Wind power blade defect detection method and device based on improved YOLO11n

Abstract

The application relates to a wind power blade defect detection method and device based on an improved YOLO11n, which comprises the following steps: training a wind power blade defect detection model by using a wind power blade defect data set; inputting image data of a wind power unit blade to be detected into the trained wind power blade defect detection model, and outputting a wind power blade defect detection result, including a defect position and a defect category; the wind power blade defect detection model is based on an improved YOLO11n model, wherein the improvement comprises: using an IDC module to improve a C3k2 module in a YOLO11n backbone network and a neck network, and forming a C3k2_IDC module. Compared with the prior art, the application has the advantages of improving the precision and efficiency of wind power blade defect detection without increasing a large amount of calculation cost.

Description

Technical Field

[0001] This invention relates to the field of automatic detection technology for power systems, and in particular to a method and device for detecting defects in wind turbine blades based on an improved YOLO11n. Background Technology

[0002] With the continuous development of the wind power industry, the condition monitoring and maintenance of wind turbine units has become a focus of industry attention. Among the components of a wind turbine unit, the health of the blades directly affects the power generation efficiency and equipment safety. Research shows that blade failure is one of the main causes of wind turbine unit malfunctions, and each blade failure requires more than 10 days to repair, resulting in serious economic losses. Therefore, developing efficient blade defect detection technology is of great significance for the safe operation of wind farms.

[0003] Traditional blade defect detection methods, such as manual inspections, involve maintenance personnel using suspended platforms for telescope observation and close-range checks. These methods suffer from low accuracy, slow speed, and high risk. Emerging technologies based on ultrasonic, acoustic emission, and fiber optic sensors have also been used for blade defect detection; however, these technologies also have limitations, such as long detection cycles, sensitivity to environmental factors and susceptibility to interference, and large scanning intervals, all of which can affect the detection results. In recent years, with the rise of deep learning in industrial vision, using deep learning for defect detection has become a powerful and effective method. Among them, the YOLO series models, with their advantages of high speed and detection accuracy, have achieved significant success in the field of target detection. In the field of wind turbine blade defect detection, the YOLO model can also leverage its advantages to accurately locate and distinguish various blade defects, and has now become the mainstream choice.

[0004] YOLO11n is the latest model in the official YOLO series, an improvement upon previous YOLO versions, further enhancing the performance and flexibility of YOLO models. However, YOLO11n also has some shortcomings in blade defect detection. The scale of defects varies significantly across different blades, making it difficult for YOLO11n to handle defects with multi-scale variations, resulting in poor detection results. Furthermore, when detecting blade images with complex backgrounds, the model's insufficient receptive field leads to inadequate feature acquisition and susceptibility to false positives due to background interference.

[0005] A search revealed Chinese invention patent application publication number CN120031861A, which discloses an optimization and detection method for a substation defect detection model based on YOLOv11. The method includes: improving YOLOv11 to obtain an improved model for substation defect detection, comprising a backbone network, a neck network, and a detection network; the backbone network is obtained by replacing the downsampling convolutional modules in the YOLOv11 backbone network with Adown downsampling modules; the neck network is obtained by replacing the C3K2 convolutional blocks in the YOLOv11 neck network with a lightweight semantic segmentation module ContextGuide; and the detection network is obtained by introducing an auxiliary detection module AUX before the detection head of the YOLOv11. Substation defect sample data is acquired, and the improved model is trained using the sample data to obtain the substation defect detection model. This existing patent application suffers from inaccurate detection of small target defects. Although it balances detection accuracy and model lightweighting, it cannot effectively handle interference from complex backgrounds and still has certain limitations in multi-target detection. Detection accuracy needs further improvement.

[0006] How to improve the accuracy of wind turbine blade defect detection based on YOLO11n has become a technical problem that needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a wind turbine blade defect detection method and device based on the improved YOLO11n.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] According to one aspect of the present invention, a method for detecting defects in wind turbine blades based on an improved YOLO11n is provided, the method comprising:

[0010] The wind turbine blade defect detection model was trained using a wind turbine blade defect dataset.

[0011] Input the image data of the wind turbine blade to be tested into the trained wind turbine blade defect detection model, and output the wind turbine blade defect detection results, including the defect location and defect category.

