Method for identifying visible light image damage features of fan blades to circumvent structural interference

Abstract

The application discloses a fan blade visible light image damage feature recognition method for avoiding structural interference, aiming at the problems that the inherent structure of the blade is easy to cause detection misjudgment and the existing model is difficult to adapt to the end side deployment of the unmanned aerial vehicle, a UAV-YoMoViT lightweight detection model is constructed, SSA-ViT is used as the backbone to extract local and global features, Light-PAN is used as the neck to fuse multi-scale features, double label allocation is combined with consistency matching measurement to be the detection head, NMS-free efficient inference is realized, and meanwhile, an independent structural interference suppression module is additionally arranged, geometric registration, prior mask, feature constraint, special discrimination, three-level classification and multi-frame time sequence verification are carried out, and structural interference such as lightning protection device holes and screw holes is accurately avoided, the application considers the accuracy and the efficiency, effectively reduces the false detection rate, and adapts to the end side deployment demand of the unmanned aerial vehicle blade inspection.

Description

Technical Field

[0001] This invention belongs to the field of visual detection technology for wind turbine blade damage, specifically involving a method for identifying visible light image damage features of wind turbine blades while avoiding structural interference. Background Technology

[0002] With the large-scale development of the wind power industry, wind turbine blades are used in complex outdoor environments for a long time. Affected by factors such as wind and sand erosion, lightning strikes, and alternating loads, the surface is prone to various damages such as scratches, sand holes, surface dirt, and gel coat peeling. If the damage is not identified and repaired in time, it will gradually expand and cause blade structural failure, resulting in huge economic losses.

[0003] Currently, damage detection solutions based on visible light images from UAVs have become the mainstream approach for wind turbine blade inspection. However, there are two major pain points in practical applications: First, the surface of wind turbine blades has a large number of inherent factory structures such as surge protector holes, screw holes, sprayed markings, and eddy current generators (VGs). These structures have many similar features to real damage in terms of texture, edge gradient, and image contrast, making it easy to make a large number of misjudgments when directly applying basic recognition algorithms. Second, mainstream visual inspection models generally suffer from high computational cost and high inference latency, making it difficult to adapt to the lightweight deployment requirements of UAVs. Moreover, most models rely on non-maximum suppression post-processing, which further reduces inference efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying damage features of wind turbine blades in visible light images that avoids structural interference, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying visible light image damage features of wind turbine blades while avoiding structural interference, comprising the following steps: S1. Acquire visible light inspection images of the wind turbine blades; S2. Construct an independent structural interference suppression module to perform pre-interference suppression processing on the visible light inspection image. This module sequentially completes blade geometric registration and coordinate system construction, multi-type structural prior mask generation, and effective region feature constraint to obtain the effective skin region image of the blade. The structural interference suppression module does not modify the detection network structure and inference logic throughout the entire process. S3. Input the effective skinned region image into the pre-trained UAV-YoMoViT damage detection model. The model extracts local and global features through the SSA-ViT backbone network, fuses multi-scale features through the Light-PAN lightweight path aggregation network, and outputs the initial damage detection result by assigning a dual-label detection head and combining it with the consistency matching metric. S4. Perform typical structural interference discrimination, three-level hierarchical classification optimization, and multi-frame temporal consistency verification on the initial damage detection results in sequence, and output the final damage classification and location results.

[0006] As a preferred implementation, in step S2, the specific process of blade geometric registration and coordinate system construction is as follows: the image pixel coordinates are mapped to the blade spanwise, chordwise, and three-dimensional physical coordinates from the blade root to the blade tip. Based on the blade root, blade leading edge, trailing edge, and lightning protection strip, the unified registration of multiple frames of images is completed by combining Hough transform, semantic segmentation, and key point detection. Based on the registration results, a digital twin template of the blade is built, the coordinate range of all fixed structures is marked and the tolerance range is set, and the core features with fixed positions are used to distinguish the inherent structural features of the blade from the design features of the blade at the factory and the non-design abnormal damage features generated during the service of the blade.

