Offshore wind turbine blade defect quantitative analysis method and system based on dynamic routing and heterogeneous expert model

Abstract

This invention discloses a method and system for quantitative analysis of defects in offshore wind turbine blades based on dynamic routing and heterogeneous expert models. The method includes: acquiring original images of the blades using a UAV; performing global noise reduction and motion blur removal using an attention-based image enhancement model; segmenting the enhanced image into non-overlapping image blocks; classifying the defect types of each image block using a routing model and dynamically distributing them to the corresponding expert models; processing discrete and continuous defects respectively using parallel instance segmentation expert models and semantic segmentation expert models; reconstructing the local mask into a global mask image and performing connected component labeling, pixel statistics, and morphological skeleton extraction to calculate the number, area, and length of defects and generate a detection report. This invention can stably acquire high-quality images in complex marine environments and can intelligently and accurately perform automated quantitative analysis of multi-morphological defects.

Description

Technical Field

[0001] This invention relates to the technical field of wind power equipment defect detection, and in particular to a quantitative analysis method and system for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models. Background Technology

[0002] Driven by both energy shortages and ecological degradation, offshore wind power is rapidly developing as an important component of clean energy. Offshore wind turbines operate in harsh environments characterized by high salinity, high humidity, and strong winds for extended periods. Their blades face various threats, including cracks, erosion, and lightning strikes. Therefore, accurate and efficient blade defect detection is crucial for ensuring power generation efficiency and equipment safety.

[0003] Currently, the operation and maintenance inspection of offshore wind turbine blades mainly relies on manual rope lowering operations after shutdown or long-distance observation via telescopes. This approach suffers from significant losses due to forced shutdowns, high labor costs, and high risks associated with high-altitude operations. While automated inspection systems based on drone platforms have emerged, two major technological bottlenecks remain:

[0004] 1. Imaging quality bottleneck: Multiple interferences such as high salt spray at sea, strong light reflection, complex shadows and drone shaking cause inspection images to generally have problems such as dynamic blurring, low contrast and uneven lighting, which seriously restrict the feature extraction ability of subsequent algorithms and result in a high false detection rate and false negative rate of defects.

[0005] 2. Identifying Model Generalization and Efficiency Bottlenecks: Existing solutions mostly rely on a single general model to handle all defect types. However, blade defects exhibit diverse morphologies, including discrete defects such as "coating peeling" and "lightning damage" (suitable for counting), as well as continuous defects such as "cracks" and "leading edge erosion" (suitable for measurement). A single model struggles to meet the demands of multiple tasks, performing poorly when accurately measuring continuous defects or counting densely overlapping discrete defects, resulting in insufficient accuracy in quantitative analysis. Some multimodal fusion solutions also fail to fundamentally address the problem of dynamic intelligent distribution for different defect morphologies.

[0006] Therefore, there is an urgent need for a technical solution that can stably acquire high-quality images in complex marine environments and can intelligently and accurately perform automated quantitative analysis of multi-morphological defects. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for quantitative analysis of defects in offshore wind turbine blades based on dynamic routing and heterogeneous expert models.

[0008] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0009] A quantitative analysis method for defects in offshore wind turbine blades based on dynamic routing and heterogeneous expert models includes:

[0010] Remotely acquire raw images of offshore wind turbine blades in operation;

[0011] An enhanced image is obtained by performing global noise reduction and motion blur removal on the original image using an image enhancement model.

[0012] The enhanced image is segmented into multiple non-overlapping image blocks;

[0013] Each image patch is classified for defect type using a routing model, generating a probability vector for the defect type of each image patch, and the image patch is dynamically distributed to the corresponding expert model based on the probability vector; the expert model includes an instance segmentation expert model and a semantic segmentation expert model set in parallel.

[0014] The distributed image patches are refined using an instance segmentation expert model and a semantic segmentation expert model, respectively. The instance segmentation expert model is used to handle discrete defects and generate instance segmentation masks, while the semantic segmentation expert model is used to handle continuous defects and generate semantic segmentation masks.

