Fan blade surface damage detection method and device based on YOLOv11

Abstract

The invention discloses a fan blade surface damage detection method and device based on YOLOv11, and the method comprises the steps: enabling a high-efficiency adaptive down-sampling module EAD to be embedded in a down-sampling layer of a YOLOv11 network, and carrying out the processing of the down-sampling layers in a backbone network and a neck network; a local cross-stage connection module C3Pk2 based on partial convolution and double convolution acceleration is embedded into a feature extraction layer of the YOLOv11 network for processing feature extraction layers in a backbone network and a neck network, and the feature extraction capability is enhanced through channel shuffling and a channel attention mechanism; after a self-attention module C2PCA based on a partial channel is added to a backbone network of the YOLOv11 network, feature enhancement is carried out by adopting a mode of combining convolution and a self-attention mechanism, and key features of long and short range dependence and highlighted damage are fused to obtain a final surface damage detection model; and inputting the fan blade surface damage image data set into the backbone network of the surface damage detection model, processing the generated three different scales of feature maps through the processed neck network and the head network of the YOLOv11 network, and outputting a damage detection result.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade surface damage detection, and more particularly to a wind turbine blade surface damage detection method based on YOLOv11.

[0002] Methods and apparatus for detecting surface damage. Background Technology

[0003] As a core component of wind power generation, the performance and reliability of wind turbine blades directly affect the power generation efficiency and economic benefits of wind farms. However, due to their long-term exposure to the natural environment, wind turbine blades are susceptible to surface damage from various factors such as wind, sand, rain, and ultraviolet radiation. This damage includes peeling, coating breakage, oil contamination, and cracks. If not detected and repaired promptly, this damage will gradually expand, eventually leading to blade breakage and causing serious safety accidents. The maintenance and replacement costs of wind turbine blades are high; therefore, regular surface damage inspection of the blades can promptly detect and repair damage, preventing further expansion, thereby extending the blade's service life and reducing maintenance costs. By detecting surface damage on wind turbine blades, the condition of the blades can be monitored in a timely manner, providing effective data for the operation and management of wind farms.

[0004] Existing conventional wind turbine blade damage detection technologies mainly include acoustic emission, ultrasonic wave, and infrared thermal imaging. While these methods can effectively detect blade damage, they still have many limitations in practical applications. On the one hand, these methods typically rely on specialized equipment and manual operation, which is not only costly but also inefficient. On the other hand, these methods have poor environmental adaptability and are easily affected by environmental conditions, leading to unstable detection accuracy. These problems make it difficult to achieve ideal results in routine wind turbine blade maintenance.

[0005] Therefore, combining target detection technology with blade surface damage detection helps to detect blade surface damage more efficiently and accurately. By combining computer vision and deep learning algorithms, damage on the blade surface can be automatically identified and located, reducing manual intervention and improving detection efficiency and accuracy. Compared with other conventional methods, target detection algorithms can work stably under various environmental conditions, avoiding interference from factors such as temperature and wind speed. They can also achieve remote monitoring and real-time alarms, avoiding downtime for detection and greatly improving detection efficiency. Summary of the Invention

[0006] This invention provides a method and apparatus for detecting surface damage on wind turbine blades based on YOLOv11. This invention offers higher accuracy in detecting surface damage on wind turbine blades and better meets the needs of monitoring and maintaining the health of wind turbine blades. Details are described below:

[0007] A first aspect is a method for detecting surface damage on wind turbine blades based on YOLOv11, the method comprising:

[0008] The efficient adaptive downsampling module (EAD) is embedded in the downsampling layer of the YOLOv11 network to process the downsampling layers in the backbone and neck network of the YOLOv11 network.

[0009] The local cross-stage connection module C3Pk2, which is accelerated by partial convolution and double convolution, is embedded into the feature extraction layer of the YOLOv11 network. It is used to process the feature extraction layers in the backbone and neck network of the YOLOv11 network, and the feature extraction capability is enhanced by channel shuffling and channel attention mechanisms.

[0010] The C2PCA self-attention module based on partial channels is added after the backbone of the YOLOv11 network. Feature enhancement is performed by combining convolution and self-attention mechanisms, fusing long and short-range dependencies, highlighting the key features of damage, and obtaining the final surface damage detection model.

[0011] The image dataset of surface damage on wind turbine blades is input into the backbone network of the surface damage detection model to generate feature maps of three different scales. The feature maps of the three different scales are then processed by the neck network and the head network built into the YOLOv11 network to output the damage detection results. The detection box, damage type and confidence level are labeled according to the damage detection results.

