Fan blade defect detection method based on improved YOLOv11

Abstract

The invention discloses a fan blade defect detection method based on an improved YOLOv11 algorithm, and aims to improve the accuracy, real-time performance and calculation efficiency of fan blade defect detection. According to the method, a multi-type fan blade defect image data set is constructed and preprocessed, and the structure of a YOLOv11 model is improved: a C3K2-WTConv module integrated with wavelet transform is adopted, so that the feature extraction depth is enhanced, and more image details are mined; a C2PSA-MLAC module fused and mixed with local channel attention is introduced, the key information expression ability is improved, and the model focuses more important features; an original bounding box regression index is replaced with a Shape-IoU loss function, and the detection effect is optimized; through improvement, the parameter scale and the calculation load are reduced, meanwhile, the robustness is effectively improved, the detection precision is improved by about 8.7%, the detection efficiency is remarkably enhanced, the technical problem of a fan blade defect detection scene is better solved, and the method has good engineering practicability and popularization value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition and detection, and particularly relates to a fan blade defect detection method based on an improved YOLOv11. BACKGROUND

[0002] In recent years, with the continuous promotion of global green energy transformation, wind power, as an important energy form of low carbon and renewable, occupies an increasingly important position in the energy structure. The stable operation of wind power generation system cannot be separated from the key component of fan blades, and its structural integrity and working state directly affect the power generation efficiency and the safety of equipment operation. However, due to the fact that fan blades are usually in outdoor environment, they are subjected to the influence of harsh natural factors such as wind erosion, rainwater erosion, temperature difference cycle, freezing snow, etc. for a long time, and are prone to produce various types of defects such as cracks, delamination, surface peeling, corrosion, etc. If these defects are not discovered and treated in time, not only will it lead to the decline of fan efficiency, but also may cause major failures or safety accidents, resulting in equipment downtime and economic losses.

[0003] At present, fan blade defect detection mainly relies on manual inspection, unmanned aerial vehicle shooting and various non-destructive testing methods such as ultrasonic and infrared imaging. Although manual inspection has certain intuitiveness, it has long detection period, high labor intensity of personnel, and great safety hazards in high tower environment. Unmanned aerial vehicle inspection improves the inspection efficiency and coverage, but usually still needs manual analysis of images, and it is difficult to accurately identify small size defects or cope with complex background. While traditional non-destructive testing technology can detect internal structural changes of the blade, it is expensive and complex to operate, and has limited ability to perceive small cracks on the surface, which is difficult to meet the growing demand for automation and intelligent detection.

[0004] In recent years, with the rapid development of artificial intelligence and computer vision technology, target detection methods based on deep learning have attracted widespread attention in the field of industrial defect recognition. Among them, the YOLO (You Only Look Once) series of target detection algorithms, with its end-to-end fast detection capability, performs well in real-time image recognition tasks, and gradually becomes a research hotspot for fan blade surface defect recognition. Although the YOLO series algorithms have evolved, such as YOLOv5, YOLOv7, YOLOv8, YOLOv9 and YOLOv10 versions have been tried to be applied in the field of wind power, but there are still problems such as complex model parameters, inaccurate small target recognition, and insufficient adaptability to multi-scale features in actual application.

[0005] In view of the above technical bottlenecks, researchers have been trying to improve the YOLO algorithm by introducing lightweight network structure, feature enhancement mechanism, attention module, multi-scale fusion strategy and other means to improve its recognition effect on the complex defects of fan blades. YOLOv11, as the latest evolution version of the YOLO series, has a more optimal structure design and higher detection accuracy, providing a technical basis for efficient and accurate identification of fan blade defects.

[0006] Under this background, the present application proposes a fan blade defect detection method based on improved YOLOv11, which comprehensively introduces up-sampling optimization strategy, attention mechanism and customized loss function, aiming to improve the detection accuracy of small target defects in complex background, reduce the computational complexity, enhance the real-time performance and adaptability, and better meet the actual needs of intelligent operation and maintenance and risk warning of wind power plants. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a fan blade defect detection method based on improved YOLOv11 algorithm, aiming to overcome the shortcomings of existing fan blade defect detection methods in terms of detection accuracy, speed and model complexity. Traditional methods such as manual inspection and unmanned aerial vehicle image acquisition rely on human experience, and have problems such as low efficiency, high misjudgment rate and unstable detection results. However, the current mainstream deep learning detection algorithms such as YOLOv5, YOLOv7, YOLOv8, YOLOv9 and YOLOv10 still face challenges such as low small target recognition accuracy, strong background interference, large model calculation amount and insufficient ability to represent edge details in the application of fan blade defect recognition, which limits their practical application value in complex environments.

[0008] To solve the above technical problems, the present application proposes a fan blade defect detection method based on improved YOLOv11, the technical scheme of which includes the following steps: Step 1: Construct an image dataset containing fan blade surface defects, and label and expand the dataset; Step 2: Improve the original YOLOv11 algorithm, and introduce the C3K2 module in the original YOLOv11 network into the WTConv (Wavelet Transform Conv, wavelet transform convolution) to construct the C3K2-WTConv module; Step 3: Introduce the MLCA (Multi-Level CrossAttention, multi-level cross-attention) attention mechanism in the C2PSA module of the original YOLOv11 algorithm, and construct the C2PSA-MLCA module; Step4: replace the CIoU (Complete Intersection over Union Loss) loss function in the original YOLOv11 algorithm with the Shape-IoU (Shape Intersection over Union Loss) loss function; Step5: train the optimized YOLOv11 algorithm on fan blade defects to obtain a trained improved YOLOv11 algorithm; Step6: use the trained optimized YOLOv11 algorithm model to detect fan blade defects.

[0009] In the preferred scheme, the defect types mentioned in Step1 include burning, cracking, holes, deformation, stains, oil stains, peeling, and rust, etc. The data set is enhanced by methods such as rotation, scaling, and cropping, and is divided into training set and test set according to the set proportion.

[0010] In the preferred scheme, after embedding WTConv in the C3K2 module in Step2, the model not only retains the advantages of the original convolution feature extraction, but also obtains more detailed information through wavelet domain decomposition, which is particularly suitable for extracting the structural features of small target defects such as cracks and fine corrosion in fan blades.

[0011] In the preferred scheme, the step of introducing the C3K2 module in the original YOLOv11 network into the wavelet transform WTConv to construct the C3K2-WTConv module in Step2 includes: Step2.1: perform 1x1 convolution operation on the input feature map to compress the channel number and extract preliminary local features, reduce redundant information and reduce the subsequent operation burden, and provide a compact structure input feature for wavelet transform; Step2.2: use the wavelet transform structure in WTConv to perform two-dimensional discrete wavelet transform on the compressed feature map, decompose the image features into low-frequency information and high-frequency information, and realize multi-scale spatial frequency separation of features; Step2.3: perform convolution enhancement processing on the multiple groups of sub-band features obtained by decomposition, fuse the key texture and edge features in each sub-band after capturing them with different convolution kernels, and reconstruct an enhanced feature map that combines edge information and context semantics; Step2.4: restore the fused sub-band features to a complete feature map through inverse wavelet transform (IWT, Inverse Wavelet Transform), realize detail preservation and feature reconstruction, and strengthen the model's expression ability for defect edges and local details while maintaining detection accuracy; Step2.5: The wavelet-enhanced feature map is connected in residual connection with the backbone output in the original path of C3K2, fusing low-frequency and high-frequency information, and finally output to the subsequent network layer, enhancing the feature robustness and detection performance of the model in complex scenes.