[0012] The wind turbine blade defect detection model is based on an improved YOLO11n model. The improvements include: using the IDC module to improve the C3k2 module in the YOLO11n backbone network and neck network to form the C3k2_IDC module.

[0013] Preferably, the C3k2_IDC module includes a first CBS module, a segmentation module, multiple C3k_IDC modules, a splicing module, and a second CBS module connected in sequence.

[0014] The feature map input to the C3k2_IDC module is first subjected to preliminary feature extraction and channel number adjustment by the first CBS module. Then, it is divided into four branches by the segmentation module. The four branches are input to different parallel branches and processed by the C3k_IDC module. Finally, the outputs of each branch are concatenated and fused. The fused features are further subjected to convolutional compression, integration and nonlinear transformation by the second CBS module, and the number of channels is adjusted to the expected output dimension.

[0015] More preferably, the C3k_IDC module is an improvement on the original C3k module, that is, the Bottleneck module in the original C3k module is replaced with the Bottleneck_IDC module.

[0016] More preferably, the Bottleneck_IDC module includes a third CBS module, an IDC module, and a fourth CBS module connected in sequence. The input data is processed by the third CBS module, the IDC module, and the fourth CBS module in sequence, and the output result is added to the input data as the final output.

[0017] More preferably, in the four branches, the first branch uses a 3×3 small kernel convolution to process local features and preserve detailed information; the second and third branches use 1×11 and 11×1 convolution kernels respectively to perform convolution operations in the width and height directions to capture global information in the width and height directions; the fourth branch is an identity mapping that does not perform any convolution operations and directly passes part of the input channels to the next layer.

[0018] Preferably, the improvement further includes adding an ASFF module in front of the detection head of the original YOLO11n model to form a Detect_ASFF module for multi-scale target detection.

[0019] More preferably, the ASFF module includes two parts: feature rescaling and adaptive fusion. In the feature rescaling part, ASFF first extracts feature maps at different levels in the network and upsamples or downsamples the feature maps at different levels to make them have the same resolution and number of channels. After feature rescaling, the ASFF module then performs adaptive fusion on the multi-scale features.

[0020] Preferably, the process of constructing the wind turbine blade defect dataset includes:

[0021] The publicly available wind turbine blade image data is processed to obtain an initial wind turbine blade defect dataset;

[0022] The initial wind turbine blade defect dataset was labeled using annotation software, and then classified according to different blade defects to obtain the final wind turbine blade defect dataset.

[0023] Preferably, the defect categories include cracks, corrosion, edge corrosion, paint peeling, and surface damage.

[0024] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1) This invention is based on an improved YOLO11n wind turbine blade defect detection method. By introducing an IDC module to design the C3k2_IDC module, the C3k2 module of the original YOLO11n model's backbone and neck network is replaced. Through the effective combination of width (IDC) and depth (C3k2), the receptive field of the model is improved without increasing the computational cost significantly. This allows the model to acquire richer feature information and effectively reduces the interference of background factors on the blade defect detection results. As a result, the efficiency of wind turbine blade defect detection and the detection accuracy in complex scenarios are improved. This provides important support for the maintenance of wind turbine blade failures, enhances the operation and maintenance capabilities of wind farms, and reduces the economic losses caused by wind turbine blade failures.

[0027] 2) This invention adds an ASFF module before the detection head to process the feature map output by the neck network, which can significantly enhance the model's multi-scale target detection capability, enabling the model to better detect small targets in complex scenes, thereby improving the detection accuracy of blade defects. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the original YOLO11n network structure;

[0029] Figure 2 This is a schematic diagram of the network structure of C3k2 in this invention;

[0030] Figure 3 This is a schematic diagram of the IDC module in this invention;

[0031] Figure 4 This is a flowchart of the ASFF module in this invention;

[0032] Figure 5 This is a schematic diagram of the improved YOLO11n network structure in this invention;

[0033] Figure 6This is a schematic diagram of the structure of the C3k2_IDC module in this invention;

[0034] Figure 7 This is a flowchart illustrating the wind turbine blade defect detection method of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0036] Example 1

[0037] This embodiment relates to a wind turbine blade defect detection method based on an improved YOLO11n, which aims to solve the problems of existing detection technologies being unable to adapt to the defect characteristics of blades at different scales and being susceptible to background interference by improving the YOLO11n model.

