Fan blade crack detection method

Abstract

The invention provides a fan blade crack detection method, which comprises the following steps of: for a fan blade image acquired by an unmanned aerial vehicle, performing deblurring processing on the image through a pre-processing algorithm designed by the invention, and overcoming the problem of motion blurring when the unmanned aerial vehicle acquires the fan blade image; and carrying out crack detection on the acquired image by adopting a crack detection neural network model improved based on YOLOv8. According to the crack detection model, self texture interference of the blade can be avoided, the crack detection precision is improved, crack characteristics of different scales are captured through a Stem module of a multi-branch structure, a CAM module is added into a C2f-1 module, crack attention is integrated, and crack related characteristics are enhanced; a residual connection mode of a residual block is designed in a C2f-2 module to avoid the problem of deep network gradient disappearance, gradient flow and effective transmission of crack characteristics are guaranteed, response of crack-related channels is enhanced through channel re-calibration, and irrelevant channels such as blade textures are inhibited.

Description

Methods for detecting cracks in wind turbine blades Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a method for detecting cracks in wind turbine blades. Background Technology

[0002] Wind turbine blades are the core components of wind power generation systems, and their health directly determines power generation efficiency and operational safety. Blade cracks are one of the main causes of blade failure. Traditional testing methods (such as acoustic emission testing and ultrasonic testing) generally suffer from high testing costs and complex operating procedures, making it difficult to meet the high-efficiency inspection needs of large-scale wind farms.

[0003] Vision-based drone inspection methods are increasingly becoming the preferred solution in the industry due to their advantages of portability and low inspection cost. However, drones are prone to image motion blur during high-speed flight, a problem that severely restricts the improvement of inspection accuracy. Specifically, in actual inspection scenarios, the drone's vibration is directly transmitted to the onboard imaging equipment through the aircraft, resulting in motion blur and loss of detail in the captured leaf images. Ultimately, this seriously affects the accuracy and reliability of subsequent image-based defect detection and analysis.

[0004] While some existing technologies attempt to improve blurring by incorporating advanced camera systems or employing image post-processing techniques, the former suffers from high costs and poor equipment compatibility, while the latter's high computational complexity makes it difficult to meet the needs of real-time on-site inspection. Furthermore, traditional visual inspection methods are susceptible to interference from the blade's own texture, leading to insufficient accuracy in crack feature extraction and consequently, false positives and false negatives. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for detecting cracks in wind turbine blades, which solves the problems of motion blur and feature recognition during image acquisition.

[0006] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0007] A method for detecting cracks in wind turbine blades, which uses the following crack detection model to detect cracks in wind turbine blade images;

[0008] The crack detection model, from front to back, includes: Stem module, C2f-1 module, 3×3 convolution, C2f-2 module, 3×3 convolution, C2f-2 module, and feature pyramid network.

[0009] Furthermore, in the Stem module:

[0010] First, features are extracted from the input image through multiple parallel branches, including standard 3×3 convolution with stride 2, standard 5×5 convolution with stride 2, dilated convolution with dilation rate of 2, and max pooling with stride 2.

[0011] Then the outputs of each branch are concatenated and the features are rearranged.

[0012] Furthermore, in the C2f-1 module:

[0013] The input features are first processed by two parallel Bottleneck modules to obtain feature maps. and In the two Bottleneck modules, one shortcut is True and the other shortcut is False.

[0014] Then proceed with the following processing:

[0015]

[0016] In the formula, This is the output feature map of the C2f-1 module. For 1×1 convolution, For feature splicing, This indicates that the data has been processed by the CAM module.

[0017] Furthermore, the processing procedure of the CAM module is as follows:

[0018]

[0019]

[0020]

[0021] In the formula, Indicates the input feature map, " indicates element-wise multiplication, For the output of channel attention, For the output of spatial attention, It is the Sigmoid activation function. This is a dimensionality reduction transformation, where the dimensionality reduction coefficient is 16. The table represents a dimensionality increase transformation. For global average pooling, For global max pooling, It is a 7×7 convolution. For average pooling, This is for max pooling.