[0012] In a preferred scheme, the C2PSA in Step3 introduces the MLCA mechanism to model the key regions in the fan blade image through cross-fusion of spatial dimensions and channel dimensions, thereby enhancing the model's perception ability for defect regions.

[0013] In a preferred scheme, the C2PSA module in the network of the YOLOv11 algorithm in Step3 introduces the MLCA attention mechanism to construct a C2PSA-MLCA module, and the specific steps are as follows: Step3.1: Perform channel-by-channel convolution operation on the input feature map to reconstruct the channel structure and extract preliminary features, providing high-quality input basis for the subsequent attention mechanism; Step3.2: Respectively construct spatial dimension attention branch and channel dimension attention branch to perform weighted calculation on the input feature map in different dimensions, improving the representation ability of features in spatial position and semantic channel; Step3.3: Cross-fuse spatial and channel attention of different levels, and realize effective integration of context features and local details through residual connection and element-by-element weighting, enhancing the response ability of the model to small target defects; Step3.4: Concatenate the fused cross-attention features and the original features and project them back to the original feature dimension, output the enhanced feature map through convolution processing, and provide it for the subsequent detection head, thereby improving the defect detection accuracy and robustness; In a preferred scheme, the Shape-IoU loss function in Step4 has good detection effect for irregularly shaped defects such as cracks and scratches in fan blades, and the traditional CIoU is limited in fitting such targets. Shape-IoU models the consistency of boundary contours through shape intersection over union, and is more suitable for fitting and detecting such actual defects.

[0014] In a preferred scheme, the specific steps for using Shape-IoU loss function to replace the CIoU loss function in the original YOLOv11 algorithm in Step4 are as follows: Step4.1: Shape-IoU extracts the center point coordinates, width, height, and other geometric parameters of the predicted bounding box and the real bounding box from the model output, providing basic information for subsequent shape difference calculation; Step4.2: Based on the traditional IoU method, calculate the intersection and union regions between the predicted box and the real box to obtain the basic overlap information; Step4.3: Construct a shape similarity constraint function through geometric indicators such as the aspect ratio of the bounding box, contour difference, etc., to comprehensively evaluate the shape structure coincidence degree of the predicted box and the real box; Step4.4: Jointly define the IoU indicator and the shape consistency indicator as a Shape-IoU loss function to enhance the sensitivity of the loss function to shape fitting errors and improve the regression accuracy; Step4.5: Based on the Shape-IoU loss, adjust the weight parameters of the regression branch in the network to improve the prediction accuracy and training efficiency of the model in positioning the complex defect shape of the fan blade.

[0015] In a preferred embodiment, the training step of the improved YOLOv11 algorithm model in Step5 includes: Step5.1: Initialize the improved YOLOv11 algorithm model by introducing C3K2-WTConv module, C2PSA-MLCA module and Shape-IoU loss function, and set the initial network weight, learning rate, batch size and training round number, etc. Training parameters; Step5.2: Input the fan blade image samples in the training set into the improved YOLOv11 model, perform forward propagation to calculate the predicted bounding box; combined with the Shape-IoU loss function, calculate the positioning and shape loss value between the predicted box and the target GT box, and perform back propagation to optimize the network parameters; Step5.3: Use the validation set to evaluate the YOLOv11 model with C3K2-WTConv and C2PSA-MLCA modules, calculate the average precision mean of each type of fan blade defect, and use it to measure the detection effect of the model after feature extraction and attention mechanism enhancement; Step5.4: Repeat Step5.2 and Step5.3 to continuously optimize the model parameters until the average precision value of the validation set converges to the preset stability standard, and complete the model training; Step5.5: Input the test set into the improved YOLOv11 algorithm model after final training, and comprehensively evaluate its detection accuracy, recall rate and real-time performance in the fan blade defect detection task, and verify the effectiveness of the three improvement measures in actual application.

[0016] In a preferred embodiment, the comprehensive detection performance is comprehensively measured by indicators such as detection accuracy, detection speed, floating point calculation amount, parameter amount and picture detection frame number per unit time.

[0017] In a preferred embodiment, the step of analyzing the detection accuracy and detection speed includes: Step5.5.1: Performance evaluation of the improved YOLOv11 algorithm model after training is completed, including calculating the changes in detection accuracy and detection speed of the original YOLOv11 model and the models after introducing C3K2-WTConv, C2PSA-MLCA and Shape-IoU loss function respectively, and analyzing the optimization effect brought by the three improvements; Step5.5.2: Evaluation to measure the detection accuracy of the model, the evaluation indexes include the average precision mean mAP of all categories, the recall R of target detection, the accuracy P, the floating-point calculation amount FLOPs of the calculation model, the parameter amount Params and the picture detection frame number FPS per second, to comprehensively evaluate the performance of the improved YOLOv11 model in detection performance and calculation efficiency; Step5.5.3: By comparing the detection results of the model after introducing C3K2-WTConv module for enhancing local details and anti-background interference ability, C2PSA-MLCA module for improving the response ability of key areas in feature extraction, Shape-IoU loss function for optimizing the stability and accuracy of defect boundary box positioning on the same data set, analyze the improvement of each module on the detection performance, and verify the effectiveness of the combined improvement scheme; Step5.5.4: Compared with existing representative detection algorithms such as YOLOv5, YOLOv7, YOLOv8, YOLOv9 and YOLOv10 under the same data set, hardware configuration and hyperparameter setting, further illustrate the superiority and actual deployment value of the improved YOLOv11 algorithm.

[0018] The fan blade defect detection method based on the improved YOLOv11 proposed by the application has the following beneficial effects: 1、The application introduces structure optimization and loss function improvement strategy based on YOLOv11, which significantly improves the detection accuracy and speed of fan blade defects; Compared with traditional methods such as manual inspection and unmanned aerial vehicle inspection, it has higher efficiency and adaptability, and is better than YOLOv5, YOLOv7, YOLOv8, YOLOv9 and YOLOv10 models in complex background and small target scene.

[0019] 2. The present invention introduces the WTConv wavelet convolution structure into the C3K2 module in the YOLOv11 backbone network to construct the C3K2-WTConv module, which effectively improves the model's ability to express details such as edges and textures, and enhances the accuracy and robustness of small target defect detection.