[0012] The step of embedding the efficient adaptive downsampling module into the downsampling layer of the YOLOv11 network is as follows:

[0013] Divided into upper and lower parts by the segmentation operation, the upper part consists of a thinning layer and a channel shuffling layer. The thinning layer is used to perform differentiated processing on different sub-channels. The lower part consists of two branches, each using a different downsampling method. The left side is the first branch, and the right side is the second branch. The first branch consists of group convolution, and the second branch consists of slicing operations and 1×1 convolution. The results of the first and second branches are summarized in the channel splicing layer as the output.

[0014] The high-efficiency adaptive downsampling module (EAD) is as follows:

[0015] For any given input feature map X∈R C×H×W The first sub-channel uses average pooling to extract global smooth features, the second sub-channel uses max pooling to enhance local features, the third sub-channel combines average pooling and max pooling to take into account both global and local features, and the fourth sub-channel retains the original features. This can be described as follows:

[0016]

[0017]

[0018] in, To divide the feature map X into n equal parts according to the channel dimension, For average pooling, For max pooling, To splice along the channel dimension, It has four channels;

[0019] The processed feature maps of the four channels are rearranged through channels to achieve cross-regional information exchange, and then re-split into two groups along the channel dimension, as described below:

[0020]

[0021]

[0022] in, For mixed washing of channels, This is the input feature map after differential processing and channel shuffling. To make feature maps Divide into two parts on average according to the channel dimension. This serves as the input for the subsequent two parallel branches;

[0023] After splitting, the first branch undergoes lightweight downsampling via group convolution, resulting in a feature map size of [size missing]. W;

[0024] The second branch uses slicing to spatially compress the feature map, resulting in a feature map of size [size missing]. W is then used to integrate cross-channel information association through a 1×1 convolution, resulting in a feature map of size W. W;

[0025] Finally, the outputs of the two branches are concatenated along the channel dimension to form a downsampled feature that combines computational efficiency and information fidelity. This downsampled feature map is then output, and its size is [size missing]. W.

[0026] The local cross-stage connection module C3Pk2, which is accelerated by partial convolution and double convolution, includes: a local cross-stage bottleneck module C3kPS based on partial convolution and channel shuffling, a bottleneck module PSBottleneck based on partial convolution and channel shuffling, and a channel attention module.

[0027] The local cross-stage bottleneck module based on partial convolution and channel shuffling consists of two branches after the segmentation operation. The first branch consists of a convolutional layer, and the second branch consists of convolutional layers and n PSBottleneck layers from top to bottom. Below the two branches are channel splicing layers and convolutional layers. The input feature map is divided into two parts along the channel dimension. One part is processed directly by a convolutional layer in the first branch, and the other part is processed by a series of PSBottleneck modules in the second branch for deeper feature extraction.

[0028] The PSBottleneck module consists of stacked "partial convolution-channel shuffle units" with optional residual connections at the end; the input feature map passes through two PConv-Shuffle units in sequence, the PConv layer only performs convolution on a portion of the channels; the Shuffle layer performs channel shuffle on the feature map;

[0029] At the end of the C3Pk2 module, the feature maps from different stages processed by the PSBottleneck or C3kPS modules are stitched together. The number of channels is adjusted through convolutional layers, the channel attention module weights the fused features, and the processing result of the C3Pk2 module is output.

[0030] The partial-channel self-attention module C2PCA is as follows:

[0031] From top to bottom, it consists of a channel segmentation module, an n-layer PCPA module, a channel concatenation module, and a convolution module. In the PCPA module, for any given input feature map X∈R... C×H×W The feature map is divided into two independent sub-feature maps of the same size through channel segmentation, and then entered into two branches, as described below:

[0032]

[0033] in, To divide the feature map X into n equal parts according to the channel dimension;

[0034] The first branch focuses on fine-grained features in local regions through partial convolution, and can be described as follows:

[0035]

[0036] in, Partial channel convolution;

[0037] The second branch utilizes single-head self-attention only on features. Figure 1 The number of half channels is processed to capture global long-range dependencies, which are expressed as follows:

[0038]

[0039] in, To perform single-head self-attention operation only on a portion of the channels;

[0040] The two-branch features are then concatenated along the channel dimension to restore the original number of channels, and cross-channel information interaction is promoted through channel shuffling operations to enhance feature diversity, as described below:

[0041]

[0042]

[0043] in, To splice along the channel dimension, For mixed washing of channels, This is the output of the C2PCA module.