[0022] Furthermore, the C2f-2 module includes the following processes in sequence:

[0023] S1 processes the input feature map through multiple parallel branches, including: two independent 3×3 grouped convolutions, a 5×5 dilated convolution, and a depthwise convolution consisting of a 1×1 convolution and a 3×3 convolution; the output feature maps of each branch are concatenated and then reduced in dimensionality through a 1×1 convolution.

[0024] S2, enhanced by the CAM module;

[0025] S3, residual connection is performed through a series of multiple residual blocks;

[0026] S4, after upsampling, is fused across layers with the input feature map;

[0027] S5, Channel recalibration;

[0028] S6, enhanced spatial attention;

[0029] S7 is sequentially processed by 1×1 convolution, batch normalization, and SiLU activation to obtain the output feature map of the C2f-2 module.

[0030] Furthermore, the processing procedure for the residual block is as follows:

[0031]

[0032] In the formula, For the first The input amount of the layer residual block, For the first Output of the layer residual block It is the ReLU activation function. For batch normalization, It is a 3×3 convolution.

[0033] Furthermore, the channel recalibration is as follows:

[0034]

[0035] In the formula, Indicates the feature map Channel recalibration processing, It is the Sigmoid activation function. The SiLU activation function is used. This is a dimensionality reduction transformation, where the dimensionality reduction coefficient is 16. The table represents a dimensionality increase transformation. For feature splicing, For global average pooling, This is for global max pooling.

[0036] Furthermore, in the feature pyramid network:

[0037] The input features are first processed by three parallel branches, where:

[0038] Branch 1, through 1×1 convolution, obtains feature map P3 with a scale of 80×80×128;

[0039] Branch 2, through a 3×3 convolution with a stride of 2, obtains a feature map P4 with a scale of 40×40×256;

[0040] Branch 3, through the SPPF module, obtains feature map P5 with a scale of 20×20×512;

[0041] Then, based on feature maps P3, P4, and P5, three feature maps of different scales are obtained through the following processing and output to the corresponding three detection heads for target detection, wherein:

[0042] After feature map P5 is upsampled, it is fused with feature map P4. After fusion, it is upsampled again and fused with feature map P3. The result is then output to the detection head.

[0043] After feature map P3 is downsampled, it is fused with feature map P4, and the result is output to the detection head;

[0044] After feature map P4 is downsampled, it is fused with feature map P5, and the result is output to the detection head.

[0045] Furthermore, before using the crack detection model, the wind turbine blade image is preprocessed as follows:

[0046] Step 1: Calculate the gradient variance of the local window region where each pixel in the image is located:

[0047]

[0048] In the formula, and All represent pixel coordinates. For the image in pixels The gradient variance of the local window region centered on the center. The number of pixels within the window. Represented by pixels A local window centered on the center. pixels within the window Sobel gradient value, For window The gradient mean within;

[0049] Step 2, based on gradient threshold and Locate the blurred and sharp regions, where: , Less than When, then the pixel It belongs to the fuzzy region; Greater than When, then the pixel It belongs to a clearly defined area;

[0050] Step 3: Solve for the fuzzy kernel by minimizing the difference between the fuzzy region edge map and the reference edge. :

[0051]

[0052]

[0053] In the formula, For the length of motion, For the direction of motion, Represents the rect function. Indicates the variable and Fuzzy kernel of influence , This is an edge map of the blurred region. For the edge map of the clear area, " represents convolution, This represents taking the square of the L2 norm;

[0054] Step 4, based on fuzzy kernel Weighted iterative deconvolution is used to deblur the blurred regions. The iterative process is as follows:

[0055]

[0056]

[0057] In the formula, Indicates the iteration number The image after that, As the initial image, and Representing images respectively and middle pixel values, and These represent the Fourier transform and its inverse transform, respectively. for conjugate, For regularization parameters;

[0058] Weights of each pixel Adjustments will be made:

[0059]

[0060]

[0061] In the formula, As the baseline regularization parameter, The baseline gradient variance;

[0062] Furthermore, the following processing is performed after step 4:

[0063] Step 5, Guided Filter Enhancement:

[0064]

[0065] In the formula, The image after deblurring. This is the image after guided filtering enhancement. Indicates guided filtering. The filter radius is... For smoothing parameters.

