Wind turbine generator blade crack defect detection method based on feature recombination network

Abstract

The invention discloses a wind turbine generator blade crack defect detection method based on a feature recombination network, and the method comprises the steps: obtaining a to-be-detected wind turbine generator blade image, inputting the to-be-detected wind turbine generator blade image to a pre-trained feature recombination network model, and obtaining a crack defect recognition result of the to-be-detected wind turbine generator blade image. According to the method, multi-scale feature extraction is carried out on the wind turbine generator blade image through the small target feature extraction capability of the backbone network layer by using an optimized and constructed feature recombination network model, and then more crack detail information in the multi-scale features is reserved by means of feature recombination and feature expansion of the neck network layer. The context information of the small target is fully acquired for feature fusion, and then the crack defect identification result of the wind turbine generator blade image is acquired through the detection head, so that the crack defect of the small target is detected more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of structural surface defect target detection, and in particular to a method for detecting crack defects in wind turbine blades based on a feature recombination network. Background Art

[0002] In the wind power industry, wind turbine blades are one of the key components that convert wind energy into mechanical energy. They are affected by the external environment, various human and non-human factors. During the operation of the blades, structural defects such as cracks, peeling and erosion are prone to occur. Existing blade inspections mainly rely on human visual observation and subjective experience to determine whether cracks exist, resulting in poor consistency in the evaluation results and difficulty in ensuring the stability of blade quality. Therefore, the introduction of a crack intelligent recognition model in the blade inspection process, the target detection of defects such as cracks, peeling and erosion, timely detection and reduction of human misjudgment and missed judgment, and accurate determination of blade quality are of great significance to ensuring the safe and reliable operation of wind turbines.

[0003] The target detection methods based on deep learning are mainly divided into two categories: two-stage target detection (Two-stage) and one-stage target detection (One-stage). Common two-stage algorithms include R-CNN, Fast R-CNN, Faster R-CNN, etc. The two-stage target detection model has certain advantages in detection accuracy, but the model is more complex and the reasoning speed is slower, which makes it difficult to meet the real-time detection requirements of crack defects in wind power scenarios. The one-stage target detection methods mainly include YOLO and SSD. Compared with the two-stage target detection method, the one-stage target detection method is faster and easier to meet the requirements of crack defect target detection in actual engineering.

[0004] In recent years, single-stage target detection methods represented by the YOLO network have been widely used in structural surface defect detection and identification. However, the existing methods simply splice the feature segments during the upsampling process without analyzing the importance of each feature segment, which easily leads to the loss of detail information and makes it difficult to fully extract the weak features of small target crack defects. There is still room for further improvement in target detection accuracy. Summary of the invention

[0005] In view of the above-mentioned deficiencies of the prior art, the present invention provides a method for detecting crack defects in wind turbine blades based on a feature recombination network, so as to improve the detection accuracy of small target crack defects on wind turbine blades.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A method for detecting crack defects in wind turbine blades based on a feature recombination network is provided. The method obtains an image of a wind turbine blade to be detected and inputs the image into a pre-trained feature recombination network model to obtain a crack defect recognition result of the image of the wind turbine blade 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 perform multi-scale feature extraction on the wind turbine blade image input into 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 fuse the multi-scale features to obtain a fused feature map; the detection head is used to obtain a crack defect recognition result of the wind turbine blade image to be detected according to the fused feature map 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, 4 convolutional fusion units, 1 large kernel selective network and 1 spatial pyramid pooling module; each convolutional fusion unit includes a cascaded convolutional layer and a C2f-FE module;

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

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

[0013] Among them, the input of the feature recombination network model is used as the input of the backbone network layer; in the backbone network layer, the outputs of the second convolutional fusion unit and the third convolutional fusion unit are also used as the input of the connection layer in the second feature recombination upsampling unit and the connection layer in the first feature recombination upsampling unit in the neck network layer respectively; the output of the spatial pyramid pooling module in the backbone network layer is used as the input of the neck network layer, and also as the input of the connection layer in the second convolutional connection fusion unit in the neck network layer; in the neck network layer, the output of the first feature recombination upsampling unit is also used as the input of the connection layer in the first convolutional connection fusion unit, and the output of the second feature recombination upsampling unit and the outputs of 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, wherein the two split outputs of the Split module are respectively used as inputs of the first fast convolutional module and the connection layer, the outputs of each fast convolutional module are also used as inputs of the connection layer, and the output of the second convolutional layer is used as the output of the C2f-FE module.

[0015] Specifically, the fast convolution attention module includes a cascaded 3×3 partial convolution layer, 2 1×1 convolution layers and an EMA efficient multi-scale attention module. The output of the EMA efficient multi-scale attention module is superimposed on the input of the fast convolution module as the output of the fast convolution module.

[0016] Specifically, the large-core selective network includes a plurality of cascaded large-core selective units; each large-core selective unit includes a first normalization layer, a large-core selection sub-block, a second normalization layer and a feedforward network sub-block connected in sequence, and the output of the large-core selection 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-core selective unit;

[0017] Among them, the large core selection sub-block includes a cascaded fully connected layer, a GELU activation function layer, an LSK large core 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.