[0044] In a second aspect, a wind turbine blade surface damage detection device based on YOLOv11 is provided, the device comprising: a processor and a memory, wherein the memory stores program instructions, and the processor invokes the program instructions stored in the memory to cause the device to perform the method described in any of the first aspects.

[0045] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.

[0046] The beneficial effects of the technical solution provided by this invention are:

[0047] 1. Compared with other surface damage detection algorithms, this invention has significant advantages in detection tasks under resource-constrained scenarios. By optimizing feature extraction, reducing the number of model parameters and improving computational efficiency, this invention balances detection accuracy and computational speed, and can still maintain high-precision detection results even under limited computational resources.

[0048] 2. To address the issue of the large number of parameters in the downsampling modules of existing algorithms, this invention proposes an efficient adaptive downsampling module (EAD). This module uses a low-parameter method to differentiate the processing of different channels of the feature map, reducing computational costs while improving the feature representation capability of the model. To address the issue of redundant computation in existing algorithms, this invention proposes a local cross-stage connection module (C3Pk2) based on partial convolution and double convolution acceleration. This module introduces partial channel convolution to reduce redundant computation and enhances feature extraction capability through channel shuffling and channel attention mechanisms.

[0049] 3. To address the problem of high computational overhead in existing self-attention algorithms, this invention proposes a partial channel-based self-attention module (C2PCA). Each layer performs single-head self-attention calculation on only a portion of the channels, followed by channel shuffling, which focuses key feature information at low cost. After stacking multiple layers, it can generate feature representations with deeper meaning and richer gradient information. This invention has demonstrated excellent detection performance on wind turbine blade surface datasets. Attached Figure Description

[0050] Figure 1 The flowchart shows a method for detecting surface damage on wind turbine blades based on YOLOv11.

[0051] Figure 2 This is a schematic diagram of the overall structure of the YOLOv11-based wind turbine blade surface damage detection method;

[0052] Figure 3 This is a schematic diagram of the structure of the high-efficiency adaptive downsampling module (EAD).

[0053] Figure 4 This is a schematic diagram of the structure of the local cross-stage connection module (C3Pk2) accelerated by partial convolution and double convolution;

[0054] Figure 5 This is a schematic diagram of a partial channel-based self-attention module (C2PCA). Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0056] Example 1

[0057] A method for detecting surface damage on wind turbine blades based on YOLOv11, see [link to relevant documentation]. Figure 1 The method includes:

[0058] S1: Collect image dataset of surface damage on wind turbine blades;

[0059] S2: The efficient adaptive downsampling module EAD is embedded into the downsampling layer of the YOLOv11 network to reduce the number of model parameters and computational overhead, while using more diverse features for subsequent feature extraction.

[0060] The existing YOLOv11 network consists of a backbone network, a neck network, and a head network. Both the backbone network and the neck network contain downsampling layers and feature extraction layers.

[0061] Therefore, step S2 processes the downsampling layers in the backbone and neck networks by embedding the Efficient Adaptive Downsampling Module (EAD) into the downsampling layers in the backbone and neck networks to reduce the number of model parameters and computational overhead.

[0062] S3: The local cross-stage connection module C3Pk2, which is accelerated by partial convolution and double convolution, is embedded into the feature extraction layer of the YOLOv11 network. The partial convolution method is used to reduce redundant computation, and the feature extraction capability is enhanced by channel shuffling and channel attention mechanisms.

[0063] Therefore, step S3 processes the feature extraction layers in the backbone and neck networks by using partial convolution to reduce redundant computation.

[0064] S4: The C2PCA self-attention module based on partial channels is added after the backbone of the YOLOv11 network. The feature enhancement is performed by combining convolution and self-attention mechanisms, which integrates long and short-range dependencies and highlights the key features of the damage.

[0065] In particular, through the above steps S2-S4, the YOLOv11 network in the prior art is improved, and the improved backbone network and neck network are obtained. Combined with the head network of the YOLOv11 network in the prior art, the improved surface damage detection model is obtained.

[0066] S5: Input the wind turbine blade surface damage image dataset into the backbone network of the surface damage detection model to generate feature maps of three different scales. The smallest scale feature map is processed by the C2PCA module. Then, the three different scale feature maps are processed by the improved neck network and the head network built into the YOLOv11 network to output the damage detection results. The detection box, damage type and confidence level are labeled according to the damage detection results.

