A method and system for characterizing and detecting multiple types of surface damage on leaves

Abstract

This invention belongs to the field of intelligent detection technology for wind turbine blades, specifically relating to a feature detection method and system for multi-category surface damage on blades. The method includes: constructing a pyramid network structure EFPN to replace the baseline model's FPN+PAN neck network. The EFPN includes a pyramid feature enhancement (PFE) module, a multi-feature fusion (MFF) module, and a dynamic interpolation fusion (DIF) module, used to integrate multi-scale damage features and suppress background noise; designing a Focaler PIoU loss function, combined with a size-adaptive penalty factor and a dynamic sample weighting mechanism, to improve the model's ability to perceive difficult samples and the accuracy of bounding box regression; and using the DepGraph structured pruning method to compress the network, combined with an HDCFR hybrid knowledge distillation strategy, to reduce model complexity while maintaining detection accuracy. This invention achieves an mAP@0.5 of 84.5% on the WTBs GLE dataset, with reduced parameter and computational costs, combining high accuracy, lightweight design, and strong robustness, making it suitable for deployment on UAV edge devices.

Description

Technical Field

[0001] This invention belongs to the field of intelligent detection technology for wind turbine blades, specifically relating to a feature detection method and system for multiple types of surface damage on blades. Background Technology

[0002] Wind energy, due to its advantages such as being renewable, low-cost, and pollution-free, has gained widespread recognition and application in green energy in recent years. To improve power generation efficiency, most wind farms are built in areas rich in wind resources, such as nearshore areas, deserts, and high-altitude regions.

[0003] Blades are one of the key components of wind turbines, responsible for converting wind energy into mechanical energy. Their operational health directly affects the power generation efficiency of the wind turbine. However, blades often face external challenges during their service life, such as salt spray, sandstorms, lightning strikes, and icing. These factors can lead to structural damage over long-term operation, and it is estimated that their maintenance costs account for approximately 15% to 20% of the total cost of a wind turbine. Therefore, efficient and accurate detection of early blade damage is crucial. This can prevent further structural deterioration and effectively reduce downtime for wind turbine maintenance, thereby lowering the operating costs of wind farms.

[0004] Traditional blade damage detection methods mainly rely on visual inspection, vibration detection, and ultrasonic and infrared detection. However, these methods are limited by environmental interference, periodic calibration, and safety risks, resulting in low detection accuracy and difficulty in meeting the high-efficiency detection requirements of increasingly larger blades. To overcome the limitations of traditional detection methods, deep learning-based UAV visual inspection has become the mainstream solution in recent years. It mainly uses YOLO series models for blade damage detection. For example, the current mainstream FE-YOLO11n model mainly includes a variable receptive field C3k2VRF module, a Sobel filter downsampling SFD module, and an inverted residual multi-scale attention C2IRMA module. Although the detection accuracy and efficiency are improved compared to traditional detection methods, this method is limited by factors such as the amplification of background noise under complex lighting conditions, the neglect of sample distribution differences, and the large number of model parameters. This leads to the dilution of acquired damage features, excessive penalty for bounding boxes of low-quality images, and low confidence in the detection of defects such as oil stains and cracks, which can easily cause missed detections or high false detection rates of leading-edge corrosion. Summary of the Invention

[0005] To address the problems existing in the prior art, the purpose of this invention is to provide a feature detection method and system for multiple types of surface damage on wind turbine blades. This method and system take into account the high-precision detection requirements for multiple types of damage such as sand holes and leading-edge corrosion under complex lighting conditions, as well as the low-resource deployment requirements for edge devices. The method and system have fewer model parameters and effectively reduce the missed detection and false detection rates of wind turbine blades.

[0006] The technical solution of this invention is: A method for feature detection of multiple types of surface damage on leaves, comprising the following steps: Acquire raw image data of the surface of the wind turbine blade to be inspected; The original image data is input into a pre-trained lightweight adaptive feature enhancement detection model, which outputs detection results including damage category, confidence level, and bounding box coordinates. The lightweight adaptive feature enhancement detection model is constructed and trained in the following manner: A baseline model based on the YOLO architecture is constructed, which includes a backbone network, a neck network, and a detection head. The backbone network receives the original image data and outputs multi-scale feature maps. The neck network is a pyramid network structure used to fuse the multi-scale feature maps at different levels and finally output the detection results through the detection head. The pyramid network structure in the baseline model is structured and pruned using the DepGraph model pruning method to obtain the pruned baseline model. The Focaler-PIoU loss function is used as the regression loss function of the detection head to optimize the pruned baseline model, obtain the optimized baseline model, and train the optimized baseline model based on the HDCFR hybrid knowledge distillation method to obtain the final lightweight adaptive feature enhancement detection model.