[0018] Specifically, the CSEAFE upsampling module is used to upsample the input features of dimension C×H×W Figure X Generate an upsampled map of dimension C×σH×σW Figure X ′ as output, σ represents the upsampling rate, and the upsampling mapping Figure X Each target position l′=(i′, j′) in ′ is in the input feature Figure X There is a corresponding source position l = (i, j) on each The processing of the CSEAFE upsampling module includes:

[0019] First, according to the input features Figure X k centered at different source positions l = (i, j) up ×k up Sub-area N(X l , k up ), the up-sampled map is generated by the kernel prediction module ψ Figure X The position perception kernel W corresponding to different target positions l′=(i′, j′) in ′ l:

[0020] W l′ =ψ(N(X l , k up ));

[0021] Among them, ψ(·) represents the kernel prediction module operation, X l Represents input features Figure X The upper position, the value at, the sub-area N(X l , k up ) is X l K up ×k up Neighborhood;

[0022] Then, the upsampled map is generated by recombining the computation through the content-aware recombinant module φ Figure X The value X′ at each different target position l′=(i′, j′) in ′ j :

[0023]

[0024] Where (n, m) represents the number of l , k up ), Represents sub-area N(X l , k up ) area radius; W l′＝(n，m) represents the corresponding position-aware kernel at the target position l′=(n, m) in the upsampled map X′; l＝(i+n，j+m) Input features Figure X The value at upper position l = (i+n, j+m);

[0025] Thus, we get the upsampled map as output Figure X ′.

[0026] Specifically, the processing process 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 gives the input features Figure X The corresponding C×H×W energy function feature map E,

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

[0029]

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

[0031]

[0032] Among them, Sigmoid(·) represents the Sigmiod activation function operation; Θ represents the element-wise multiplication operation.

[0033] Specifically, the processing process of the GSConv mixed convolution module includes:

[0034] First, the input feature map is convolved and downsampled, and then the output of the convolution downsampling is further processed by deep convolution. After the output of the deep convolution processing is spliced ​​with the output of the convolution downsampling, the spliced ​​result is shuffled to obtain the output feature map.

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

[0036] S101: preparing a wind turbine blade crack defect sample image dataset, which is divided into a training image dataset and a test image dataset, wherein the wind turbine blade crack defect sample images in the wind turbine blade crack defect sample image dataset are all pre-marked with crack defect locations and types;

[0037] S102: inputting the training image data set into the feature recombination network model for training, and performing parameter optimization training on the feature recombination network model with the goal of minimizing the loss function until the feature recombination network model converges, thereby obtaining a trained feature recombination network model;

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

[0039] Specifically, in the loss function, the classification loss selects the total loss function composed of the sum of the CIoU loss function and the DFL loss function, and the positioning loss selects the binary cross entropy BCE loss function.

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

[0041] 1. The wind turbine blade crack defect detection method based on the feature recombination network of the present invention utilizes the optimized feature recombination network model to extract multi-scale features of the wind turbine blade image through the small target feature extraction capability of the backbone network layer, and then retains more crack detail information in the multi-scale features with the help of feature recombination and feature expansion of the neck network layer, fully obtains the context information of the small target for feature fusion, and then obtains the crack defect recognition result of the wind turbine blade image through the detection head, thereby realizing more accurate detection of small target crack defects.

[0042] 2. In the feature recombination 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 the backbone network layer and the neck network layer at the same time. In addition, the large core 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 the present invention, a lightweight CARAFE upsampling module is designed in the neck network layer. More crack detail information is retained through feature recombination and feature expansion to improve the upsampling effect. The SimAM attention mechanism is used to enhance attention to the contextual information of crack defect features. The GSConv mixed convolution module is also used to reduce the size of the model. While making it lightweight, it also helps to reduce the missed detection rate and false detection rate of small target crack defects. Finally, more accurate detection of small target crack defects in wind turbine blades is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0045] Figure 1 Schematic diagram of the architecture of the feature recombination network model used in the method of the present invention;

[0046] Figure 2 It is a schematic diagram of the structure of the C2f-FE module;

[0047] Figure 3 Schematic diagram of the structure of the fast convolutional attention module FEblock in the C2f-FE module;

[0048] Figure 4 This is a schematic diagram of the structure of the EMA efficient multi-scale attention module;

[0049] Figure 5 This is the processing flow chart of the EMA efficient multi-scale attention module;

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

[0051] Figure 7 Processing flow chart for the large core selection sub-block LKSelection;

[0052] Figure 8 It is the processing flow chart of the CSEAFE upsampling module;

[0053] Fig. 9 This is the processing flow chart of SimAM’s attention module;

[0054] Fig.10 This is the processing flow chart of the GSConv mixed convolution module;

[0055] Fig.11 This is a comparison chart of the thermal map detection effects of different models in the embodiment;

[0056] Fig.12 4 is a comparison chart of the detection effects of different models in the embodiments. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0058] The present invention will be further described in detail below in conjunction with the accompanying drawings.

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

[0060] The present invention discloses a method for detecting crack defects in wind turbine blades based on a feature recombination network. The method utilizes an optimized feature recombination network model to extract multi-scale features of wind turbine blade images through the small target feature extraction capability of a backbone network layer. The method then uses feature recombination and feature expansion of a neck network layer to retain more crack detail information in the multi-scale features, fully acquires context information of small targets for feature fusion, and then obtains crack defect recognition results of wind turbine blade images through a detection head, thereby achieving more accurate detection of small target crack defects.

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

[0062] 1 Feature Recombination Network Model

[0063] In view of the problem of low accuracy of small target crack detection during the operation of wind turbine blades, the present invention uses a pre-trained feature recombination network model to detect crack defects in wind turbine blade images. The feature recombination network model used in the method of the present invention is obtained by optimizing and improving the YOLOv8 network as the 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 includes sequentially connected convolutional layers, 4 convolutional fusion units, 1 large kernel selective network and 1 spatial pyramid pooling module; each convolutional fusion unit includes cascaded convolutional layers and C2f-FE modules.

