Wind power generation blade surface defect detection method based on YOLO algorithm

Abstract

The invention discloses a wind power generation blade surface defect detection method based on a YOLO algorithm, and aims to solve the problems of low precision, complex model, poor robustness and the like in the existing wind power generation blade surface defect detection method. YOLOv8 is used as a reference model for improvement, a small target detection layer is added, an up-sampling operator is replaced, a P-ECSA module, a PMFusion module and an LF-Detect detection head are designed, and a new frame loss function W-IShipIoU is designed; detecting the surface defects of the wind power generation blade by using the trained model; the detection precision and speed are obviously improved, and the method is suitable for real-time monitoring and fault diagnosis of surface defects of the wind power generation blade; by introducing multi-scale feature fusion and an attention mechanism, the capability of capturing details by the model is enhanced, the calculation amount is reduced, and the method has a remarkable effect in the aspect of guaranteeing safe and stable operation of a wind power generation system.

Description

Technical Field

[0001] The present invention relates to the technical field of object detection in computer vision technology, and particularly to a method for detecting surface defects of wind turbine blades based on the YOLO algorithm. Background Art

[0002] As an important part of renewable energy, wind power plays an increasingly important role in the global energy structure. However, the stable operation and power generation efficiency of wind power systems are directly affected by the state of wind turbine blades. As a key component of wind power systems, wind turbine blades are exposed to harsh natural environments for a long time and are extremely vulnerable to various damages, such as surface damage, surface cracks, and paint peeling. These defects not only reduce the power generation efficiency of wind power systems and increase maintenance costs, but may also damage the structural integrity of the blades, leading to serious consequences such as shutdowns and equipment damage, and even threatening the safe operation of wind farms.

[0003] Currently, there are mainly two methods for detecting surface defects of wind turbine blades: one is the traditional detection method based on machine vision, and the other is object detection based on deep learning.

[0004] Traditional detection methods mainly rely on manual visual inspection and image processing techniques. Although this method can detect surface defects of blades to a certain extent, there are many deficiencies. First, manual inspection is not only time-consuming and laborious with low efficiency, but is also easily affected by the experience and subjective judgment of inspectors, resulting in insufficient detection accuracy. Second, traditional image processing techniques have limited recognition capabilities for complex backgrounds and diverse defect types, and it is difficult to achieve comprehensive and accurate detection of surface defects of blades.

[0005] With the rapid development of artificial intelligence technology, especially the wide application of deep learning technology in the field of object detection, the method for detecting surface defects of wind turbine blades based on deep learning has gradually become a research hotspot. Such methods achieve automatic recognition and positioning of defects by constructing deep learning models and training and learning a large number of images of surface defects of blades. However, there are still some problems with existing algorithms for detecting surface defects of wind turbine blades based on deep learning. For example, due to the complex model structure of some algorithms, the computational load is large and the detection speed is slow, making it difficult to meet the requirements of real-time detection. At the same time, the detection accuracy of some algorithms still needs to be improved, especially the detection effect for small targets and defects in complex backgrounds is not ideal.

[0006] For example, CN118735900A discloses a method for detecting surface defects of wind turbine blades based on improved YOLOv8n. It attempts to improve the backbone part of YOLOv8n by integrating StarNet_Blocks into the C2f module. However, this improvement method does not significantly improve the detection accuracy and still has the problem of a large number of parameters. Therefore, how to further improve the YOLOv8n network structure, improve the detection accuracy of wind turbine blade surface defects and reduce the number of parameters has become a technical problem that needs to be solved urgently.

[0007] Object detection technology based on deep learning has made remarkable progress in the field of image processing. Among them, the YOLO series of algorithms has received extensive attention for its high efficiency and accuracy. As one of the latest versions of the YOLO series, YOLOv8n performs well in object detection tasks. However, when applying it to the detection of wind turbine blade surface defects, there are still problems of insufficient detection accuracy and a large number of parameters. Specifically, when the existing YOLOv8n network processes the wind turbine blade surface defect dataset, due to the limitation of the network structure, it is difficult to fully capture the fine features on the blade surface, resulting in low detection accuracy. At the same time, the large number of parameters also increases the complexity of the model and the demand for computing resources, which is not conducive to deployment and promotion in practical applications.

