A method for detecting crack defects in wind turbine blades based on feature recombination networks

Abstract

This invention discloses a method for detecting crack defects in wind turbine blades based on a feature reconstruction network. The method acquires an image of the wind turbine blade to be detected, inputs it into a pre-trained feature reconstruction network model, and obtains the crack defect identification result of the wind turbine blade image. The method utilizes an optimized feature reconstruction network model to extract multi-scale features from the wind turbine blade image through the small target feature extraction capability of the backbone network layer. Then, it uses feature reconstruction and feature expansion in the neck network layer to retain more detailed crack information from the multi-scale features, fully acquiring the contextual information of small targets for feature fusion. Finally, it obtains the crack defect identification result of the wind turbine blade image through a detection head, achieving more accurate detection of small target crack defects.

Description

Technical Field

[0001] This invention relates to the field of structural surface defect target detection technology, specifically to a method for detecting crack defects in wind turbine blades based on feature recombination networks. Background Technology

[0002] In the wind power industry, wind turbine blades are one of the key components for converting wind energy into mechanical energy, and they are affected by the external environment and various human and non-human factors. During blade operation, structural defects such as cracks, peeling, and erosion are prone to occur. Current blade inspection mainly relies on human observation and subjective experience to determine the presence of cracks, resulting in poor consistency in evaluation results and difficulty in ensuring the stability of blade quality. Therefore, introducing an intelligent crack recognition model into the blade inspection process to target defects such as cracks, peeling, and erosion, timely detection and reduction of human misjudgment and omissions, and accurate assessment of blade quality are of great significance for ensuring the safe and reliable operation of wind turbines.

[0003] Deep learning-based object detection methods are mainly divided into two categories: two-stage object detection and one-stage object detection. Common two-stage algorithms include R-CNN, Fast R-CNN, and Faster R-CNN. Two-stage object detection models have certain advantages in detection accuracy, but they are more complex and have slower inference speeds, making them difficult to meet the real-time detection requirements of cracks and defects in wind power scenarios. Single-stage object detection methods mainly include YOLO and SSD. Compared to two-stage object detection methods, single-stage object detection methods are faster and more likely to meet the needs of crack and defect detection in practical engineering.

[0004] In recent years, single-stage target detection methods, represented by YOLO networks, have been widely used for the detection and identification of surface defects in structures. However, existing methods simply stitch together feature segments during the upsampling process without analyzing the importance of each segment, which can easily lead to the loss of detailed information and make it difficult to fully extract the weak features of small target cracks and defects. There is still room for improvement in target detection accuracy. Summary of the Invention

[0005] To address the shortcomings of the existing technology, this invention provides a method for detecting crack defects in wind turbine blades based on feature recombination networks, thereby improving the detection accuracy of small target crack defects on wind turbine blades.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A method for detecting crack defects in wind turbine blades based on feature recombination networks is proposed. The method involves acquiring an image of the wind turbine blade to be detected, inputting it into a pre-trained feature recombination network model, and obtaining the crack defect identification result of the wind turbine blade image to be detected.

[0008] The feature reconstruction network model includes a backbone network layer, a neck network layer, and a detection head. The backbone network layer is used to extract multi-scale features from the wind turbine blade image input to the feature reconstruction network to obtain a multi-scale feature map. The neck network layer is used to perform feature reconstruction and feature expansion processing on the multi-scale feature map, and to fuse the multi-scale features to obtain a fused feature map. The detection head is used to obtain the crack defect identification result of the wind turbine blade image to be detected based on the fused feature map, which is used as the output of the feature reconstruction network model.

[0009] Specifically, in the feature recombination network model:

[0010] The backbone network layer includes sequentially connected convolutional layers, four convolutional fusion units, one large kernel selective network, and one spatial pyramid pooling module; each convolutional fusion unit includes cascaded convolutional layers and a C2f-FE module;

[0011] The neck network layer includes two feature reconstruction upsampling units and two convolutional connection fusion units connected in sequence; wherein, the feature reconstruction upsampling unit includes a CSEAFE upsampling module, a SimAM attention module, a connection layer and a C2f-FE module connected in sequence; the convolutional connection fusion unit includes a GSConv hybrid convolution module, a connection layer and a C2f-FE module connected in sequence.

[0012] The detection head is used to obtain the crack defect identification result of the wind turbine blade image based on the output of the neck network layer, and as the output of the feature reconstruction network model;

[0013] In this system, the input of the feature reconstruction network model serves as the input to the backbone network layer. Within the backbone network layer, the outputs of the second and third convolutional fusion units also serve as the inputs to the connection layers of the second and first feature reconstruction upsampling units in the neck network layer, respectively. In addition to serving as the input to the neck network layer, the output of the spatial pyramid pooling module in the backbone network layer also serves as the input to the connection layer of the second convolutional connection fusion unit in the neck network layer. Furthermore, the output of the first feature reconstruction upsampling unit in the neck network layer also serves as the input to the connection layer of the first convolutional connection fusion unit. The outputs of the second feature reconstruction upsampling unit and the two convolutional connection fusion units are all output to the detection head.

[0014] Specifically, the C2f-FE module includes a cascaded first convolutional layer, a split module, several fast convolutional attention modules, a connection layer, and a second convolutional layer. The two split outputs of the split module serve as the inputs of the first fast convolutional module and the connection layer, respectively. The outputs of each fast convolutional module also serve as the inputs of the connection layer. The output of the second convolutional layer serves as the output of the C2f-FE module.

[0015] Specifically, the fast convolutional attention module includes a cascaded 3×3 partial convolutional layer, two 1×1 convolutional layers, and an EMA efficient multi-scale attention module. The output of the EMA efficient multi-scale attention module is superimposed with the input of the fast convolutional module to serve as the output of the fast convolutional module.

[0016] Specifically, the large kernel selectivity network includes multiple cascaded large kernel selectivity units; each large kernel selectivity unit includes a first normalization layer, a large kernel selectivity sub-block, a second normalization layer, and a feedforward network sub-block connected in sequence, and the output of the large kernel selectivity sub-block is superimposed with the input of the first normalization layer and then used as the input of the second normalization layer, and the output of the feedforward network sub-block is superimposed with the input of the second normalization layer and then used as the output of the large kernel selectivity unit;

[0017] The large kernel selection sub-block consists of a cascaded fully connected layer, a GELU activation function layer, an LSK large kernel selection module, and another fully connected layer; the feedforward network sub-block consists of a cascaded fully connected layer, a deep convolutional layer, a GELU activation function layer, and another fully connected layer.