[0020] 3. The present invention integrates the MLCA attention mechanism into the C2PSA module of YOLOv11. By guiding the model to focus on key areas and suppress background interference, it achieves efficient encoding and semantic enhancement of feature maps, thereby improving detection performance with low computational overhead.

[0021] 4. This paper adopts the Shape-IoU loss function to replace the CIoU loss function originally used by YOLOv11, and introduces shape constraints and contour overlap information in bounding box regression, making positioning more accurate and gradient propagation smoother, effectively improving the target positioning quality under complex defect contours.

[0022] 5. The dataset construction and preprocessing strategy adopted by the present invention fully considers the diversity of wind turbine blade defect types, and combines high-quality labeling and sample balancing methods to improve the generalization ability of the model and provide a guarantee for the subsequent algorithm performance.

[0023] 6. The training strategy of the present invention adopts a multi-stage, stable convergence approach, combined with weight warm-up and dynamic learning rate adjustment mechanism to ensure the stable performance of each module and achieve collaborative optimization in multiple key indicators (accuracy, speed, and computational complexity).

[0024] 7. Through algorithm improvement and module fusion, this invention significantly reduces the total number of parameters and model calculation complexity; experimental results show that compared with the original YOLOv11 model, the improved model's mAP is improved by about 8.7%, verifying its high performance and lightweight characteristics.

[0025] 8. The improved solution proposed in the present invention is comprehensively optimized on multiple key submodules, and overall achieves high precision, high efficiency and strong adaptability in wind turbine blade defect detection, and has good practical application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 It is a technical flow chart of the present invention; Figure 2 This is the structural diagram of the WTConv module of the present invention; Figure 3 This is a structural diagram of the MLCA module of the present invention; Figure 4 This is the improved YOLOv11 structure diagram of the present invention; Figure 5 Part defect detection graph before improvement of the present application; Figure 6 Part defect detection graph after improvement of the present application. DETAILED DESCRIPTION

[0027] The technical solutions in the present application are further described below in combination with examples and drawings: Example 1 As shown in the figure, the present embodiment provides a fan blade defect detection method based on improved YOLOv11 algorithm, covering data construction, model optimization and detection process whole cycle content, the specific steps are as follows: Figure 1 Step1: Construct fan blade defect image dataset and perform standardized preprocessing Step1.1, data collection: Collect fan blade images taken during the operation of wind power plant, and combine with network public defect samples to form image basic set covering multiple types of defects, such as cracks, erosion, holes, wear, burning, damage and dirt, etc. Step1.2, data annotation:

[0028] Use Labelimg and other annotation software to annotate the defects frame of the image, and generate standard format annotation file containing defect label and coordinate data. Step1.3, data enhancement and division:

[0029] Use affine transformation, brightness disturbance, Gaussian noise injection and other methods to perform image enhancement operation to expand sample diversity; after enhancement, the data is divided into training set and verification set according to 8:2 to ensure the consistency of training and verification sample distribution. Step2: Improve YOLOv11 algorithm and introduce C3K2-WTConv module to replace C3K2 module in the original backbone network

[0030] Step2.1, lightweight convolution kernel replacement: In the YOLOv11 backbone network, introduce C3K2-WTConv module, use deformable wide-height convolution kernel to replace traditional convolution structure, to reduce model parameter quantity and improve feature extraction flexibility. Step2.2, structure fusion mechanism:

[0031] The module processes different scale information through double-channel path respectively, effectively improving the size adaptability of defect target. Step2.3, spatial dynamic receptive field expansion:

[0032] ​Through dynamically learnable spatial receptive kernel weights, fine perception of complex structure regions in blade images is achieved, enhancing the model's ability to recognize irregular defects.

[0033] Step2.4, Channel compression and expansion strategy: Using the strategy of combining 1x1 convolution and group convolution, the information expression efficiency is improved while the channel is compressed, optimizing the use efficiency of computing resources.

[0034] Step3: Introducing MLCA hybrid local channel attention in YOLOv11 algorithm C2PSA module Step3.1, Enhance fine-grained feature extraction ability: MLCA can weight process the local channel of image feature map, emphasizing the response degree of local key areas on different channels, thus improving the recognition effect of fine-grained defects such as fan blade micro-cracks and edge peeling.

[0035] Step3.2, Improve the information interaction efficiency between channels: Traditional channel attention focuses on global average weight distribution, while MLCA focuses attention in a local range, making the information flow between adjacent channels more accurate, which helps to preserve structural features and texture differences.

[0036] Step3.3, Reduce background interference and improve detection robustness: MLCA has an inhibitory effect on background noise in fan blade images. By focusing on effective channel areas, it reduces the model's misjudgment of irrelevant features, thus improving the model's detection stability in complex backgrounds.

[0037] Step3.4, Consider local and global feature fusion: Under the MLCA mechanism, local channel attention is combined with global spatial attention in the C2PSA module, achieving "channel-space" dual-dimensional enhancement, effectively improving the model's generalization ability in multi-scale and multi-shape defect detection.

[0038] Step3.5, Lightweight design, low computational overhead: MLCA uses lightweight operations such as groupable convolution and region normalization to improve model performance while controlling computational resource consumption, making it suitable for fan blade online real-time defect detection scenarios.

[0039] Step4: Replace the original CIoU loss function used by YOLOv11 with Shape-IoU loss function Step4.1, More accurately align the shape difference between the predicted box and the real box: The Shape-IoU loss function not only considers the degree of positional overlap, but also introduces sensitivity modeling for target aspect ratio and shape distribution, allowing for more precise evaluation of the differences between target frame shapes, thereby improving the fitting accuracy of the bounding box, especially for long and irregular targets such as fan blades.

[0040] Step 4.2, alleviate the positioning deviation caused by inconsistent length and width: Traditional CIoU only measures the center point offset through Euclidean distance and introduces an aspect ratio penalty term, while Shape-IoU explicitly models shape consistency in its design, effectively addressing the error accumulation caused by large differences in the aspect ratio between predicted and true frames, improving the stability of model positioning.

[0041] Step 4.3, improve the robustness of target overlap rate evaluation: Shape-IoU places more emphasis on the matching degree of overlapping areas in space shape in loss calculation, rather than just area overlap, providing more discriminative regression signals in complex scenarios such as occlusion, rotation, and asymmetry, helping to improve the detection performance of fan blade edge defects.

[0042] Step 4.4, optimize regression convergence speed and detection accuracy: Since Shape-IoU has a more reasonable gradient distribution, its loss function provides smoother feedback on bounding box prediction errors, significantly accelerating model training convergence speed while improving the final bounding box regression accuracy and reducing the rate of missed and false detections.

[0043] Step 4.5, better adapt to high aspect ratio and multi-scale target detection tasks: Fan blade defects often manifest as small cracks, narrow scratches, and other structures with large spatial scale variations. Shape-IoU performs better in fitting different scale target shapes, significantly improving the generalization ability of YOLOv11 in complex industrial images.