[0067] Among them, the backbone network improved by steps S2-S4 serves as the first part of the surface damage detection model, receiving the wind turbine blade surface damage image dataset and generating feature maps of three different scales.

[0068] In summary, based on the existing target detection model YOLOv11, this invention proposes a high-efficiency surface damage detection model based on partial channel processing, taking into full account the damage characteristics of wind turbine blade surfaces and the task scenario. The overall structure of the model is described in [link to model description]. Figure 2This model employs a novel, highly efficient adaptive downsampling module (EAD), a local cross-stage connection module (C3Pk2) accelerated by partial convolution and double convolution, and a self-attention module (C2PCA) based on partial channels. This approach highlights important feature information while significantly reducing computational overhead and improving the model's ability to correctly identify samples. The surface damage detection model of this invention achieved optimal detection performance on a wind turbine blade surface damage dataset.

[0069] Example 2

[0070] The following section uses specific calculation formulas. Figures 3-5 The solution in Example 1 will be further described below:

[0071] In step S2, the specific method of the efficient adaptive downsampling module EAD is as follows:

[0072] S211: As Figure 3 As shown, the EAD module is divided into two parts, upper and lower, with the segmentation operation as the boundary. The upper part consists of a thinning layer and a channel shuffling layer, where the thinning layer performs differential processing on different sub-channels. The lower part consists of two branches, each employing a different downsampling method; the left side is the first branch, and the right side is the second branch. The first branch consists of group convolutions, and the second branch consists of slicing operations and 1×1 convolutions. Finally, the results from the first and second branches are combined in the channel concatenation layer as the output of the EAD module. For any given input feature map X∈R... C×H×W First, the channel is divided into four sub-channels along the channel dimension. Each sub-channel is processed differently using different strategies: the first sub-channel uses average pooling to extract global smooth features; the second sub-channel uses max pooling to enhance local features; the third sub-channel combines average pooling and max pooling to balance global and local features; and the fourth sub-channel retains the original features to maintain the integrity of details. The specific process can be described as follows:

[0073]

[0074]

[0075] in, To divide the feature map X into n equal parts according to the channel dimension, For average pooling, For max pooling, To splice along the channel dimension, It has four channels.

[0076] S212: The processed feature maps of the four channels are rearranged via channels to achieve cross-regional information interaction, and then re-split into two groups along the channel dimension. The specific process can be described as follows:

[0077]

[0078]

[0079] in, For mixed washing of channels, This is the input feature map after differential processing and channel shuffling. To make feature maps Divide into two parts on average according to the channel dimension. This serves as the input for the subsequent two parallel branches.

[0080] S213: After splitting, the first branch undergoes lightweight downsampling via group convolution, resulting in a feature map size of [size missing]. W represents the number of parameters while preserving the spatial structure, where C is the number of channels, H is the height, and W is the width.

[0081] S214: The second branch uses a slicing operation to spatially compress the feature map, resulting in a feature map of size [size missing]. W is then used to integrate cross-channel information association through a 1×1 convolution, resulting in a feature map of size W. W.

[0082] S215: Finally, the outputs of the two branches are concatenated along the channel dimension to form a downsampled feature that combines computational efficiency and information fidelity. This downsampled feature is then used as the output of the EAD module, and the size of the output feature map is [size missing]. W.

[0083] In step S3, the specific method for the local cross-stage connection module C3Pk2, which is accelerated by partial convolution and double convolution, is as follows:

[0084] S311: As Figure 4 As shown in (a), the C3Pk2 module mainly consists of a convolution module, a separation and splicing module, a local cross-stage bottleneck module (C3kPS) based on partial convolution and channel shuffling, a bottleneck module (PSBottleneck) based on partial convolution and channel shuffling, and a channel attention module.

[0085] S312: As Figure 4As shown in (b), the C3kPS module consists of two branches after the segmentation operation. The first branch consists of a single convolutional layer, and the second branch, from top to bottom, consists of convolutional layers and n PSBottleneck layers, where the size of n is related to the model type. Below the two branches are channel concatenation layers and convolutional layers. In the C3kPS module, the input feature map is evenly divided into two parts along the channel dimension. One part is directly processed by a convolutional layer in the first branch, while the other part undergoes deeper feature extraction through a series of PSBottleneck modules in the second branch.