[0018] Specifically, the CSEAFE upsampling module is used to process input features of dimension C×H×W. Figure X Generate an upsampling map of dimension C×σH×σW. Figure X As the output, σ represents the upsampling rate, and the upsampling mapping. Figure X Each target position l′=(i′,j′) in the input features Figure X Each of them has a corresponding source position l = (i, j), and The processing steps of the CSEAFE upsampling module include:

[0019] First, based on input features Figure X k centered at different source locations l = (i, j) up ×k up Subregion N(X) l k up The upsampling map is generated by the kernel prediction module ψ. Figure X The position sensing kernel W corresponding to different target positions l′=(i′,j′) in the matrix. l:

[0020] W l′ =ψ(N(X) l k up ));

[0021] Where ψ(·) represents the kernel prediction module operation, X l Representing input features Figure X The position, the value at that position, and the subregion N(X) l k up That is, X l k up ×k up Neighborhood;

[0022] Then, the upsampling map is generated through the content-aware recombination module φ. Figure X The values ​​of X′ at different target positions l′=(i′,j′) in the matrix. j :

[0023]

[0024] Where (n, m) represents the subregion N(X) l k up The position coordinates in ) Represents subregion N(X) l k up The radius of the region; W l′＝(n，m) This represents the position-aware kernel corresponding to the target position l′=(n,m) in the upsampling mapping X′; X l＝(i+n，j+m) Input features Figure X The value at position l = (i+n, j+m);

[0025] Thus, the upsampling map as the output is obtained. Figure X ′.

[0026] Specifically, the processing procedure of the SimAM attention module includes:

[0027] For input features of dimension C×H×W Figure X For the value x at each position in each channel of the graph c，i Calculate the corresponding energy functions respectively This yields the input features. Figure X The corresponding energy function characteristic map E of C×H×W,

[0028] in, Representing input features Figure X The value x at position i in any c-th channel c,iThe corresponding energy function is c = 1, 2, ..., C, i = 1, 2, ..., M, M = W × H; λ is the regularization term; Input features Figure X The variance of the values ​​at each position in each channel c. Input features Figure X The mean of the values ​​at each position on each of the c-th channels:

[0029]

[0030] Then, the output features of the SimAM attention module are calculated using the following formula. Figure X SimAM :

[0031]

[0032] Where Sigmoid(·) represents the Sigmoid activation function operation; Θ represents the element-wise multiplication operation.

[0033] Specifically, the processing procedure of the GSConv hybrid convolution module includes:

[0034] First, the input feature map is convolutionally downsampled. Then, the output of the convolutionally downsampled map is further processed by depthwise convolution. The output of the depthwise convolution is then concatenated with the output of the convolutionally downsampled map. Finally, the concatenated result is shuffled to obtain the output feature map.

[0035] Specifically, the feature recombination network model is trained in the following manner:

[0036] S101: Prepare a sample image dataset of wind turbine blade crack defects, which is divided into a training image dataset and a test image dataset. The wind turbine blade crack defect sample images in the wind turbine blade crack defect sample image dataset are all pre-marked with the location and type of crack defects.

[0037] S102: Input the training image dataset into the feature reconstruction network model for training, and optimize the parameters of the feature reconstruction network model with the goal of minimizing the loss function until the feature reconstruction network model converges, thus obtaining the trained feature reconstruction network model.

[0038] S103: Test the feature reconstruction network model using the test image dataset to confirm the recognition performance of the trained feature reconstruction network model; if the recognition performance meets the requirements, end the training of the feature reconstruction network model; otherwise, return to step S102.

[0039] Specifically, in the loss function, the classification loss is the total loss function formed by the sum of the CIoU loss function and the DFL loss function, and the localization loss is the binary cross-entropy (BCE) loss function.

[0040] Compared with the prior art, the present invention has the following technical effects:

[0041] 1. The present invention provides a method for detecting crack defects in wind turbine blades based on feature recombination networks. This method utilizes an optimized feature recombination network model to extract multi-scale features from wind turbine blade images through the small target feature extraction capability of the backbone network layer. Then, by leveraging feature recombination and feature expansion in the neck network layer, more detailed crack information from the multi-scale features is retained. This fully acquires the contextual information of the small target for feature fusion, and then obtains the crack defect identification result of the wind turbine blade image through the detection head, achieving more accurate detection of small target crack defects.

[0042] 2. In the feature reorganization network model of the present invention, the C2f-Faster module is combined with the EMA efficient multi-scale attention mechanism to design the C2f-FE module, which is introduced into both the backbone network layer and the neck network layer. Furthermore, the large kernel selective network LSKNet is introduced into the backbone network layer to dynamically adjust the receptive field of the backbone network layer for feature extraction, thereby improving the model's feature extraction capability.

[0043] 3. In the feature recombination network model of this invention, a lightweight CARAFE upsampling module is designed in the neck network layer. Through feature recombination and feature expansion, more crack detail information is retained, improving the upsampling effect. The SimAM attention mechanism is used to enhance the attention to the contextual information of crack defect features. In addition, the GSConv hybrid convolution module is used to reduce the size of the model. While being lightweight, it also helps to reduce the false negative rate and false positive rate of small target crack defects. Finally, more accurate detection of small target crack defects in wind turbine blades is achieved. Attached Figure Description

[0044] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0045] Figure 1 This is a schematic diagram of the architecture of the feature recombination network model used in the method of the present invention;

[0046] Figure 2 This is a schematic diagram of the C2f-FE module.

[0047] Figure 3 This is a schematic diagram of the fast convolutional attention module FEblock in the C2f-FE module;

[0048] Figure 4 A schematic diagram of the structure of the EMA's efficient multiscale attention module;

[0049] Figure 5 The processing flowchart for the EMA's efficient multi-scale attention module;

[0050] Figure 6 This is a schematic diagram of the structure of the large kernel selective unit LSK block in the large kernel selective network LSKNet.

[0051] Figure 7 Flowchart of LKSelection for selecting sub-blocks for the large kernel;

[0052] Figure 8 This is a flowchart of the CSEAFE upsampling module.

[0053] Figure 9 This is a flowchart of the SimAM attention module's processing flow.

[0054] Figure 10 This is a flowchart of the GSConv hybrid convolution module's processing flow.

[0055] Figure 11 This is a comparison chart of the heatmap detection effects of different models in the embodiments;

[0056] Figure 12 This is a comparison chart of the detection performance of different models in the embodiments. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0058] The present invention will now be described in further detail with reference to the accompanying drawings.