[0008] To sum up, the detection of wind turbine blade surface defects is an important link to ensure the safe and stable operation of the wind power generation system. Traditional detection methods have problems of low efficiency and insufficient accuracy, while existing deep learning-based detection methods have limitations such as complex models and slow detection speed. Therefore, developing a fast and high-precision wind turbine blade surface defect detection algorithm is of great significance for improving the operation efficiency and safety of the wind power generation system. The purpose of the present invention is to provide a method for detecting wind turbine blade defects based on the YOLO algorithm to overcome the deficiencies of the existing technology and achieve fast and accurate detection of blade surface defects. Summary of the Invention

[0009] The technical content solved by the present invention is to provide a method for detecting wind turbine blade surface defects based on the YOLO algorithm, which solves the problems of insufficient detection accuracy and excessive consumption of computing resources in the field of wind turbine blade surface defect detection.

[0010] To solve the above technical problems, the technical solution adopted by the present invention is as follows: A method for detecting wind turbine blade surface defects based on the YOLO algorithm, comprising the following steps: Step1: Establish a wind turbine blade image dataset; Step2: Improve based on the YOLOv8 as the benchmark model; Step3: Set hyperparameters and train the improved model; Step 4: Use the trained model to detect the surface defects of wind turbine blades, and determine the types and location information of the defects.

[0011] In the preferred solution, the dataset in Step 1 includes three types of defect pictures: surface damage, surface cracks, and paint peeling, and the dataset is divided into a training set, a validation set, and a test set.

[0012] In the preferred solution, in Step 1, the dataset is a publicly available dataset collected from the Kaggle network platform, annotated using the open-source software LabelImg, and randomly divided into a training set, a validation set, and a test set according to a set ratio.

[0013] In the preferred solution, the specific steps of Step 2 include: Step 2.1: Add an additional small object detection layer to the original backbone network, and replace the original upsampling operator with an ATupsample (Attention Transformed Upsampling) module. Step 2.2: Add a P-ECSA (Parallel Enhanced Channel-Spatial Attention) module to the backbone network. Step 2.3: Replace the original C2f module in the YOLOv8 neck with a lightweight parallel multi-scale fusion module (Parallel Multi-Scale Fusion, PMFusion). Step 2.4: Replace the original decoupled detection head of YOLOv8 with an LF-Detect (Lightweight Feature-Fusion Detection) head. Step 2.5: Design a W-IShipIoU (Weighted Intersection over Shape IoU) border loss function to replace the original conventional loss function.

[0014] In the preferred solution, the P-ECSA module in Step 2.2 combines channel, spatial, and self-attention mechanisms to achieve dynamic feature weighting and multi-scale feature fusion, and uses grouped convolution and RepConv to reduce the model's computational overhead.

[0015] In a preferred solution, the detection head LF-Detect in Step 2.4 uses serial PConv and grouped volumes to form a new feature fusion structure to reduce the computational amount and the number of parameters of the model.

[0016] In a preferred solution, the bounding box loss function W-IShipIoU in Step 2.5 combines the center point position of the Wise-IoU loss, the width-to-height ratio and shape matching aspect ratio of the Shape-IoU, and introduces the position correction factor of the Wise-IoU to improve the positioning accuracy.

[0017] In a preferred solution, the expression of the bounding box loss function W-IshipIoU in Step 2.5 is as follows: (1) (2) (3) (4) (5) (6) In the formula, represents the W-IShipIoU loss function; represents the overlap degree between the predicted box and the ground truth box; is the correction factor for measuring the Euclidean distance between the center point of the predicted box and the center point of the ground truth box; is the correction factor for measuring the matching degree between the predicted box and the ground truth box in terms of width and height; is the position correction factor of the Wise-IoU, where is the normalization factor; and represent the width and height of the ground truth box respectively; 、 represent the horizontal and vertical coordinates of the center point of the predicted box respectively; and represent the horizontal and vertical coordinates of the center point of the ground truth box respectively; represents the width of the predicted box; represents the height of the predicted box; represents the diagonal distance; represents the exponential term, controlling the growth trend of the loss; represents the influence factor for controlling the shape error through the exponential function; and represent the weight coefficients in the horizontal and vertical directions respectively; 、 represent the width penalty term and the height penalty term respectively.

[0018] In a preferred solution, the hyperparameters in Step 3 include the number of training times, BatchSize, the maximum learning rate, and the cosine annealing algorithm is used to dynamically adjust the learning rate.