[0066] Step 6, residual edge overlay:

[0067]

[0068] In the formula, This is the image after residual edge overlay. This is the enhancement coefficient.

[0069] The beneficial effects of this invention are as follows:

[0070] (1) This invention provides a method for detecting cracks in wind turbine blades, wherein the crack detection model based on YOLOv8 is used to detect cracks in wind turbine blade images, which can avoid interference from the blade's own texture and improve the accuracy of crack detection.

[0071] (2) In the detection model of the present invention, the Stem module with a multi-branch structure captures crack features at different scales. Specifically, 3×3 convolution is used to extract local detail features, 5×5 convolution is used to extract global features, dilated convolution is used to expand the receptive field to capture long cracks, max pooling is used to preserve texture information, and finally, multi-scale feature fusion is used to enhance the sensitivity to small cracks. The above improvements only increase the number of parameters by 15% in terms of computational efficiency, but can greatly make up for the shortcomings of traditional single convolution operation in extracting fine crack features and significantly improve the feature extraction capability.

[0072] (3) In the detection model C2f-1 module of the present invention, the CAM module is used to pay attention to both the channel dimension and the spatial dimension at the same time, and when the two are fused, the spatial attention is ensured to act only on the enhanced crack channel; then, crack attention is integrated on the basis of the original C2f module to enhance crack-related features and suppress irrelevant background.

[0073] (4) In the detection model C2f-2 module of the present invention, the residual connection method of the residual block can avoid the gradient vanishing problem of deep network and ensure the gradient flow and effective transmission of crack features; then, through the channel recalibration operation, the response of crack-related channels is enhanced and irrelevant channels such as blade texture are suppressed.

[0074] (5) The wind turbine blade crack detection method of the present invention uses a newly designed preprocessing algorithm to deblur the image, overcoming the motion blur problem that exists when the UAV collects wind turbine blade images; wherein the gradient variance is used as the detection basis to locate the blurred area in the image, and then the blur kernel is estimated by minimizing the difference between the blurred edge and the reference edge. Finally, the clear image is recovered by weighted iterative deconvolution based on the blur kernel, avoiding the problem of noise amplification by traditional deconvolution.

[0075] (6) In the preprocessing algorithm of the present invention, the image after deblurring is enhanced by guided filtering and residual edge superposition in turn to specifically enhance the crack edge (slender feature) and overcome the edge blurring or noise residue problems that may exist after deblurring. Attached Figure Description

[0076] Figure 1 is a diagram of the improved crack detection neural network model based on YOLOv8 in this invention.

[0077] Figure 2 is an architecture diagram of the Stem module in the crack detection model of the present invention;

[0078] Figure 3 is an architecture diagram of the C2f-1 module in the crack detection model of the present invention;

[0079] Figure 4 is an architecture diagram of the C2f-2 module in the crack detection model of the present invention;

[0080] Figure 5 is a diagram of the feature pyramid network architecture in the crack detection model of this invention. Detailed Implementation

[0081] Embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein similar or identical reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0082] Drones are used to capture aerial images of wind turbine blades for vision-based crack detection. Specifically, the images of wind turbine blades captured by the drone include:

[0083] I. Image Preprocessing

[0084] In the image, the cracked area may become blurred due to jitter, while the remaining static background area may be partially clear. Therefore, image preprocessing is required to reduce the blur.

[0085] 1. Locate the blurred areas in the image.

[0086] This embodiment specifically uses gradient variance detection to determine the blurred region, where the gradient value of the blurred region is lower and the variance is smaller, while the gradient and variance of the clear region are larger.

[0087] 1.1 Using a sliding window, calculate the gradient variance of the local window region where each pixel in the image is located:

[0088]

[0089] In the formula, and All represent pixel coordinates. For the image in pixels The gradient variance of the local window region centered on the center. This represents the number of pixels within the window (e.g., 49 pixels for a 7x7 window). Represented by pixels A local window centered on the center. pixels within the window Sobel gradient value, For window The gradient mean within.