[0015] The instance segmentation mask and semantic segmentation mask output by each expert model are recombined into a global segmentation mask map. Based on the global segmentation mask map, connected component labeling, pixel statistics and morphological skeleton extraction operations are performed to calculate the number, area and length of defects and generate a structured defect detection report.

[0016] Furthermore, the image enhancement model adopts an encoder-decoder structure based on an attention mechanism, which reconstructs the local texture details of offshore wind turbine blades lost due to motion blur and noise through gated linear unit operations and channel attention mechanisms, while preserving the global contextual information of the image.

[0017] Furthermore, the routing model extracts multi-level feature maps of image patches through a feature extraction network, compresses them into feature vectors through a global average pooling layer, and inputs them into a fully connected classification head. The output is a logical score vector containing the original score of each defect category, and the Sigmoid activation function is applied independently to each defect category to generate the probability vector. The maximum value in the probability vector is compared according to a preset threshold. If the maximum value corresponds to a discrete defect, the image patch is distributed to the instance segmentation expert model. If the maximum value corresponds to a continuous defect, the image patch is distributed to the semantic segmentation expert model.

[0018] Furthermore, the instance segmentation expert model is based on a single-stage instance segmentation architecture. After the input image block extracts features through the backbone network, the detection head predicts the bounding box and mask coefficients. The mask coefficients Prototype mask shared globally Linear combination generates instance segmentation mask :

[0019] ;

[0020] in, For the number of prototypes, For the Sigmoid function:

[0021] .

[0022] Furthermore, the instance segmentation expert model employs a soft nonmaximum suppression algorithm to handle overlap defects, penalizing the confidence score of each detection box using the following penalty formula:

[0023] ;

[0024] in, For the first The updated confidence score for each detection box; For the first The original confidence score of each detection box before the update; This is the intersection-union ratio (IUU) between the highest-scoring bounding box and the current bounding box. This is the Gaussian variance hyperparameter, used to control the severity of the penalty.

[0025] Furthermore, the semantic segmentation expert model is based on an encoder-decoder architecture, which uses skip connections to splice and fuse high-level semantic information from the deep layers of the network with precise localization and edge information from the shallow layers of the network, and outputs a pixel-level semantic segmentation mask.

[0026] Furthermore, the defect area The calculation formula is:

[0027] ;

[0028] Defect length The calculation formula is:

[0029] ;

[0030] in, This represents the total number of defective pixels. This represents the total number of pixels in the skeleton lines. The ground sampling distance is calculated using the drone's camera parameters and flight altitude.

[0031] Furthermore, to achieve the above objectives, the present invention also provides a quantitative analysis system for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models, used to implement the above-mentioned quantitative analysis method for offshore wind turbine blade defects, including:

[0032] The image acquisition module is used to acquire raw image data of offshore wind turbine blades in operation;

[0033] The image pre-enhancement module is used to perform global noise reduction and motion blur removal on the original image using an image enhancement model to obtain an enhanced image.

[0034] The image slicing module is used to slice the enhanced image into multiple non-overlapping image blocks;

[0035] The dynamic routing module is used to classify the defect type of each image patch using a routing model and dynamically distribute the image patch to the corresponding expert model according to the probability vector.

[0036] The heterogeneous expert processing module includes parallel instance segmentation expert models and semantic segmentation expert models, which are used to handle discrete defects and continuous defects, respectively.

[0037] The results summary and quantitative analysis module is used to reconstruct the local masks output by each expert model into a global segmentation mask image, and calculate the number of defects, total area and total length, and generate a structured defect detection report.

[0038] Furthermore, the dynamic routing module and the heterogeneous expert processing module communicate using asynchronous message queues, supporting real-time distribution of image blocks and parallel inference of expert models.