[0007] Preferably, the pyramid network structure includes a pyramid feature enhancement module, a multi-feature fusion module, and a dynamic interpolation fusion module; The pyramid feature enhancement module is used to receive feature maps of different levels output by the backbone network. By stitching the feature maps of different levels together, and capturing global context information in the horizontal and vertical directions based on the rectangular self-calibration module, the key damage features of the feature maps of different levels are integrated to obtain the key damage features after integration of different levels. Based on the key damage features integrated at different levels, the multi-feature fusion module and the dynamic interpolation fusion module work together to align and fuse the channels of high-level semantic features and low-level detail features in the key damage features across different levels.

[0008] Preferably, the Focaler-PIoU loss function is constructed by combining the PIOU loss term and the Focaler weight term, and is used to dynamically adjust the regression weights for samples of different difficulty levels, determined according to the following formula: , in, , , , , , In the formula: The Focaler-PIoU loss function is used. For PIOU loss term, Let be the intersection-union ratio (IoU) of the predicted bounding box and the ground truth bounding box. For Focaler weight terms reconstructed based on dynamic thresholds, A penalty factor adapted to the target size is introduced into the PIOU loss term. , , and This represents the absolute value of the distance between corresponding edges of the predicted bounding box and the target bounding box. Indicates the threshold for identifying difficult samples. , This represents the threshold for simple sample recognition. .

[0009] Preferably, the DepGraph model pruning method includes the following steps: decomposing the optimized baseline model into a layer sequence consisting of several input-output pairs, identifying inter-layer dependencies and intra-layer dependencies, constructing a dependency graph that can display the dependency relationships based on the identified inter-layer dependencies and intra-layer dependencies, and performing grouped synchronous pruning of the basic dependencies, residual dependencies, and splicing dependencies in the network according to the dependency graph.

[0010] Preferably, the dependency graph is determined according to the following formula: , In the formula: For dependency graph, For the first i The input of the layer sequence, For the first j The output of the layer sequence, For connections between two adjacent layers, For logical OR operation, For logic and operations, It is "true" when the condition is met. For the first i The pruning scheme for the layer input, For the first j The pruning scheme output by the layer.

[0011] Preferably, the hybrid knowledge distillation method combines feature space-based mask generation distillation and output probability distribution-based soft label distillation in a weighted manner, enabling the student model to simultaneously learn the intermediate layer feature representations and the final output decision distribution of the teacher model. Specifically, the mask generation distillation uses a randomly generated binary mask matrix to filter key regions in the feature map and guides the student model's feature map to align with the teacher model's feature map using a channel adaptation layer and a feature projection layer. The soft label distillation uses the difference in probability output between the teacher and student models after activation function processing as importance weights, and guides the student model's prediction distribution to approximate the teacher model's through a weighted bidirectional cross-entropy loss.

[0012] Preferably, the total loss function of the HDCFR hybrid knowledge distillation method is determined according to the following formula: , In the formula: Let be the total loss function of the hybrid knowledge distillation method. The original detection loss for the student model, To balance the hyperparameters of the loss, To generate distillation loss based on feature space masking, This is the soft-label distillation loss based on the output probability distribution.

[0013] Preferably, the teacher model is a YOLO11m model, and the student model is a pruned baseline model; during the distillation process, the parameters of the teacher model are fixed, and the parameters of the student model are updated by minimizing the total loss function.

[0014] Preferably, the dataset used for training the lightweight adaptive feature enhancement detection model includes: surface images of wind turbine blades collected under various lighting conditions, wherein the surface images are labeled with category labels and bounding box information for four types of damage: sand holes, leading edge corrosion, oil stains, and surface cracks.

[0015] A feature detection system for multiple types of surface damage on leaves includes: The image acquisition unit is used to acquire raw image data of the surface of the wind turbine blade to be inspected; The processing unit is communicatively connected to the image acquisition unit and is equipped with a trained lightweight adaptive feature enhancement detection model. It is used to receive the original image data and process the original image data to obtain damage detection results. The damage detection results include: damage category, confidence level, and bounding box coordinates.