[0065] In the backbone network layer, the wind turbine blade image input into the feature reconstruction network model first passes through the convolution layer in the backbone network layer for initial convolution, and then passes through 4 convolution fusion units containing C2f-FE modules layer by layer for multi-level feature extraction. It then passes through the large kernel selective network LSKNet to adjust the receptive field of the backbone network layer for feature extraction to adapt to crack defect characteristics of different scales and improve the model's feature extraction capability. Finally, it passes through the spatial pyramid pooling module to generate a multi-scale feature map of a specified size.

[0066] The neck network layer includes two feature reorganization upsampling units and two convolution connection fusion units connected in sequence; among them, the feature reorganization upsampling unit includes a CSEAFE upsampling module, a SimAM attention module, a connection layer and a C2f-FE module connected in sequence; the convolution connection fusion unit includes a GSConv mixed convolution module, a connection layer and a C2f-FE module connected in sequence.

[0067] The multi-scale feature map enters the neck network layer and first passes through two feature recombination upsampling units. In the feature recombination upsampling unit, it is first upsampled by the CSEAFE upsampling module, and more crack detail information is retained through feature recombination and feature expansion to improve the upsampling effect. The SimAM attention module is introduced in the feature recombination upsampling unit, and the SimAM attention mechanism is used to enhance the attention to the contextual information of the crack defect feature. Then, after being processed by the connection layer and the C2f-FE module, it is output to two convolution connection fusion units. In the convolution connection fusion unit, the convolution operation is optimized by the GSConv mixed convolution module. The GSConv mixed convolution module is a lightweight convolution module that helps to reduce the size of the model, and while being lightweight, it also helps to reduce the missed detection rate and false detection rate of small target crack defects. Then, it is processed by the connection layer and the C2f-FE module and output. After passing through two convolution connection fusion units, a fused feature map as the output of the neck network layer is obtained.

[0068] The detection head is used to obtain the crack defect recognition result of the wind turbine blade image according to the output of the neck network layer as the output of the feature reconstruction network model. The crack defect recognition result finally output by the detection head can present the category, location and confidence of the crack defect target in the wind turbine blade image. In the specific implementation, the detection head part can adopt the mainstream decoupled head structure (Decoupled Head) to reduce the number of parameters and computational complexity, and perform the classification task and training regression task of crack defect recognition at the same time.

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

[0070] The feature recombination network model used in the method of the present invention is mainly optimized and improved based on the YOLOv8 network in that: firstly, the C2f-Faster module is combined with the EMA efficient multi-scale attention mechanism to design the C2f-FE module, which is introduced into the backbone network layer and the neck network layer at the same time, and the large core 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 feature extraction capability of the model; in the neck network layer, a lightweight CARAFE upsampling module is designed to retain more crack detail information through feature recombination and feature expansion, thereby improving the upsampling effect, and using the SimAM attention mechanism to enhance the attention to the contextual information of the crack defect feature, and also combining with the GSConv mixed convolution module to reduce the size of the model, and while lightweighting, it also helps to reduce the missed detection rate and false detection rate of small target crack defects; finally, more accurate detection of small target crack defects in wind turbine blades is achieved.

[0071] Next, several major optimized and improved module units in the feature recombination network model are introduced in detail.

[0072] 1.1C2f-FE module

[0073] In order to improve the detection performance of small target cracks and realize real-time intelligent detection of cracks, the EMA attention mechanism is combined with the Fasterblock structure and added into the C2f magic puppet 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, wherein the two split outputs of the Split module are respectively used as the input of the first fast convolutional module and the connection layer, the output of each fast convolutional module is also used as the input of the connection layer, and the output of the second convolutional layer is used as the output of the C2f-FE module.

[0075] The C2f-FE module is designed based on this, which enhances the model's multi-scale feature extraction capability through efficient convolution operations and feature fusion mechanisms, significantly improves the accuracy and speed of small target detection, reduces the computational cost, and can detect small target crack defects more quickly and accurately.

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

[0077] The first half of the 3×3 partial convolutional layer and two 1×1 convolutional layers in the fast convolutional attention module FEblock is a Fasterblock structure. The PConv operation of the partial convolutional layer reduces redundant information and memory usage in the calculation and improves the learning ability of the network. 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 of all channels can be fully and effectively utilized, and the diversity of features extracted by the model can be maintained.

[0078] Since some cracks occupy a small proportion in the image and the target pixels are low, the crack quality detection effect is not ideal. Therefore, by introducing the Efficient Multi-scale Attention (EMA) module that learns efficient multi-scale attention mechanism across space in the C2f-FE module, it can better focus on the salient features of the target object to improve the feature extraction ability of the model.

[0079] EMA efficient multi-scale attention module is a commonly used multi-scale feature extraction unit. For any given input feature Figure X ∈R C×H×W , X is divided into g sub-features in the cross-channel dimension direction for learning different semantic features. Its module structure is as follows Figure 4 As shown in the figure, the EMA efficient multi-scale attention module reconstructs some channels in batches while retaining the information on the channel and reducing the computational overhead. It further aggregates the output features of the two parallel branches in a cross-dimensional interactive manner to capture the pairwise relationship at the pixel level, so that the model can achieve higher accuracy with fewer parameters.

[0080] The processing flow of the EMA efficient multi-scale attention module on the input feature map is as follows Figure 5As shown in the figure. The size of the input feature map is defined as C / / G×H×W, where two parallel paths are 1×1 branches that perform global average pooling operations in the horizontal and vertical dimensions, and the third parallel path is a 3×3 branch for feature extraction. On this basis, the output features of the two branches are adjusted using the Sigmoid function and normalization operations, and different cross-channel interaction features are implemented between the two parallel paths to obtain 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. In order to improve computational efficiency, the nonlinear function Softmax is used to fit the linear transformation; then the matrix multiplication operation is used to obtain the spatial attention map under different scales of spatial information, as well as the attention map of accurate spatial position information.