[0038] Using YOLO11n as the base model, the C3k2 module in the YOLO11n backbone and neck network is improved using the Inception Depthwise Convolution (IDC) module, forming the C3k2_IDC module. This module improves the network's detection capability without significantly increasing computational cost. The Detect_ASFF module is improved using the Adaptively Spatial Feature Fusion (ASFF) module, aiming to enhance the model's multi-scale defect detection capability. Details are as follows.

[0039] The network structure diagram of YOLO11n is as follows: Figure 1 As shown, the YOLO11n network structure includes a backbone, a neck, and a detector. Compared to older YOLO models, YOLO11n introduces the C3k2 module, which is derived from the C2f module. YOLO11n sets a parameter `c3k` in the code to determine the structure of the C3k2 module. When `c3k` is True, it's a normal C3k2; when `c3k` is False, the C3k module in C3k2 is replaced by the Bottleneck, and C3k2 is effectively replaced by C2f. A schematic diagram of the structures of C3k2, C2f, C3k, and the Bottleneck is shown below. Figure 2 As shown.

[0040] The network structure of the IDC module is as follows: Figure 3 As shown, the IDC module decomposes large kernel convolutions into multiple parallel branches, achieving a larger receptive field and retaining richer feature information while avoiding the high computational overhead of directly using large kernel convolutions. The IDC module consists of three parts: first, the input tensor is divided into four groups along the channel dimension; then, the four groups are input into different parallel branches; and finally, the outputs of each branch are concatenated.

[0041] Specifically, for input X, the IDC module divides it into four groups along the channel dimension:

[0042] X hw ,X w ,X h ,X id =Split(X)

[0043] =X :,:g ,X :,g:2g ,X :,2g:3g ,X :,3g: ;

[0044] g = r g C;

[0045] Where X is the input tensor, g is the number of branch channels, and a ratio r is set. g To determine, C is the number of channels for input X, and X... hw ,X w ,X h ,X id These are the four outputs after being divided by channel.

[0046] The partitioned data is then fed into four parallel branches for processing, such as... Figure 3 The first branch uses a 3×3 small kernel convolution to process local features and preserve detailed information; the second and third branches use 1×11 and 11×1 convolution kernels respectively, performing convolution operations in the width and height directions, which can efficiently capture global information in the width and height directions; the fourth branch is an identity mapping that does not perform any convolution operations and directly passes part of the input channels to the next layer, thereby reducing unnecessary calculations and improving computational efficiency.

[0047]

[0048] X ′ id =X id ;

[0049] Where, k s kb This is the size of the convolution kernel, with default values ​​of 3 and 11 respectively; This represents a k×k depthwise convolution on X, with both input and output channels being g; X ′ hw X ′ w X ′ h X ′ id These are the outputs after branch processing.

[0050] Finally, the outputs of each branch are concatenated to obtain the final output result X. ′ .

[0051] X′=Concat(X′ hw ,X′ w ,X′ h ,X′ id );

[0052] Concat() represents a concatenation operation on the channel dimension.

[0053] The C3k2 module of YOLO11n is improved using the IDC module, forming the C3k2_IDC module. In the C3k2_IDC module, the C3k_IDC module replaces the original C3k module in the C3k2 module, which improves model performance with minimal increase in computational overhead. A schematic diagram of the C3k2_IDC network structure is shown below. Figure 6 As shown. Similarly, the C3k_IDC module is an improvement on the original C3k module, that is, the Bottleneck module in the original C3k module is replaced with the Bottleneck_IDC module.

[0054] like Figure 6 The C3k2_IDC module includes a first CBS module CBS-1, a splitting module Split, multiple C3k_IDC modules, a splicing module Concat, and a second CBS module CBS-2, which are connected in sequence.

[0055] The feature map input to the C3k2_IDC module first undergoes preliminary feature extraction and channel number adjustment in the first CBS module to prepare for subsequent branch processing. Then, it is divided into four branches by the segmentation module. The four branches are input to different parallel branches and processed by the C3k_IDC module. Finally, the outputs of each branch are concatenated and fused. The fused features are further processed by the second CBS module through convolutional compression, integration, and nonlinear transformation to extract more effective and compact feature representations, and the number of channels is adjusted to the expected output dimension.