[0016] Compared with the prior art, the feature detection method and system for multiple types of surface damage on blades of the present invention has the following advantages: This invention effectively addresses the problems of background information enhancement and damage feature dilution caused by the traditional FPN+PAN structure's indiscriminate use of features at the same level and adjacent scales, as well as features at the same scale in adjacent levels, during multi-scale feature fusion. By constructing a baseline model based on the YOLO architecture and replacing its neck network with a pyramidal network structure (EFPN), this invention significantly alleviates the problem of excessive penalty imposed on low-quality image bounding box regression due to ignoring sample distribution differences. This improves the bounding box localization accuracy and the ability to perceive difficult samples, such as oil stains and cracks. Furthermore, by combining the DepGraph structured pruning method with the HDCFR hybrid knowledge distillation strategy, the invention retains the key feature expression ability while compressing the number of model parameters, thus solving the problem of accuracy degradation during the lightweighting process. The resulting lightweight adaptive feature enhancement detection model significantly reduces the number of parameters and computational cost on the dataset compared to mainstream models. It balances the high-precision detection requirements of various types of damage such as sand holes and leading-edge corrosion under complex lighting conditions with the low-resource deployment requirements of edge devices, effectively reducing the missed detection and false detection rates of wind turbine blades. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the EFPN structure in an embodiment of the present invention; Figure 2 This is a schematic diagram of the HDCFR mixed knowledge distillation method in an embodiment of the present invention; Figure 3 This is a schematic diagram of the YOLO11n structure of the lightweight adaptive feature enhancement detection model in an embodiment of the present invention; Figure 4 This is a schematic diagram showing the performance comparison structure between the model and the baseline model and YOLO11n in an embodiment of the present invention; Figure 5 This is a flowchart of the detection method in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0020] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0021] Currently, for high-quality blade damage images periodically collected by drones at wind farms, the mainstream detection model FE-YOLO11n can improve the detection accuracy of irregularly shaped damage such as oil stains and surface cracks. However, because drones typically conduct regular inspections under sunny or cloudy conditions, the collected wind turbine blade surface damage images exhibit more complex background information under varying light intensities. This presents numerous challenges to current damage detection models, which still face issues such as low detection accuracy, false positives, and false negatives.

[0022] This invention, through the design of an efficient feature pyramid network neck structure EFPN and a Focaler-PIoU regression loss function, improves the mean accuracy (mAP@0.5) of detecting pinholes, leading-edge corrosion, oil stains, and surface cracks under different illumination intensities by 2.5%, reaching 82.1%. Furthermore, by introducing the DepGraph model pruning method and the proposed HDCFR hybrid knowledge distillation method, it reduces the increased complexity caused by replacing the neck structure and the decreased detection accuracy caused by network structure pruning, respectively. Based on these improved strategies, experiments on the WTBs-GLE dataset show that the lightweight adaptive feature enhancement detection model LAFE-YOLO reduces the number of parameters and computational cost by 24.8% and 20.0%, respectively, compared to FE-YOLO11n, while improving mAP@0.5 by 4.9%, reaching 84.5%. Furthermore, compared with current mainstream detection methods such as YOLO11s, YOLOv12s and Hyper-YOLOs, LAFE-YOLO achieves a better balance in detection accuracy, model complexity and computational cost, making it more suitable for deployment on edge devices. It also has stronger comprehensive detection performance for sand holes, leading edge corrosion, oil stains and crack damage under different lighting conditions.

[0023] To solve the following problem: (1) Solve the problem of partial background information feature enhancement caused by fusing neighboring scale features of the same level and the same scale features of adjacent levels in the existing neck network structure FPN+PAN.

[0024] (2) Solve the problem that the existing regression loss function CIoU is too heavily penalized for low-quality wind turbine blade damage images due to regression optimization based only on the geometric factors of the bounding box, resulting in a decrease in the consistency between the predicted bounding box and the actual damage location.

[0025] (3) Solve the problems of low detection accuracy, false detection and missed detection in the existing damage detection model for identifying common leaf surface damage under different lighting conditions.

