 31 times, and 25 times faster than MobileVitXXS on GPUs, CPUs, and ARM processors, respectively, while achieving a 29% improvement in accuracy.

[0116] The novel convolution operation reduces computational effort and improves detection efficiency;

[0117] Fastnetblock contains 3×3 convolution, 1×1 convolution layer, batch normalization and activation function;

[0118] The input image is converted into a suitable grid format through the embedding layer;

[0119] After the image passes through four Fastnet blocks, it is processed by global pooling and fully connected layers;

[0120] The final detection result is output, providing efficient target detection performance.

[0121] Based on the demand analysis, the system architecture of the deep learning wind blade health monitoring system was designed. The whole system consists of two parts: the server and the client. The server and the client need to communicate in the same local area network. The whole system module block diagram is as follows Figure 8 As shown:

[0122] System users collect defect data of wind turbine blades through interface operations;

[0123] After data collection, the data processing module is used for preprocessing. The data collection flow chart and data analysis flow chart in the system are as follows: Figure 9 As shown;

[0124] Data Storage Module: This module is responsible for storing key data required for system operation. The stored data primarily consists of two parts: original images of wind turbine blades and trained object detection weights. Historical data is stored with labels, including the region and turbine number. If the data comes from multiple wind farms, the region of the stored data is the wind farm name. When used within a single wind farm, this can be replaced with a general region divided by azimuth, such as the east or south of the wind farm.

[0125] Data processing module: Diagnose and read the PNG file at the current moment, then use the improved YOLOv11 model to detect and classify it, output the deep learning detection and classification results, and display the final results on the human-computer interaction interface;

[0126] User interaction module: User interaction includes two aspects, namely the system user input interface and the calculation result display interface. The user input interface is responsible for accepting user input requests and sending them to the server. The user input interface mainly includes the data upload interface and the data selection interface. The calculation result display interface is responsible for displaying the calculation results returned by the user request to the user. The functional flow chart of data analysis is as follows Figure 9As shown in the figure, this function uses the improved YOLOv11 model to automatically annotate the generated analysis results and use boxes to mark the damage information in the analysis results, which can quickly analyze the health status of wind turbine blades;

[0127] User Management Module: The system verifies user identities through account and password authentication and allows them to access specific resources and functions in the system to ensure system security and controllable permissions. When a user requests access to the wind turbine blade fault detection system, they are required to enter their account and password for verification. If the verification information entered is correct, the system will authorize the user to access the resources and functions within their specific permissions; if the user enters incorrect verification information, the system will deny access.

[0128] Data preprocessing breaks down the video into individual image frames;

[0129] Applying an improved generative adversarial network to preprocess image data;

[0130] Further processing of images using multi-exposure image fusion technology;

[0131] The processed images are annotated, and the annotation results are used for subsequent target detection;.

[0132] Use the improved object detection model to detect and classify the annotated images;

[0133] The object detection model provides accurate defect identification during the inspection process;

[0134] The detection results are saved in the data storage module for subsequent analysis and decision-making.

[0135] Example:

[0136] This example verifies the accuracy of wind turbine blade defect detection in this method;

[0137] This example includes four types of defects: gel coat damage, cracks, blisters, and oil stains. There are 683 training data and 172 test data.