[0056] like Figure 6 The Bottleneck_IDC module is an improvement upon the original Bottleneck module, replacing the CBS module with the IDC module. The Bottleneck_IDC module consists of a first CBS module (CBS-1), an IDC module, and a second CBS module (CBS-2), connected sequentially. Input data is processed sequentially by the first CBS module, the IDC module, and the second CBS module. The output result is then added to the input data to obtain the final output.

[0057] The network structure of the ASFF module is as follows: Figure 4 As shown, it mainly consists of two parts: feature rescaling and adaptive fusion. In the feature rescaling part, ASFF first extracts feature maps from different levels of the network. Since the obtained feature maps have different resolutions and different numbers of channels, ASFF upsamples or downsamples the feature maps of different levels to make them have the same resolution and number of channels. For upscaling, a 1×1 convolutional layer is first used to compress the number of feature channels, and then the resolution is increased by interpolation. For downscaling at a ratio of 1 / 2, a 3×3 convolution with a stride of 2 is used to modify both the number of channels and the resolution. For downscaling at a ratio of 1 / 4, a 3×3 max pooling layer is added before the 3×3 convolution. Taking Level 3 as an example, a 1x1 convolution is performed on the feature map of Level 2, and then it is scaled to twice the resolution of the original image to obtain X. l2→l3 Perform a 1x1 convolution on the Level 1 feature map and scale it to four times the original image resolution to obtain X. l1→l3 After feature rescaling, the ASFF module then performs adaptive multi-feature fusion. Taking the fusion of Level 3 feature maps as an example, the ASFF module achieves multi-scale feature fusion by combining Level 1 and Level 2 feature maps. Ultimately, the fused Level 3 features... The calculation is as follows:

[0058]

[0059] in, and Let be the feature vectors at position (i,j) from the feature maps of Level 1, Level 2, and Level 3 after feature rescaling. and These are the spatial importance weights of Level 3, adaptively learned by the network, and satisfy the following:

[0060]

[0061] Improve the detection head of the YOLO11n model: Add an ASFF module before the detection head of the original YOLO11n model to form the Detect_ASFF module. Process the feature map output by the neck network, which can significantly improve the multi-scale target detection capability of the model, enabling the model to better detect small targets in complex scenes, thereby improving the overall detection performance.

[0062] Improved YOLO11n model: Replace the C3k2 module in the original YOLO11n model with the improved C3k2_IDC module, and replace the Detect head with the improved Detect_ASFF, resulting in the improved YOLO11n model. For example... Figure 5 As shown, the network architecture of the improved YOLO11n model includes a backbone network, a neck network, and a detection head (Detect_ASFF). The backbone network is responsible for feature extraction; the neck network, located between the backbone network and the detection head, performs feature fusion and enhancement; and the detection head is the decision-making part of the improved YOLO11n model, responsible for generating the final detection results.

[0063] A wind turbine blade defect detection method based on the improved YOLO11n, such as Figure 7 This includes the following steps:

[0064] S1, Construct a wind turbine blade defect dataset:

[0065] Based on publicly available wind turbine blade image data, an initial wind turbine blade defect dataset was obtained by processing the data. The initial wind turbine blade defect dataset was then labeled using the annotation software X-AnyLabeling, and classified according to different blade defects to obtain the final wind turbine blade defect dataset. The wind turbine blade defects can be categorized into the following five types: crack, erosion, LE-Erosion, paint peel off, and surface damage.

[0066] S2, Construct a wind turbine blade defect detection model based on the improved YOLO11n:

[0067] Using the YOLO11n model as the base model, the original YOLO11n C3k2 module was replaced with the C3k2_IDC module (an improvement of the IDC module), and the original YOLO11n Detect module was replaced with the Detect_ASFF module (an improvement of ASFF). The final result is the improved YOLO11n model. A schematic diagram of the improved YOLO11n model is shown below. Figure 5 As shown.

[0068] S3, Training the wind turbine blade defect detection model:

[0069] This example was trained on an Ubuntu 20.04.6 operating system, with an Intel(R) Xeon(R) Silver4214R CPU@2.40GHz, 128GB of memory, an NVIDIA GeForce RTX 3090 graphics card, Python version 3.12.7, PyTorch version 2.6.0, and CUDA version 12.4.