[0044] Step 5: Use the optimized YOLOv11 algorithm model to detect and train fan blade defects, obtaining a trained YOLOv11 model Step 5.1, model initialization: Initialize the optimized YOLOv11 algorithm model, set the network weights, learning rate, batch size, and the number of training iterations.

[0045] Step 5.2, model pre-training: Input the training set samples into the improved YOLOv11 algorithm model for pre-training, calculate the loss between the predicted frame and the target GT frame, and optimize the network weights through backpropagation.

[0046] Step 5.3, Model Evaluation: Evaluate the model's performance using the validation set and calculate the mean average precision (mAP) value for the surface defect classes.

[0047] Step 5.4, Model Iterative Training: Repeat Step 5.2 and Step 5.3 until the mAP value converges to a stable state, completing the training of the improved YOLOv11 algorithm model.

[0048] Step 5.5, Model Testing: Test the trained model using the test set to evaluate its comprehensive detection performance, including detection accuracy, detection speed, floating-point computation, parameter quantity, and image detection frame rate per unit time.

[0049] Step 6: Use the trained improved YOLOv11 algorithm model to detect fan blade defects Use the trained improved YOLOv11 algorithm model for fan blade defect detection and apply it to actual fan blade defect recognition and positioning tasks to ensure efficient and accurate defect detection.

[0050] Through the detailed description of the above embodiments, the improved YOLOv11-based fan blade defect detection method proposed by the present application is fully demonstrated, including data preparation, model improvement, training, and evaluation. This method introduces WTConv into the C3K2 module to form the C3K2-WTConv module, adds the MLCA mixed local channel attention mechanism to the C2PSA module to form the C2PSA-MLCA, and uses the Shape-IoU loss function. This method improves the accuracy and speed of fan blade defect detection, providing strong support for wind farm operation and maintenance.

[0051] Example 2 In another preferred embodiment, this embodiment further details the improved YOLOv11-based fan blade defect detection method of the present application based on Example 1, providing a specific embodiment to fully demonstrate the technical solution of the present application and its practical application effect.

[0052] Step 1: Construct a fan blade defect dataset Data collection: Collect fan blade images taken during wind farm operation and combine publicly available defect samples to form a multi-type defect coverage image base set, such as cracks, erosion, holes, wear, burning, damage, and dirt. Data labeling: Use Labelimg and other labeling software to label defects in images, generate standard format annotation files containing defect labels and coordinate data; Data augmentation and division: Use affine transformation, brightness disturbance, Gaussian noise injection, etc. to perform image enhancement operations to expand sample diversity. The enhanced data is divided into training set and validation set according to 8:2 to ensure the consistency of training and validation sample distribution.

[0053] Step 2: Optimize YOLOv11 algorithm model Step 2.1, introduce wavelet transform WTConv in C3K2 module, build C3K2-WTConv to replace C3k2 module in YOLOv11: Lightweight convolution kernel replacement: In the main network of YOLOv11, introduce C3K2-WTConv module, replace traditional convolution structure with deformable width and height convolution kernel to reduce model parameter quantity and improve feature extraction flexibility; Structural fusion mechanism: The module processes different scale information through double-channel paths, effectively improving the size adaptability of the defect target; Spatial dynamic receptive field expansion: Through dynamically learnable spatial receptive kernel weights, fine perception of complex structure regions in leaf image is realized, and the recognition ability of the model for special-shaped defects is enhanced; Channel compression and expansion strategy: Use 1x1 convolution and group convolution combination strategy to compress channels while improving information expression efficiency and optimizing computing resource utilization efficiency.

[0054] Step 2.2, introduce MLAC mixed local channel attention in C2PSA to form C2PSA-MLAC module Enhance fine-grained feature extraction capability: MLCA can weight process local channels of image feature map, emphasize the response degree of local key area on different channels, and improve the recognition effect of small cracks, edge peeling and other fine-grained defects on fan blades; Improve the efficiency of information interaction between channels: Traditional channel attention focuses on global average weight distribution, while MLCA focuses attention in local range, making the information flow between adjacent channels more accurate, which helps to retain structural features and texture differences; Reduce background interference and improve detection robustness: MLCA has an inhibitory effect on background noise in fan blade images. By focusing on effective channel areas, it reduces the model's misjudgment of irrelevant features, thereby improving the model's detection stability in complex backgrounds; Balancing local and global feature fusion: Under the MLCA mechanism, local channel attention is combined with global spatial attention in the C2PSA module, achieving "channel-space" dual-dimensional enhancement, effectively improving the model's generalization ability in multi-scale and multi-shape defect detection; Lightweight design, low computational overhead: MLCA uses lightweight operations such as group convolution and region normalization to control computational resource consumption while improving model performance, making it suitable for online real-time defect detection scenarios for fan blades.

[0055] Step 2.3, replace the loss function Replace the CIoU loss function in the original YOLOv11 algorithm with the Shape-IoU loss function; Shape-IoU enhances the sensitivity to target shape geometric differences while maintaining IoU consistency, providing smoother and discriminative gradient feedback, which helps improve bounding box fitting accuracy. Especially for long cracks or irregular shape defects in fan blades, Shape-IoU can effectively improve the regression quality of the model, reduce the miss and false detection rate, and speed up the convergence of model training.

[0056] Step 3: Model training and evaluation Experimental environment configuration: In the Windows11 system, use the PyTorch deep learning framework and Python development language for model training, configure the relevant environment versions such as Python3.8, PyTorch2.3.0, CUDA12.1, etc.; Model initialization: Initialize the network weights, learning rate, batch size, and training iteration number of the optimized YOLOv11 algorithm model, set the initial learning rate to 0.01, the batch size to 4, and the iteration number to 150 times; Model training: input the samples in the training set into the improved YOLOv11 algorithm model for pre-training, calculate the loss value between the predicted box and the target box GT (Ground Truth Box), and propagate the loss value back to optimize the network weights; Model evaluation: evaluate the performance of the improved YOLOv11 algorithm model using the validation set, calculate the average precision value (mAP) of the surface defect class to measure the detection performance; Model testing: test the trained improved YOLOv11 algorithm model using the test set to evaluate its comprehensive detection performance, evaluation indicators include mAP, recall rate (Recall), average IoU of bounding box regression, and model floating point computation (FLOPs), parameter quantity (Params), and picture detection frame rate per unit time (FPS); Ablation experiment: In order to further verify the feasibility and effectiveness of simultaneously applying the C3K2-WTConv module, the C2PSA-MLCA module and the Shape-IoU loss function, an ablation experiment is performed to analyze the improvement of each module on the overall detection performance by comparing the detection performance when each module is applied alone.

[0057] Through the above detailed steps, the specific application and actual effect of the application in the field of fan blade defect detection are fully demonstrated.

[0058] Embodiment 3 In another preferred embodiment, based on embodiment 2, the embodiment provides a fan blade defect detection method based on improved YOLOv11, and the specific implementation steps are as follows.