[0086] S313: As Figure 4 As shown in (c), the PSBottleneck module consists of stacked "dual PConv-Shuffle (partial convolution-channel shuffling) units," with an optional residual connection at the end. The input feature map passes through two PConv-Shuffle units sequentially. The PConv (partial convolution) layer convolves only some channels, reducing redundant computation; the Shuffle (channel shuffling) layer shuffles the feature map by channels. When the Shortcut value is set to True, the feature map after this series of processing is added to the original input feature map to form a residual connection.

[0087] S314: The input features are first processed through convolutional layers. The resulting feature map is then divided into two parts: one part is input into subsequent modules to extract deeper features, and the other part is directly used for subsequent concatenation operations. After feature map segmentation, the module selects different processing paths based on the preset parameter C3k. If C3k is True, the C3kPS module is used; if C3k is False, the PSBottleneck module is used. The C3kPS module or PSBottleneck module is repeated n times to further extract and process features, where n is related to the model type.

[0088] S315: At the end of the C3Pk2 module, feature maps from different stages processed by the PSBottleneck or C3kPS modules are concatenated. Then, convolutional layers adjust the number of channels to enhance feature fusion and effectively integrate information from different paths. Finally, a channel attention module weights the fused features, suppresses redundant information, highlights more valuable features, and outputs the processing result from the C3Pk2 module.

[0089] In step S4, the specific method of C2PCA based on partial channel self-attention module is as follows:

[0090] S411: The C2PCA module is a highly efficient feature extraction architecture combining convolution and self-attention mechanisms. Its core consists of multiple stacked partially convolutional self-attention modules (PCPAs). The number of stacked PCPA modules can be adjusted according to the specific task to enhance the model's ability to capture key features. Figure 5 As shown, the C2PCA module consists of a channel segmentation module, an n-layer PCPA module, a channel concatenation module, and a convolution module from top to bottom. The value of n can be adjusted according to the task requirements; in this embodiment, n is 2. The PCPA module consists of a segmentation module, two parallel branches, a channel concatenation module, and a channel shuffling module from top to bottom. The two parallel branches are the first branch and the second branch from left to right. The first branch consists of partial convolutions, and the second branch consists of single-head self-attention.

[0091] S412: In the PCPA module, for any given input feature map X∈R C×H×W First, the feature map is divided into two independent sub-feature maps of the same size through channel segmentation, and then each sub-map enters one of the two branches. The specific process can be described as follows:

[0092]

[0093] in, To divide the feature map X into n equal parts according to the channel dimension.

[0094] S413: The first branch focuses on fine-grained features in local regions through partial convolution. The specific process can be described as follows:

[0095]

[0096] in, This is a partial channel convolution.

[0097] S414: The second branch utilizes single-head self-attention only on features. Figure 1 Half-channel processing is used to capture global long-range dependencies. The specific process can be described as follows:

[0098]

[0099] in, This is to perform single-head self-attention operation only on a portion of the channels.

[0100] S415: The two-branch features are then concatenated along the channel dimension to restore the original number of channels, and cross-channel information interaction is promoted through channel shuffling to enhance feature diversity. The specific process can be described as follows:

[0101]

[0102]

[0103] in, To splice along the channel dimension, For mixed washing of channels, This is the output of the C2PCA module.

[0104] In summary, the embodiments of the present invention provide faster and more accurate detection of surface damage on wind turbine blades, while reducing computational overhead, thus better meeting the needs of monitoring and maintaining the health of wind turbine blades in industrial production.

[0105] Example 3

[0106] To comprehensively evaluate the accuracy of the model, this embodiment of the invention uses the most commonly used metrics such as precision (P), recall (R), average precision (AP), and mean average precision (mAP). Average precision (AP) is the average of the precision under different recall rates and is generally used to evaluate the detection accuracy of a certain class. Mean average precision (mAP) is the average detection precision of all target classes and is usually used to evaluate the overall performance of the network model.

[0107] The above indicators are defined as follows:

[0108]

[0109]

[0110]

[0111]

[0112] in, FP is the number of positive samples correctly identified by the model; FP is the number of negative samples that the model classifies as positive; FN is the number of positive samples that the model classifies as negative. Precision is expressed as a function of recall, where n represents the number of categories. This represents the average detection accuracy for the i-th class. Furthermore, to evaluate the model's size and computational complexity, the number of parameters and GFLOPS (1 billion floating-point operations per second) are used to reflect the model's size and computational load. Larger parameter counts and GFLOPS indicate greater hardware resource consumption and higher configuration requirements for practical applications.