[0081] 1.2LSKNet Module

[0082] In order to enhance the multi-scale information and detail retention capabilities of the feature map, a large kernel selective network LSKNet (Localization-Sensitive Keypoints Network) is introduced in the backbone network layer. The feature map is optimized in the feature extraction stage to enhance the global and local information, so that the feature pyramid can more effectively integrate these features, which helps to improve the effect of small target detection. Since the detection of small target cracks is easily affected by background information and the feature information is not rich enough, it is very easy to be ignored or misclassified. Therefore, the designed LSKNet dynamically adjusts the receptive field in the feature extraction module, selects the appropriate convolution kernel size and expansion rate when detecting small targets, so as to better capture the characteristic information of small target crack defects.

[0083] The core of LSKNet is to combine large convolution kernels and selective mechanisms. Different categories of objects have different background information requirements, which requires the model to adaptively select them. Therefore, by decoupling a series of depth-wise convolutions with large convolution kernels and continuous expansion, a network with a larger receptive field is constructed. At the same time, on this basis, a spatial selection mechanism is used to spatially select feature maps from large convolution kernels of different scales, so that the model pays more attention to the key background information of the target in space.

[0084] The large core selective network LSKNet includes a plurality of cascaded large core selective units LSK block, and the specific number of large core selective units included can be selected and set according to needs, usually 2 to 4.

[0085] The structure of each large kernel selection unit LSK block mainly includes two residual sub-blocks, namely the large kernel selection sub-block (Large Kernel Selection, LKSelection) and the feed-forward network sub-block (Feed-forward Network, FFN). Specifically, its structure is as follows Figure 6 As shown, it includes a first normalization layer, a large core selection sub-block, a second normalization layer and a feedforward network sub-block connected in sequence, and the output of the large core selection 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 core selection unit. Among them, the large core selection sub-block includes a cascade of a fully connected layer, a GELU activation function layer, an LSK large core selection module and another fully connected layer; the feedforward network sub-block includes a cascade of a fully connected layer, a deep convolutional layer, a GELU activation function layer and another fully connected layer. Among them, the large core 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 mixing and feature refinement.

[0086] The processing of the large core selection sub-block LKSelection is as follows Figure 7 As shown in the figure, in the large kernel selection sub-block LKSelection, the input feature map first goes through the large kernel decomposition step and is processed using different large convolution kernels to capture spatial information of different scales; the outputs of two different convolution kernels are combined through channel splicing so that different spatial features can be used simultaneously in subsequent steps; the spliced ​​feature map undergoes a combination of average pooling and maximum pooling to further strengthen the key areas of the feature map; then, a new feature map is generated through self-attention (SA) and convolution operations, and then the focus on the target area is further enhanced through the spatial selection mechanism; finally, a mask is generated using the Sigmoid function, and then the mask is element-wise multiplied with the feature map to obtain the final output feature map. This step is used to weight more important areas in the feature map to enhance the network's ability to detect specific objects.

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

[0088] 1.3 CARAFE Module

[0089] In order to retain more detail information during the upsampling process, a lightweight CARAFE upsampling module is designed. It recombines features in a content-aware way, effectively retains detail information, makes the feature expression of small targets richer and more accurate, and significantly improves the detection accuracy of small target crack defects. Existing upsampling methods cannot effectively retain and reorganize detail information, resulting in insufficient feature expression of small targets, and detail information is easily lost during the upsampling process, affecting the detection effect of small targets. Therefore, the designed lightweight content-aware CSEAFE upsampling module can generate adaptive reorganization weights according to the content of the input features, enhancing the model's ability to capture small target details.

[0090] The CSEAFE upsampling module is mainly composed of two key modules: the kernel prediction module and the content-aware reconstruction module. The kernel prediction module is responsible for generating upsampling kernels for feature reconstruction, predicting the attention weight of each upsampling position based on the mapping relationship between the downsampling feature map and the upsampling position, and maintaining spatial details and contextual information during the feature reconstruction process. The role of the content-aware reconstruction module is to retain as much spatial information as possible during the upsampling process to better maintain the accuracy of object boundaries.

[0091] In the processing of CARAFE, if an input feature with dimensions C×H×W is given Figure X , setting the upsampling rate σ, CARAFE maps the upsampling to the dimension C×σH×σW Figure X 'As output, upsample the map Figure X Each target position l′=(i′,j′) in ′ is in the input feature Figure X There is a corresponding source position l = (i, j) on each

[0092] Specifically, if Figure 8 As shown, the processing process of the CSEAFE upsampling module includes:

[0093] N(X l ,k up ) is represented as input feature Figure X The source position l = k centered at (i, j) up ×k up Sub-region, namely X l K up ×k up Neighborhood; then, the CSEAFE upsampling module firstly calculates the input features Figure X k centered at different source locations l=(i,j) up ×k up Sub-area N(X l ,k up), the up-sampled map is generated by the kernel prediction module ψ Figure X The position perception kernel W corresponding to different target positions l′=(i′,j′) in ' l' , which is calculated as:

[0094] W l' =ψ(N(X l ,k up ));

[0095] Among them, ψ(·) represents the kernel prediction module operation, X l Represents input features Figure X The value at position l on the sub-region N(X l ,k up ) is X l K up ×k up Neighborhood.