[0059] This invention discloses a method for detecting crack defects in wind turbine blades based on a feature reconstruction network. This method acquires an image of the wind turbine blade to be detected, inputs it into a pre-trained feature reconstruction network model, and obtains the crack defect identification result of the wind turbine blade image. The feature reconstruction network model used in this invention is obtained by optimizing and improving the YOLOv8 network as the basic model framework. Figure 1 As shown, the system includes a backbone network layer, a neck network layer, and a detection head. The backbone network layer is used to extract multi-scale features from the wind turbine blade image input to the feature reconstruction network, resulting in a multi-scale feature map. The neck network layer is used to perform feature reconstruction and feature expansion processing on the multi-scale feature map, and to fuse the multi-scale features to obtain a fused feature map. The detection head is used to obtain the crack defect identification result of the wind turbine blade image to be detected based on the fused feature map, which serves as the output of the feature reconstruction network model.

[0060] This invention presents a method for detecting crack defects in wind turbine blades based on feature recombination networks. It utilizes an optimized feature recombination network model to extract multi-scale features from wind turbine blade images through the small target feature extraction capability of the backbone network layer. Then, by leveraging feature recombination and feature expansion in the neck network layer, more detailed crack information from the multi-scale features is retained. This allows for the full acquisition of contextual information of small targets for feature fusion, and finally, the detection head obtains the crack defect identification results of the wind turbine blade image, achieving more accurate detection of small target crack defects.

[0061] The following section provides a more detailed description of the wind turbine blade crack defect detection method based on feature recombination network and the feature recombination network model used in this invention.

[0062] 1 Feature Recombination Network Model

[0063] To address the issue of low accuracy in detecting small-target cracks during wind turbine blade operation, this invention employs a pre-trained feature reconstruction network model for crack defect detection in wind turbine blade images. The feature reconstruction network model used in this invention is obtained by optimizing and improving the YOLOv8 network as its basic model framework. Figure 1 As shown, it includes a backbone network layer, a neck network layer, and a detection head.

[0064] The backbone network layer consists of sequentially connected convolutional layers, four convolutional fusion units, one large kernel selective network, and one spatial pyramid pooling module; each convolutional fusion unit includes cascaded convolutional layers and a C2f-FE module.

[0065] In the backbone network layer, the wind turbine blade image input to the feature recombination network model is first initially convolved through the convolutional layers in the backbone network layer, and then passed through four convolutional fusion units containing C2f-FE modules for multi-level feature extraction. After that, it is passed through the large kernel selective network LSKNet to adjust the receptive field of the backbone network layer for feature extraction, so as to adapt to crack defect features of different scales and improve the model's feature extraction capability. Finally, it is passed through the spatial pyramid pooling module to generate multi-scale feature maps of a specified size.

[0066] The neck network layer comprises two feature reconstruction upsampling units and two convolutional connection fusion units connected in sequence. The feature reconstruction upsampling unit comprises a CSEAFE upsampling module, a SimAM attention module, a connection layer, and a C2f-FE module connected in sequence. The convolutional connection fusion unit comprises a GSConv hybrid convolution module, a connection layer, and a C2f-FE module connected in sequence.

[0067] The multi-scale feature map enters the neck network layer, first passing through two feature reconstruction and upsampling units. In these units, the CSEAFE upsampling module performs upsampling, preserving more crack detail information through feature reconstruction and expansion, thus improving the upsampling effect. Furthermore, a SimAM attention module is introduced within this unit to enhance attention to the contextual information of crack defect features. After processing by the connection layer and C2f-FE module, the output is sent to two convolutional connection fusion units. In these units, the convolution operation is optimized using the GSConv hybrid convolution module, a lightweight module that helps reduce model size and, while also reducing false negatives and false negatives for small target crack defects. The output then passes through the connection layer and C2f-FE module before being processed again. Finally, after passing through the two convolutional connection fusion units, the fused feature map is obtained as the output of the neck network layer.

[0068] The detection head is used to obtain the crack defect identification results of the wind turbine blade image based on the output of the neck network layer, and serves as the output of the feature reconstruction network model. The final crack defect identification results output by the detection head can present information such as the category, location, and confidence level of the crack defect targets in the wind turbine blade image. In specific implementations, the detection head can adopt the mainstream decoupled head structure to reduce the number of parameters and computational complexity, while simultaneously performing the classification task and training regression task for crack defect identification.

[0069] Among them, such as Figure 1As shown, the input of the feature reconstruction network model is used as the input of the backbone network layer. In the backbone network layer, the outputs of the second and third convolutional fusion units are also used as the inputs of the connection layers in the second and first feature reconstruction upsampling units of the neck network layer, respectively. In addition to being used as the input of the spatial pyramid pooling module in the backbone network layer, the output of the spatial pyramid pooling module in the backbone network layer is also used as the input of the connection layer in the second convolutional connection fusion unit of the neck network layer. In the neck network layer, the output of the first feature reconstruction upsampling unit is also used as the input of the connection layer in the first convolutional connection fusion unit. The outputs of the second feature reconstruction upsampling unit and the two convolutional connection fusion units are all output to the detection head.

[0070] The feature reconstruction network model used in this invention, based on the YOLOv8 network, has the following main optimizations and improvements: First, a C2f-FE module is designed by combining the C2f-Faster module with the EMA efficient multi-scale attention mechanism and simultaneously introducing it into both the backbone and neck layers. Furthermore, a large-kernel selective network LSKNet is introduced into the backbone layer to dynamically adjust the receptive field for feature extraction, thereby enhancing the model's feature extraction capabilities. In the neck layer, a lightweight CARAFE upsampling module is designed to retain more crack detail information through feature reconstruction and expansion, improving the upsampling effect. The SimAM attention mechanism is used to enhance attention to the contextual information of crack defect features. Additionally, a GSConv hybrid convolution module is employed to reduce the model's size, which, while lightweight, also helps reduce the false negative and false positive rates for small target crack defects. Ultimately, this achieves more accurate detection of small target crack defects in wind turbine blades.

[0071] Next, we will introduce in detail the main optimized and improved module units in the feature recombination network model.

[0072] 1.1C2f-FE module

[0073] To improve the detection performance of small target cracks and achieve real-time intelligent crack detection, the EMA attention mechanism is combined with the Fasterblock structure and added to the C2f magic block to construct the C2f-FE module.

[0074] The structure of the C2f-FE module is as follows: Figure 2 As shown, it includes a cascaded first convolutional layer, a split module, several fast convolutional attention modules, a connection layer, and a second convolutional layer. The two split outputs of the split module serve as the inputs to the first fast convolutional module and the connection layer, respectively. The outputs of each fast convolutional module also serve as the inputs to the connection layer. The output of the second convolutional layer serves as the output of the C2f-FE module.