[0070] The training parameters in this embodiment are set as follows: 300 training iterations, batch size of 8, AdamW optimizer, initial learning rate of 0.001111, momentum of 0.9, number of threads of 4, weight decay of 0.0005, and image size of 640×640.

[0071] The specific training process is as follows: The constructed wind turbine blade defect dataset is divided into training, validation, and test sets in a 3:1:1 ratio. First, the improved YOLO11n model is trained using the constructed training set. During training, the model weights are adjusted using backpropagation and gradient descent to optimize model performance. After each training cycle, the model is evaluated on the validation set, and then the AdamW optimizer is used to tune the model's hyperparameters. After training, the optimal model weights are saved for testing.

[0072] S4, Testing of the wind turbine blade defect detection model:

[0073] The improved YOLO11n model was tested using the above test set. The saved optimal model weights were used to output the wind turbine blade defect detection results, including the defect location and defect category.

[0074] To accurately evaluate the performance of the improved YOLO11n model, several detection performance metrics were used, including precision (P), recall (R), mean average precision (mAP50, mAP50-95), F1 score (F1), GFLOPS, frames per second (Fps), and number of parameters. The formulas for calculating the evaluation metrics are as follows:

[0075]

[0076] Where TP stands for True Positive Instance, representing the number of positive samples correctly detected. FP stands for False Positive Instance, representing the number of negative samples incorrectly detected as positive samples. FN stands for False Negative Instance, representing the number of positive samples incorrectly detected as negative samples. mAP stands for Mean Precision, where m is the total number of classes; AP iis the average precision of class i; mAP50 refers to the average precision when the IOU threshold is 0.5, and mAP50-95 refers to the average precision when the IOU threshold is between 0.5 and 0.95. GFLOPS represents the number of floating-point operations per second. Pre, Infer, and Post represent the preprocessing time, inference time, and postprocessing time, respectively.

[0077] Example 2

[0078] This embodiment also relates to the verification of a wind turbine blade defect detection method based on the improved YOLO11n.

[0079] To verify the effectiveness of the improved YOLO11n algorithm in wind turbine blade defect detection, the improved YOLO11n algorithm was compared with YOLOv9, YOLOv10, YOLO11n, and YOLO11s under the same experimental configuration. The comparison results are shown in Table 1.

[0080] Table 1

[0081] algorithm P R mAP50 mAP50-95 F1 Fps GFLOPS Params(M) YOLOv9t 0.878 0.825 0.871 0.47 0.846 100 7.6 1.97 YOLOv10n 0.825 0.653 0.808 0.456 0.69 133 6.5 2.26 YOLO11n 0.869 0.778 0.862 0.452 0.799 113 6.3 2.58 YOLO11s 0.863 0.869 0.909 0.514 0.857 101 21.3 9.41 This invention 0.895 0.88 0.931 0.528 0.883 107 9.8 4.27

[0082] As shown in Table 1, the present invention outperforms other algorithms in terms of precision, recall, mean precision, and F1 score. Compared to the original YOLO11n, the present invention improves precision by 2.6 percentage points. In terms of mAP50 and mAP50-95, the present invention improves by 6.9 and 7.6 percentage points, respectively, compared to the original YOLO11n. The present invention also significantly outperforms the original YOLO11n model in recall (R) and F1 score. Although the present invention is slightly inferior to the original YOLO11n model in Fps, GFLOPS, and parameter count, the performance of the present invention in Fps, GFLOPS, and parameter count is within an acceptable range because the requirements for detection precision and its importance in wind turbine blade defect detection far outweigh these three metrics.

[0083] To better evaluate the effectiveness of the two improved modules in the improved YOLO11n algorithm, ablation experiments were conducted on the test set for the two improved modules. The experimental results are shown in Table 2:

[0084] Table 2

[0085]

[0086]

[0087] Table 2 shows that, firstly, using ASFF alone to improve the detection head, compared with the original YOLO11n model, all three parameters (precision, mAP, and F1 score) are improved, but the recall rate is slightly decreased. Secondly, using IDC alone to improve C3k2, compared with the original YOLO11n model, all three parameters (recall, mAP, and F1 score) are improved, while GFLOPS and the number of parameters remain unchanged, but the precision rate is slightly decreased. Finally, integrating the two modules, compared with the original YOLO11n model, all three parameters (precision, recall, mAP, and F1 score) are significantly improved. This indicates that the addition of each module has a positive effect on the detection results, reflecting the effectiveness of the improvements made in this invention.