[0096] The kernel prediction module is responsible for generating position-aware kernels in a content-aware manner. Figure X Each source position on corresponds to the upsampled map Figure X ' 2 target locations. Each target location requires a k up ×k up The position-aware kernel of is used as the reorganization weight. Therefore, this module will output a size of C up ×H×W location-aware kernel, where C up =σ 2 k up 2 .

[0097] The kernel prediction module consists of three submodules, namely, channel compressor, content encoder and kernel normalizer, as Figure 8 As shown in Figure 2. The channel compressor reduces the number of channels of the input feature map. Then, the content encoder takes the compressed feature map as input and encodes the content to generate position-aware kernels. 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 channel from C to C m ,Reducing the number of channels of the input feature map can reduce the number of parameters and computational cost in subsequent steps, making CARAFE more efficient. Under the same budget, a larger kernel size can also be used for the content encoder. Experimental results show that reducing the feature channels within an acceptable range does not harm the performance.

[0099] In the content encoder, use a size of k encoderThe convolutional layer generates a position-aware kernel based on the content of the input features. The encoder parameter is k encoder ×k encoder ×C m ×C up Intuitively, increasing k encoder It can enlarge the encoder's receptive field and exploit contextual information in a larger area, which is important for predicting position-aware kernels. However, the computational complexity grows with the square of the kernel size, while the benefits from larger kernel sizes do not. Therefore, k is usually taken as encoder =k up -k, the value of k is usually an integer from 0 to 2, which can achieve a good balance between performance and efficiency; when k = 0, that is, k encoder =k up .

[0100] In the kernel normalizer, before re-weighting the input feature map, each k up ×k up All position-aware kernels are spatially normalized by a Softmax function. The normalization step forces the kernel values ​​to sum to 1, which is soft-selective within a local region. Due to the kernel normalizer, CARAFE does not perform any rescaling and changes the mean of the feature maps.

[0101] Then, the upsampled map is generated by recombining the computation through the content-aware recombinant module φ Figure X The value X at each different target position l'=(i',j') in ' l ”, which is calculated as

[0102]

[0103] Where (n,m) represents the number of l ,k up ), Represents sub-area N(X l ,k up ) area radius; W l′＝(n,m) represents the corresponding position-aware kernel at the target position l'=(n,m) in the upsampled map X'; l＝(i+n,j+m) Input features Figure X The value at upper position l = (i+n, j+m);

[0104] In this process, each location-aware kernel W is used l' , the content-aware reassembly module will reassemble the features in the local area. The content-aware reassembly module can be regarded as a weighted summation operator. For the target location l'=(n,m) and k centered at the source location l=(i,j), up ×kup Sub-area N(X l ,k up ), the upsampling mapping is calculated by the above formula Figure X The value X′ at each different target position l′=(i′,j′) in ' l′ , so that the region sub-region N(X l ,k up ) will be based on the content of the feature rather than the distance of the position, and the value X′ at the upsampled target position l′ Produce different contributions. Since the information of related points in the local area can be paid more attention, the semantics of the recombined feature map may be stronger than the original feature map.

[0105] Thus, the value X′ at each up-sampled target position is determined l′ After that, we get the upsampled map as output Figure X '.

[0106] Compared with conventional upsampling convolution methods, the CSEAFE upsampling module can upsample each feature point in the image through adaptively generated position-aware kernels, so that these feature points gather contextual semantic information, thereby effectively reducing the loss of image features in the upsampling process and enhancing the network's feature extraction capabilities.

[0107] 1.4SimAM Attention Module

[0108] In order to further enhance the expression of detailed features and contextual information and thus improve the model detection effect, the SimAM attention module is introduced into the model. The SimAM attention mechanism of the SimAM attention module is a three-dimensional attention mechanism that combines spatial information and channel information. It can better focus on the key information of the input features, thereby improving the feature expression ability of the network. Figure X ∈R C×H×W , the SimAM attention module defines the energy function Evaluate the value x at each position in each channel c,i The specific processing process is 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 gives the input features Figure X The corresponding C×H×W energy function feature map E,

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

[0111]

[0112] When the energy function The lower the value of , the c,i The more different a value is from the surrounding values, the more important it is. c,i The importance of To express.

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

[0114]

[0115] Among them, Sigmoid(·) represents the Sigmiod activation function operation; Θ represents the element-wise multiplication operation.

[0116] The SimAM attention mechanism in the SimAM attention module can be expressed as Fig. 9 As shown in the figure, C represents the number of channels of the input feature map, H represents the height, and W represents the width. The SimAM attention mechanism generates three-dimensional attention mechanism weights through neuron perception to more accurately extract small target features.

[0117] 1.5GSConv hybrid convolution module

[0118] In order to reduce the false detection and missed detection of small target crack defects, the GSConv hybrid convolution module is designed to reduce the time of standard convolution to extract image features. While maintaining the feature extraction capability, it significantly reduces the number of model parameters and the amount of calculation. The GSConv hybrid convolution module is a new type of lightweight convolution composed of standard convolution and depth-separable convolution. Its structure is as follows: Fig.10 The processing process of the GSConv hybrid convolution module is as follows: first, the input feature map is convolved and downsampled, and then the output of the convolution downsampled is further processed by deep convolution, and then the output of the deep convolution processing is concat-joined with the output of the convolution downsampled, and the concatenation result is shuffled to obtain the output feature map.

[0119] Although the depth-wise separable convolution can effectively reduce the number of network parameters, it will also lose inter-channel feature information. GSConv, on the other hand, can maximize the connection between feature maps while reducing the number of parameters, and enhance the ability to extract small target information. While lightweighting the network, it can effectively reduce the missed detection rate and false detection rate of small targets, thereby further improving the overall detection performance of the network. Therefore, through the lightweight model processing of the GSConv hybrid convolution module, the target detection task can be completed more efficiently while ensuring the accuracy of the model.