[0113] To illustrate the accuracy of the method, the prediction results were compared with those of other models on a wind turbine blade surface damage dataset, resulting in the model prediction results shown in Table 1. The last row presents the experimental results of the model proposed in this embodiment, while the others show experimental results of some common and recent network models. mAP50 is typically used to evaluate the overall performance of a network model. It can be seen that the model proposed in this embodiment has a significant advantage in overall performance, achieving a 6.7% improvement compared to YOLOv11n, while also having the lowest parameter count and computational complexity, making it more suitable for real-world wind turbine blade surface damage detection. EP-YOLO improves detection speed by reducing redundant computations and simplifying the downsampling module, and enhances detection accuracy by optimizing feature extraction, significantly improving the overall performance of the model while maintaining its lightweight nature.

[0114]

[0115] To better illustrate the contribution of each improvement to the performance of the model proposed in this embodiment, Table 2 focuses on the three modules EAD, C2PCA, and C3Pk2, discussing their independent and combined performance enhancements. It can be seen that both individual modules and combinations of multiple modules improve mAP50 to varying degrees, fully demonstrating the effectiveness of each module proposed in this embodiment in enhancing model performance.

[0116] Table 2 Ablation experiment results on the wind turbine blade surface damage dataset.

[0117]

[0118] Example 4

[0119] A wind turbine blade surface damage detection device based on YOLOv11, the device includes: a processor and a memory, the memory storing program instructions, and the processor calling the program instructions stored in the memory to cause the device to execute the following method steps in Embodiment 1:

[0120] The efficient adaptive downsampling module (EAD) is embedded in the downsampling layer of the YOLOv11 network to process the downsampling layers in the backbone and neck network of the YOLOv11 network.

[0121] The local cross-stage connection module C3Pk2, which is accelerated by partial convolution and double convolution, is embedded into the feature extraction layer of the YOLOv11 network. It is used to process the feature extraction layers in the backbone and neck network of the YOLOv11 network, and the feature extraction capability is enhanced by channel shuffling and channel attention mechanisms.

[0122] The C2PCA self-attention module based on partial channels is added after the backbone of the YOLOv11 network. Feature enhancement is performed by combining convolution and self-attention mechanisms, fusing long and short-range dependencies, highlighting the key features of damage, and obtaining the final surface damage detection model.

[0123] The image dataset of surface damage on wind turbine blades is input into the backbone network of the surface damage detection model to generate feature maps of three different scales. The feature maps of the three different scales are then processed by the neck network and the head network built into the YOLOv11 network to output the damage detection results. The detection box, damage type and confidence level are labeled according to the damage detection results.

[0124] The step of embedding the efficient adaptive downsampling module into the downsampling layer of the YOLOv11 network is as follows:

[0125] Divided into upper and lower parts by the segmentation operation, the upper part consists of a thinning layer and a channel shuffling layer. The thinning layer is used to perform differentiated processing on different sub-channels. The lower part consists of two branches, each using a different downsampling method. The left side is the first branch, and the right side is the second branch. The first branch consists of group convolution, and the second branch consists of slicing operations and 1×1 convolution. The results of the first and second branches are summarized in the channel splicing layer as the output.

[0126] The high-efficiency adaptive downsampling module (EAD) is as follows:

[0127] For any given input feature map X∈R C×H×W The first sub-channel uses average pooling to extract global smooth features, the second sub-channel uses max pooling to enhance local features, the third sub-channel combines average pooling and max pooling to take into account both global and local features, and the fourth sub-channel retains the original features. This can be described as follows:

[0128]

[0129]

[0130] in, To divide the feature map X into n equal parts according to the channel dimension, For average pooling, For max pooling, To splice along the channel dimension, It has four channels;

[0131] The processed feature maps of the four channels are rearranged through channels to achieve cross-regional information exchange, and then re-split into two groups along the channel dimension, as described below:

[0132]

[0133]

[0134] in, For mixed washing of channels, This is the input feature map after differential processing and channel shuffling. To make feature maps Divide into two parts on average according to the channel dimension. This serves as the input for the subsequent two parallel branches;

[0135] After splitting, the first branch undergoes lightweight downsampling via group convolution, resulting in a feature map size of [size missing]. W;

[0136] The second branch uses slicing to spatially compress the feature map, resulting in a feature map of size [size missing]. W is then used to integrate cross-channel information association through a 1×1 convolution, resulting in a feature map of size W. W;

[0137] Finally, the outputs of the two branches are concatenated along the channel dimension to form a downsampled feature that combines computational efficiency and information fidelity. This downsampled feature map is then output, and its size is [size missing]. W.