[0026] See Figures 1 to 5 As shown, this embodiment provides a method and system for feature detection of multiple types of surface damage on blades, including the following core steps: S1. Obtain the raw image data of the surface of the wind turbine blade to be inspected; S2. Input the original image data into a pre-trained lightweight adaptive feature enhancement detection model. The forward propagation of this detection model outputs detection results containing damage category, confidence level, and bounding box coordinates. The lightweight adaptive feature enhancement detection model was constructed and trained in the following manner: S21. Construct a baseline model based on the YOLO architecture. The baseline model includes the backbone network, the neck network, and the detection head in sequence. S22. Replace the neck network of the baseline model with a pyramid network structure (EFPN) and use the Focaler-PIoU loss function as the regression loss function of the detection head to obtain the optimized baseline model. S23. Use the DepGraph model pruning method to perform structured pruning on the optimized baseline model to obtain the pruned baseline model; S24. The pruned baseline model is trained using a hybrid knowledge distillation method to obtain the final lightweight adaptive feature enhancement detection model.

[0027] Specifically, the above method will be further explained in detail below with reference to the accompanying drawings: (1) First, a baseline model (FE-YOLO11n) based on the YOLO architecture is constructed. YOLO (You Only Look Once) is a real-time object detection algorithm based on convolutional neural networks (CNNs). It takes the entire image as input and directly outputs the location and category information of the target. The baseline model of the conventional YOLO architecture mainly includes a backbone network, a neck network, and a head network (i.e., the detection head). The backbone network is responsible for extracting and processing multi-scale features, converting the original image data into multi-scale feature maps, and providing basic representations for subsequent detection tasks. For example, the backbone system of YOLOv12 consists of Conv (convolutional layers), C3k2, and A2C2f modules, with the core being the Residual Efficient Layer Aggregation Network (R-ELAN). The neck network is used to aggregate and refine the features extracted by the backbone network. The head network is used to generate the final detection prediction results.

[0028] This invention addresses the problem of damage feature dilution easily caused by the neck network structure FPN+PAN in the baseline model FE-YOLO11n during the feature fusion stage. To address this issue, a highly efficient Feature Pyramid Network (EFPN) structure is designed, such as... Figure 1 As shown, the core modules of the EFPN structure are the Pyramid Feature Enhancement (PFE) module, the Multi-Feature Fusion (MFF) module, and the Dynamic Interpolation Fusion (DIF) module. The PFE module is designed based on the Rectangular Self-calibration Model (RCM), which consists of a Rectangular Self-calibration Attention Mechanism (RCA), Batch Normalization (BN), and a Multilayer Perceptron (MLP). PFE first concatenates feature maps from different levels, then uses RCA in the RCM to capture global contextual information in the horizontal and vertical directions to integrate key damage features from each level. Finally, BN and MLP normalize the key damage features and calibrate the shape in the vertical direction, respectively. The MFF and DIF modules are used for multi-scale feature fusion. Through a combination of dynamic interpolation and 1×1 Conv (two-dimensional convolution), low-level and high-level feature maps are channel-aligned and fused to integrate damage feature information from different network layers and improve the model's ability to adapt to complex backgrounds. By combining the PFE, MFF, and DIF modules, EFPN enables the neck network to fuse damage features more efficiently and reduces erroneous enhancement of image background information.

[0029] (2) A Focaler-PIoU loss function is proposed. The Focaler-PIoU loss function combines the PIoU loss function with the Focaler-IoU loss function, improving the dynamic perception capability of the head network (detection head) for difficult samples while accelerating the network convergence process. The technical highlights of Focaler-PIoU mainly focus on two aspects: thresholding... and Dynamically assigning sample weights helps reduce the problem of baseline model CIoU loss functions ignoring differences in sample distribution; based on size-adaptive penalty factors... Achieving convergence of the straight path helps reduce unnecessary expansion of the predicted bounding box. The mathematical expression of the Focaler-PIoU loss function is shown below:

[0030] , in, , , , , , In the formula, The Focaler-PIoU loss function is used. For PIOU loss term, Let be the intersection-union ratio (IoU) of the predicted bounding box and the ground truth bounding box. for Loss item, This refers to the Focaler weights reconstructed based on dynamic thresholds. This represents the penalty factor for adapting to the target size. , , and This represents the absolute value of the distance between corresponding edges of the predicted bounding box and the target bounding box. and These represent the width and height of the predicted bounding box, respectively. and These represent the width and height of the target bounding box, respectively. and These represent the width and height of the intersection of the predicted bounding box and the target bounding box, respectively. This represents the reconstructed Focaler-IoU. Indicates the threshold for identifying difficult samples. . This represents the threshold for simple sample recognition.