[0120] 2 Training of feature recombination network model

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

[0122] S101: preparing a wind turbine blade crack defect sample image dataset, which is divided into a training image dataset and a test image dataset, wherein the wind turbine blade crack defect sample images in the wind turbine blade crack defect sample image dataset are all pre-marked with crack defect locations and types;

[0123] S102: inputting the training image data set into the feature recombination network model for training, and performing parameter optimization training on the feature recombination network model with the goal of minimizing the loss function until the feature recombination network model converges, thereby obtaining a trained feature recombination network model;

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

[0125] Among the loss functions used in the training process, the classification loss selects the total loss function consisting of the sum of the CIoU loss function and the DFL loss function, and the positioning loss selects 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 regression speed and training accuracy of the model training.

[0126] 3. Embodiment

[0127] In this embodiment, a data set is used to test and verify the method of the present invention and other methods to better reflect the technical advantages of the method of the present invention.

[0128] 3.1 Dataset and Experimental Parameters

[0129] In this experiment, 192 wind turbine blade crack images were collected on site as a data set, covering different angles, and divided into training sets, validation sets, and test sets. The Labelimg library was used to annotate the data set with a rectangular box, which contains one category: crack (liewen).

[0130] The operating system of the experimental computer is 64-bit Windows 10 Professional Edition, the CPU model is Intel (R) Core (TM) i7-12700, the GPU is NVIDIA GeForce RTX4060, and the memory size is 8G. The integrated compiler software is Pychram2023.1.3, the CUDA version is 12.1, the deep learning framework is Pytorch 2.0, and Python 3.11. During the experimental training process, the training cycle (epoch) is 300, the threads (workers) are 4, the batch size (batchsize) is 8, and the initial learning rate (lr0) is 0.01.

[0131] 3.2 Comparison of experimental results

[0132] 3.2.1 Ablation Experiment

[0133] In order to verify the effectiveness of the feature reconstruction network model proposed in the present invention in crack detection tasks in actual scenarios, an ablation experiment was designed for the improved method. Based on YOLOv8, the C2f-FE module, LSKNet module, lightweight CARAFE upsampling module, SimAM attention mechanism, and hybrid convolution GSConv were designed. Each module was accumulated in turn. 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 the present invention has different degrees of improvement in precision, recall, and mAP indicators, and the number of model parameters and the amount of calculation are significantly reduced, proving that the feature recombination network model is effective. Among them, experiment ① designed a C2f-FE module, and the overall parameter and computational complexity of the model decreased; based on experiment ①, experiment ② introduced GSConv on the Neck end to further reduce the parameter volume of the model and improve the model reasoning speed; based on experiment ②, experiment ③ introduced LSKNet to enhance the feature extraction capability of the model and improve the detection accuracy of the model; based on experiment ③, experiment ④ designed a lightweight CARAFE upsampling module in the Neck module. Although the model precision and recall rate increased compared with experiment ③, the parameter volume and computational complexity of the model increased slightly compared with experiment ③, with the parameter volume increasing by 8.7% and the computational complexity increasing by 7.7% compared with experiment ③; the last group is the experimental results based on the proposed FRNet model. Based on experiment ④, the SimAM attention mechanism is added to enhance the contextual information of small targets. The model's detection accuracy of cracks increased by 0.9% compared with experiment ④. Compared with the YOLOv8 model, the feature recombination network model improved mAP@0.5 by 2.9%, and the number of parameters and the amount of computation were reduced by 16.7% and 13.6%, respectively, which further proved the effectiveness of the proposed method.

[0137] Experiment ① designed the C2f-FE module to better capture important information of the image when extracting features, and optimized the convolution operation to reduce unnecessary calculations and information loss, thereby improving the overall performance of the model; Experiment ② shows that the model can effectively reduce the amount of network parameters and calculations through the Shuffle operation and depthwise separable convolution, and speed up the inference speed of the model; Experiment ③ shows that large-kernel convolution can capture a larger receptive field, so that the model can better capture the small-scale features of small crack defects; The slight increase in the amount of calculation and parameters in Experiment ④ shows that the model complexity will be slightly increased through feature recombination and feature expansion; The last group of experiments shows that the SimAM attention mechanism improves the expression ability of the feature map, so that the model can better focus on the crack defect area and improve the detection accuracy of the model; Therefore, the feature recombination network model significantly improves the model feature extraction ability, dynamically adjusts the receptive field, and retains the small target crack detail information through modules such as C2f-FE, LSKNet, and CARAFE, and has significant improvements in detection accuracy and lightweight, especially in small target crack defect detection, which meets the needs of actual engineering applications.

[0138] 3.2.2 Comparative Experiment

[0139] In order to further verify the detection effect of the feature recombination network model of the present invention, a comparative experiment is conducted with some target detection algorithms, including YOLOv3, YOLOv5, and RT-DETR-1. At the same time, some YOLOv8 algorithms based on lightweight improvements are compared. The comparative experimental results are shown in Table 2.

[0140] Table 2. Experimental results of model algorithm comparison

[0141]

[0142] As shown in Table 2, compared with YOLOv3-tiny, YOLOv5n, and RT-DETR-1, the method proposed in the present invention improves mAP@0.5 by 39.3%, 14.2%, and 34.4% respectively; YOLOv8n-EfficientNet, YOLOv8n-MobileVit, and YOLOv8n-MobileNetV3 are lightweight improvements of the Backbone part of YOLOv8. It can be seen that the three have significant effects in reducing the number of model parameters and the amount of calculation, but are far from the feature recombination network model in terms of accuracy, which further proves that the feature recombination network model is effective. The feature recombination network model of the present invention dynamically adjusts the receptive field and efficient convolution operation through the backbone network layer, improves the feature extraction capability and detection accuracy of small targets; through the neck network layer, more crack detail information is retained, the missed detection rate of the feature recombination network model is reduced, and crack defects can be more accurately located, and the detection accuracy is improved.