[0138] The local cross-stage connection module C3Pk2, which is accelerated by partial convolution and double convolution, includes: a local cross-stage bottleneck module C3kPS based on partial convolution and channel shuffling, a bottleneck module PSBottleneck based on partial convolution and channel shuffling, and a channel attention module.

[0139] The local cross-stage bottleneck module based on partial convolution and channel shuffling consists of two branches after the segmentation operation. The first branch consists of a convolutional layer, and the second branch consists of convolutional layers and n PSBottleneck layers from top to bottom. Below the two branches are channel splicing layers and convolutional layers. The input feature map is divided into two parts along the channel dimension. One part is processed directly by a convolutional layer in the first branch, and the other part is processed by a series of PSBottleneck modules in the second branch for deeper feature extraction.

[0140] The PSBottleneck module consists of stacked "partial convolution-channel shuffle units" with optional residual connections at the end; the input feature map passes through two PConv-Shuffle units in sequence, the PConv layer only performs convolution on a portion of the channels; the Shuffle layer performs channel shuffle on the feature map;

[0141] At the end of the C3Pk2 module, the feature maps from different stages processed by the PSBottleneck or C3kPS modules are stitched together. The number of channels is adjusted through convolutional layers, the channel attention module weights the fused features, and the processing result of the C3Pk2 module is output.

[0142] Among them, the self-attention module C2PCA based on partial channels is:

[0143] From top to bottom, it consists of a channel segmentation module, an n-layer PCPA module, a channel concatenation module, and a convolution module. In the PCPA module, for any given input feature map X∈R... C×H×W The feature map is divided into two independent sub-feature maps of the same size through channel segmentation, and then entered into two branches, as described below:

[0144]

[0145] in, To divide the feature map X into n equal parts according to the channel dimension;

[0146] The first branch focuses on fine-grained features in local regions through partial convolution, and can be described as follows:

[0147]

[0148] in, Partial channel convolution;

[0149] The second branch utilizes single-head self-attention only on features. Figure 1 The number of half channels is processed to capture global long-range dependencies, which are expressed as follows:

[0150]

[0151] in, To perform single-head self-attention operation only on a portion of the channels;

[0152] The two-branch features are then concatenated along the channel dimension to restore the original number of channels, and cross-channel information interaction is promoted through channel shuffling operations to enhance feature diversity, as described below:

[0153]

[0154]

[0155] in, To splice along the channel dimension, For mixed washing of channels, This is the output of the C2PCA module.

[0156] In summary, the embodiments of the present invention provide faster and more accurate detection of surface damage on wind turbine blades, while reducing computational overhead, thus better meeting the needs of monitoring and maintaining the health of wind turbine blades in industrial production.

[0157] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0158] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0159] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0160] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0161] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0162] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0163] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0164] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0165] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0166] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0167] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting surface damage on wind turbine blades based on YOLOv11, characterized in that, The method includes: The efficient adaptive downsampling module (EAD) is embedded in the downsampling layer of the YOLOv11 network to process the downsampling layers in the backbone and neck network of the YOLOv11 network. The local cross-stage connection module C3Pk2, which is accelerated by partial convolution and double convolution, is embedded into the feature extraction layer of the YOLOv11 network. It is used to process the feature extraction layers in the backbone and neck network of the YOLOv11 network, and the feature extraction capability is enhanced by channel shuffling and channel attention mechanisms. The C2PCA self-attention module based on partial channels is added after the backbone of the YOLOv11 network. Feature enhancement is performed by combining convolution and self-attention mechanisms, fusing long and short-range dependencies, highlighting the key features of damage, and obtaining the final surface damage detection model. The image dataset of surface damage on wind turbine blades is input into the backbone network of the surface damage detection model to generate feature maps of three different scales. The feature maps of the three different scales are then processed by the neck network and the head network built into the YOLOv11 network to output the damage detection results. The detection box, damage type and confidence level are labeled according to the damage detection results.