[0039] Furthermore, the result summarization and quantitative analysis module employs a two-pass scanning method to perform a connected component labeling algorithm and a Zhang-Suen thinning algorithm to perform morphological skeleton extraction. The quantitative analysis results are stored in a database and compared with historical data to generate trend analysis charts.

[0040] Compared with existing technologies, the principles and advantages of this technical solution are as follows:

[0041] 1. An encoder-decoder structure based on an attention mechanism is adopted, which can simultaneously preserve global context information and reconstruct local texture details. It effectively overcomes the dynamic blurring problems caused by high salt spray at sea, strong light reflection, complex shadows and drone shaking. The peak signal-to-noise ratio of the output image is not less than 30dB and the structural similarity index is not less than 0.9, laying a high-quality image foundation for subsequent defect recognition.

[0042] 2. A real-time, adaptive image block distribution mechanism is implemented through a lightweight convolutional neural network. Based on the local features of the input image block, dynamic decisions are made to distribute different types of defects to the expert model that is best suited for them. This effectively solves the problems of insufficient generalization ability and decreased accuracy of a single model when dealing with multi-form defects. The processing time for each image block does not exceed 10 milliseconds, ensuring high efficiency in real-time inspection.

[0043] 3. The instance segmentation expert employs a soft nonmaximum suppression algorithm, which effectively solves the problems of overlap and missed detection in high-density defects; the semantic segmentation expert fuses high-level semantic information and shallow localization information through skip connections to achieve pixel-level accurate segmentation of continuous defects such as cracks. The two expert models are processed in parallel, optimizing the recognition accuracy and quantitative analysis capability of discrete and continuous defects respectively.

[0044] 4. Through local mask reconstruction, connected component marking and morphological skeleton extraction, the number, area and length of defects are accurately calculated. The ground sampling distance GSD is innovatively incorporated into the area and length conversion formula to ensure the consistency between the quantitative analysis results and the actual physical dimensions. It supports trend comparison with historical data and can be used for defect evolution prediction and maintenance decision support. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the principle of the quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models, as described in this embodiment of the invention.

[0047] Figure 2 A flowchart illustrating the workflow of the routing model;

[0048] Figure 3 Flowchart for parallel processing of the heterogeneous expert processing module;

[0049] Figure 4 Flowchart for results summarization and quantitative analysis;

[0050] Figure 5 This is a connection block diagram of the offshore wind turbine blade defect quantitative analysis system based on dynamic routing and heterogeneous expert models, according to an embodiment of the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to specific embodiments:

[0052] like Figure 1 As shown in this embodiment, the quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models includes the following steps:

[0053] S1. Remotely acquire raw images of offshore wind turbine blades in operation;

[0054] In this step,

[0055] Raw image data was collected by drones while offshore wind turbine blades were in operation. The raw image data came from a high-resolution optical camera mounted on the drone, with a resolution of no less than 40 million pixels, and the collection range covered the entire surface area of ​​the blade.

[0056] S2. Use an image enhancement model to perform global noise reduction and motion blur removal on the original image to obtain an enhanced image;

[0057] In this step,

[0058] An attention-based image enhancement model is used to process the original image data. This model employs an encoder-decoder structure to preserve the global contextual information of the image and reconstructs local texture details lost due to motion blur and noise through gated linear unit operations and channel attention mechanisms. The gated linear unit module performs gated linear unit operations to filter effective features, and the channel attention module enhances the weights of important features through channel attention mechanisms. The image enhancement model performs multi-scale convolution and deconvolution operations on the original image and outputs a high-resolution enhanced image with the same size as the input. The peak signal-to-noise ratio of the enhanced image is not less than 30dB, and the structural similarity index is not less than 0.9. The enhanced image is used as the input for the image slicing step.