[0075] Therefore, the C2f-FE module was designed. Through efficient convolution operations and feature fusion mechanisms, it enhances the model's multi-scale feature extraction capabilities, significantly improves the accuracy and speed of small target detection, and reduces computational costs, enabling faster and more accurate detection of small target cracks and defects.

[0076] The structure of the fast convolutional attention module FEblock in the C2f-FE module is as follows: Figure 3 As shown, it includes a cascaded 3×3 partial convolutional layer, two 1×1 convolutional layers, and an EMA efficient multi-scale attention module. The output of the EMA efficient multi-scale attention module is superimposed with the input of the fast convolutional module and used as the output of the fast convolutional module.

[0077] The FasterBlock (FEblock) fast convolutional attention module consists of a 3×3 partial convolutional layer and two 1×1 convolutional layers in its first half, forming a FasterBlock structure. The PConv operation of the partial convolutional layers reduces redundant information and memory usage in computation, while improving the network's learning ability. Then, two 1×1 convolutions are performed on the feature map to extract effective information. In this way, while reducing the number of model parameters, the information from all channels can be fully and effectively utilized, and the diversity of features extracted by the model can be maintained.

[0078] Because some cracks occupy a small proportion of the image and have low target pixel counts, the crack quality detection effect is not ideal. Therefore, by introducing the Efficient Multi-scale Attention (EMA) module, which utilizes cross-space learning efficiency multi-scale attention mechanism, into the C2f-FE module, we can better focus on the salient features of the target object and improve the model's feature extraction capability.

[0079] The EMA (Efficient Multi-Scale Attention Module) is a commonly used multi-scale feature extraction unit, capable of handling any given input features. Figure X ∈R C×H×W X is divided into g sub-features across the channel dimension for learning different semantic features. Its module structure is as follows: Figure 4 As shown, the EMA efficient multi-scale attention module, while preserving information on the channels and reducing computational overhead, further aggregates the output features of two parallel branches through batch reconstruction of some channels in a cross-dimensional interactive manner to capture pixel-level pairwise relationships, enabling the model to achieve higher accuracy with fewer parameters.

[0080] The processing flow of the EMA efficient multi-scale attention module for input feature maps is as follows: Figure 5As shown. The size of the input feature map is defined as C / / G×H×W, where two parallel paths are 1×1 branches performing global average pooling operations in the one-dimensional horizontal and vertical dimensions, and a third parallel path is a 3×3 branch for feature extraction. Based on this, the output features of the two branches are adjusted using the Sigmoid function and normalization operations to achieve different cross-channel interactive features between the two parallel paths, thereby obtaining pixel-level pairwise relationships. EMA encodes the global spatial information of the 1×1 branch output and the 3×3 branch through two-dimensional global average pooling. To improve computational efficiency, a nonlinear function Softmax is used to fit a linear transformation; then, matrix multiplication is used to obtain spatial attention maps at different scales of spatial information, as well as attention maps with accurate spatial location information.

[0081] 1.2 LSKNet Module

[0082] To enhance the multi-scale information and detail preservation capabilities of feature maps, a large kernel selective network, LSKNet (Localization-Sensitive Keypoints Network), is introduced into the backbone network layer. This network optimizes the feature maps during the feature extraction stage, enhancing both global and local information. This allows the feature pyramid to more effectively fuse these features, improving the detection performance of small targets. Since the detection of cracks in small targets is easily influenced by background information and lacks sufficient feature information, they are prone to being overlooked or misclassified. Therefore, the designed LSKNet dynamically adjusts the receptive field in the feature extraction module, selecting appropriate convolutional kernel size and dilation rate when detecting small targets, thereby better capturing the feature information of crack defects in small targets.

[0083] The core of LSKNet lies in combining large convolutional kernels with a selection mechanism. Different object categories require different background information, necessitating adaptive selection by the model. Therefore, a network with a larger receptive field is constructed by decoupling a series of continuously expanding depth-wise convolutions with large kernels. Furthermore, a spatial selection mechanism is employed to spatially select feature maps from large convolutional kernels of different scales, allowing the model to focus more on the key background information of the target in space.

[0084] The large kernel selective network LSKNet consists of multiple cascaded large kernel selective units (LSK blocks). The specific number of large kernel selective units can be selected and set as needed, usually 2 to 4.

[0085] Each Large Kernel Selection (LSK) block primarily consists of two residual sub-blocks: the Large Kernel Selection (LKSelection) sub-block and the Feed-forward Network (FFN) sub-block. Specifically, its structure is as follows: Figure 6 As shown, the network comprises a first normalization layer, a large kernel selection sub-block, a second normalization layer, and a feedforward network sub-block connected in sequence. The output of the large kernel selection sub-block is superimposed with the input of the first normalization layer and then used as the input of the second normalization layer. The output of the feedforward network sub-block is superimposed with the input of the second normalization layer and then used as the output of the large kernel selection unit. The large kernel selection sub-block includes a cascaded fully connected layer, a GELU activation function layer, an LSK large kernel selection module, and another fully connected layer. The feedforward network sub-block includes a cascaded fully connected layer, a deep convolutional layer, a GELU activation function layer, and another fully connected layer. The large kernel selection sub-block (LKSelection) is used to dynamically adjust the receptive field of the network as needed, and the feedforward network sub-block (FFN) is used for channel blending and feature refinement.

[0086] The processing procedure of the large kernel selector block LKSelection is as follows: Figure 7 As shown, in the large kernel selection sub-block LKSelection, the input feature map first undergoes a large kernel decomposition step, using different large convolutional kernels to capture spatial information at different scales. The outputs of two different convolutional kernels are concatenated together via channels to utilize different spatial features simultaneously in subsequent steps. The concatenated feature map then undergoes a combination of average pooling and max pooling to further enhance key regions. Next, a new feature map is generated through self-attention (SA) and convolution operations, and then a spatial selection mechanism further enhances the focus on target regions. Finally, a mask is generated using the sigmoid function, and this mask is then element-wise multiplied with the feature map to obtain the final output feature map. This step is used to weight more important regions in the feature map to enhance the network's ability to detect specific objects.

[0087] The entire large kernel selective network LSKNet enhances the ability to extract local features of the target through a local perception mechanism, which helps to capture detailed information of small targets.

[0088] 1.3 CARAFE Module

[0089] To preserve more detailed information during upsampling, a lightweight CARAFE upsampling module was designed. This module recombines features in a content-aware manner, effectively preserving detailed information and resulting in richer and more accurate feature representations of small targets, thus significantly improving the detection accuracy of crack defects in small targets. Existing upsampling methods cannot effectively preserve and reassemble detailed information, leading to insufficient feature representation of small targets. Furthermore, detailed information is easily lost during upsampling, affecting the detection performance of small targets. Therefore, the designed lightweight content-aware CARAFE upsampling module can generate adaptive recombination weights based on the content of the input features, enhancing the model's ability to capture details of small targets.