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

[0144] 3.2.3 Visualization Verification Experiment

[0145] In order to verify the superiority of the feature recombination network model in crack defect detection, the detection effects of different models are compared using heat maps. Heat maps can convert crack defect datasets into intuitive, colorful images, which are intuitively represented by changes in color temperature: warm tones represent higher values, while cool tones represent lower values. The darker the color, the better the detection effect of the model. The comparison of heat map detection effects is shown in the figure below. Fig.11 shown.

[0146] Depend on Fig.11 It can be seen that compared with YOLOv5 and YOLOv8, the feature recombination network model has the best detection effect. Fig.11(a) It can be seen that YOLOv5 can successfully identify crack defects, but the dark part of the heat map is small, and YOLOv8 has missed detections and false detections. The feature recombination network model of the present invention can not only better identify crack defects, but also has a darker color, indicating that the detection of small target crack defects has been further improved; Fig.11 (b) Fig.11 (c) It can be seen that YOLOv5 and YOLOv8 do not focus on the crack defect itself, and dark areas appear around the defects. However, the color of the feature reconstruction network model is darker at the crack defect, indicating that the crack defect recognition is more accurate and the detection effect is better than the previous two. The feature reconstruction network model improves the model feature extraction capability through the design of C2f-FE, GSConv, LSKNet, CARAFE and other modules, through efficient convolution operations and feature fusion mechanisms, especially in capturing the details of small target crack defects. It can more accurately identify small target crack defects and avoid missed detection and false detection, indicating that the feature reconstruction network model has significantly improved the detection accuracy and reliability.

[0147] In order to verify the effectiveness of the feature recombination network model in the actual crack target detection scenario, this study selected some images from the crack dataset for detection, and compared the detection results. Fig.12 As shown. Fig.12 It can be seen that the green marked box can intuitively find that the YOLOv8 model has insufficient crack detection capabilities for some small targets, and there is a phenomenon of missed detection. The method proposed in the present invention uses modules such as C2f-FE, LSKNet, and CARAFE to enable the feature recombination network model to dynamically adjust the receptive field, which can better adapt to crack characteristics of different scales. Through feature recombination and feature expansion, more crack detail information is retained, and the model's attention to small target crack characteristics is enhanced. In actual scenarios, more small target crack defects can be detected more accurately, reducing the missed detection rate of small targets. Therefore, the FRNet model proposed in the present invention helps to improve the accuracy of small target crack detection while achieving a lightweight effect, meeting the actual needs of intelligent crack detection.

[0148] 4 Overview

[0149] In view of the problems that small target crack defect detection is prone to false detection and missed detection during the operation of wind turbine blades, the present invention proposes a wind turbine blade crack defect detection method based on a feature recombination network. The method designs a C2f-FE feature extraction module and introduces it into the backbone network layer and the neck network layer at the same time. In addition, a large core 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 feature extraction capability. In the neck network layer, a lightweight CARAFE upsampling module is designed to retain more crack detail information through feature recombination and feature expansion, thereby improving the upsampling effect. The SimAM attention mechanism is used to enhance the attention to the contextual information of the crack defect feature. The GSConv mixed convolution module is also used to reduce the size of the model. While lightweighting, it also helps to reduce the missed detection rate and false detection rate of small target crack defects. Finally, more accurate detection of small target crack defects in wind turbine blades is achieved.

[0150] The experimental results show that the mAP, parameter amount and computational complexity of the proposed method for crack defect detection are 78.6%, 2.5M and 7.0G respectively. Compared with the original YOLOv8 network model, mAP@0.5 is increased by 2.9%, and the parameter amount and computational complexity are reduced by 16.7% and 13.6% respectively. Compared with YOLOv3-tiny, YOLOv5n, DERT, EfficientNet, MobileVit, MobileNetV3 and other algorithms, the feature reconstruction network model has lower parameter amount and computational complexity while maintaining high accuracy, and the missed detection rate and false detection rate are significantly reduced, which meets the application of practical engineering.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described with reference to the preferred embodiments of the present invention, it should be understood by those skilled in the art that various changes may be made in form and details without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A method for detecting crack defects in wind turbine blades based on feature recombination network, characterized in that: Obtaining an image of a wind turbine blade to be inspected, inputting the image into a pre-trained feature recombination network model, and obtaining a crack defect recognition result of the image of the wind turbine blade to be inspected; 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 perform multi-scale feature extraction on the wind turbine blade image input into 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 fuse the multi-scale features to obtain a fused feature map; the detection head is used to obtain a crack defect recognition result of the wind turbine blade image to be detected according to the fused feature map as the output of the feature reconstruction network model.