[0007] As a preferred implementation, in step S2, the multi-type structural prior mask generation automatically identifies interference areas and generates shielding masks before feature extraction. Three different schemes are used to adapt to different data conditions. All masks are expanded outwards by 5-15 pixels to prevent structural edge shadows from being misjudged as scratches. (1) CAD / Operation and Maintenance Ledger Template Mask: Based on the original blade drawings and operation and maintenance ledger, the hole positions, eddy current generator layout, spray markings, and crossbeam areas are directly marked, resulting in the highest mask accuracy; (2) Weakly supervised detection mask: Four types of detection sub-models, namely holes, eddy current generators, spray marks, and beams, are trained separately. During training, they are defined as normal structure categories and corresponding masks are automatically generated. (3) Geometric rule filtering mask: In the case of no drawings, the geometric features are automatically filtered: roundness greater than 0.85 and area within a fixed range are judged as screw holes / lightning arrester holes; aspect ratio greater than 8 and the angle between the direction and the main beam is less than 15° are judged as crossbeams; significant frequency domain period peaks are judged as eddy current generator arrays; HSV high saturation color gamut + character outline are judged as spray markings.

[0008] As a preferred implementation, in step S2, the specific rules for the effective region feature constraints are as follows: the model feature extraction, fusion, and detection processes are only performed in the effective skin region of the leaf outside the mask, while multiple feature constraints are added: First, grayscale normalization is performed on local image stripes to eliminate the large-scale grayscale gradient interference caused by the beam; Second, a scale discrimination rule is introduced. The inherent structural dimensions of the blade are fixed, while the actual damage scale is random. The scale inconsistency is used as an auxiliary discrimination criterion. Third, morphological topology verification is added, relying on the roundness of connected domains, aspect ratio, and arrangement rules to distinguish target morphologies such as holes, scratches, and eddy current generators.

[0009] In a preferred embodiment, in step S3, the SSA-ViT sparse spatial attention visual transformer serves as the backbone network, combining the positional induction bias performance of CNNs and the global recognition advantages of ViTs, and is composed of a local feature extraction unit and a global feature extraction unit. The local feature extraction unit is a CNN structure, consisting of two parts: depthwise convolution and pointwise convolution. Depthwise convolution is a grouped convolution where each input channel is convolved with a separate convolutional kernel, and then the outputs of the different channels are concatenated together. Pointwise convolution is achieved by setting the convolution kernel to 1. The model first performs depthwise convolution on each channel of the input image, and then integrates the information between the channels through pointwise convolution to form the local features of the image; The global feature extraction unit consists of four parts: unfolding, SSA sparse space attention module, feedforward neural network FFN, and folding. The module design incorporates a fold-Transformer-unfold operation, transforming standard convolution into a stack of three consecutive operations: unfolding, matrix multiplication, and folding. It replaces local matrix multiplication in convolution with a stack of Transformer layers, giving the module both convolutional characteristics and global processing capabilities. SSA uses a separable self-attention mechanism to replace the multi-head attention mechanism in ViT, selecting only pixels with the same position for dot product operation.

[0010] In a preferred implementation, in step S3, the Light-PAN lightweight path aggregation network serves as the feature fusion neck, aggregating features from different levels to improve the model's accuracy in recognizing small targets. It mainly consists of three parts: (1) The bottom-up pathway is responsible for transmitting high-resolution but low-level information from shallow convolutional layers; (2) A top-down pathway carries a data stream with strong representational power extracted from deep layers but with a small spatial size; (3) Bidirectional connection mechanism, through lateral jump links, enables features at different scales to complement and support each other; The Light-PAN boasts three technological advantages: first, its lightweight design reduces the additional overhead of drone aerial photography, balancing accuracy and efficiency; second, it employs a weighted summation method to synthesize the inputs from each branch, capturing the correlation characteristics between different levels; and third, it allows for the setting of a dynamically adjustable scaling factor, enabling flexible configuration of weight allocation modes according to different task requirements, thereby enhancing the model's ability to adapt to diverse scenarios.