[0031] (3) Introducing the DepGraph model pruning method (DependencyGraph: a general-purpose structured pruning method for architectures). This method addresses the problem of increased model complexity caused by replacing the EFPN structure by performing structured pruning on the network structure. For the three common dependency relationships—basic dependency, residual dependency, and concatenation dependency—the DepGraph pruning method first decomposes the deep neural network and then re-represents the network. The implementation process of this operation is as follows:

[0032] , , In the formula, and These represent the input and output of parameterized layers such as convolution or non-parameterized operations such as residual addition, respectively. For the first i The input of the layer sequence, For the first i The output of the layer sequence, symbol " "" indicates a connection between two adjacent layers.

[0033] Based on the above formula, DepGraph can identify two types of dependencies: when the network topology is known, and when the network has... In the connection layer, because, and When the corresponding intermediate features are the same, inter-layer dependencies can be identified; when the input and output of a network layer share the same pruning scheme, they can be pruned simultaneously, as shown below. At this point, intra-layer dependencies can be identified. Based on these two dependency rules, the final mathematical expression for dependency modeling is as follows:

[0034] , In the formula, For dependency graph, For the first i The input of the layer sequence, For the first j The output of the layer sequence, For connections between two adjacent layers, For logical OR operation, For logic and operations, It is "true" when the condition is met. For the first i The pruning scheme for the layer input, For the first j The pruning scheme output by the layer.

[0035] The DepGraph pruning method constructs a dependency graph to display the dependencies between modeling network layers, thereby simultaneously checking inter-layer dependencies caused by deep neural network connections and intra-layer dependencies between network layer inputs and outputs. It then identifies all input-output pairs to construct a dependency graph for grouped synchronous pruning. Compared with traditional pruning techniques, this method can solve the structural legality problem of pruning complex networks while maintaining a high compression ratio.

[0036] (4) A Hybrid Knowledge Distillation (HDCFR) method is proposed. This method combines feature-based distillation (MGD) and soft-label-based distillation (BCKD) to achieve hybrid distillation. Its goal is to simultaneously improve the student model's ability to represent features and learn from the teacher model's experience in handling difficult samples under the guidance of the teacher model. The teacher model is the YOLO11m model from the YOLO11 series. The YOLO11m model is between the lightweight YOLO11s and the high-performance YOLO11l. It uses the new C3k2 (Cross Stage Partial with kernel size 2) block to replace the traditional C2f module, which significantly enhances the model's spatial feature extraction capability and computational efficiency without increasing the number of parameters. The YOLO11m model adds a C2PSA (Cross Stage Partial with Spatial Attention) module to the backbone network. This spatial attention mechanism allows the model to "focus" on key regions in the image, greatly improving the detection of small targets in complex backgrounds and handling situations where objects occlude each other.

[0037] like Figure 2 As shown. The mathematical expression for the HDCFR method is as follows:

[0038] , , , , , , , , , In the formula, Let be the original detection loss of the student model, representing the original loss. The hyperparameter representing the balancing loss, This represents the total loss obtained through the hybrid distillation strategy, i.e., the total loss function of the hybrid knowledge distillation method. To generate distillation loss based on feature space masking, This is the soft-label distillation loss based on the output probability distribution. The horizontal coordinates of the feature map The vertical coordinates represent the feature map. This represents a random number in the interval (0,1). Hyperparameter representing the mask ratio, Representing the l Random mask for the layer, and The first two represent students and teachers respectively. Layer feature map, The representative adaptation layer (implemented by a 1×1 Conv) and These represent 3×3 Conv layers. Represents the projection layer. Representative characteristics. This represents the total number of distillation layers. , and These represent the number of channels, height, and width of the feature map, respectively. Represents the number of categories. Represents the binary classification score. This represents the importance weight between binary classification scores. express In the The position and the first Values ​​in each category express In the The position and the first Values ​​in each category This represents the binary cross-entropy loss.