[0090] 1.2, Set two gradient thresholds and ,and .like Less than Then mark the pixel. It belongs to the fuzzy region; if Greater than Then mark the pixel. This is a clear region. Gradient threshold in this embodiment. Take 1 / 3 of the global gradient variance of the image. Values 1.1 times.

[0091] 2. Blurred image It can be represented as a clear image. With fuzzy kernel Convolution and then superimposed noise :

[0092]

[0093] For linear motions such as translational jitter, the fuzzy kernel is a rectangular pulse kernel:

[0094]

[0095] In the formula, For the length of motion, For the direction of motion, Represents the rect function. Indicates the variable and Fuzzy kernel of influence .

[0096] Using the edge map of the clear region as a reference, the solution is obtained by minimizing the difference between the edge map of the blurred region and the reference edge. and The value of is used to estimate the fuzzy kernel. :

[0097]

[0098] In the formula, This is an edge map of the blurred region. An edge map of a clear region. This represents taking the square of the L2 norm.

[0099] 3. Fuzzy kernel based on estimation Weighted iterative deconvolution is used to recover (i.e., deblur) the blurred regions. The iterative process is as follows:

[0100]

[0101]

[0102] In the formula, Indicates the iteration number The image after that, As the initial image, and Representing images respectively and middle pixel values, and These represent the Fourier transform and its inverse transform, respectively. for The conjugate of the denominator Used for normalization to avoid overall image brightness shift. This is a regularization parameter used to suppress noise.

[0103] Weights of each pixel Adjustments will be made:

[0104]

[0105]

[0106] In the formula, The preset baseline regularization parameters; The preset baseline gradient variance is used to control the weight decay rate.

[0107] 4. After deblurring, there may be blurred edges or residual noise. Therefore, it is necessary to specifically enhance the crack edges. To address this:

[0108] 4.1 Guided Filter Enhancement: Guided filtering is used to enhance the deblurred image. (The entire image, including the clear areas that have not been deblurred), undergoes edge-preserving smoothing:

[0109]

[0110] In the formula, This is the image after guided filtering enhancement. Indicates guided filtering. The filter radius (3~5 in this embodiment) For the smoothing parameter (in this embodiment, we take...) ).

[0111] 4.2 Residual Edge Overlay: Calculate the residual (highlighting edges) between the deblurred image and the guided filtering result, and overlay them back into the enhanced image.

[0112]

[0113] In the formula, This is the image after residual edge overlay (i.e., the image output after the entire preprocessing). The enhancement factor (valued at 1.2 to 1.5 in this embodiment) is used to strengthen the crack edge.

[0114] II. Crack Detection

[0115] For the preprocessed image, the following crack detection neural network model based on YOLOv8 is used for crack detection.

[0116] 1. Overall Architecture

[0117] As shown in Figure 1, in the crack detection model, for the input image (scale 640×640×3):

[0118] S1 captures crack features at different scales through the Stem module and outputs a feature map with a scale of 320×320×64.

[0119] S2, after being processed by the C2f-1 module, yields a feature map with a scale of 320×320×128, which is then scaled to 160×160×128 by a 3×3 convolution with a stride of 2.

[0120] S3, after being processed by the C2f-2 module, yields a feature map with a scale of 160×160×256, which is then scaled to 80×80×256 by a 3×3 convolution with a stride of 2.

[0121] S4, after being processed again by the C2f-2 module, yields a feature map with a scale of 80×80×512;

[0122] S5, finally fed into the feature pyramid network, outputs the crack detection results.

[0123] 2. Stem module

[0124] As shown in Figure 2, the Stem module employs a multi-branch structure to simultaneously capture crack features at different scales, thus overcoming the limitations of traditional single convolution operations in extracting fine crack features. Specifically, it first sets up four parallel branches to process the input image (640×640×3):

[0125] Branch 1 uses a standard 3×3 convolution with a stride of 2 to extract local detail features;

[0126] Branch 2 uses a standard 5×5 convolution with a stride of 2 to extract global features;

[0127] Branch 3 uses a dilated convolution with an expansion rate of 2 to expand the receptive field and capture long cracks;

[0128] Branch 4 uses max pooling with a stride of 2 + 1×1 convolution to preserve texture information;

[0129] Then, the output feature maps of the above four branches are spliced ​​and rearranged to output a feature map with a scale of 320×320×64.