[0019] In a preferred solution, in Step 4, when using the trained model to detect the surface defects of wind turbine blades, the performance evaluation index MAP commonly used in object detection algorithms is adopted for evaluation.

[0020] In a preferred solution, the method further includes real-time feedback of the detected surface defect information of the wind turbine blade to the user so that the user can take corresponding maintenance or repair measures in time.

[0021] A method for detecting surface defects of wind turbine blades based on the YOLO algorithm provided by the present invention has the following beneficial effects: 1. By introducing the P-ECSA attention module, the present invention solves the technical problem that existing object detection algorithms are difficult to accurately identify small target defects under complex backgrounds and variable lighting conditions, significantly enhances the model's ability to capture feature information, improves the sensitivity to small target defects, and thus achieves high-precision detection in complex environments.

[0022] 2. The present invention adds two small target detection layers to the original backbone network of YOLOv8, increases the detection accuracy of the model and the perception ability of small targets, and combines the original upsampling with the attention mechanism so that the model can capture key target information at different scales during the upsampling process, thereby improving the detection accuracy.

[0023] 3. The present invention uses the PMFusion module to achieve effective fusion of multi-scale features, solves the technical problem that it is difficult to effectively fuse feature information at different scales and affects the detection accuracy, further improves the detection accuracy of the model, and enables the model to better handle defect detection tasks of various scales.

[0024] 4. The lightweight detection head LF-Detect designed by the present invention solves the problem that the detection process has a high demand for computing resources, which limits the wide application of the model, significantly reduces the consumption of computing resources and memory, makes the model more efficient and feasible in practical applications, reduces the hardware requirements, and improves the applicability of the model.

[0025] 5. The present invention proposes a new method for calculating the IoU loss function, namely W-IShipIoU. By combining the advantages of Wise-IoU and Shape-IoU, the W-IShipIoU loss function solves the technical problem that the traditional IoU loss function has deficiencies in evaluating the overlap degree between the predicted bounding box and the ground truth bounding box, which affects the positioning accuracy. It realizes a more accurate evaluation of the overlap degree between the predicted bounding box and the ground truth bounding box, optimizes the position and shape of the predicted bounding box, improves the positioning accuracy and shape matching degree, and thus enhances the accuracy of the detection result.

[0026] 6. For the YOLOv8 model, the present invention makes innovative improvements such as adding a small object detection layer, combining upsampling and attention mechanism, and using grouped convolution and RepConv to reduce the computational burden, which solves the deficiencies of the YOLOv8 model in the detection of surface defects of wind turbine blades, such as technical problems in detection accuracy and the number of parameters, improves the detection accuracy of the model and the perception ability of small objects, reduces the model parameters and computational amount at the same time, and enhances the performance of the model in the hardware environment.

[0027] 7. The present invention combines channel, spatial and self-attention mechanisms, uses grouped convolution and RepConv to reduce the computational burden, and enhances the ability of the model to capture details in complex scenarios through residual connections.

[0028] 8. By using multiple groups of PConv, the present invention designs a lightweight and multi-scale structure PMFusion to replace the C2f module in the original Neck of YOLOv8, greatly reducing the model parameters and computational amount, and further optimizing the performance in the hardware environment.

[0029] 8. The present invention designs a lightweight and multi-scale feature fusion detection head LF-Detect to replace the original decoupled detection head of YOLOV8, improves the feature fusion ability of the YOLOV8 detection head, and significantly reduces the computational amount and the number of parameters of the model.

[0030] 10. By using the center point position of the Wise-IoU loss, combining the width-height ratio and shape matching aspect ratio of Shape-IoU, and introducing the position correction factor of Wise-IoU to penalize the deviation of the center point of the box to improve the positioning accuracy, and ensuring the matching of the predicted bounding box and the ground truth bounding box in width, height and shape through the shape correction term of Shape-IoU.