2. The method for detecting crack defects in wind turbine blades based on feature recombination network according to claim 1 is characterized in that: In the feature recombination network model: The backbone network layer includes sequentially connected convolutional layers, 4 convolutional fusion units, 1 large kernel selective network and 1 spatial pyramid pooling module; each convolutional fusion unit includes a cascaded convolutional layer and a C2f-FE module; The neck network layer includes two feature reorganization upsampling units and two convolution connection fusion units connected in sequence; wherein the feature reorganization upsampling unit includes a CSEAFE upsampling module, a SimAM attention module, a connection layer and a C2f-FE module connected in sequence; and the convolution connection fusion unit includes a GSConv mixed convolution module, a connection layer and a C2f-FE module connected in sequence; The detection head is used to obtain the crack defect recognition result of the wind turbine blade image according to the output of the neck network layer as the output of the feature reconstruction network model; Among them, the input of the feature recombination network model is used as the input of the backbone network layer; in the backbone network layer, the outputs of the second convolutional fusion unit and the third convolutional fusion unit are also used as the input of the connection layer in the second feature recombination upsampling unit and the connection layer in the first feature recombination upsampling unit in the neck network layer respectively; the output of the spatial pyramid pooling module in the backbone network layer is used as the input of the neck network layer, and also as the input of the connection layer in the second convolutional connection fusion unit in the neck network layer; in the neck network layer, the output of the first feature recombination upsampling unit is also used as the input of the connection layer in the first convolutional connection fusion unit, and the output of the second feature recombination upsampling unit and the outputs of the two convolutional connection fusion units are all output to the detection head.

3. The method for detecting crack defects in wind turbine blades based on feature recombination network according to claim 2 is 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, wherein the two split outputs of the Split module are respectively used as inputs of the first fast convolutional module and the connection layer, the outputs of each fast convolutional module are also used as inputs of the connection layer, and the output of the second convolutional layer is used as the output of the C2f-FE module.

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

5. The method for detecting crack defects in wind turbine blades based on feature recombination network according to claim 2 is characterized in that: The large-core selective network includes a plurality of cascaded large-core selective units; each large-core selective unit includes a first normalization layer, a large-core selection sub-block, a second normalization layer and a feedforward network sub-block connected in sequence, and the output of the large-core selection 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-core selective unit; Among them, the large core selection sub-block includes a cascaded fully connected layer, a GELU activation function layer, an LSK large core 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.

6. The method for detecting crack defects in wind turbine blades based on feature recombination network according to claim 2 is characterized in that: The CSEAFE upsampling module is used to generate an upsampled map X' with dimensions C×σH×σW as output for an input feature map X with dimensions C×H×W, where σ represents the upsampling rate, and each target position l'=(i',j') in the upsampled map X' has a corresponding source position l=(i,j) on the input feature map X, and The processing of the CSEAFE upsampling module includes: First, according to the k centered at different source locations l = (i, j) on the input feature map X up ×k up Sub-area N(X l ,k up ), the kernel prediction module ψ predicts and generates the position-aware kernel W corresponding to different target positions l'=(i',j') in the upsampled map X' l' : W l' ψ(N(X l ,k up )): Among them, ψ(·) represents the kernel prediction module operation, X l Represents the value at position l on the input feature map X, sub-region N(X l ,k up ) is X l K up ×k up Neighborhood; Then, the content-aware reorganization module φ is used to reorganize and calculate the value X at each different target position l'=(i',j') in the upsampled map X' l ”: Where (n,m) represents the number of l ,k up ), Represents sub-area N(X l ,k up ) area radius; W l'＝(n,m) represents the corresponding position-aware kernel at the target position l'=(n,m) in the upsampled map X'; l＝(i+n,j+m) represents the value at position l = (i + n, j + m) on the input feature map X; Thus, an up-sampled map X' is obtained as output.

7. The method for detecting crack defects in wind turbine blades based on feature recombination network according to claim 2 is characterized in that: The processing of the SimAM attention module includes: For an input feature map X of dimension C×H×W, the value x at each position in each channel of the map is c,i Calculate the corresponding energy functions respectively Thus, we get the C×H×W energy function feature map corresponding to the input feature map X. in, Represents the value x of the i-th position on any c-th channel in the input feature map X c,i The corresponding energy function is c = 1, 2, ..., C, i = 1, 2, ..., M, M = W × H; λ is the regularization term; is the variance of the values ​​at each position on each channel of the cth input feature map X, is the mean of the values ​​at each position on the cth channel in the input feature map X: Then, the output feature map X of the SimAM attention module is calculated by the following formula: SimAM : Among them, Sigmoid(·) represents the Sigmiod activation function operation; Θ represents the element-wise multiplication operation.

8. The method for detecting crack defects in wind turbine blades based on feature recombination network according to claim 2 is characterized in that: The processing process of the GSConv mixed convolution module includes: First, the input feature map is convolved and downsampled, and then the output of the convolution downsampling is further processed by deep convolution. After the output of the deep convolution processing is spliced ​​with the output of the convolution downsampling, the spliced ​​result is shuffled to obtain the output feature map.

9. The method for detecting crack defects in wind turbine blades based on feature recombination network according to claim 1, characterized in that: The feature recombination network model is trained in the following way: S101: preparing a wind turbine blade crack defect sample image dataset, which is divided into a training image dataset and a test image dataset, wherein the wind turbine blade crack defect sample images in the wind turbine blade crack defect sample image dataset are all pre-marked with crack defect locations and types; S102: inputting the training image data set into the feature recombination network model for training, and performing parameter optimization training on the feature recombination network model with the goal of minimizing the loss function until the feature recombination network model converges, thereby obtaining a trained feature recombination network model; S103: Test the feature recombination network model through the test image data set to confirm the recognition performance of the trained feature recombination network model; if the recognition performance meets the requirements, end the training of the feature recombination network model; otherwise, return to step S102.

10. The method for detecting crack defects in wind turbine blades based on feature recombination network according to claim 9, characterized in that: In the loss function, the classification loss selects the total loss function composed of the sum of the CIoU loss function and the DFL loss function, and the positioning loss selects the binary cross entropy BCE loss function.

Images