[0011] In a preferred implementation, in step S3, the dual-label allocation strategy uses two label branches—one-to-many allocation and one-to-one allocation—for target classification and regression. Each label branch consists of a classifier and a locator. During the training phase, the two branches are optimized together with the model backbone and neck. One-to-many assignment assigns multiple positive samples to each real damage instance, generating rich monitoring signals to ensure the model optimization effect; One-to-one assignment assigns only one prediction to each true label. The one-to-one head retains the same structure and optimization objective as the one-to-many assignment. Label assignment is obtained by using one-to-one matching, allowing the backbone and neck to enjoy rich supervision information from the one-to-many assignment. During the inference phase, the one-to-many distribution branch is discarded, and only the one-to-one head is used for prediction, enabling end-to-end deployment of the model without the need for non-maximum suppression post-processing and without additional inference cost.

[0012] In a preferred embodiment, in step S3, the consistency matching metric is used to achieve predictive matching of the dual-label allocation strategy, and the specific process is as follows: The predicted matching degree is calculated using the matching score formula, where the matching score is calculated by the classification score, the intersection-union ratio of the predicted target bounding box and the ground truth bounding box, and the spatial prior. The spatial prior is used to indicate whether the predicted anchor point is within the instance. The value is 1 when the anchor point completely covers the target, and less than 1 when it partially covers or does not cover the target. Two hyperparameters are set to balance the influence of semantic prediction and location regression tasks. The two branches are initialized in the same way. Based on the maximum IoU prediction result, the maximum scores of one-to-many and one-to-one matching are obtained respectively, and the alignment loss between the classification target and the task is calculated. The supervision gap between the two branches is calculated by the Wasserstein distance between different classification targets. As the matching score increases, the gap between the two branches decreases. When the distance reaches the minimum value, the best positive sample for a one-to-many head corresponds to the best positive sample for a one-to-one head, and the two branches reach consistent optimization.

[0013] As a preferred implementation, in step S4, the three-level hierarchical classification optimization abandons the global direct classification mode and strengthens the interference discrimination ability from the perspective of training and inference logic, specifically divided into three layers: First layer: Complete the initial structural / anomaly assessment, prioritizing the identification of the inherent structure of the blade; The second layer: Only for valid areas that are judged to be abnormal, detailed damage classification such as scratches, pinholes, dirt, and gel coat peeling is carried out; The third layer assesses the extent and severity of identified damage; during the training phase, sufficient negative samples containing various types of interference, boundary samples where interference and damage are adjacent, and multi-light and multi-angle samples are added as training datasets to ensure that the model learns pure damage features.

[0014] In a preferred embodiment, in step S4, the multi-frame temporal consistency verification uses multiple frames of visible light images continuously inspected by the UAV for post-verification. Structures such as blade crossbeams, holes, and eddy current generators have fixed positions in all frames, and damage mostly occurs randomly and develops gradually over time. A multi-frame voting mechanism is set up so that only when anomalies are detected simultaneously in multiple consecutive frames are they judged as valid damage, thus filtering out temporary false detections caused by shadows and noise in a single frame.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention adds an independent structural interference suppression module before the detection network, without altering the original network structure and reasoning logic throughout the process. Through a six-layer mechanism of geometric registration, prior masking, feature constraint, specific discrimination, three-level classification, and time-series verification, it accurately distinguishes between the inherent structure of the blade and service damage, effectively reducing misjudgments caused by structures such as surge protector holes, screw holes, spray markings, and eddy current generators, and significantly improving the detection accuracy.

[0016] 2. The UAV-YoMoViT lightweight model constructed in this invention adopts the SSA-ViT backbone to fuse the advantages of CNN local features and ViT global modeling capabilities. It replaces the traditional multi-head attention with a separable self-attention mechanism, which greatly reduces the amount of computation and parameters, significantly improves the inference speed, and is suitable for the deployment needs of edge devices such as drones.

[0017] 3. This invention employs Light-PAN lightweight feature fusion for the neck area, and efficiently aggregates multi-scale features through a bidirectional pathway and weighted summation fusion strategy, thereby improving the recognition accuracy of small-sized damage such as sand holes; the dynamic scaling factor design enhances the model's adaptability to complex inspection scenarios such as ambiguity and scale changes.