[0031] 11. The improvement points of the present invention effectively solve the key problems in the field of wind power blade surface defect detection, significantly improve the detection accuracy and efficiency, reduce the consumption of computing resources, enhance the applicability and stability of the model, provide a new solution for the blade surface defect detection in the wind power industry, and promote the technological progress and development in this field. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The present invention will be further described below in conjunction with the drawings and embodiments: Figure 1 It is the detection flow chart of the detection method of the present invention; Figure 2 It is the module structure diagram of ATupsample of the detection method of the present invention; Figure 3 It is the structure diagram of the P-ECSA module of the detection method of the present invention; Figure 4 It is the structure diagram of the PMFusion model network of the detection method of the present invention; Figure 5 It is the structure diagram of the LF-Detect module of the detection method of the present invention; Figure 6 It is the structure diagram of the improved model of the detection method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The technical solutions in the present invention will be further described below in conjunction with the drawings and embodiments: Embodiment 1 As Figures 1 to 6 shown, a wind power blade surface defect detection method based on the YOLO algorithm includes the following steps: Step 1. Collect the publicly available wind power blade surface defect dataset from the Kaggle network platform. The defect types include 3 categories: surface damage, surface crack, and paint peeling, with a total of 2700 images. Use the open-source software LabelImg to label the collected wind power blade data. Randomly divide the processed dataset into a training set, a validation set, and a test set according to the ratio of 8:1:1. Among them, the training set contains 2160 images, the validation set contains 270 images, and the test set contains 270 images.

[0034] Step 2: Network improvement of the model. Add a small object detection layer to the original YOLOv8 model backbone network, and replace the original upsampling operator with a newly designed ATupsample module; add the module P-ECSA that combines channel, spatial, and self-attention to the original backbone network; design a new lightweight and multi-scale PMFusion to replace the original C2f module in the YOLOv8 Neck; design a lightweight detection head LF-Detect, and replace the original loss function with the bounding box loss function W-IShipIoU to obtain the improved model.

[0035] The improvement of the YOLOv8 model is as Figure 6 shown, and the specific measures are as follows: (1) Add an additional upsampling layer to the original YOLOv8 backbone network to upsample the feature maps at lower levels to generate multi-scale feature maps of 320×320 and 160×160. Based on this, on the basis of the original 3 object detection layers in YOLOv8, two additional small object detection layers are further added, corresponding to the scales of 320×320 and 160×160 respectively, to enhance the model's detection ability for small objects. At the same time, replace the original upsampling operator in YOLOv8 with the ATupsample module. As Figure 2 shown, this module combines the original upsampling with the attention mechanism, enabling the model to capture key object information at different scales during the upsampling process and forming a residual structure, thereby improving the detection accuracy.

[0036] (2) As Figure 3As shown in the figure, it is the structural diagram of the P-ECSA module. After adding the P-ECSA attention module to the original backbone network SPPF (Spatial Pyramid Pooling Fast) of YOLOv8, through the combination of channel attention, spatial attention, and self-attention mechanisms, the dynamic weighting and fusion of multi-scale features are achieved. In this module, the channel attention mechanism helps to adaptively weight the features of different channels, highlight the important channels, and suppress the irrelevant channels, thereby improving the expression ability of the network; the spatial attention mechanism weights the features at different positions in the spatial dimension to capture the important information at different spatial positions and further enhances the spatial perception ability of the features; the self-attention mechanism, during the multi-scale feature fusion process, helps the model focus on the important regions by calculating the correlation between the input features and optimizes the feature representation. In terms of convolution operations, the module uses grouped convolution instead of traditional convolution to reduce the computational amount and improve the efficiency. At the same time, the RepConv (Repeated Convolution) structure is adopted to improve the expressiveness of the convolution kernel and the parameter utilization rate of the network. The P-ECSA attention module of the present invention can process multi-scale information more effectively, enhance the performance of the network in complex tasks, reduce the computational overhead while improving the accuracy, and has stronger robustness and adaptability.

[0037] As shown in Figure 4 the figure, it is the structural diagram of the PMFusion module. The PMFusion module effectively retains the advantages of multi-scale feature fusion while significantly reducing the consumption of computing resources and memory by introducing multiple convolution operations of different scales, PConv (Partial Convolution), and grouped convolution, and combining average pooling. Convolutions of different scales can capture feature information at different levels in the image. The use of PConv and grouped convolution further reduces the computational amount and improves the processing efficiency. The average pooling layer (Avgpool2d) helps to retain the key spatial information while reducing the redundant features and further optimizes the performance of the model. Replacing C2f in the original neck network of YOLOv8 with the PMFusion module enables the improved network to significantly improve the detection speed while ensuring high accuracy in the task of detecting surface defects of wind turbine blades, enhances the ability to capture details, and effectively supports real-time monitoring and fault diagnosis applications, improving the performance and stability of the overall model.