[0059] S3. Divide the enhanced image into multiple non-overlapping image blocks;

[0060] In this step, the resolution of the enhanced image is 4096×3072 pixels. The system divides the enhanced image into N non-overlapping image blocks, each with a size of 224×224 pixels. The segmentation process is implemented using a sliding window algorithm, with the window step size equal to the image block size to ensure no overlap. The number of image blocks N is calculated as follows: Where W and H are the width and height of the enhanced image, w and h are the width and height of the image patch, each image patch contains a pixel matrix of a local region, and the image patch serves as the input data set for the dynamic routing step. The system assigns a unique identifier to each image patch and stores its coordinate information in the original image to facilitate subsequent mask reassembly operations.

[0061] S4. Use the routing model to classify the defect type of each image block, generate a probability vector for the defect type of each image block, and dynamically distribute the image blocks to the corresponding expert models according to the probability vectors; the expert models include instance segmentation expert models and semantic segmentation expert models set in parallel.

[0062] In this step,

[0063] A routing model based on a lightweight convolutional neural network is used to classify defect types for each image patch. The routing model extracts multi-level feature maps of the image patches through its feature extraction network. The feature maps are compressed into feature vectors by a global average pooling layer. The feature vectors are then input into a fully connected classification head to output the original logical score vector. , The system contains raw scores for each defect category, including cracks, bubbles, erosion, and normal defects, and a logical score vector. Each defect category is activated independently using the Sigmoid activation function to generate a probability vector. probability vector The system represents the confidence score of an image patch belonging to each defect category. It compares the maximum value in the probability vector based on a preset threshold. If the maximum value corresponds to a discrete defect, the image patch is assigned to the instance segmentation expert model; if the maximum value corresponds to a continuous defect, the image patch is assigned to the semantic segmentation expert model. The routing decision signal serves as the input to the heterogeneous expert processing steps. The workflow diagram of the routing model is shown below. Figure 2 As shown.

[0064] Unlike existing technologies that use fixed rules or a single model to handle all defects, this routing model dynamically decides based on the local features of the input image patch, distributing different types of defects to the most suitable expert model. This design effectively solves the problems of insufficient generalization ability and accuracy degradation of a single model when dealing with multi-morphological defects. The routing model processes each image patch in less than ten milliseconds during the inference phase, ensuring the system's high efficiency in real-time inspection, while achieving highly reliable routing decisions through comparison of probability vectors and thresholds.

[0065] The training process of the routing model uses the cross-entropy loss function to optimize the parameters. The loss function L is calculated as follows: ,in It is the number of defect categories. It is the one-hot encoding of the real label. It predicts probabilities, and the training data includes a set of image patches labeled with defect types. Data augmentation techniques include rotation, scaling, and color jitter to improve the model's generalization ability. The routing model uses batch processing to accelerate computation during the inference phase, and the classification time for each image patch does not exceed ten milliseconds, ensuring real-time routing decisions.

[0066] S5. The distributed image blocks are refined using the instance segmentation expert model and the semantic segmentation expert model, respectively.

[0067] in,

[0068] 1) Instance segmentation expert models are used to handle discrete defects and generate instance segmentation masks;

[0069] The instance segmentation expert model is based on a single-stage instance segmentation architecture. The input image patch first extracts features through a backbone network, then predicts bounding boxes and mask coefficients using a detection head. The mask coefficients... Prototype mask shared globally Linear combination generates instance segmentation mask ,

[0070] ;

[0071] in It is the number of prototypes. It is the Sigmoid function.

[0072] ;

[0073] To accurately depict a specific defect, such as a bubble, multiple basic prototype masks, such as prototype A representing a circle and prototype B representing a highlight feature, need to be combined with specific coefficients, such as 0.8 times prototype A plus 0.3 times prototype B, to efficiently generate a high-resolution bubble instance segmentation mask.