[0039] The HDCFR hybrid knowledge distillation method, which combines feature-based MGD distillation and soft-label-based BCKD distillation, aims to reduce the following problems associated with single distillation: feature-based distillation alone struggles to accurately constrain the final prediction distribution of the student model, especially for complex detection categories; while soft-label-based distillation alone fails to learn rich intermediate layer information from the teacher model. In summary, HDCFR enables the student model to learn rich feature information from the teacher model in the intermediate layers and further refines the output decision.

[0040] (5) Based on the EFPN structure, Focaler-PIoU loss function, DepGraph model pruning method, and HDCFR hybrid knowledge distillation method, the final network structure diagram of the lightweight adaptive feature enhancement detection model LAFE-YOLO is as follows: Figure 3 As shown. Addressing the issue that FE-YOLO11n is easily affected by background information when detecting low-light and high-brightness images, leading to low damage recognition rates and missed detections, LAFE-YOLO significantly reduces model complexity and computational costs while maintaining detection accuracy, making it more suitable for deployment on edge devices.

[0041] Furthermore, to facilitate the implementation and application of the above method, this invention provides a system capable of implementing the above method, specifically: a feature detection system for multiple types of surface damage on blades, comprising: An image acquisition unit, configured on a drone or fixed monitoring device, is used to acquire raw image data of the surface of the wind turbine blades to be inspected. The processing unit is communicatively connected to the image acquisition unit and is equipped with a trained lightweight adaptive feature enhancement detection model for receiving the original image data and processing the original image data. The output unit is communicatively connected to the processing unit and is used to output the damage detection results obtained after the processing unit processes the original image data. The damage detection results include: damage category, confidence level, and bounding box coordinates.

[0042] Based on the above method and system design, the present invention has the following technical effects: (1) Based on the comparative experiment of detection performance, from Figure 4 It can be seen that the LAFE-YOLO model proposed in this invention has a stronger comprehensive detection capability for pinholes, leading edge corrosion, oil stains and surface cracks, while the baseline model FE-YOLO11n and the mainstream detection model YOLO11n have a worse ability to identify damage, especially oil stains and surface crack damage that are easily interfered with by background information. This is one of the advantages of this invention.

[0043] (2) By combining the designed EFPN structure and Focaler-PIoU loss function with the introduced DepGraph model pruning method and the proposed HDCFR hybrid knowledge distillation method, the final LAFE-YOLO detection model reduces the number of parameters and computational cost by at least 68.5% and 76.5% respectively compared with the mainstream detection models YOLOv8s, YOLOv9s, YOLO11s, YOLOv12s and Hyper-YOLOs, and the average accuracy mAP@0.5 is improved by at least 1.9%, as shown in Table 1. This achieves a better balance between detection accuracy, model complexity and computational cost, and is more robust to complex background information. It is also easier to deploy on edge devices such as UAVs to reduce computational resource consumption, which is one of the advantages of this invention.

[0044] Table 1. Performance comparison results with mainstream non-lightweight detection methods Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting the characteristics of multiple types of surface damage on leaves, characterized in that, Includes the following steps: Acquire raw image data of the surface of the wind turbine blade to be inspected; The original image data is input into a pre-trained lightweight adaptive feature enhancement detection model, which outputs detection results including damage category, confidence level, and bounding box coordinates. The lightweight adaptive feature enhancement detection model is constructed and trained in the following manner: A baseline model based on the YOLO architecture is constructed, which includes a backbone network, a neck network, and a detection head. The backbone network receives the raw image data and outputs feature maps at different levels. The neck network is a pyramid network structure used to fuse feature maps at different levels and finally output the detection results through the detection head. The pyramid network structure in the baseline model is structured and pruned using the DepGraph model pruning method to obtain the pruned baseline model. The Focaler-PIoU loss function is used as the regression loss function of the detection head to optimize the pruned baseline model, obtain the optimized baseline model, and train the optimized baseline model based on the HDCFR hybrid knowledge distillation method to obtain the final lightweight adaptive feature enhancement detection model.

2. The method for feature detection of multiple types of surface damage on blades according to claim 1, characterized in that, The pyramid network structure includes a pyramid feature enhancement module, a multi-feature fusion module, and a dynamic interpolation fusion module; The pyramid feature enhancement module is used to receive feature maps of different levels output by the backbone network. By stitching the feature maps of different levels together, and capturing global context information in the horizontal and vertical directions based on the rectangular self-calibration module, the key damage features of the feature maps of different levels are integrated to obtain the key damage features after integration of different levels. Based on the key damage features integrated at different levels, the multi-feature fusion module and the dynamic interpolation fusion module work together to align and fuse the channels of high-level semantic features and low-level detail features in the key damage features across different levels.