[0130] 3. C2f-1 module

[0131] 3.1, As shown in Figure 3, the C2f-1 module, for the input feature map (320×320×64):

[0132] First, the input features are processed by two parallel Bottleneck modules (originally modules of YOLOv8 itself). In the two Bottleneck modules, one shortcut (residual connection) is True (meaning that the input features are added to the output), and the other shortcut (residual connection) is False (meaning that the input features are not added to the output).

[0133] The output of each Bottleneck module is then augmented using the CAM module.

[0134] Finally, the original output features of each Bottleneck module and the features enhanced by the CAM module (all four feature maps are 320×320×64 in scale) are concatenated and then fused by 1×1 convolution to obtain the output feature map of this module (320×320×128).

[0135] The above process can be expressed as a function expression as follows:

[0136]

[0137] In the formula, This is the output feature map of the C2f-1 module. Represents a 1×1 convolution. Indicates feature splicing, This indicates that the data has been processed by the CAM module. and These represent the outputs of the two Bottleneck modules, respectively.

[0138] 3.2 The CAM module described above includes channel attention and spatial attention to simultaneously focus on the channel and spatial dimensions of the feature map, wherein:

[0139] 1) Channel attention is achieved by first processing the input feature map using Global Average Pooling (GAP) and Global Max Pooling (GMP) respectively, concatenating the results, and then sequentially performing dimensionality reduction, dimensionality increase, and Sigmoid activation.

[0140]

[0141] In the formula, ( , , (representing the image's height, width, and number of channels, respectively) represents the input feature map. For the output of channel attention, This represents the Sigmoid activation function. This indicates a dimensionality increase transformation (i.e., the feature map channel dimension is increased from 1). Become ), This indicates a dimensionality reduction transformation (i.e., the feature map channel dimension is reduced from...). Become In this embodiment, the dimensionality reduction coefficient =16), Indicates global average pooling. Indicates global max pooling;

[0142] 2) Spatial attention is achieved by first processing the input feature map using average pooling (AvgPool) and max pooling (MaxPool), concatenating the results, and then using a 7×7 convolution to capture spatial dependencies to highlight the continuous crack region. Finally, sigmoid activation is applied.

[0143]

[0144] In the formula, For the output of spatial attention, Represents a 7×7 convolution. Indicates average pooling. This indicates max pooling;

[0145] 3) The fusion relationship between channel attention and spatial attention is as follows:

[0146]

[0147] In the formula, " indicates element-wise multiplication (i.e., Hadamard product).

[0148] 4. C2f-2 module

[0149] The C2f-2 module is shown in Figure 4:

[0150] 1) First, the input feature map is processed by four parallel branches. The four branches are two independent 3×3 grouped convolutions, a 5×5 dilated convolution, and a depthwise convolution consisting of a 1×1 convolution and a 3×3 convolution. The output feature maps of the four branches are concatenated and then reduced in dimensionality by a 1×1 convolution.

[0151] 2) The dimensionality-reduced feature map is enhanced by the CAM module.

[0152] 3) Residual connection processing is performed on the enhanced feature maps, specifically using a series of concatenated residual blocks. In each residual block, the input features are first subjected to two consecutive 3×3 convolutions, batch normalization (BN), and ReLU activation. The result is then added to the input and used as the output of the residual block.

[0153]

[0154] In the formula, For the first The input of the layer residual block (i.e., the first layer residual block) The output of the layer residual block), where the initial input of the first layer is the feature enhanced by the CAM module in the previous step. For the first Output of the layer residual block Represents the ReLU activation function. Indicates batch normalization, This represents a 3×3 convolution.

[0155] 4) For the feature map after residual connection processing, after upsampling, it is fused across layers with the input feature map of the C2f-2 module through 1×1 convolution.