[0074] When dealing with high-density defects, the intersection-union ratio (CUI) of the bounding boxes of two or more real, independent defect targets may exceed a conventional threshold. If a traditional non-maximum suppression algorithm is used, it may incorrectly treat one of the real defects as a duplicate detection and delete it, resulting in an underestimation of the final defect count. To address this issue, the instance segmentation expert model employs a soft non-maximum suppression algorithm to handle overlapping defects, penalizing the confidence score of each detection box using the following formula:

[0075] ;

[0076] in, For the first The updated confidence score for each detection box; For the first The original confidence score of each detection box before the update; This is the intersection-union ratio (IUU) between the highest-scoring bounding box and the current bounding box. This is the Gaussian variance hyperparameter, used to control the severity of the penalty.

[0077] Unlike the hard deletion strategy of standard nonmaximum suppression (NMS), soft NMS applies a penalized confidence decay to other detection boxes that overlap with the highest-scoring box, rather than directly zeroing them out. The technical advantage is that if an overlapping box is itself a high-confidence true defect, even though its score is penalized due to overlap, it may still be higher than the final confidence threshold, thus preserving the true defect. This design makes the counting model of this invention far more robust and accurate than the standard NMS method when dealing with dense, overlapping defects.

[0078] In the instance segmentation expert model, the number of prototype masks, k, is set to 32. The prototype masks are learned through model training, and the mask coefficients are predicted by a fully connected layer. The resolution of the mask after linear combination is the same as that of the input image patch. The Gaussian variance hyperparameter of the soft nonmaximum suppression algorithm is also considered. The confidence penalty threshold is set to 0.5 and 0.25 to ensure accurate counting of high-density defects. The skip connection feature map of the instance segmentation expert model comes from the corresponding layer of the encoder and is fused with deep features through channel concatenation. The decoder performs upsampling operation using transposed convolution with an upsampling factor of 2. Finally, the output mask is converted into a binary image by the Argmax function.

[0079] 2) Semantic segmentation expert models are used to handle continuous defects and generate semantic segmentation masks;

[0080] The semantic segmentation expert model is based on an encoder-decoder architecture. This network, through its unique skip connection design, concatenates and fuses the deep layers of the network responsible for understanding high-level semantic information ("what") with the shallow layers responsible for locating precise edge information ("where"). Its input, image patches labeled as cracks and leading-edge erosion, is downsampled through the encoder path to extract high-level semantic features, and then upsampled through the decoder path to restore spatial resolution. The decoder then... The layer output is:

[0081] ;

[0082] in It comes from the depths of the network. The feature map of a deep layer contains high-level semantic information, but its resolution is very low. Upsampling operations will transform the deep, low-resolution feature map into a more comprehensive feature map. Magnify it to have the same size as the shallow feature map to prepare for stitching. It is the feature map of the encoder skip connections, directly from the first layer of the shallow network. The layer is passed horizontally, containing precise location and edge information. The concatenation operation stacks the upsampled deep information and the skipped shallow information together along the channel dimension. The convolution operation is used for feature fusion, and the semantic segmentation expert model outputs a pixel-level semantic segmentation mask, where a pixel value of 1 represents a crack and 0 represents the background.

[0083] In this step,

[0084] It receives distributed image patches from the dynamic routing step as input, and executes instance segmentation expert models and semantic segmentation expert models in parallel. Instance segmentation masks and semantic segmentation masks serve as inputs for the result summarization and quantitative analysis steps. Figure 3 The diagram shown is a flowchart of the parallel processing.

[0085] By employing parallel instance segmentation and semantic segmentation expert models, targeted processing of both discrete and continuous defects is achieved. The instance segmentation expert model utilizes a soft nonmaximum suppression algorithm, effectively addressing the overlap and missed detection issues in high-density defects. This is achieved through the use of the Gaussian variance hyperparameter. By controlling the confidence penalty, true defect instances are preserved. The semantic segmentation expert, through skip connections in the encoder-decoder structure, fuses high-level semantic information with shallow localization information, achieving pixel-level accurate segmentation of continuous defects such as cracks. Compared to existing single-model or simple multi-model fusion schemes, this step significantly improves the recall rate and segmentation accuracy of defect identification through the heterogeneous design of the expert model and the collaborative scheduling of dynamic routing.