[0090] The CSEAFE upsampling module mainly consists of two key modules: a kernel prediction module and a content-aware reconstruction module. The kernel prediction module is responsible for generating the upsampling kernel for feature reconstruction, predicting the attention weights at each upsampling location based on the mapping between the downsampling feature map and the upsampling location, thus preserving spatial details and contextual information during feature reconstruction. The content-aware reconstruction module aims to retain as much spatial information as possible during upsampling, better maintaining the accuracy of object boundaries.

[0091] In the CARAFE processing, if an input feature with dimensions C×H×W is given... Figure X Given an upsampling rate σ, CARAFE maps upsampling data of dimension C×σH×σW. Figure X As output, upsampling mapping Figure X Each target position l′=(i′,j′) in the input features Figure X Each of them has a corresponding source position l = (i,j), and

[0092] Specifically, such as Figure 8 As shown, the processing procedure of the CSEAFE upsampling module includes:

[0093] N(X) l ,k up ) represents the input features Figure X k is centered at the source position l=(i,j). up ×k up Subregion, i.e. X l k up ×k up Neighborhood; therefore, the CSEAFE upsampling module first determines the neighborhood based on the input features. Figure X k centered at different source locations l=(i,j) up ×k up Subregion N(X) l ,k upThe upsampling map is generated by the kernel prediction module ψ. Figure X The position sensing kernel W corresponding to different target positions l′=(i′,j') l' The calculation method is as follows:

[0094] W l' =ψ(N(X) l ,k up ));

[0095] Where ψ(·) represents the kernel prediction module operation, X l Representing input features Figure X The value at position l, subregion N(X) l ,k up That is, X l k up ×k up Neighborhood.

[0096] The kernel prediction module is responsible for generating location-aware kernels in a content-aware manner. This is done in conjunction with the input features. Figure X Each source location on corresponds to an upsampling map. Figure X 'on σ 2 There are several target locations. Each target location requires a k... up ×k up The position-aware kernel is used as the reassembly weight. Therefore, this module will output a size of C. up ×H×W position sensing kernel, where C up =σ 2 k up 2 .

[0097] The kernel prediction module consists of three sub-modules: a channel compressor, a content encoder, and a kernel normalizer, such as... Figure 8 As shown, the channel compressor reduces the number of channels in the input feature map. Then, the content encoder takes the compressed feature map as input and encodes the content to generate a position-aware kernel. Finally, the kernel normalizer applies the Softmax function to each position-aware kernel.

[0098] In the channel compressor, a 1×1 convolutional layer is used to compress the input feature channels from C to Cn. m Reducing the number of channels in the input feature map can reduce the number of parameters and computational cost in subsequent steps, making CARAFE more efficient. A larger kernel size can also be used for the content encoder within the same budget. Experimental results show that reducing the number of feature channels within an acceptable range does not harm performance.

[0099] In the content encoder, a size of k is used. encoderThe convolutional layers generate a position-aware kernel based on the input features. The encoder has parameters k. encoder ×k encoder ×C m ×C up Intuitively, increasing k encoder This allows for the expansion of the encoder's receptive field and the utilization of contextual information over a larger area, which is crucial for predicting position-aware kernels. However, computational complexity increases quadratically with kernel size, while the benefits from larger kernel sizes are negligible. Therefore, k is typically chosen as the optimal value. encoder =k up -k, where k is typically an integer between 0 and 2, achieving a good balance between performance and efficiency; when k = 0, i.e., k encoder =k up .

[0100] In the kernel normalizer, it is used before performing recombination weighting operations on the input feature map, for each k... up ×k up The position-aware kernels are spatially normalized using the Softmax function. This normalization step forces the kernel values ​​to sum to 1, representing a soft selection within local regions. Due to the kernel normalizer, CARAFE does not perform any rescaling that alters the mean of the feature maps.

[0101] Then, the upsampling map is generated through the content-aware recombination module φ. Figure X The values ​​X at different target positions l' = (i', j') l The calculation method is as follows:

[0102]

[0103] Where (n,m) represents the subregion N(X) l ,k up The position coordinates in ) Represents subregion N(X) l ,k up The radius of the region; W l′＝(n,m) This represents the position-aware kernel corresponding to the target position l' = (n, m) in the upsampling mapping X'; X l＝(i+n,j+m) Input features Figure X The value at position l = (i+n, j+m);

[0104] In this process, each location sensing kernel W is used. l' The content-aware recombining module reassembles features within a local region. The content-aware recombining module can be viewed as a weighted summation operator. For the target position l' = (n, m) and k centered at the source position l = (i, j), up ×kup Subregion N(X) l ,k up The upsampling mapping is calculated using the above formula. Figure X The values ​​X′ at different target positions l' = (i', j') l′ Such that the subregion N(X) l ,k up Each pixel in the dataset will be upsampled to the value X′ at the target location based on the content of the features rather than the distance to the location. l′ They produce different contributions. Since more information about relevant points within a local region can be focused on, the semantics of the recombined feature map may be stronger than that of the original feature map.

[0105] Therefore, the value X′ at each upsampling target location is determined. l′ Then, the upsampling map is obtained as the output. Figure X '.

[0106] Compared to conventional upsampling convolution methods, the CSEAFE upsampling module can perform upsampling operations on each feature point in the image through adaptively generated position-aware kernels. This allows these feature points to incorporate semantic information from the context, effectively reducing the loss of image features during the upsampling process and enhancing the network's feature extraction capabilities.

[0107] 1.4 SimAM Attention Module

[0108] To further enhance the representation of detailed features and contextual information, thereby improving the model's detection performance, a SimAM attention module is introduced into the model. The SimAM attention mechanism is a three-dimensional attention mechanism that combines spatial and channel information, enabling it to better focus on the key information of the input features, thus improving the network's feature representation capability. For input features... Figure X ∈R C×H×W SimAM attention module defines an energy function. Evaluate the value x at each position in each channel. c,i The importance of this, and its specific processing procedure are as follows:

[0109] For input features of dimension C×H×W Figure X For the value x at each position in each channel of the graph c,i Calculate the corresponding energy functions respectively This yields the input features. Figure X The corresponding energy function characteristic map E of C×H×W,

[0110] in, Representing input features Figure X The value x at position i in any c-th channel c,i The corresponding energy function is c = 1, 2, ..., C, i = 1, 2, ..., M, M = W × H; λ is the regularization term; Input features Figure X The variance of the values ​​at each position in each channel c. Input features Figure X The mean of the values ​​at each position on each of the c-th channels:

[0111]

[0112] When the energy function The lower the value of x, the better. c,i The greater the difference between the value and its surrounding values, the higher its importance. Therefore, the value x c,i The importance can be determined by To express.