[0038] As shown in Figure 5As shown in the figure, it is the structural diagram of the LF-Detect module. The LF-Detect module realizes lightweight feature fusion and efficient detection by combining PConv (partial convolution), grouped convolution, average pooling, and max pooling. PConv reduces the computational complexity during the feature extraction stage through learning in local convolution regions and enhances the network's sensitivity to local information. Grouped convolution effectively reduces the number of parameters by processing input channels in groups, improving the computational efficiency of the convolutional layer. The combination of average pooling and max pooling (Maxpool2d) compresses and aggregates features at different scales, preserving both global information and highlighting local important features, thus enhancing the model's spatial perception ability and detail capture ability. By replacing the original detection head of YOLOV8 with the new lightweight detection head LF-Detect, the present invention not only significantly reduces the memory footprint of the model but also improves the detection speed, especially suitable for tasks with high real-time and accuracy requirements for the surface defects of wind turbine blades. The improved module can ensure higher computational efficiency and lower hardware requirements while maintaining high detection accuracy, adapting to edge computing devices and large-scale deployment scenarios.

[0039] (5)According to the ideas of Wise-IoU (Weighted Intersection over Union) and Shape-IoU (Shape Intersection over Union) losses, a new calculation method for the IoU loss function is proposed, named W-IShipIoU. The W-IShipIoU designed in the present invention combines the center point position of the Wise-IoU loss with the width-height ratio and shape matching aspect ratio of Shape-IoU, and introduces the position correction factor of Wise-IoU to improve the localization accuracy by penalizing the deviation of the center point of the bounding box and through the shape correction term of Shape-IoU and to ensure the matching of the predicted bounding box and the ground truth bounding box in terms of width, height, and shape; respectively represent the width penalty term and the height penalty term; where the W-ShipIoU loss, 、 、 、 and The specific formulas are as follows: (1) (2) (3) (4) (5) (6) In the formula, represents the W-IShipIoU loss function; represents the overlap degree between the predicted bounding box and the ground truth bounding box; is a correction factor for measuring the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box. By introducing the shape correction factor, it ensures the normalization of bounding boxes with different aspect ratios; is a correction factor for measuring the matching degree between the predicted bounding box and the ground truth bounding box in terms of width and height. It performs a non-linear scaling on the size difference to reduce the impact of smaller errors and gives a stronger penalty to larger errors; is the position correction factor of Wise-IoU, where is the normalization factor, usually the generalized diagonal distance; and represent the width and height of the ground truth bounding box respectively; 、 represent the horizontal and vertical coordinates of the center point of the predicted bounding box respectively; and represent the horizontal and vertical coordinates of the center point of the ground truth bounding box respectively; represents the width of the predicted bounding box; represents the height of the predicted bounding box; represents the diagonal distance; represents the exponential term, which controls the growth trend of the loss, usually taking 4; represents the influence factor for controlling the shape error through the exponential function; and represent the weight coefficients in the horizontal and vertical directions respectively, and their values are related to the shape of the ground truth bounding box; 、 represent the width penalty term and the height penalty term respectively.

[0040] Step 3: Set the following hyperparameters: The number of training epochs is 240, the BatchSize (batch size) is set to 16, the maximum learning rate is set to 0.01, the minimum is set to 0, and the cosine annealing algorithm is used to dynamically adjust the learning rate.

[0041] Step 4: After completing the training of the model, use the trained model to detect the surface defects of wind turbine blades, so as to determine the types of defects and the location information of the defects.

[0042] This model is evaluated using the commonly used performance evaluation index MAP (Mean Average Precision) in object detection algorithms.

[0043] The experimental hardware configuration is a Xeon(R) Silver 4214R processor, the graphics card is an NVIDIA GeForce RTX3080 Ti, the development language is python3.8, the CUDA version is 11.3, and the Pytorch version is 1.11.

[0044] Using the improved YOLOV8 model designed by the present invention, as long as the user gives an image with defects on the surface of the wind turbine blade, the system can detect the defect information on the blade surface according to the trained model.

[0045] Example 2 In another preferred embodiment, on the basis of Embodiment 1, this embodiment details the complete process of the method for detecting surface defects of wind turbine blades based on the YOLO algorithm.