[0138] In terms of mean average precision (mAP5) across all defect categories reached 96%, a 3 percentage point improvement over the original YOLOv11 model (93%). Specifically for each damage type, the model achieved a top-tier accuracy rate of 982% in crack detection, with near-perfect accuracy for pinholes (995%) and oil stains (994%). Furthermore, it achieved a reliable recognition accuracy of 866% for gel coat damage detection, a feature with more complex surface deformation characteristics. Experimental data confirmed that the improved model possesses excellent detection capabilities for the differentiated characteristics of different types of blade damage in complex industrial scenarios.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method and system for detecting defects in wind turbine blades based on a target detection model, characterized in that: Includes the following modules: (1) Data acquisition module: A six-rotor drone is used to collect images and videos of wind turbine blades. The drone decomposes the videos into image data at a rate of 2 frames per second. (2) Data enhancement module: An improved HP-VAE-GAN is used to generate defect images of wind turbine blades to expand the training set. Convolutional blocks are added to the encoder of the HP-VAE-GAN and CBAM (Convolutional Block Attention Module) is integrated to improve feature extraction capabilities. (3) Image fusion module: It fuses multiple exposure images based on the spatial frequency integration framework (MEF-SFI), including spatial fusion branch, frequency fusion branch and attention-based fusion block to generate a high dynamic range fused image; (4) Object detection module: Lightweight improvements are made to the YOLOv11 model, including replacing the backbone network with the EMO model, adding the EMA attention mechanism before the detection head, combining C3K2 with the lightweight attention network LANet, and using the new operator PConv to improve the efficiency of spatial feature extraction; (5) Edge computing module: Develop the user interface through PyQt5 to realize automatic data collection, storage and intelligent diagnosis and analysis.

2. The system for detecting defects in wind turbine blades based on a target detection model according to claim 1, characterized in that: The improved HP-VAE-GAN specifically includes: (a) A single-sample generation model consisting of Patch-VAE and Patch-GAN; (b) The encoder consists of four convolutional blocks, among which the newly added convolutional block integrates the CBAM mechanism to enhance feature expression through channel and spatial attention weight distribution; (c) The input image is divided into multiple image blocks, each of which contains local features and texture information.

3. The system for detecting defects in wind turbine blades based on a target detection model according to claim 1, characterized in that: The fusion method of the spatial frequency integration framework (MEF-SFI) includes: (a) Convolution operations are performed on overexposed and underexposed input images respectively; (b) The image is converted to the frequency domain by fast Fourier transform (FFT) to extract global features, and then converted back to the spatial domain by inverse fast Fourier transform (IFFT) to preserve local features; (c) Optimizing feature fusion using channel attention and spatial attention mechanisms.

4. The system for detecting defects in wind turbine blades based on a target detection model according to claim 1, characterized in that: The lightweight improvements of the YOLOv11 model include: (a) The backbone network of the EMO model is based on an improved inverted residual block design, which includes a 1×1 convolutional layer, a 3×3 depthwise convolutional layer, and an attention matrix; (b) Combining C3K2 in the model with the lightweight attention network LANet effectively speeds up the defect recognition process; (c) An EMA attention mechanism is added before the detection head to enhance the feature extraction of small object defects through an efficient multi-scale attention mechanism based on cross-space learning; (d) Use the Partial_C_v1Detect operator to perform convolution on specific areas of the feature map to reduce the amount of computation.

5. The system for detecting defects in wind turbine blades based on a target detection model according to claim 2, characterized in that: The edge computing module further includes: (a) Data processing unit: performs image enhancement, filtering, segmentation and feature extraction preprocessing on the collected images; (b) Data storage unit: stores marked defect types and detection results, wherein the defect types include cracks, corrosion, and deformation; (c) Intelligent diagnosis unit: Outputs blade health status detection results based on the trained lightweight YOLOv11 model.

6. The method for detecting blade damage of a wind turbine generator set according to any one of claims 1 to 5, wherein: The following steps are involved: (1) Collect image and video data of wind turbine blades using a six-rotor drone and decompose the video into image sequences; (2) Use the improved HP-VAE-GAN to generate defect images to expand the training set, and fuse multiple exposure images through the MEF-SFI framework to optimize data quality; (3) Inputting the labeled data set into the lightweight improved YOLOv11 model for training, wherein the labeling includes defect area selection and type labeling; (4) The collected data is processed in real time through edge computing devices, and the blade damage detection results are output and stored in the database.

7. The wind turbine blade damage detection method according to claim 6, characterized in that: The training process of the YOLOv11 model includes: (a) Divide the input image into multiple image blocks and then divide it into multiple pre-selected boxes; (b) After training, the best pre-selected box is obtained, which is the marked defect area, and the defect type is marked.

Images