[0113] Then, the output features of the SimAM attention module are calculated using the following formula. Figure X SimAM :

[0114]

[0115] Where Sigmoid(·) represents the Sigmoid activation function operation; Θ represents the element-wise multiplication operation.

[0116] The SimAM attention mechanism in the SimAM attention module can be represented as follows: Figure 9 As shown in the diagram. C represents the number of channels in the input feature map, H represents the height, and W represents the width. The SimAM attention mechanism generates three-dimensional attention weights through neuronal perception, enabling more accurate extraction of features from small targets.

[0117] 1.5GSConv Hybrid Convolutional Module

[0118] To reduce false positives and false negatives in small-target crack defects, a GSConv hybrid convolution module was designed to reduce the time required for standard convolution to extract image features. While maintaining feature extraction capabilities, it significantly reduces the number of model parameters and computational cost. The GSConv hybrid convolution module is a novel lightweight convolutional module composed of standard convolution and depthwise separable convolution, and its structure is as follows: Figure 10 As shown, the processing procedure of the GSConv hybrid convolution module is as follows: first, the input feature map is convolutionally downsampled; then, the output of the convolutionally downsampled map is further processed by depthwise convolution; then, the output of the depthwise convolution is concatenated with the output of the convolutionally downsampled map; finally, the concatenated result is shuffled to obtain the output feature map.

[0119] While depthwise separable convolutions can effectively reduce the number of network parameters, they also lose inter-channel feature information. GSConv, however, maximizes the preservation of connectivity between feature maps while reducing the number of parameters, enhancing the extraction capability of small target information. By reducing the network weight, it effectively lowers the false negative and false positive rates for small targets, thereby further improving the overall detection performance of the network. Therefore, the model lightweighting through GSConv's hybrid convolution module can more efficiently complete target detection tasks while maintaining model accuracy.

[0120] 2. Training of the Feature Recombination Network Model

[0121] The feature recombination network model is trained in the following manner:

[0122] S101: Prepare a sample image dataset of wind turbine blade crack defects, which is divided into a training image dataset and a test image dataset. The wind turbine blade crack defect sample images in the wind turbine blade crack defect sample image dataset are all pre-marked with the location and type of crack defects.

[0123] S102: Input the training image dataset into the feature reconstruction network model for training, and optimize the parameters of the feature reconstruction network model with the goal of minimizing the loss function until the feature reconstruction network model converges, thus obtaining the trained feature reconstruction network model.

[0124] S103: Test the feature reconstruction network model using the test image dataset to confirm the recognition performance of the trained feature reconstruction network model; if the recognition performance meets the requirements, end the training of the feature reconstruction network model; otherwise, return to step S102.

[0125] In the loss functions used during training, the classification loss is the sum of the CIoU loss function and the DFL loss function, while the localization loss is the binary cross-entropy (BCE) loss function. This is similar to the training loss function in the YOLOv8 network, which helps to balance the convergence and regression speed and training accuracy of the model.

[0126] 3 Examples

[0127] In this embodiment, a dataset is used to test and verify the method of the present invention and other methods to better demonstrate the technical advantages of the method of the present invention. This will be described in detail below.

[0128] 3.1 Dataset and Experimental Parameters

[0129] This experiment used 192 images of wind turbine blade cracks collected on-site as the dataset, covering different angles, and divided into training, validation, and test sets. The Labelimg library was used to annotate the dataset with bounding boxes, containing one category: crack.

[0130] The experimental computer ran a 64-bit Windows 10 Professional operating system, with an Intel(R) Core(TM) i7-12700 CPU, an NVIDIA GeForce RTX4060 GPU, and 8GB of RAM. The integrated compiler was PyTorch 2023.1.3, CUDA version 12.1, the deep learning framework PyTorch 2.0, and Python 3.11. During training, the number of epochs was 300, the number of workers was 4, the batch size was 8, and the initial learning rate (lr0) was 0.01.

[0131] 3.2 Comparison of Experimental Results

[0132] 3.2.1 Ablation Experiment

[0133] To verify the effectiveness of the proposed feature reorganization network model in crack detection tasks in real-world scenarios, an ablation experiment was designed based on the improved method. The C2f-FE module, LSKNet module, lightweight CARAFE upsampling module, SimAM attention mechanism, and hybrid convolution GSConv were designed on the basis of YOLOv8. Each module was accumulated sequentially, and the experimental results are shown in Table 1.

[0134] Table 1 Ablation Experiment Results

[0135]

[0136] As shown in Table 1, under the same experimental conditions, the FRNet model in the feature recombination network model of this invention has improved the precision, recall and mAP to varying degrees, and the number of model parameters and computational cost have decreased significantly, proving that the feature recombination network model is effective. Experiment 1 introduced the C2f-FE module, which reduced the overall number of parameters and computational cost of the model. Experiment 2, based on Experiment 1, introduced GSConv at the Neck end to further reduce the number of parameters and improve the inference speed. Experiment 3, based on Experiment 2, introduced LSKNet to enhance the model's feature extraction capability and improve its detection accuracy. Experiment 4, based on Experiment 3, designed a lightweight CARAFE upsampling module within the Neck module. Although the model's precision and recall increased compared to Experiment 3, the number of parameters and computational cost increased slightly compared to Experiment 3, with the number of parameters increasing by 8.7% and the computational cost increasing by 7.7%. The last group of experiments, based on the proposed FRNet model, added the SimAM attention mechanism to Experiment 4 to enhance the contextual information of small targets, resulting in a 0.9% increase in the model's crack detection accuracy compared to Experiment 4. Compared to the YOLOv8 model, the feature recombination network model improves mAP@0.5 by 2.9%, and reduces the number of parameters and computation by 16.7% and 13.6%, respectively, further demonstrating the effectiveness of the proposed method.

[0137] Experiment 1 demonstrates how the C2f-FE module, through feature extraction, better captures important image information while optimizing convolution operations, reducing unnecessary computation and information loss, thus improving the overall model performance. Experiment 2 shows that the model effectively reduces network parameters and computational cost through Shuffle operations and depthwise separable convolutions, accelerating inference. Experiment 3 illustrates that large kernel convolutions can capture a larger receptive field, enabling the model to better capture small-scale features of small crack defects. Experiment 4 shows that the slight increase in computational cost and parameters indicates that feature recombination and feature expansion slightly increase model complexity. The final set of experiments demonstrates that the SimAM attention mechanism improves the expressive power of feature maps, enabling the model to better focus on crack defect regions and improve detection accuracy. Therefore, the feature recombination network model, through modules such as C2f-FE, LSKNet, and CARAFE, significantly improves the model's feature extraction capabilities, dynamically adjusts the receptive field, and preserves detailed information of small target cracks, resulting in significant improvements in detection accuracy and lightweight design. It performs particularly well in the detection of small target crack defects, meeting the needs of practical engineering applications.