[0059] Step 1: In this embodiment, fan blade defect images are obtained by combining field collection with public data, and more advanced data enhancement methods are used, such as image random flipping, noise disturbance, brightness adjustment, etc., to improve the model's adaptability to complex scenes. The constructed dataset contains a total of 8614 images, which are divided into training set and validation set according to the ratio of 8:2, to ensure the stability and universality of training and evaluation.

[0060] Step 2: YOLOv11 detection algorithm structure optimization Step 2.1: Introducing wavelet transform to construct C3K2-WTConv module In embodiment 3, to solve the problem of limited expression ability of the original feature extraction module of YOLOv11, wavelet transform convolution (WTConv) is introduced into the C3K2 module to construct the C3K2-WTConv structure. This module can extract more rich frequency and spatial features, significantly improve the network's response ability to fan blade cracks, scratches and other detailed targets, and enhance the multi-scale detection effect.

[0061] Step 2.2: Fusion of MLCA mechanism to construct C2PSA-MLCA module In embodiment 3, the mixed local channel attention (MLCA) mechanism is embedded in the channel attention structure C2PSA of YOLOv11 to construct the C2PSA-MLCA module. This module takes into account local detail perception and channel-dependent modeling, further improving the recognition accuracy and positioning precision of the model in processing complex textures and small defects.

[0062] Step 2.3: Use Shape-IoU instead of CIoU to optimize the bounding box regression To improve the boundary box fitting quality and training stability, the original CIoU loss function of YOLOv11 is replaced with Shape-IoU in this scheme. This loss function is more sensitive to the shape difference of the target box during regression, and can provide smooth and discriminative gradient feedback, thereby improving detection accuracy and accelerating model convergence.

[0063] Step 3: Multi-stage training strategy Unlike Example 2, in Example 3, the pre-implementation strategy is different. In this experiment, a multi-stage progressive training method is used. In the initial stage, the basic structure is used for fast warm-up training, then the C3K2-WTConv module and the C2PSA-MLCA module are gradually introduced, and the Shape-IoU loss function is enabled in the subsequent stage to continuously optimize network performance. The introduction order and weight configuration of each module and loss term are verified by experiments to ensure stable convergence and optimal performance.

[0064] By constructing a more rich data set and introducing C3K2-WTConv module, C2PSA-MLCA module and Shape-IoU loss function, and at the same time cooperating with multi-stage training strategy, this embodiment significantly enhances the performance of YOLOv11 in fan blade defect detection task, especially in identifying small and various shaped surface defects, showing higher precision and stability, and having good industrial application prospect.

[0065] Example 4 In another preferred embodiment, based on Examples 1 and 2, this embodiment provides a fan blade defect detection method based on improved YOLOv11, the specific implementation steps are as follows: Step 1: Construct a data set containing fan blade defects, label the image defects, and divide the data set into training and validation sets; Step 1, the data used in this study is derived from images taken by fan blade image acquisition equipment in wind farms and related image resources obtained from the network. After completing the initial data collection, in order to improve the generalization ability and robustness of the YOLOv11 model, the images were subjected to data enhancement processing. The enhancement operation covers a variety of image transformation methods such as random rotation, scaling, cropping and flipping, aiming to enrich sample diversity and alleviate model overfitting problems. Finally, the enhanced data set contains 8614 images of 7 defects including cracks, erosion, holes, wear, burning, damage and dirt, and is divided into training and validation sets in the ratio of 8:2 for model training and performance evaluation.

[0066] Step 2: Construct an improved YOLOv11 algorithm model and obtain an optimized YOLOv11 algorithm model; Step 2.1: YOLOv11 is a target detection algorithm model released by the Ultralytics team in 2024. Compared with the previous version, YOLOv11 has made many optimizations in the overall architecture, especially in feature extraction efficiency and detection accuracy. The model uses C3k2 modules instead of the original C2f structure to enhance the adaptability of the module. Then the C2PSA module is introduced in the feature fusion stage, which combines the improved C2f structure and the PSA attention mechanism, thereby enhancing the model's ability to extract key features. The C2PSA module is designed to integrate multi-head attention mechanism and feedforward neural networks (FFN), which can effectively integrate multi-dimensional feature information and preserve the complex relationship between space and semantics, improving the accuracy of feature representation. In order to further improve the inference speed and detection performance, YOLOv11 uses a decoupled structure in the detection head and introduces the SIoU loss function to optimize the positioning accuracy and convergence speed. Overall, YOLOv11 significantly improves the detection robustness and speed while maintaining the advantage of lightweight computation; Step 2.2: In the YOLOv11 structure, the original C3K2 module has good performance in multi-scale feature extraction, but when dealing with fine-grained targets such as fan blade surface cracks, scratches and other defects, edge information is easily lost by shallow convolution, affecting detection accuracy. To improve the model's ability to extract edge contours and local details, a wavelet transform is introduced into the C3K2 module to construct the C3K2-WTConv module. Wavelet transform is a multi-scale analysis method that has both time-frequency domain local analysis capabilities, which can decompose images into different frequency bands of information, making it easier to extract edge, texture and structural features of images. Unlike traditional convolution, wavelet transform has strong directionality and spatial locality, especially suitable for detecting image mutations and structural abnormalities, improving the sensitivity of defect detection. The WTConv module decomposes the image feature map by introducing discrete wavelet transform, splitting it into low-frequency and high-frequency subbands, and combining convolution operations to extract features of different frequency components. The low-frequency part retains the main structure of the image, and the high-frequency part highlights the edge and texture details, and the fusion of the two can significantly improve the perception ability of small targets. WTConv does not need to significantly increase the number of parameters, and has good computational efficiency and lightweight characteristics. The module structure of WTConv is shown in Figure 2 .

[0067] The specific operation process of C3K2-WTConv is as follows: In the C3K2-WTConv module, Haar wavelet transform (Haar WT, Haar Wavelet Transform) is used as a feature decomposition method. For the input image X, a layer of Haar wavelet transform along one spatial dimension can be achieved through a specific depth convolution, and the convolution kernel used is and , followed by a standard 2x downsampling operation. To complete a two-dimensional wavelet transform, the above process needs to be applied in the horizontal and vertical directions respectively, ultimately forming four filter channels with a convolution stride of 2. It is the overall information extraction of the image corresponding to the low-pass filter, 、 、 It is a set of high-pass filters that capture the edge features of the image in the horizontal, vertical and diagonal directions, respectively. See formula (1). For each input channel, the output after convolution processing is shown in formula (2): (1) (2) The system has four channels, each with X Half the resolution. represent X The low frequency of represents the image structure, while They are X High-frequency components in the horizontal, vertical, and diagonal directions retain edge information. The Inverse Wavelet Transform (IWT) performs a restoration operation through transposed convolution.