[0156] 5) The fused feature maps undergo channel recalibration. The process is as follows: first, channel weights are generated with reference to the channel attention in the CAM module; then, the channel weights are multiplied element-wise by the input features for recalibration.

[0157]

[0158] In the formula, Indicates the feature map (i.e., channel recalibration processing corresponding to the fused feature map) The SiLU activation function is used. and Similar to the dimensionality-up transformation and dimensionality-down transformation of channel attention in the CAM module above.

[0159] 6) The recalibrated feature map is enhanced by spatial attention, which is consistent with that in the CAM module.

[0160] 7) The feature map after spatial attention enhancement is then sequentially processed by 1×1 convolution, batch normalization, and SiLU activation to serve as the output feature map of the C2f-2 module.

[0161] 5. Feature Pyramid Network

[0162] As shown in Figure 5, the feature pyramid network processes the feature map (80×80×512) output by the second C2f-2 module through three parallel branches, where:

[0163] Branch 1 uses a 1×1 convolution to obtain feature map P3 with a scale of 80×80×128;

[0164] Branch 2 uses a 3×3 convolution with a stride of 2 to obtain a feature map P4 with a scale of 40×40×256;

[0165] Branch 3 obtains a feature map P5 with a scale of 20×20×512 through the SPPF module;

[0166] Small crack detection: After feature map P5 is upsampled, it is fused with feature map P4. After fusion, it is upsampled again and fused with feature map P3. The result is used as the final extracted large-scale feature map and output to the detection head to detect small cracks.

[0167] Medium-sized crack detection: After downsampling feature map P3, it is fused with feature map P4. The result is used as the final extracted medium-scale feature map and output to the detection head to detect medium-sized cracks.

[0168] Large crack detection: After feature map P4 is downsampled, it is fused with feature map P5. The result is used as the final extracted small-scale feature map and output to the detection head to detect large cracks.

[0169] 6. Loss Function

[0170] The above-mentioned crack detection neural network model, based on the improved YOLOv8, combines Focal-EioU loss and SloU loss with the original CLS loss during training:

[0171]

[0172] In the formula, The total loss of the crack detection neural network model is... For Focal-EioU loss, For SloU loss, For CLS loss, , , These are the weighting coefficients.

[0173] This invention is not limited to the above-described embodiments. Any obvious improvements, substitutions, or modifications that can be made by those skilled in the art without departing from the essence of this invention are within the scope of protection of this invention.

Claims

1. A method for detecting cracks in wind turbine blades, characterized in that: The following crack detection model is used to detect cracks in wind turbine blade images; the crack detection model includes, from front to back, the following modules: Stem module, C2f-1 module, 3×3 convolution, C2f-2 module, 3×3 convolution, C2f-2 module, and feature pyramid network.

2. The method for detecting cracks in wind turbine blades according to claim 1, characterized in that: In the Stem module: features are first extracted from the input image through multiple parallel branches, including standard 3×3 convolution with stride of 2, standard 5×5 convolution with stride of 2, dilated convolution with dilation of 2, and max pooling with stride of 2; then the outputs of each branch are concatenated and the features are rearranged.

3. The method for detecting cracks in wind turbine blades according to claim 1, characterized in that: In the C2f-1 module: the input features are first processed by two parallel Bottleneck modules to obtain the feature map. and In the two Bottleneck modules, one shortcut is True and the other is False; then the following processing is performed: In the formula, This is the output feature map of the C2f-1 module. For 1×1 convolution, For feature splicing, This indicates that the data has been processed by the CAM module.

4. The method for detecting cracks in wind turbine blades according to claim 3, characterized in that: The processing procedure of the CAM module is as follows: In the formula, Indicates the input feature map, " indicates element-wise multiplication, For the output of channel attention, For the output of spatial attention, It is the Sigmoid activation function. This is a dimensionality reduction transformation, where the dimensionality reduction coefficient is 16. The table represents a dimensionality increase transformation. For global average pooling, For global max pooling, It is a 7×7 convolution. For average pooling, This is for max pooling.