[0018] 4. This invention adopts a dual-label allocation strategy combined with consistency matching metric. During the training phase, it uses one-to-many branches to provide rich supervision signals to ensure accuracy. During the inference phase, it uses one-to-one branches to achieve end-to-end detection without NMS, which balances detection accuracy and inference efficiency and meets the needs of real-time inspection of UAVs. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall framework of the UAV-YoMoViT invention. Figure 2 This is a schematic diagram of the SSA-ViT image encoder of the present invention; Figure 3This is a schematic diagram of the SSA and the process of solving the context factor in the SSA of the present invention; Figure 4 This is a schematic diagram of the Light-PAN module structure of the present invention. Figure 5 This is a schematic diagram of the dual-label allocation strategy of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to embodiments.

[0021] The following embodiments are used to illustrate the present invention, but should not be used to limit the scope of protection of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions, and simple improvements to the method of the present invention under the premise of the concept of the present invention are all within the scope of protection claimed by the present invention.

[0022] Please see Figure 1 This invention provides a method for identifying visible light image damage features of wind turbine blades while avoiding structural interference, comprising the following steps: Step 1: Visible light inspection image acquisition By using a drone equipped with a high-definition visible light camera, the drone flies at a constant speed along the span of the wind turbine blade to collect full-coverage visible light images of the front, leading edge, and trailing edge of the blade. This ensures clear imaging of the entire area from the blade root to the blade tip, and the image resolution meets the requirements for detailed identification of minor damage, serving as input data for subsequent detection.

[0023] Step 2: Pre-structural interference suppression processing This step is performed before the image input detection model, without altering the network structure and inference logic of the subsequent UAV-YoMoViT model, and sequentially completes the three pre-interference suppression steps: (1) Blade geometric registration and coordinate system construction The input image undergoes blade semantic segmentation to extract key features such as blade contour, blade root, leading edge, trailing edge, and lightning protection strip. Combining Hough transform and keypoint detection algorithms, the image pixel coordinates are mapped to the blade's true spanwise, chordwise, and three-dimensional physical coordinates from blade root to tip, completing the geometric calibration of a single frame image. Multiple inspection images of the same blade are uniformly registered, and a digital twin template for the corresponding blade model is built. The template is marked with the coordinate ranges of all factory-fixed structures, such as surge arrester holes, screw holes, eddy current generators, and painted markings, and corresponding tolerance intervals are set. Relying on the core feature of fixed structural positions, the inherent structural features of the blade at the factory design level are distinguished from random abnormal damage generated during service.

[0024] (2) Generation of prior masks for multiple structural types Based on the available data at the site, a corresponding solution was selected to generate a shielding mask for structural interference. All masks were expanded outward by 5-15 pixels to prevent the shadow gradient at the structural edges from being misjudged as scratch-like damage. If the original CAD drawings and maintenance logs of the blades are available, the CAD / maintenance log template masking solution can be used to directly mark the coordinates of the hole positions, eddy current generator layout, spray markings, and crossbeam area to generate the most accurate mask. If sufficient historical inspection data is available but no original factory drawings are available, a weakly supervised detection mask scheme is adopted. Four types of detection sub-models, namely holes, eddy current generators, spray marks, and beams, are pre-trained separately. During training, they are defined as normal structural categories. During inference, the target is automatically identified and the corresponding mask is generated. If no prior data is available, a geometric rule-based filtering masking scheme is adopted: calculate the roundness of connected components, and targets with a roundness greater than 0.85 and an area within a preset fixed range are identified as screw holes / surge arrester holes; targets with an aspect ratio greater than 8 and an angle less than 15° with the main beam are identified as crossbeams; perform frequency domain analysis on the image, and regions with significant periodic peaks are identified as eddy current generator arrays; extract high-saturation regions in the HSV color space, and combine them with character outline features to identify spray-painted markings.

[0025] (3) Effective region feature constraints The masked area is marked as a non-detection area. Subsequent feature extraction, feature fusion, and detection processes are performed only within the effective skin area of ​​the leaf outside the mask, while applying triple feature constraints: ① Perform grayscale normalization on the local image strips of the blade to eliminate the large-scale grayscale gradient interference caused by the crossbeam; ② Introduce scale discrimination rules. Based on the characteristics of fixed inherent structural size of blades and random actual damage scale, scale inconsistency is used as an auxiliary basis for damage discrimination. ③ Add morphological topology verification. By calculating the roundness, aspect ratio, and neighborhood arrangement of the target's connected domains, distinguish the morphological features of different targets such as holes, scratches, and eddy current generator arrays.