[0046] First, collect the publicly available dataset of surface defects of wind turbine blades from the Kaggle network platform. The defect types include three categories: surface damage, surface cracks, and paint peeling, with a total of 2700 pictures. Use the open-source software LabelImg to label the collected wind turbine blade data to clarify the position and category of each defect. Subsequently, randomly divide the processed dataset into a training set, a validation set, and a test set according to a ratio of 8:1:1. The training set contains 2160 pictures for model training; the validation set contains 270 pictures for model validation and adjustment; the test set contains 270 pictures for final evaluation of the model's performance.

[0047] Next, improve the YOLOv8 model. Add a P-ECSA attention module after the original backbone network SPPF to enhance the model's ability to process multi-scale information. Then, replace the C2f module in the original neck network of YOLOv8 with a PMFusion module to retain the advantages of multi-scale feature fusion while reducing computational complexity and memory consumption. In addition, replace the original decoupled detection head of YOLOv8 with a new lightweight detection head LF-Detect to further improve the detection speed and accuracy. At the same time, design a new calculation method for the IoU loss function, W-IShipIoU, to improve the positioning accuracy and shape matching degree.

[0048] In the model training stage, set the number of training times to 240, the BatchSize to 16, the maximum learning rate to 0.01, and use the cosine annealing algorithm to dynamically adjust the learning rate. After training is completed, use the trained model to detect the surface defects of wind turbine blades to determine the type and location information of the defects. In this embodiment, the mean average precision (MAP) is used as a performance evaluation index to evaluate the performance of the model.

[0049] Example 3 In another preferred embodiment, based on Embodiment 2, this embodiment focuses on describing the improvement details of the YOLOv8 model and its effects.

[0050] When improving the YOLOv8 model, special attention was paid to the detection accuracy and perception ability of the model for small targets. Therefore, two small target detection layers were added to the original backbone network, and the original upsampling operator was replaced with the ATupsample module to achieve target information capture at different scales. At the same time, the P-ECSA module combines channel attention, spatial attention, and self-attention mechanisms to achieve dynamic feature weighting and multi-scale feature fusion, further improving the detection accuracy of the model.

[0051] In addition, the PMFusion module effectively retains the advantages of multi-scale feature fusion and significantly reduces the consumption of computing resources and memory by introducing convolutional operations, PConv, and group convolution at multiple different scales. The LF-Detect module achieves lightweight feature fusion and efficient detection by combining PConv, group convolution, average pooling, and max pooling. These improvements enable the model to significantly improve the detection speed while ensuring high accuracy in the task of detecting surface defects of wind turbine blades.

[0052] Embodiment 4 In another preferred embodiment, based on Embodiments 1, 2, and 3, this embodiment takes a specific application scenario as an example to demonstrate the effects of the present invention in practical applications.

[0053] Taking a wind farm as an example, the wind turbine blades often have surface damage, cracks, and paint peeling and other defects. In order to detect and repair these defects in a timely manner and ensure the safe and stable operation of the wind power generation system, this wind farm adopts the method for detecting surface defects of wind turbine blades based on the YOLO algorithm proposed by the present invention.

[0054] In practical applications, first, the dataset is collected and preprocessed according to the methods of Embodiments 1 and 2, and then the YOLOv8 model is improved and trained. After training is completed, the trained model is deployed to the monitoring system of the wind farm. When the monitoring system captures an image of the wind turbine blade, the model can automatically detect the types and location information of the defects on the blade surface and feedback the results to the relevant personnel for processing in real time.

[0055] Through practical applications, the method of the present invention significantly improves the detection accuracy and speed of surface defects of wind turbine blades, providing a strong guarantee for the safe operation of wind farms. At the same time, this method also has good robustness and generalization ability, and can adapt to defect detection tasks in different environments and conditions.

[0056] In a preferred solution, the dataset in Step 1 includes three types of defect pictures: surface damage, surface cracks, and paint peeling, and the dataset is divided into a training set, a validation set, and a test set; the above settings are aimed at comprehensively covering common object surface defect types, and by scientifically dividing the dataset, it is ensured that the model can effectively learn features during the training process and accurately evaluate the generalization ability, laying a solid foundation for subsequent defect recognition and classification tasks.

[0057] In a preferred solution, in Step 1, the dataset is a publicly available dataset collected from the Kaggle network platform, labeled using the open-source software LabelImg, and randomly divided into a training set, a validation set, and a test set according to a set ratio; the above settings ensure the diversity and quality of the data. At the same time, the application of the open-source software LabelImg improves the labeling efficiency, and the random division strategy helps to improve the stability and generalization ability of model training.