[0138] 3.2.2 Comparative Experiment

[0139] To further verify the detection performance of the feature recombination network model of this invention, it was compared with some target detection algorithms, including YOLOv3, YOLOv5, and RT-DETR-1, and also compared with some YOLOv8 algorithms based on lightweight improvements. The results of the comparison experiments are shown in Table 2.

[0140] Table 2. Experimental Results Comparing Model Algorithms

[0141]

[0142] As shown in Table 2, compared with YOLOv3-tiny, YOLOv5n, and RT-DETR-l, the method proposed in this invention improves mAP@0.5 by 39.3%, 14.2%, and 34.4%, respectively. YOLOv8n-EfficientNet, YOLOv8n-MobileVit, and YOLOv8n-MobileNetV3 are lightweight improvements to the backbone of YOLOv8. It can be seen that these three are significantly effective in reducing the number of model parameters and computational cost, but their accuracy is far inferior to the feature reconstruction network model, further proving the effectiveness of the feature reconstruction network model. The feature reconstruction network model of this invention dynamically adjusts the receptive field and uses efficient convolution operations through the backbone network layer, improving the feature extraction capability and detection accuracy for small targets. The neck network layer retains more crack detail information, reducing the false negative rate of the feature reconstruction network model, enabling more accurate crack defect localization and improved detection accuracy.

[0143] Therefore, the feature recombination network model of the present invention has significant advantages. It improves accuracy while being lightweight, and meets the application value of actual crack target detection. It is suitable for deployment on devices with limited computing resources.

[0144] 3.2.3 Visualization Verification Experiment

[0145] To verify the superiority of the feature reorganization network model in crack defect detection, heatmaps were used to compare the detection performance of different models. Heatmaps transform crack defect datasets into intuitive, colorful images, visually representing changes in color temperature: warmer tones indicate higher values, while cooler tones indicate lower values. The darker the color, the better the model's detection performance. A comparison of heatmap detection performance is shown below. Figure 11 As shown.

[0146] Depend on Figure 11 It can be seen that the feature recombination network model has the best detection performance compared to YOLOv5 and YOLOv8. Among them, the feature recombination network model... Figure 11(a) It can be seen that YOLOv5 can successfully identify crack defects, but the dark areas in the heatmap are fewer. YOLOv8 showed missed detections and false detections. The feature reconstruction network model of this invention can not only identify crack defects better, but also has darker colors, indicating a higher improvement in the detection of small target crack defects; Figure 11 (b) Figure 11 (c) It can be seen that YOLOv5 and YOLOv8 do not focus on the crack defect itself, and dark areas appear around the defect. In contrast, the feature reconstruction network model shows a darker color at the crack defect, indicating that it is more accurate in identifying crack defects and its detection performance is better than the former two. The feature reconstruction network model improves the model's feature extraction capability by designing modules such as C2f-FE, GSConv, LSKNet, and CARAFE, and through efficient convolution operations and feature fusion mechanisms. It performs particularly well in capturing details of small target crack defects, and can more accurately identify small target crack defects, avoiding missed detections and false detections. This shows that the feature reconstruction network model has significantly improved detection accuracy and reliability.

[0147] To verify the effectiveness of the feature recombination network model in real-world crack detection scenarios, this study selected a subset of images from a crack dataset for detection, comparing the detection results with those of other models. Figure 12 As shown. By Figure 12 As can be seen from the green highlighted boxes, the YOLOv8 model has insufficient ability to detect some small target cracks and exhibits missed detections. The method proposed in this invention uses modules such as C2f-FE, LSKNet, and CARAFE to dynamically adjust the receptive field of the feature recombination network model, enabling it to better adapt to crack features of different scales. Through feature recombination and feature expansion, more crack detail information is preserved, enhancing the model's focus on small target crack features. In real-world scenarios, it can more accurately detect more small target crack defects, reducing the missed detection rate for small targets. Therefore, the FRNet model proposed in this invention helps improve the accuracy of small target crack detection while achieving a lightweight effect, meeting the practical needs of intelligent crack detection.

[0148] 4 Overview

[0149] To address the issues of false positives and false negatives in the detection of small target cracks and defects during wind turbine blade operation, this invention proposes a crack and defect detection method for wind turbine blades based on a feature recombination network. This method designs a C2f-FE feature extraction module, which is incorporated into both the backbone and neck layers. Furthermore, a large-kernel selective network, LSKNet, is introduced into the backbone layer to dynamically adjust the receptive field for feature extraction, thereby enhancing the model's feature extraction capabilities. In the neck layer, a lightweight CARAFE upsampling module is designed to retain more crack detail information through feature recombination and feature expansion, improving the upsampling effect. The SimAM attention mechanism is used to enhance the attention to the contextual information of crack and defect features. Additionally, a GSConv hybrid convolution module is employed to reduce the model size, which, while lightweight, also helps reduce the false positive and false negative rates for small target cracks and defects. Ultimately, this method achieves more accurate detection of small target cracks and defects in wind turbine blades.

[0150] Experimental results show that the proposed method for crack defect detection achieves 78.6% lower mAP, 2.5M fewer parameters, and 7.0G less computational cost. Compared with the original YOLOv8 network model, mAP@0.5 is improved by 2.9%, while the number of parameters and computational cost are reduced by 16.7% and 13.6%, respectively. Compared with algorithms such as YOLOv3-tiny, YOLOv5n, DERT, EfficientNet, MobileVit, and MobileNetV3, the feature recombination network model maintains high accuracy while having lower parameter and computational costs, and significantly reduces the false negative and false positive rates, meeting the requirements for practical engineering applications.