[0088] In summary, this invention proposes a wind turbine blade defect detection method based on an improved YOLO11n. By introducing the IDC and ASFF modules, it enhances the detection capability of multi-scale blade defects in the model, effectively reduces the interference of environmental factors on the blade defect detection results, and thus improves the detection accuracy. Finally, related experiments verify the effectiveness of the proposed method from multiple dimensions. Therefore, the proposed method can provide important support for the maintenance of wind turbine blade faults, improve the operation and maintenance capabilities of wind farms, and reduce the economic losses caused by wind turbine blade faults.

[0089] Example 3

[0090] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0091] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0092] The processing unit performs the various methods and processes described above. For example, in some embodiments, the methods may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the methods by any other suitable means (e.g., by means of firmware).

[0093] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0094] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0095] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0096] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting defects in wind turbine blades based on an improved YOLO11n, characterized in that, The method includes: The wind turbine blade defect detection model was trained using a wind turbine blade defect dataset. Input the image data of the wind turbine blade to be tested into the trained wind turbine blade defect detection model, and output the wind turbine blade defect detection results, including the defect location and defect category. The wind turbine blade defect detection model is based on an improved YOLO11n model. The improvements include: using the IDC module to improve the C3k2 module in the YOLO11n backbone network and neck network to form the C3k2_IDC module. The C3k2_IDC module includes a first CBS module, a segmentation module, multiple C3k_IDC modules, a splicing module, and a second CBS module connected in sequence. The feature map input to the C3k2_IDC module is first subjected to preliminary feature extraction and channel number adjustment by the first CBS module. Then, it is divided into four branches by the segmentation module. The four branches are input to different parallel branches and processed by the C3k_IDC module. Finally, the outputs of each branch are concatenated and fused. The fused features are further subjected to convolutional compression, integration and nonlinear transformation by the second CBS module, and the number of channels is adjusted to the expected output dimension. Of the four branches, the first branch uses a 3×3 small kernel convolution to process local features and preserve detailed information; the second and third branches use 1×11 and 11×1 convolution kernels respectively to perform convolution operations in the width and height directions to capture global information in the width and height directions; the fourth branch is an identity mapping that does not perform any convolution operations and directly passes part of the input channels to the next layer.

2. The wind turbine blade defect detection method based on the improved YOLO11n according to claim 1, characterized in that, The C3k_IDC module is an improvement on the original C3k module, that is, the Bottleneck module in the original C3k module is replaced with the Bottleneck_IDC module.

3. The wind turbine blade defect detection method based on the improved YOLO11n according to claim 2, characterized in that, The Bottleneck_IDC module includes a third CBS module, an IDC module, and a fourth CBS module connected in sequence. The input data is processed by the third CBS module, the IDC module, and the fourth CBS module in sequence, and the output result is added to the input data as the final output.

4. The wind turbine blade defect detection method based on the improved YOLO11n according to claim 1, characterized in that, The improvements also include adding an ASFF module in front of the detection head of the original YOLO11n model to form the Detect_ASFF module for multi-scale target detection.

5. The wind turbine blade defect detection method based on the improved YOLO11n according to claim 4, characterized in that, The ASFF module consists of two parts: feature rescaling and adaptive fusion. In the feature rescaling part, ASFF first extracts feature maps from different levels of the network and upsamples or downsamples the feature maps from different levels to make them have the same resolution and number of channels. After feature rescaling, the ASFF module then adaptively fuses the multi-scale features.

6. The wind turbine blade defect detection method based on the improved YOLO11n according to claim 1, characterized in that, The process of constructing the wind turbine blade defect dataset includes: The publicly available wind turbine blade image data is processed to obtain an initial wind turbine blade defect dataset; The initial wind turbine blade defect dataset was labeled using annotation software, and then classified according to different blade defects to obtain the final wind turbine blade defect dataset.

7. The wind turbine blade defect detection method based on the improved YOLO11n according to claim 1, characterized in that, The defect categories include cracks, corrosion, edge corrosion, paint peeling, and surface damage.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

Images