[0026] Step 3: Damage detection of the UAV-YoMoViT model The pre-processed image of the effective skinned region is input into the pre-trained UAV-YoMoViT damage detection model. The overall framework of UAV-YoMoViT is as follows: Figure 2 As shown, it consists of three parts connected in sequence: SSA-ViT backbone, Light-PAN neck, and dual-label assignment detection head, which sequentially complete feature extraction, feature fusion, and damage prediction.

[0027] (1) Feature extraction of SSA-ViT backbone network The structure of the SSA-ViT image encoder is as follows: Figure 3 As shown, lightweight edge image encoding is achieved by combining the positional induction bias performance of CNN and the global recognition advantage of ViT.

[0028] Local feature extraction: The input image is first subjected to depthwise convolution, and each input channel is grouped and convolved with an independent convolution kernel. Then, it is combined with the information of different channels through 1×1 pointwise convolution to output the local features of the image.

[0029] Global Feature Extraction: Local features, after being unfolded and serialized, enter the SSA sparse spatial attention module. SSA uses a separable self-attention mechanism to replace the multi-head attention mechanism in traditional ViT, selecting only pixels with the same position for dot product operation, reducing information redundancy and computational load.

[0030] The input image X is weighted. The keyword branch k is linearly projected onto the d-dimensional space and produces the corresponding output. Then use context factors To calculate the context vector This is used as a weighted sum of input X, where:

[0031] Context vector Encode all image embeddings in input X, in The context factor encoded in X shares label information with all images in X. Value branch weight The input x is linearly projected into a d-dimensional space, and the output is generated using the Softmax activation function. Then, using broadcast multiplication in NumPy, The context vector in the middle is passed to Finally, the result is fed back to another weight. A linear layer to generate the final output. .

[0032] The formula for calculating SSA is: in, This represents the softmax operation. Let y be the input image and y be the output image. These represent the weights of branches I, K, and V, respectively. These are the context factor and the context vector, respectively. Because MHA does not consider the computational cost of its implementation when calculating time complexity, its computation is resource-intensive. The SSA method proposed in this paper fully considers the computational resources of edge and mobile devices. According to laboratory test results, the computational cost is significantly reduced, and the inference speed is increased by 3 times.

[0033] (2) Light-PAN neck feature fusion The structure of the Light-PAN module is as follows: Figure 4 As shown, a bidirectional feature pyramid architecture is used to aggregate multi-scale features, which improves the model's accuracy in identifying damage to small targets.

[0034] Light-PAN incorporates bottom-up, top-down, and bidirectional connection mechanisms: the bottom-up path transmits high-resolution detailed features from shallow networks, the top-down path transmits high-semantic features from deep networks, and lateral skip connections enable the mutual complementarity of features at different scales, solving the problems of gradient vanishing and multi-scale recognition difficulties in traditional unidirectional feature pyramids.

[0035] The fusion method uses weighted summation instead of simple splicing to better capture the correlation characteristics of cross-level features; at the same time, a dynamically adjustable scaling factor is set, which can flexibly configure the weight allocation mode according to task requirements and improve the model's adaptability to diverse scenarios such as fuzziness and scale changes.

[0036] (3) Dual-label assignment head detection Dual-label allocation strategy such as Figure 5 As shown, a dual-branch structure with one-to-many and one-to-one parallel implementations is adopted, combined with a consistency matching metric to achieve damage detection. During the training phase, the two branches, along with the backbone and neck, are optimized together: the one-to-many branch assigns multiple positive samples to each real damage instance, providing rich supervision signals to ensure model convergence and detection accuracy; the one-to-one head retains the same structure and optimization objective as the one-to-many branch, assigning only one positive sample to each real instance.

[0037] The target location is predicted using a consistency matching metric, and the prediction result is shown in the following formula:

[0038] Where p is the categorical score, Let 'b' and 'b' represent the predicted bounding box and the ground truth bounding box of the target, respectively. 's' is the spatial prior, indicating whether the predicted anchor point is within the instance. In other words, the spatial prior is used to determine whether the anchor point reasonably covers the object. If an anchor point completely covers a target, the spatial prior 's' is 1, indicating that the anchor point is valid. If the frame only partially covers or does not cover the target at all, 's' will be less than 1, indicating that the anchor point is inappropriate. and These are two important hyperparameters used to balance the impact of semantic prediction and location regression tasks.