2. The method for detecting surface damage of wind turbine blades based on YOLOv11 according to claim 1, characterized in that, The step of embedding the high-efficiency adaptive downsampling module into the downsampling layer of the YOLOv11 network is as follows: Divided into two parts by the segmentation operation, the upper part consists of a refinement layer and a channel washing layer. The refinement layer is used to perform differentiated processing on different sub-channels. The lower part consists of two branches, each using a different downsampling method. The left side is the first branch, and the right side is the second branch. The first branch consists of group convolutions, and the second branch consists of slicing operations and 1×1 convolutions; the results of the first and second branches are aggregated in the channel splicing layer as the output.

3. The method for detecting surface damage of wind turbine blades based on YOLOv11 according to claim 2, characterized in that, The high-efficiency adaptive downsampling module is: For any given input feature map X∈R C×H×W The first sub-channel uses average pooling to extract global smooth features, the second sub-channel uses max pooling to enhance local features, the third sub-channel combines average pooling and max pooling to take into account both global and local features, and the fourth sub-channel retains the original features. This can be described as follows: ； ； in, To divide the feature map X into n equal parts according to the channel dimension, For average pooling, For max pooling, To splice along the channel dimension, It has four channels; The processed feature maps of the four channels are rearranged through channels to achieve cross-regional information exchange, and then re-split into two groups along the channel dimension, as described below: ； ； in, For mixed washing in the channel, This is the input feature map after differential processing and channel shuffling. To make feature maps Divide into two parts on average according to the channel dimension. This serves as the input for the subsequent two parallel branches; After splitting, the first branch undergoes lightweight downsampling via group convolution, resulting in a feature map size of [size missing]. W; The second branch uses slicing to spatially compress the feature map, resulting in a feature map of size [size missing]. W is then used to integrate cross-channel information association through a 1×1 convolution, resulting in a feature map of size W. W; Finally, the outputs of the two branches are concatenated along the channel dimension to form a downsampled feature that combines computational efficiency and information fidelity. This downsampled feature map is then output, and its size is [size missing]. W.

4. The method for detecting surface damage of wind turbine blades based on YOLOv11 according to claim 2, characterized in that, The local cross-stage connection module C3Pk2, which is accelerated by partial convolution and double convolution, includes: a local cross-stage bottleneck module C3kPS based on partial convolution and channel shuffling, a bottleneck module PSBottleneck based on partial convolution and channel shuffling, and a channel attention module. The local cross-stage bottleneck module based on partial convolution and channel shuffling consists of two branches after the segmentation operation. The first branch consists of a convolutional layer, and the second branch consists of convolutional layers and n PSBottleneck layers from top to bottom. Below the two branches are channel splicing layers and convolutional layers. The input feature map is divided into two parts along the channel dimension. One part is processed directly by a convolutional layer in the first branch, and the other part is processed by a series of PSBottleneck modules in the second branch for deeper feature extraction. The PSBottleneck module consists of stacked "partial convolution-channel shuffle units" with optional residual connections at the end; the input feature map passes through two PConv-Shuffle units in sequence, the PConv layer only performs convolution on a portion of the channels; the Shuffle layer performs channel shuffle on the feature map; At the end of the C3Pk2 module, the feature maps from different stages processed by the PSBottleneck or C3kPS modules are stitched together. The number of channels is adjusted through convolutional layers, the channel attention module weights the fused features, and the processing result of the C3Pk2 module is output.

5. The method for detecting surface damage of wind turbine blades based on YOLOv11 according to claim 1, characterized in that, The C2PCA self-attention module based on partial channels is: From top to bottom, it consists of a channel segmentation module, an n-layer PCPA module, a channel concatenation module, and a convolution module. In the PCPA module, for any given input feature map X∈R... C×H×W The feature map is divided into two independent sub-feature maps of the same size through channel segmentation, and then entered into two branches, as described below: ； in, To divide the feature map X into n equal parts according to the channel dimension; The first branch focuses on fine-grained features in local regions through partial convolution, and can be described as follows: ； in, Partial channel convolution; The second branch utilizes single-head self-attention to process only half of the feature map's channels, capturing global long-range dependencies, expressed as: ； in, To perform single-head self-attention operation only on a portion of the channels; The two-branch features are then concatenated along the channel dimension to restore the original number of channels, and cross-channel information interaction is promoted through channel shuffling operations to enhance feature diversity, as described below: ； ； in, To splice along the channel dimension, For mixed washing of channels, This is the output of the C2PCA module.

6. A wind turbine blade surface damage detection device based on YOLOv11, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-5.

Images