[0086] S6. Reassemble the instance segmentation masks and semantic segmentation masks output by each expert model into a global segmentation mask image. Based on the global segmentation mask image, perform connected component labeling, pixel statistics, and morphological skeleton extraction operations to calculate the number, area, and length of defects, and generate a structured defect detection report. For example... Figure 4 The diagram shown is a flowchart of the results summary and quantitative analysis.

[0087] In this step, the local masks (including instance segmentation masks and semantic segmentation masks) are first reconstructed into a global instance segmentation mask map based on the coordinate information of the image patches in the original image. With global semantic segmentation mask image The reassembly process is achieved through a pixel-level stitching algorithm, ensuring seamless alignment of mask boundaries. Subsequently, the system calls a connected component marking algorithm to traverse the array. The number of independent connected regions is counted as the defect count. Connected component labeling adopts the eight-neighbor connection rule, and each connected region is assigned a unique identifier. The system calculates based on a pixel summation algorithm. The total number of defective pixels, combined with the ground sampling distance GSD, is used to convert the pixel count into the actual physical area. The area calculation formula is as follows:

[0088] ;

[0089] For linear defect systems, morphological skeleton extraction algorithms are applied to thin the defect region into single-pixel-wide skeleton lines. Skeleton extraction employs a thinning iterative algorithm, and the total number of skeleton line pixels, combined with the Gaussian Dispersion Score (GSD), is converted into the actual physical length. The length calculation formula is as follows: Finally, the system integrates defect identifiers, quantities, areas, and lengths to generate a structured defect detection report, which is presented on the display terminal in tabular or graphical format.

[0090] The results aggregation and quantitative analysis steps, through local mask reconstruction, connected component marking, and morphological skeleton extraction, achieve accurate calculation of the number, area, and length of defects. This invention incorporates the ground sampling distance (GSD) into the area and length conversion formula, ensuring consistency between the quantitative analysis results and actual physical dimensions. Compared to traditional methods relying on manual measurement or simple pixel statistics, this invention achieves standardized output of defect parameters through automated algorithms and supports trend comparison with historical data. The output of this step not only provides a structured report but can also be used for defect evolution prediction and maintenance decision support, greatly improving the intelligence level of offshore wind turbine blade operation and maintenance.

[0091] In the results summary and quantitative analysis step, the ground sampling distance (GSD) is calculated using the UAV camera parameters and flight altitude. The GSD formula is: ,in It is the flight altitude. It refers to the sensor pixel size. The focal length and connected component labeling algorithm employ a two-pass scanning method. The first pass assigns temporary labels, and the second pass parses equivalent labels. The morphological skeleton extraction algorithm uses the Zhang-Suen thinning algorithm, iteratively removing boundary pixels until the skeleton is stable. The quantitative analysis results are stored in the database and compared with historical data to generate trend analysis charts for predicting defect evolution and maintenance decisions.

[0092] Specifically, this embodiment also includes a quantitative analysis system for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models, used to implement the aforementioned quantitative analysis method for offshore wind turbine blade defects, such as... Figure 5 As shown, it includes:

[0093] The image acquisition module is used to acquire raw image data of offshore wind turbine blades in operation;

[0094] The image pre-enhancement module is used to perform global noise reduction and motion blur removal on the original image using an image enhancement model to obtain an enhanced image.

[0095] The image slicing module is used to slice the enhanced image into multiple non-overlapping image blocks;

[0096] The dynamic routing module is used to classify the defect type of each image patch using a routing model and dynamically distribute the image patch to the corresponding expert model according to the probability vector.

[0097] The heterogeneous expert processing module includes parallel instance segmentation expert models and semantic segmentation expert models, which are used to handle discrete defects and continuous defects, respectively.