[0058] In a preferred solution, the P-ECSA module in Step 2.2 combines channel, spatial, and self-attention mechanisms to achieve dynamic feature weighting and multi-scale feature fusion, and uses grouped convolution and RepConv to reduce the model's computational overhead; the above settings effectively improve the model's adaptability to complex scenarios while maintaining high computational efficiency; in addition, by introducing residual connections, the gradient flow is further optimized, ensuring the stability and convergence speed of the training process.

[0059] In a preferred solution, the detection head LF-Detect in Step 2.4 uses a serial combination of PConv and grouped convolution to form a new feature fusion structure to reduce the computational amount and the number of parameters of the model; the above settings not only significantly improve the speed of object detection but also ensure that the detection accuracy is not affected; at the same time, by introducing the attention mechanism, the feature fusion process is further optimized, enhancing the model's ability to recognize objects in complex scenarios.

[0060] In a preferred solution, the bounding box loss function W-IShipIoU in Step 2.5 combines the center point position of the Wise-IoU loss, the width-to-height ratio of the Shape-IoU, and the aspect ratio of shape matching, and introduces the position correction factor of the Wise-IoU to improve the positioning accuracy; the above settings further enhance the model's ability to detect ship targets with complex shapes, especially in challenging scenarios such as target occlusion and overlap, and can still accurately identify and locate, effectively improving the robustness and accuracy of ship detection.

[0061] In a preferred solution, the hyperparameters in Step 3 include the number of training times, BatchSize, the maximum learning rate, and the cosine annealing algorithm is used to dynamically adjust the learning rate; the above settings are aimed at achieving more efficient model convergence and performance improvement by finely regulating the model training process; at the same time, the application of the cosine annealing algorithm can effectively prevent the model from falling into local optima and further improve the generalization ability of the model.

[0062] In a preferred solution, in Step 4, when using the trained model to detect the surface defects of wind turbine blades, the mean average precision MAP, a commonly used performance evaluation index in object detection algorithms, is used for evaluation; the above settings ensure the accuracy and reliability of the detection results; at the same time, in order to further improve the detection efficiency, parallel processing technology is also introduced, enabling the model to run efficiently on a multi-core CPU (Central Processing Unit) or GPU (Graphic Processing Unit), greatly shortening the detection time.

[0063] In a preferred solution, the method further includes real-time feedback of the detected surface defect information of the wind turbine blade to the user so that the user can take corresponding maintenance or repair measures in a timely manner; the above settings not only improve the operation and maintenance efficiency of the wind power system, but also ensure that the blade operates in the best state, effectively extending the service life of the blade, while reducing the potential safety risks caused by defects.

[0064] In a preferred solution, the small target detection layers added to the original backbone network in Step 2.1 correspond to scales of 320×320 and 160×160 respectively to enhance the model's detection ability for small targets; the above settings enable the model to show higher flexibility and accuracy when processing targets of different sizes, especially when facing small-sized and detail-rich targets, the detection effect has been significantly improved.

[0065] In summary, the present invention proposes a method for detecting surface defects of wind turbine blades based on an improved YOLOv8 algorithm, aiming to solve the problems of insufficient detection accuracy and excessive consumption of computing resources in this field. The object detection algorithms in the prior art often have difficulty in accurately identifying small target defects in the face of complex backgrounds and variable lighting conditions, and the detection process has a high demand for computing resources. Therefore, the present invention effectively improves the detection accuracy and perception ability of the model for small target defects by introducing the P-ECSA attention module, the PMFusion module, and the lightweight detection head LF-Detect, while reducing the consumption of computing resources and memory, realizing efficient detection. In addition, the present invention also proposes a new method for calculating the IoU loss function, W-IShipIoU, which can more accurately evaluate the overlapping degree between the predicted box and the ground truth box, further improving the localization accuracy and shape matching degree of the model. The scheme also emphasizes the processing and annotation process of the dataset, ensuring the high-quality data required for model training, thereby improving the generalization ability and robustness of the model. Through these innovative improvements and designs, the present invention not only improves the detection accuracy and efficiency of the surface defects of wind turbine blades, but also provides new ideas and methods for improving the performance of object detection algorithms, bringing new breakthroughs and developments to the field of detecting surface defects of wind turbine blades.