[0039] In dual-label assignment, the same initialization operation is first performed on both branches, resulting in the same predicted classification score p. , will be the largest The prediction results are expressed as The maximum one-to-many and one-to-one matching scores are respectively represented as and Assume that a one-to-many branch generates positive samples. and one-to-one branch selection A prediction, using To measure this, the alignment damage between the classification target and the task can be obtained as follows:

[0040] The supervision gap between the two branches can be calculated by the Wasserstein distance between the different classification objectives, i.e.:

[0041] along with As the value increases, the gap between the two decreases, that is... exist The ranking is higher in China. When it reaches its minimum value, it means yes The sample with the highest score, and This means Therefore, the optimal positive sample for a one-to-many head relationship corresponds to the optimal positive sample for a one-to-one head relationship, thus achieving consistent optimization for the two branches.

[0042] During the inference phase, the one-to-many branches are discarded directly, and the prediction result is output only through the one-to-one head. No non-maximum suppression post-processing is required, thus achieving efficient end-to-end inference.

[0043] Step 4: Post-interference suppression and result optimization The initial detection results output by the model are processed sequentially, including typical interference discrimination, three-level hierarchical classification optimization, and multi-frame temporal consistency verification, to further filter out false detections and output the final results.

[0044] (1) Typical interference identification Dedicated discrimination rules are applied to each of the five core structural interferences to further reduce the probability of confusion with actual damage: Surge arrester holes and screw holes: Hough circle / ellipse fitting is used to locate the target holes, and the shielding is completed by matching the template position; the characteristics of closed ring edge and symmetrical gray level inside and outside are used to distinguish holes from solid sand hole damage; Crossbeam region: A linear enhancement algorithm is used to extract long strip-shaped regions and generate a mask to distinguish the low-frequency smooth gray-scale changes of the crossbeam from the high-frequency linear features of the scratches; Spray-painted markings: High-saturation colors and character topological features are extracted to distinguish them from natural dirt, irregular fading from gel coat peeling, and random color blocks; during the model training phase, spray-painted markings have been separately labeled as normal categories to avoid feature confusion; Eddy current generator: Its periodic arrangement characteristics are identified by autocorrelation and projection algorithms, and the shadow patterns are matched by the direction of illumination to distinguish it from discrete scratches and sand hole arrays.

[0045] (2) Three-level hierarchical classification optimization The model employs a three-tiered classification logic, abandoning the global direct classification mode and strengthening its ability to distinguish interference from the training and inference logic: The first layer performs binary classification of structure and anomalies, prioritizing the identification and exclusion of inherent blade structural targets, thus narrowing the subsequent detection range; the second layer performs fine-grained damage classification only on valid areas determined to be abnormal, outputting four types of damage results: scratches, pinholes, dirt, and gelcoat peeling; the third layer calculates parameters such as damage range and size for each type of identified damage, completing the damage severity assessment. During the model training phase, sufficient negative samples containing various types of interference, boundary samples between interference and damage, and multi-light and multi-angle samples have been added to ensure that the model learns pure damage features and improves the accuracy of classification boundaries.

[0046] (3) Multi-frame timing consistency check Multiple consecutive frames of inspection images of the same blade are retrieved, and the detected abnormal targets are matched and verified. Since the inherent structure of the blade is fixed in position in consecutive frames, false detections caused by shadows and noise in a single frame only occur in that single frame, while real damage exists stably in consecutive frames. Therefore, a multi-frame voting mechanism is set up: only when the same type of abnormality is detected in multiple consecutive frames at the same physical coordinate position is it determined to be valid damage; otherwise, it is determined to be a false detection and filtered out.