[0151] 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 it. Although the present invention has been described with reference to preferred embodiments, those skilled in the art should understand that various changes in form and detail can be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for detecting crack defects in wind turbine blades based on feature recombination networks, characterized in that, The image of the wind turbine blade to be detected is acquired and input into a pre-trained feature reconstruction network model to obtain the crack defect identification result of the wind turbine blade image to be detected. The feature reconstruction network model includes a backbone network layer, a neck network layer, and a detection head. The backbone network layer is used to extract multi-scale features from the wind turbine blade image input to the feature reconstruction network to obtain a multi-scale feature map. The neck network layer is used to perform feature reconstruction and feature expansion processing on the multi-scale feature map, and to fuse the multi-scale features to obtain a fused feature map. The detection head is used to obtain the crack defect identification result of the wind turbine blade image to be detected based on the fused feature map, which is used as the output of the feature reconstruction network model. In the feature recombination network model: The backbone network layer includes sequentially connected convolutional layers, four convolutional fusion units, one large kernel selective network, and one spatial pyramid pooling module; each convolutional fusion unit includes cascaded convolutional layers and a C2f-FE module; The neck network layer includes two feature reconstruction upsampling units and two convolutional connection fusion units connected in sequence; wherein, the feature reconstruction upsampling unit includes a CARAFE upsampling module, a SimAM attention module, a connection layer and a C2f-FE module connected in sequence; the convolutional connection fusion unit includes a GSConv hybrid convolution module, a connection layer and a C2f-FE module connected in sequence. The detection head is used to obtain the crack defect identification result of the wind turbine blade image based on the output of the neck network layer, and as the output of the feature reconstruction network model; In this system, the input of the feature reconstruction network model serves as the input to the backbone network layer. Within the backbone network layer, the outputs of the second and third convolutional fusion units also serve as the inputs to the connection layers of the second and first feature reconstruction upsampling units in the neck network layer, respectively. In addition to serving as the input to the neck network layer, the output of the spatial pyramid pooling module in the backbone network layer also serves as the input to the connection layer of the second convolutional connection fusion unit in the neck network layer. Furthermore, the output of the first feature reconstruction upsampling unit in the neck network layer also serves as the input to the connection layer of the first convolutional connection fusion unit. The outputs of the second feature reconstruction upsampling unit and the two convolutional connection fusion units are all output to the detection head.

2. The method for detecting crack defects in wind turbine blades based on feature recombination networks according to claim 1, characterized in that, The C2f-FE module includes a cascaded first convolutional layer, a split module, several fast convolutional attention modules, a connection layer, and a second convolutional layer. The two split outputs of the split module serve as the inputs of the first fast convolutional module and the connection layer, respectively. The outputs of each fast convolutional module also serve as the inputs of the connection layer. The output of the second convolutional layer serves as the output of the C2f-FE module.

3. The method for detecting crack defects in wind turbine blades based on feature recombination networks according to claim 2, characterized in that, The fast convolutional attention module includes a cascaded 3×3 partial convolutional layer, two 1×1 convolutional layers, and an EMA efficient multi-scale attention module. The output of the EMA efficient multi-scale attention module is superimposed with the input of the fast convolutional module to serve as the output of the fast convolutional module.

4. The method for detecting crack defects in wind turbine blades based on feature recombination networks according to claim 1, characterized in that, The large kernel selectivity network includes multiple cascaded large kernel selectivity units; each large kernel selectivity unit includes a first normalization layer, a large kernel selectivity sub-block, a second normalization layer, and a feedforward network sub-block connected in sequence, and the output of the large kernel selectivity sub-block is superimposed with the input of the first normalization layer and then used as the input of the second normalization layer, and the output of the feedforward network sub-block is superimposed with the input of the second normalization layer and then used as the output of the large kernel selectivity unit; The large kernel selection sub-block consists of a cascaded fully connected layer, a GELU activation function layer, an LSK large kernel selection module, and another fully connected layer; the feedforward network sub-block consists of a cascaded fully connected layer, a deep convolutional layer, a GELU activation function layer, and another fully connected layer.

5. The method for detecting crack defects in wind turbine blades based on feature recombination networks according to claim 1, characterized in that, The CARAFE upsampling module is used for upsampling of dimension 1. Input feature map Generate a dimension of Upsampling Map As output Indicates the upsampling rate, upsampling map. Each target location Input feature map Each of them has a corresponding source location. ,and , ; The processing steps of the CARAFE upsampling module include: First, based on the input feature map The above are different source locations Centered subregion Through the kernel prediction module Predict and generate upsampling mapping Different target locations Corresponding position sensing kernel : ； in, This indicates the operation of the kernel prediction module. Represents the input feature map Upper position Value at, sub-region That is of Neighborhood; Then, through the content-aware reorganization module Recombination calculation generates upsampling mapping map Different target locations value at : ； in, Indicates in sub-region Position coordinates in Subregion The radius of the region; Indicates upsampling mapping Target position The corresponding location sensing kernel at the location; Input feature map Upper position The value at; Thus, the upsampling map is obtained as the output. .

6. The method for detecting crack defects in wind turbine blades based on feature recombination networks according to claim 1, characterized in that, The processing procedure of the SimAM attention module includes: For dimension Input feature map For the values ​​at each position in each channel of the graph Calculate the corresponding energy functions respectively Thus, the input feature map is obtained. corresponding Energy function characteristic diagram , : ； in, Represents the input feature map Any of the first The first one on each channel Position value The corresponding energy function, , , ; It is a regular term; Input feature map The Middle The variance of the values ​​at each position on each channel Input feature map The Middle The mean of the values ​​at each position on each channel: ， ； Then, the output feature map of the SimAM attention module is calculated using the following formula. : ； in, This represents the Sigmoid activation function operation; This indicates element-wise multiplication.

7. The method for detecting crack defects in wind turbine blades based on feature recombination networks according to claim 1, characterized in that, The processing steps of the GSConv hybrid convolution module include: First, the input feature map is convolutionally downsampled. Then, the output of the convolutionally downsampled map is further processed by depthwise convolution. The output of the depthwise convolution is then concatenated with the output of the convolutionally downsampled map. Finally, the concatenated result is shuffled to obtain the output feature map.

8. The method for detecting crack defects in wind turbine blades based on feature recombination networks according to claim 1, characterized in that, The feature recombination network model is trained in the following manner: S101: Prepare a sample image dataset of wind turbine blade crack defects, which is divided into a training image dataset and a test image dataset. The wind turbine blade crack defect sample images in the wind turbine blade crack defect sample image dataset are all pre-marked with the location and type of crack defects. S102: Input the training image dataset into the feature reconstruction network model for training, and optimize the parameters of the feature reconstruction network model with the goal of minimizing the loss function until the feature reconstruction network model converges, thus obtaining the trained feature reconstruction network model. S103: Test the feature reconstruction network model using the test image dataset to confirm the recognition performance of the trained feature reconstruction network model; if the recognition performance meets the requirements, end the training of the feature reconstruction network model; otherwise, return to step S102.

9. The method for detecting crack defects in wind turbine blades based on feature recombination networks according to claim 8, characterized in that, In the loss function, the classification loss is the total loss function formed by the sum of the CIoU loss function and the DFL loss function, and the localization loss is the binary cross-entropy (BCE) loss function.

Images