[0068] (3) By recursively decomposing the low-frequency components, cascade wavelet decomposition is achieved. The decomposition of each layer can be expressed as (4) Where, 、 、 、 Respectively expressed in The four sub-band images after the second wavelet transform are X The low frequency of represents the image structure, X High-frequency components in horizontal, vertical, and diagonal directions; It is the previous layer (i.e. The four sub-band images (the Layer wavelet transform X The low frequency of represents the image structure,X High-frequency components in the horizontal, vertical and diagonal directions) as the input of the wavelet transform of the current layer; It is a wavelet transform operation, which decomposes the input image into the four sub-bands mentioned above. Figure 2 middle, 、 、 、 Represent the first layer of wavelet transform X The low frequency of represents the image structure, X High-frequency components in the horizontal, vertical and diagonal directions; among them, is further decomposed into the input of the second layer wavelet transform; 、 、 、 Represent the second layer wavelet transform X The low frequency of represents the image structure, X High-frequency components in horizontal, vertical, and diagonal directions.

[0069] Compared to the C3K2 module in the original YOLOv11, C3K2-WTConv maintains feature expressiveness while maintaining a more gradual parameter growth, scaling only logarithmically rather than rapidly increasing with increasing receptive field. Furthermore, leveraging hierarchical wavelet decomposition, the WTConv module significantly expands the model's receptive range while maintaining minimal parameter overhead, effectively enhancing its ability to perceive the image's macroscopic structure. This performance is particularly strong in low-frequency response, improving the network's ability to discern shape and structural defects while reducing reliance on texture redundancy.

[0070] Step 2.3: In YOLOv11, the MLCA (Mixed Local Channel Attention) mechanism is introduced and integrated with the original C2PSA (Cross-Stage Partial Squeeze-and-Attention) module to build the C2PSA-MLCA module, which significantly improves the model's feature modeling ability and detail perception ability in fan blade defect detection. Traditional attention mechanisms often focus on one dimension of channel or spatial information, making it difficult to balance the local structure and inter-channel fine-grained dependence. The MLCA mechanism, by fusing local attention and inter-channel interaction, can capture small abnormal areas in images more carefully, especially suitable for detecting small-scale defects such as cracks, peeling, and rust in fan blades. The C2PSA-MLCA module combines the advantages of cross-stage partial connection and multi-dimensional attention fusion in structure. The C2PSA module enhances feature flow through partial residual connection to reduce information loss, while the MLCA further enhances the selective attention ability to local details and edge features. This fusion structure can effectively improve the accuracy and stability of feature extraction while maintaining computational efficiency. In addition, MLCA has a lightweight structure that can be flexibly embedded in the YOLOv11 network without significantly increasing model complexity, meeting the dual requirements of detection speed and accuracy in actual industrial scenarios.

[0071] The C2PSA-MLCA module improves the comprehensive ability of YOLOv11 in multi-scale feature perception, target region focusing, and small target recognition by jointly modeling in channel and local spatial dimensions, providing strong support for improving the accuracy and robustness of fan blade defect detection. Figure 3 The module structure diagram of MLCA, Input is the input, Output is the output, LAP is the local average pooling, GAP is the global average pooling, Reshape is the feature rearrangement operation, Conv1d is the feature conversion, UNAP is the de-pooling operation.

[0072] MLCA is an attention mechanism used to enhance the feature representation capability of deep learning models. Its core idea is to enable the model to capture key features in the input data more effectively through multi-level and multi-channel information fusion. In the MLCA structure, the input data is first processed by the LAP (Local Average Pooling) layer for local average pooling, which helps to extract local feature information of the input data. Next, the processed feature map enters two branches: one branch directly performs global average pooling (GAP) to obtain global features; the other branch first reshapes and then performs one-dimensional convolution through the Conv1d layer to further extract features. These two branches represent different levels of understanding of the features. The global average pooled feature map also needs to be reshaped and processed by the UNAP (Unpooling and Non-linear ActivationPooling) layer for unpooling and nonlinear activation to restore some spatial information and enhance feature expression capability.

[0073] Subsequently, the output results of the two branches are again converged at the LAP layer and fused through point multiplication. This fusion method effectively combines local features with global features, enhancing the model's perception of features at different scales. Finally, the fused feature map is weighted and fused with the original input data to generate the final output result. The entire MLCA attention mechanism significantly improves the model's recognition accuracy and robustness for complex data patterns through multi-level and multi-channel feature extraction and fusion.

[0074] Step 2.4: Replace the original CIoU loss function of YOLOv11 with Shape-IoU; Replacing the original CIoU (Complete IoU) loss function in YOLOv11 with the Shape-IoU loss function can more effectively improve the model's sensitivity to target shape matching and positioning accuracy. Shape-IoU further considers the similarity of the shape structure between the predicted box and the real box based on the traditional IoU metric, especially in handling elongated, curved, or irregular boundary targets (such as cracks and peeling in fan blades). Compared to CIoU, which only relies on center point distance, aspect ratio, and overlap area for optimization, Shape-IoU can more accurately reflect the true shape of the target, prompting the model to better fit the target contour when regressing the boundary box, thereby improving overall detection performance and detection stability for small targets.

[0075] The Shape-IoU loss function is defined as follows: (5) (6) (7) (8) (9) (10) where IoU (Intersection over Union) is a basic measure to evaluate the overlap between two bounding boxes, given a predicted bounding box B and a ground truth bounding box B gt , the area of their union. represents the intersection area between the predicted and ground truth bounding boxes, and a higher IoU value indicates a closer prediction. and are used to calculate the weight coefficients, which adjust the differences in horizontal and vertical directions according to the aspect ratio of the ground truth bounding box, and represent the width and height of the ground truth bounding box, respectively, is a factor to adjust the influence of target size in the dataset. is a measure of the shape deviation between the predicted and ground truth bounding boxes, and are the center coordinates of the predicted bounding box, and are the center coordinates of the ground truth bounding box, is a normalization constant to ensure the scale of the shape distance is reasonable. the shape penalty term, is a power parameter, usually set to 4, to adjust the influence of deviation and enhance the sensitivity to shape differences; and are the shape difference weights in width and height directions, and are the width and height of the predicted bounding box; is the expression of the Shape-IoU loss function; is a weight or bias term; is the negative power form of the natural exponential function e , and as increases, the value of

[0076] Figure 4For the improved YOLOv11 network structure, Backbone is the backbone network; Neck is the neck; Head is the head; Conv module is mainly used to extract local feature information in the image, and performs convolution operation to enhance the feature expression ability; C3k2-WTConv module is a structure based on the traditional C3K2 structure and fused with wavelet transform, which has stronger structure feature extraction ability and larger receptive field, while suppressing invalid texture and improving the sensitivity to target edges; SPPF (Spatial Pyramid Pooling-Fast) module realizes multi-scale feature fusion through different scale maximum pooling operations, thereby enhancing the context understanding ability of the network; C2PSA-MLCA module combines hybrid local channel attention (MLCA) to highlight key areas in the feature fusion stage, effectively improving the detection performance of small targets; Upsample module is used for up-sampling processing, which restores low-resolution feature maps to high-resolution for fusion with shallow features; Concat module realizes the splicing and fusion of feature maps of different scales, which is beneficial to the integration and sharing of semantic information; V11Detect module as the detection head, respectively, for different scale fusion feature map target classification and boundary box regression, complete the final defect detection task. The overall module cooperates, effectively improves the accuracy and robustness of YOLOv11 in fan blade defect detection.