3. The method for feature detection of multiple types of surface damage on blades according to claim 1, characterized in that, The Focaler-PIoU loss function is constructed by combining the PIOU loss term and the Focaler weight term. It is used to dynamically adjust the regression weights for samples of different difficulty levels and is determined according to the following formula: ， in, , , , , , In the formula: The Focaler-PIoU loss function is used. For PIOU loss term, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. for Loss item, For Focaler weight terms reconstructed based on dynamic thresholds, A penalty factor adapted to the target size is introduced into the PIOU loss term. , , and This represents the absolute value of the distance between corresponding edges of the predicted bounding box and the target bounding box. and These represent the width and height of the predicted bounding box, respectively. and These represent the width and height of the target bounding box, respectively. and These represent the width and height of the intersection of the predicted bounding box and the target bounding box, respectively. Indicates the threshold for identifying difficult samples. , This represents the threshold for simple sample recognition. .

4. The method for feature detection of multiple types of surface damage on blades according to claim 1, characterized in that, The DepGraph model pruning method includes the following steps: decomposing the optimized baseline model into a layer sequence consisting of several input-output pairs, identifying inter-layer dependencies and intra-layer dependencies, constructing a dependency graph that can display the dependency relationships based on the identified inter-layer dependencies and intra-layer dependencies, and performing grouped synchronous pruning of basic dependencies, residual dependencies, and splicing dependencies in the network according to the dependency graph.

5. The method for feature detection of multiple types of surface damage on blades according to claim 4, characterized in that, The dependency graph is determined according to the following formula: ， In the formula: For dependency graph, For the first i The input of the layer sequence, For the first j The output of the layer sequence, For connections between two adjacent layers, For logical OR operation, For logic and operations, It is "true" when the condition is met. For the first i The pruning scheme for the layer input, For the first j The pruning scheme output by the layer.

6. The method for feature detection of multiple types of surface damage on blades according to claim 1, characterized in that, The HDCFR hybrid knowledge distillation method combines feature space-based mask generation distillation with output probability distribution-based soft label distillation in a weighted manner, enabling the student model to simultaneously learn the intermediate layer feature representations and final output decision distribution of the teacher model. Specifically, the mask generation distillation uses randomly generated binary mask matrices to filter key regions in the feature map and guides the student model's feature map to align with the teacher model's using a channel adaptation layer and a feature projection layer. The soft label distillation uses the difference in probability output between the teacher and student models after activation function processing as importance weights, and guides the student model's prediction distribution to approximate the teacher model's through a weighted bidirectional cross-entropy loss.

7. The method for feature detection of multiple types of surface damage on blades according to claim 6, characterized in that, The total loss function of the hybrid knowledge distillation method is determined according to the following formula: ， In the formula: Let be the total loss function of the hybrid knowledge distillation method. The original detection loss for the student model, To balance the hyperparameters of the loss, To generate distillation loss based on feature space masking, This is the soft-label distillation loss based on the output probability distribution.

8. The method for feature detection of multiple types of surface damage on blades according to claim 7, characterized in that, The teacher model is a YOLO11m model, and the student model is a pruned baseline model. During the distillation process, the parameters of the teacher model are fixed, and the parameters of the student model are updated by minimizing the total loss function.

9. The method for feature detection of multiple types of surface damage on blades according to claim 1, characterized in that, The dataset used for training the lightweight adaptive feature enhancement detection model includes: surface images of wind turbine blades collected under various lighting conditions, with the surface images labeled with category labels and bounding box information for four types of damage: sand holes, leading edge corrosion, oil stains, and surface cracks.

10. A feature detection system for multiple types of surface damage on blades, used to implement the method described in any one of claims 1-9, characterized in that, include: The image acquisition unit is used to acquire raw image data of the surface of the wind turbine blade to be inspected; The processing unit is communicatively connected to the image acquisition unit and is equipped with a trained lightweight adaptive feature enhancement detection model. It is used to receive the original image data and process the original image data to obtain damage detection results. The damage detection results include: damage category, confidence level, and bounding box coordinates.

Images