[0047] The final output includes a complete identification result containing damage category, location coordinates, size parameters, and severity level.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying damage features of wind turbine blades using visible light images while avoiding structural interference, characterized in that, Includes the following steps: S1. Acquire visible light inspection images of the wind turbine blades; S2. Construct an independent structural interference suppression module to perform pre-interference suppression processing on the visible light inspection image. Sequentially complete the blade geometric registration and coordinate system construction, multi-type structural prior mask generation, and effective region feature constraint to obtain the effective skin region image of the blade. The structural interference suppression module does not change the detection network structure and inference logic throughout the process. S3. Input the effective skinned region image into the pre-trained UAV-YoMoViT damage detection model. The model extracts local and global features through the SSA-ViT backbone network, fuses multi-scale features through the Light-PAN lightweight path aggregation network, and outputs the initial damage detection result by assigning a dual-label detection head and combining it with the consistency matching metric. S4. Perform typical structural interference discrimination, three-level hierarchical classification optimization, and multi-frame temporal consistency verification on the initial damage detection results in sequence, and output the final damage classification and location results.

2. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S2, the specific process of blade geometric registration and coordinate system construction is as follows: the image pixel coordinates are mapped to the three-dimensional physical coordinates of the blade spanwise, chordwise, and blade root to blade tip. Based on the blade root, blade leading edge, trailing edge, and lightning protection strip, the unified registration of multiple frames of images is completed by combining Hough transform, semantic segmentation, and key point detection. Based on the registration results, a digital twin template of the blade is built, the coordinate range of all fixed structures is marked and the tolerance range is set, and the core features with fixed positions are used to distinguish the inherent structural features of the blade from the design features of the blade at the factory and the non-design abnormal damage features generated during the service of the blade.

3. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S2, the multi-type structure prior mask generation automatically identifies interference areas and generates shielding masks before feature extraction. Three schemes are used to adapt to different data conditions. All masks are expanded outwards by 5-15 pixels to prevent structural edge shadows from being misjudged as scratches. (1) CAD / Operation and Maintenance Ledger Template Mask: Based on the original blade drawings and operation and maintenance ledger, the hole positions, eddy current generator layout, spray markings, and crossbeam areas are directly marked, resulting in the highest mask accuracy; (2) Weakly supervised detection mask: Four types of detection sub-models, namely holes, eddy current generators, spray marks, and beams, are trained separately. During training, they are defined as normal structure categories and corresponding masks are automatically generated. (3) Geometric rule filtering mask: In the case of no drawings, the geometric features are automatically filtered: roundness greater than 0.85 and area within a fixed range are judged as screw holes / lightning arrester holes; aspect ratio greater than 8 and the angle between the direction and the main beam is less than 15° are judged as crossbeams; significant frequency domain period peaks are judged as eddy current generator arrays; HSV high saturation color gamut + character outline are judged as spray markings.

4. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S2, the specific rules for the effective region feature constraints are as follows: the model feature extraction, fusion, and detection processes are only performed in the effective skin region of the leaf outside the mask, while multiple feature constraints are added: First, grayscale normalization is performed on local image stripes to eliminate the large-scale grayscale gradient interference caused by the beam; Second, a scale discrimination rule is introduced. The inherent structural dimensions of the blade are fixed, while the actual damage scale is random. The scale inconsistency is used as an auxiliary discrimination criterion. Third, morphological topology verification is added, relying on the roundness of connected domains, aspect ratio, and arrangement rules to distinguish target morphologies such as holes, scratches, and eddy current generators.

5. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S3, the SSA-ViT sparse spatial attention visual transformer serves as the backbone network, combining the positional induction bias performance of CNN and the global recognition advantage of ViT, and is composed of a local feature extraction unit and a global feature extraction unit. The local feature extraction unit is a CNN structure, consisting of two parts: depthwise convolution and pointwise convolution. Depthwise convolution is a grouped convolution where each input channel is convolved with a separate convolutional kernel, and then the outputs of the different channels are concatenated together. Pointwise convolution is achieved by setting the convolution kernel to 1. The model first performs depthwise convolution on each channel of the input image, and then integrates the information between the channels through pointwise convolution to form the local features of the image; The global feature extraction unit consists of four parts: unfolding, SSA sparse space attention module, feedforward neural network FFN, and folding. The module design incorporates a fold-Transformer-unfold operation, transforming standard convolution into a stack of three consecutive operations: unfolding, matrix multiplication, and folding. It replaces local matrix multiplication in convolution with a stack of Transformer layers, giving the module both convolutional characteristics and global processing capabilities. SSA uses a separable self-attention mechanism to replace the multi-head attention mechanism in ViT, selecting only pixels with the same position for dot product operation.

6. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S3, the Light-PAN lightweight path aggregation network serves as the feature fusion neck, aggregating features from different levels to improve the model's accuracy in recognizing small targets. It mainly consists of three parts: (1) The bottom-up pathway is responsible for transmitting high-resolution but low-level information from shallow convolutional layers; (2) A top-down pathway carries a data stream with strong representational power extracted from deep layers but with a small spatial size; (3) Bidirectional connection mechanism, through lateral jump links, enables features at different scales to complement and support each other; The Light-PAN boasts three technological advantages: first, its lightweight design reduces the additional overhead of drone aerial photography, balancing accuracy and efficiency; second, it employs a weighted summation method to synthesize the inputs from each branch, capturing the correlation characteristics between different levels; and third, it allows for the setting of a dynamically adjustable scaling factor, enabling flexible configuration of weight allocation modes according to different task requirements, thereby enhancing the model's ability to adapt to diverse scenarios.

7. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S3, the dual-label allocation strategy uses two label branches—one-to-many allocation and one-to-one allocation—for target classification and regression. Each label branch consists of a classifier and a locator. During the training phase, the two branches are optimized together with the model backbone and neck. One-to-many assignment assigns multiple positive samples to each real damage instance, generating rich monitoring signals to ensure the model optimization effect; One-to-one assignment assigns only one prediction to each true label. The one-to-one head retains the same structure and optimization objective as the one-to-many assignment. Label assignment is obtained by using one-to-one matching, allowing the backbone and neck to enjoy rich supervision information from the one-to-many assignment. During the inference phase, the one-to-many distribution branch is discarded, and only the one-to-one head is used for prediction, enabling end-to-end deployment of the model. No non-maximum suppression post-processing is required, and there is no additional inference cost.

8. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S3, the consistency matching metric is used to achieve predictive matching of the dual-label allocation strategy. The specific process is as follows: The predicted matching degree is calculated using the matching score formula, where the matching score is calculated by the classification score, the intersection-union ratio of the predicted target bounding box and the ground truth bounding box, and the spatial prior. The spatial prior is used to indicate whether the predicted anchor point is within the instance. The value is 1 when the anchor point completely covers the target, and less than 1 when it partially covers or does not cover the target. Two hyperparameter settings are used to balance the effects of semantic prediction and location regression tasks; The two branches are initialized in the same way. Based on the maximum IoU prediction result, the maximum scores of one-to-many and one-to-one matching are obtained respectively, and the alignment loss between the classification target and the task is calculated. The supervision gap between the two branches is calculated by the Wasserstein distance between different classification targets. As the matching score increases, the gap between the two branches decreases. When the distance reaches the minimum value, the best positive sample for a one-to-many head corresponds to the best positive sample for a one-to-one head, and the two branches reach consistent optimization.

9. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S4, the three-level hierarchical classification optimization abandons the global direct classification mode and strengthens the interference discrimination ability from the perspective of training and inference logic, specifically divided into three layers: First layer: Complete the initial structural / anomaly assessment, prioritizing the identification of the inherent structure of the blade; The second layer: Only for valid areas that are judged to be abnormal, detailed damage classification such as scratches, pinholes, dirt, and gel coat peeling is carried out; The third layer: assessing the extent and severity of identified damage; During the training phase, a sufficient number of negative samples containing various types of interference, boundary samples where interference and damage are adjacent, and multi-light and multi-angle samples are added as training datasets to ensure that the model learns pure damage features.

10. The method for identifying visible light image damage features of wind turbine blades to avoid structural interference according to claim 1, characterized in that, In step S4, the multi-frame temporal consistency verification uses multiple frames of visible light images continuously inspected by the UAV for post-verification. Structures such as blade crossbeams, holes, and eddy current generators have fixed positions in all frames, and damage mostly occurs randomly and develops gradually over time. A multi-frame voting mechanism is set up so that only when anomalies are detected simultaneously in multiple consecutive frames are they judged as valid damage, thus filtering out temporary false detections caused by shadows and noise in a single frame.

Images