[0077] Step 3: Based on the optimized YOLOv11 algorithm model, train using the fan blade defect dataset to obtain the trained improved YOLOv11 algorithm model Step 3.1: The experiment was carried out in the Windows 11 operating system environment, using the PyTorch deep learning framework and Python programming language. The environment configuration includes Python 3.8, PyTorch 2.3.0, and CUDA 12.1 version. The hardware used in the experiment is a 13th generation Intel Core i5-13400F (2.50GHz) processor and a NVIDIA GeForce RTX 4060 graphics card. The input image resolution is set to 640x640, which effectively balances the consumption of computing resources while ensuring sufficient detail expression. The batch size is set to 4, which maintains the stability of the training process and controls the use of video memory; the initial learning rate is 0.01, which helps the model quickly enter the convergence stage in the early stage of training and avoids large fluctuations. The entire training process is performed for 150 iterations; Step 3.2: In this experiment, multiple performance indicators are selected to evaluate the model performance, including floating-point operations (FLOPs), recall (Recall, including overall recall Rall), average precision (AP), precision (Precision, P), mean average precision (mAP), and total number of model parameters. mAP includes two levels: mAP@0.5 represents the average precision at an IoU threshold of 0.5, while mAP@0.5:0.95 represents the average precision when taking multiple thresholds within the range of 0.5 to 0.95, reflecting the detection stability of the model under different judgment criteria. Accuracy is used to measure the proportion of correctly identified positive samples, recall represents the proportion of actual positive samples identified by the model, and average precision is calculated by the area under the precision-recall curve, which is used to comprehensively evaluate the model performance. is the average value of all classes Average Precision (AP) value, representing the overall detection ability. FLOPs (Floating-point operations per second) measures the computational resources required by the model, reflecting its running efficiency, while the parameter quantity reflects the structural complexity of the model, i.e., the total number of parameters to be trained in the model, with the following calculation formula: (8) (9) (10) (11) In the formula, True Positives represents the number of correctly classified positive samples; False Positives represents the number of incorrectly classified positive samples; False Negatives represents the number of samples that failed to be correctly classified as positive. Precision at a given recall value, is the total number of classes, is the average precision of the th class.

[0078] To further verify the feasibility and effectiveness of the optimized YOLOv11, an ablation experiment is conducted, as shown in Table 1: Table 1

[0079] According to the table, after the ablation experiment, the basic performance of YOLOv11 without introducing any module is P 72.3%, R 73.9%, mAP@0.5 73.8%, and the parameter amount is 19.2M. After introducing the C3K2-WTConv module alone, the mAP of the model is improved to 75.3%, indicating that the wavelet transform enhances the feature expression ability; after introducing the C2PSA-MLCA module, the mAP is increased to 76.3%, and the parameter amount is reduced to 18.8M, which reflects that the module has higher computing efficiency while improving the detection performance; replacing the Shape-IoU loss function makes the mAP reach 75.0%, which shows that it is effective for the optimization of target positioning. After combining any two modules, the detection performance is further enhanced, among which the combination of C2PSA-MLCA and Shape-IoU performs particularly outstanding, with mAP reaching 78.6%. When all three are integrated, the model achieves the best performance in all indicators, with P 78.3%, R 80.3%, and mAP@0.5 82.5%, which is significantly improved compared with the original YOLOv11, verifying the effectiveness and synergistic gain advantage of the proposed improvement method in the fan blade defect detection task.

[0080] Step 4: Evaluate the trained YOLOv7 algorithm.

[0081] In order to comprehensively verify the superiority of the method proposed in the present application in detecting fan blade defects, comparative experiments were conducted with the current mainstream target detection algorithms, including YOLOv5, YOLOv7, YOLOv8, YOLOv9 and YOLOv10. The comparative experiments were conducted under the same configuration, same data set and same parameters, and the experimental results are shown in Table 2. Table 2

[0082] Table 2 shows the results of various wind turbine blade defect detection algorithms. The table shows that the proposed improved algorithm achieves superior detection performance while maintaining a low parameter count. The proposed algorithm achieves a mAP@0.5 of 82.5%, significantly outperforming YOLOv5 (72.4%), YOLOv8 (72.1%), and YOLOv10 (72.4%). Compared to YOLOv5 (72.4M, 22.5M, and 47.1M parameters, respectively), its parameters are only 21.2M, demonstrating higher detection efficiency and lightweight design. While YOLOv7 achieves a mAP of 77.4%, higher than the original YOLOv11 architecture's 73.8%, its parameter count is as high as 74.9M, far exceeding the proposed algorithm. When compared with the YOLOv11 baseline model (73.8%, 19.2M), the algorithm of the present invention improves the accuracy to 82.5% with only a slight increase in parameters, verifying the significant performance gain after the introduction of C3K2-WTConv, C2PSA-MLCA modules and Shape-IoU loss function, and demonstrating better comprehensive detection capabilities and practical application value.

[0083] Figure 6 This is the detection result of the improved YOLOv11 algorithm. Figure 5 This is the original YOLOv11 algorithm in the model detection result. The detection results are displayed by selecting some pictures, where Cracking is crack, Edge_Erosion is erosion, and Tape_Erosion is wear. The comparison results are as follows Figure 5 and Figure 6 As shown in the figure, the experimental results show that the improved algorithm has a higher accuracy than the YOLOv11 algorithm in detecting defects, which also proves that the improvement of the algorithm has a good effect.

[0084] The present invention provides a method for detecting defects in wind turbine blades based on an improved YOLOv11 algorithm, which proposes an innovative solution to the problems of low detection accuracy, slow detection speed, and high model complexity in the field of wind turbine blade defect detection. Compared with traditional detection methods such as manual inspections and drone inspections, the method of the present invention not only has higher efficiency and lower costs, but can also better adapt to complex environmental changes. At the same time, in response to the problems faced by existing technologies such as YOLOv5, YOLOv7, YOLOv8, YOLOv9, YOLOv10 and other algorithms in the task of detecting defects in wind turbine blades, such as insufficient small target defect recognition capabilities, large background noise interference, and excessive model calculation complexity, the present invention has been deeply customized and optimized.

[0085] This paper deeply optimizes the YOLOv11 algorithm and proposes the following three key improvements: (1) Introduce wavelet transform WTConv to replace C3K2 module, construct C3K2-WTConv module The present application first introduces wavelet transform (WTConv) in the C3K2 module of YOLOv11. Wavelet transform can effectively extract frequency domain features and enhance the model's ability to capture the details of fan blade surface defects. Through this transformation, the model can better handle subtle defect information, thereby improving the detection accuracy of fan blade defects.