Claims

1. A method for detecting surface defects of wind turbine blades based on the YOLO algorithm, characterized in that, It includes the following steps: Step1: Establish a wind power blade image dataset; Step2: Improve based on the YOLOv8 as the benchmark model; Step3: Set hyperparameters and train the improved model; Step4: Use the trained model to detect the surface defects of wind power blades and determine the types and location information of the defects.

2. The method for detecting surface defects of a wind turbine blade based on the YOLO algorithm according to claim 1, wherein: In the Step1, the dataset includes three types of defect pictures: surface damage, surface cracks, and paint peeling, and the dataset is divided into a training set, a validation set, and a test set.

3. A method for detecting surface defects of wind turbine blades based on the YOLO algorithm according to claim 2, characterized in that: In the Step1, the dataset is a publicly available dataset collected from the Kaggle network platform, annotated using the open-source software LabelImg, and randomly divided into a training set, a validation set, and a test set according to a set ratio.

4. A method for detecting surface defects of a wind turbine blade based on the YOLO algorithm according to claim 3, characterized in that, The specific steps of the Step2 include: Step2.1: Add an additional small target detection layer to the original backbone network and replace the original upsampling operator with the ATupsample module; Step2.2: Add the P-ECSA module to the backbone network; Step2.3: Replace the original C2f module in the YOLOv8 neck with the lightweight and multi-scale PMFusion module; Step2.4: Replace the original decoupled detection head of YOLOv8 with the lightweight and multi-scale feature fusion detection head LF-Detect; Step2.5: Design a weighted shape intersection over union bounding box loss function W-IShipIoU to replace the original conventional loss function.

5. The method for detecting surface defects of a wind turbine blade based on the YOLO algorithm according to claim 4, wherein: The P-ECSA module in the Step2.2 combines channel, spatial, and self-attention mechanisms to achieve dynamic feature weighting and multi-scale feature fusion, and uses grouped convolution and RepConv to reduce the model's computational overhead.

6. The method for detecting surface defects of wind turbine blades based on the YOLO algorithm according to claim 5, wherein: The detection head LF-Detect in the Step2.4 uses a serial PConv and grouped convolution to form a new feature fusion structure to reduce the model's computational amount and number of parameters.

7. The method for detecting surface defects of wind turbine blades based on the YOLO algorithm according to claim 6, wherein: The bounding box loss function W-IShipIoU in the Step2.5 combines the center point position of the Wise-IoU loss, the width-height ratio and shape matching aspect ratio of the Shape-IoU, and introduces the position correction factor of the Wise-IoU to improve the positioning accuracy.

8. A method for detecting surface defects of wind turbine blades based on the YOLO algorithm according to claim 7, characterized in that: The expression of the bounding box loss function W-IshipIoU in the Step2.5 is as follows: （1）； （2）； （3）； （4）； （5）； （6）； In the formula, represents the W-IShipIoU loss function; represents the overlap degree between the predicted bounding box and the ground truth bounding box; is a correction factor for measuring the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box; is a correction factor for measuring the matching degree between the predicted bounding box and the ground truth bounding box in terms of width and height; is the position correction factor of Wise-IoU, where is the normalization factor; and represent the width and height of the ground truth bounding box respectively; 、 represent the horizontal and vertical coordinates of the center point of the predicted bounding box respectively; and represent the horizontal and vertical coordinates of the center point of the ground truth bounding box respectively; represents the width of the predicted bounding box; represents the height of the predicted bounding box; represents the diagonal distance; represents the exponential term; represents the influence factor for controlling the shape error through the exponential function; and represent the weight coefficients in the horizontal and vertical directions respectively; 、 represent the width penalty term and the height penalty term respectively.

9. The method for detecting surface defects of a wind turbine blade based on the YOLO algorithm according to claim 8, wherein: The hyperparameters in the Step3 include the number of training times, BatchSize, and maximum learning rate, and the cosine annealing algorithm is used to dynamically adjust the learning rate.

10. A method for detecting surface defects of a wind turbine blade based on the YOLO algorithm according to claim 9, characterized in that: In the Step4, when using the trained model to detect the surface defects of wind power blades, the mean average precision, a commonly used performance evaluation metric in object detection algorithms, is used for evaluation.

Images