5. The method for detecting cracks in wind turbine blades according to claim 1, characterized in that: The C2f-2 module includes the following processing steps in sequence: S1, processing the input feature map through multiple parallel branches, including: two independent 3×3 grouped convolutions, a 5×5 dilated convolution, and a depthwise convolution consisting of a 1×1 convolution and a 3×3 convolution; the output feature maps of each branch are concatenated and then dimensionality reduced by a 1×1 convolution; S2, enhancement by the CAM module; S3, residual connection by continuous multi-layer residual blocks; S4, upsampling and cross-layer fusion with the input feature map; S5, channel recalibration; S6, spatial attention enhancement; S7, sequentially passing through a 1×1 convolution, batch normalization, and SiLU activation to obtain the output feature map of the C2f-2 module.

6. The method for detecting cracks in wind turbine blades according to claim 5, characterized in that: The processing procedure for the residual block is as follows: In the formula, For the first The input amount of the layer residual block, For the first Output of the layer residual block It is the ReLU activation function. For batch normalization, It is a 3×3 convolution.

7. The method for detecting cracks in wind turbine blades according to claim 5, characterized in that: The channel recalibration is as follows: In the formula, Indicates the feature map Channel recalibration processing, It is the Sigmoid activation function. The SiLU activation function is used. This is a dimensionality reduction transformation, where the dimensionality reduction coefficient is 16. The table represents a dimensionality increase transformation. For feature splicing, For global average pooling, This is for global max pooling.

8. The method for detecting cracks in wind turbine blades according to claim 1, characterized in that: In the feature pyramid network: the input features are first processed by three parallel branches, where: branch 1, through 1×1 convolution, obtains feature map P3 with a scale of 80×80×128; branch 2, through 3×3 convolution with a stride of 2, obtains feature map P4 with a scale of 40×40×256; branch 3, through the SPPF module, obtains feature map P5 with a scale of 20×20×512; then, based on feature maps P3, P4, and P5, three feature maps of different scales are obtained through the following processing and output to the corresponding three detection heads for target detection, where: feature map P5 is upsampled and fused with feature map P4, then upsampled again and fused with feature map P3, and the result is output to the detection head; feature map P3 is downsampled and fused with feature map P4, and the result is output to the detection head; feature map P4 is downsampled and fused with feature map P5, and the result is output to the detection head.

9. The method for detecting cracks in wind turbine blades according to claim 1, characterized in that: Before using the crack detection model, the wind turbine blade image is preprocessed as follows: Step 1, calculate the gradient variance of the local window region where each pixel in the image is located: In the formula, and All represent pixel coordinates. For the image in pixels The gradient variance of the local window region centered on the center. The number of pixels within the window. Represented by pixels A local window centered on the center. pixels within the window Sobel gradient value, For window Step 2, based on the gradient mean; Step 3, based on the gradient threshold. and Locate the blurred and sharp regions, where: ， Less than When, then the pixel It belongs to the fuzzy region; Greater than When, then the pixel It belongs to the clear region; Step 3, solve for the fuzzy kernel by minimizing the difference between the edge map of the fuzzy region and the reference edge. ： In the formula, For the length of motion, For the direction of motion, Represents the rect function. Indicates the variable and Fuzzy kernel of influence ， This is an edge map of the blurred region. For the edge map of the clear region, " represents convolution, This indicates taking the square of the L2 norm; Step 4, based on the fuzzy kernel Weighted iterative deconvolution is used to deblur the blurred regions. The iterative process is as follows: In the formula, Indicates the iteration number The image after that, As the initial image, and Representing images respectively and middle pixel values, and These represent the Fourier transform and its inverse transform, respectively. for conjugate, For regularization parameters; Weights of each pixel Adjustments will be made: In the formula, As the baseline regularization parameter, The baseline gradient variance is given.

10. The method for detecting cracks in wind turbine blades according to claim 9, characterized in that: Following step 4, the following processing is performed: Step 5, guided filter enhancement: In the formula, The image after deblurring. This is the image after guided filtering enhancement. Indicates guided filtering. The filter radius is... For smoothing parameters. Step 6, residual edge overlay: In the formula, This is the image after residual edge overlay. This is the enhancement coefficient.

Images