[0086] (2) Introduce MLAC mixed local channel attention mechanism in C2PSA module to form C2PSA-MLAC module To further improve the model's feature extraction ability in fan blade defect detection, the present application introduces MLAC mixed local channel attention mechanism (MLAC) in the C2PSA module. This mechanism can adaptively adjust the weight relationship between channels, allowing the model to focus on key areas when dealing with defects of various scales, improving the recognition accuracy of small target defects while reducing computational burden.

[0087] (3) Use Shape-IoU loss function to replace CIOU loss function The present application also uses Shape-IoU loss function to replace CIOU loss function in YOLOv11. Shape-IoU loss function considers the influence of the shape of the bounding box, allowing the model to more accurately fit the morphological characteristics of the fan blade during detection, thereby improving the positioning accuracy of the detection.

[0088] In terms of data set construction, the present application constructs a high-quality data set containing various fan blade defects (such as cracks, burns, holes, deformities, etc.), and enhances the diversity of the data set through data augmentation techniques such as rotation, scaling, cropping, etc., further improving the model's generalization ability.

[0089] Through these innovative algorithm improvements and module optimizations, the present application effectively improves the precision and speed of fan blade defect detection, while reducing computational complexity and parameter quantity, enhancing the model's practicality. In practical applications, the method of the present application has been verified in wind farm environments and has achieved significant detection results, proving its effectiveness and reliability.

[0090] In summary, the fan blade defect detection method based on improved YOLOv11 algorithm of the present application significantly improves detection performance through the introduction of wavelet transform, mixed local channel attention mechanism and Shape-IoU loss function, etc. innovative technologies, has high academic value and engineering application potential, and can provide efficient and reliable defect detection solutions for intelligent operation and maintenance of wind farms.

Claims

1. A fan blade defect detection method based on improved YOLOv11 is characterized by: The following steps are involved: Step 1: Construct an image dataset containing surface defects of wind turbine blades, and annotate and preprocess the dataset; Step 2: Improve the YOLOv11 algorithm by introducing the wavelet transform WTConv into the C3K2 module to form the C3K2-WTConv module, replacing the original C3K2 module in the YOLOv11 backbone network; Step 3: Based on the C2PSA attention module in the improved YOLOv11, the MLCA hybrid local channel attention mechanism is introduced to form the C2PSA-MLCA module to enhance the feature extraction capability of the model; Step 4: Use the Shape-IoU loss function to replace the CIoU loss function in the original YOLOv11 algorithm; Step 5: Use the improved YOLOv11 algorithm model to train the fan blade defects to obtain a trained improved YOLOv11 algorithm model; Step 6: Use the trained improved YOLOv11 algorithm model to detect wind turbine blade defects.

2. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 1 is characterized in that: The types of surface defects of the wind turbine blades in the image dataset in Step 1 include cracks, erosion, holes, wear, burns, damage and dirt. The dataset is expanded by rotation, scaling and cropping, and divided into a training set and a validation set according to a predetermined ratio.

3. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 2 is characterized in that: The C3K2-WTConv module in Step 2 is combined with wavelet transform to enhance the extraction of multi-scale features, effectively retaining texture information and edge structure, and improving the feature expression ability and spatial perception ability of wind turbine blade defect detection.

4. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 3 is characterized in that: The steps of replacing the C3K2 module in the YOLOv11 backbone network with the C3K2-WTConv module in Step 2 include: Step 2.1: Perform channel compression on the input feature map and fuse local information through convolution; Step 2.2: Introduce wavelet transform operation to perform frequency domain enhancement on the feature map to improve the expression ability of texture and edges; Step 2.3: Fuse the wavelet features with the original feature map and preserve the context information through residual connections; Step 2.4: Combine the nonlinear activation function with the normalization operation to output the enhanced multi-scale feature map to enhance the overall representation ability of the model.

5. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 4 is characterized in that: The C2PSA-MLCA module in Step 3 integrates channel attention and local channel interaction mechanisms to improve the model's responsiveness to small targets and edge defects. The steps are as follows: Step 3.1: Introduce the C2PSA attention module to perform channel compression and spatial reconstruction on the input features; Step 3.2: Combined with the MLCA mechanism, local channel interactions are introduced to significantly enhance fine-grained features; Step 3.3: Normalize the fused attention weights to improve the model's focus on key areas; Step 3.4: Feed the enhanced features back to the backbone network to improve the detection accuracy and robustness of wind turbine blade defects.

6. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 5 is characterized in that: The Shape-IoU loss function in Step 4 introduces a target contour shape matching mechanism to improve the ability of the target box to fit the real boundary and enhance the positioning accuracy of complex defect structures.

7. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 6 is characterized in that: The specific steps of replacing the original CIoU in YOLOv11 with the Shape-IoU loss function in Step 4 include: Step 4.1: Construct a loss function based on the boundary shape similarity between the predicted box and the true box; Step 4.2: Fusion of IoU and boundary contour differences to form a composite loss that includes geometric structure error; Step 4.3: Introduce a shape penalty term to optimize the irregular target boundary and improve the accuracy of defect location; Step 4.4: Optimize the model regression capability through gradient backpropagation to improve the overall detection performance and convergence speed of the model.

8. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 7 is characterized in that: The training steps for improving the YOLOv11 algorithm model in Step 5 include: Step 5.1: Set parameters, including network initial weights, learning rate, and batch size; Step 5.2: Input the training set image, forward propagate the predicted target box, and calculate the loss through Shape-IoU and back propagate to optimize the network; Step 5.3: Evaluate the model performance on the validation set and calculate the evaluation accuracy including the mAP indicator; Step 5.4: Iterate the training until the model converges and output the optimal weight model; Step 5.5: Use the test set to evaluate the detection capability and stability of the model under different defect categories.

9. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 8 is characterized in that: The model performance evaluation includes the following indicators: detection accuracy, detection speed, FLOPs, number of model parameters and number of frames processed per unit time FPS.

10. The fan blade defect detection method based on the improved YOLOv11 algorithm according to claim 9 is characterized in that: The evaluation model includes the following steps: Step 5.5.1: Compare the detection indicators of the YOLOv11 algorithm before and after improvement on the same dataset, including mAP, R and IoU; Step 5.5.2: Evaluate the changes in detection speed and complexity of the model after introducing the C3K2-WTConv, C2PSA-MLCA modules and Shape-IoU loss function; Step 5.5.3: Quantitatively analyze the contribution of each module in improving small target recognition, edge contour extraction, and false detection rate control; Step 5.5.4: Compare and verify with other advanced target detection algorithms under the same experimental conditions to prove the superiority and application potential of the method of the present invention in the task of wind turbine blade defect detection.

Images