[0098] The results summary and quantitative analysis module is used to reconstruct the local masks output by each expert model into a global segmentation mask image, and calculate the number of defects, total area and total length, and generate a structured defect detection report.

[0099] The dynamic routing module and the heterogeneous expert processing module communicate asynchronously via message queues, supporting real-time distribution of image blocks and parallel inference of expert models.

[0100] This embodiment constructs a complete quantitative analysis system for defects in offshore wind turbine blades through four core modules: image pre-enhancement, dynamic routing, heterogeneous expert processing, and result aggregation and quantitative analysis. Compared with existing technologies, this application is reflected in three aspects: First, it adopts an image enhancement model based on an attention mechanism, which significantly improves image quality in complex environments; second, it introduces a dynamic routing mechanism to realize intelligent distribution of defect images and collaborative scheduling of expert models; third, through parallel processing by a heterogeneous expert processing module, it optimizes the identification accuracy and quantitative analysis capability of discrete and continuous defects respectively. These modules jointly solve the core problems of poor imaging quality, insufficient model generalization ability, and low quantitative analysis accuracy in existing technologies, providing reliable technical support for the intelligent operation and maintenance of offshore wind turbine blades.

[0101] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A quantitative analysis method for defects in offshore wind turbine blades based on dynamic routing and heterogeneous expert models, characterized in that, include: Remotely acquire raw images of offshore wind turbine blades in operation; An enhanced image is obtained by performing global noise reduction and motion blur removal on the original image using an image enhancement model. The enhanced image is segmented into multiple non-overlapping image blocks; Each image patch is classified for defect type using a routing model, generating a probability vector for the defect type of each image patch, and the image patch is dynamically distributed to the corresponding expert model based on the probability vector; the expert model includes an instance segmentation expert model and a semantic segmentation expert model set in parallel. The distributed image patches are refined using an instance segmentation expert model and a semantic segmentation expert model, respectively. The instance segmentation expert model is used to handle discrete defects and generate instance segmentation masks, while the semantic segmentation expert model is used to handle continuous defects and generate semantic segmentation masks. The instance segmentation mask and semantic segmentation mask output by each expert model are recombined into a global segmentation mask map. Based on the global segmentation mask map, connected component labeling, pixel statistics and morphological skeleton extraction operations are performed to calculate the number, area and length of defects and generate a structured defect detection report.

2. The quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models according to claim 1, characterized in that, The image enhancement model adopts an encoder-decoder structure based on an attention mechanism. It reconstructs the local texture details of offshore wind turbine blades lost due to motion blur and noise through gated linear unit operations and channel attention mechanism, while preserving the global context information of the image.

3. The quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models according to claim 1, characterized in that, The routing model extracts multi-level feature maps of image patches through a feature extraction network, compresses them into feature vectors through a global average pooling layer, and then inputs them into a fully connected classification head. The output is a logical score vector containing the original score of each defect category. The Sigmoid activation function is applied independently to each defect category to generate the probability vector. The maximum value in the probability vector is compared according to a preset threshold. If the maximum value corresponds to a discrete defect, the image patch is distributed to the instance segmentation expert model. If the maximum value corresponds to a continuous defect, the image patch is distributed to the semantic segmentation expert model.

4. The quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models according to claim 1, characterized in that, The instance segmentation expert model is based on a single-stage instance segmentation architecture. After the input image block extracts features through the backbone network, the detection head predicts the bounding box and mask coefficients. The mask coefficients Prototype mask shared globally Linear combination generates instance segmentation mask : ； in, For the number of prototypes, For the Sigmoid function: 。 5. The quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models according to claim 4, characterized in that, The instance segmentation expert model employs a soft nonmaximum suppression algorithm to handle overlap defects, penalizing the confidence score of each detection box using the following penalty formula: ； in, For the first The updated confidence score for each detection box; For the first The original confidence score of each detection box before the update; This is the intersection-union ratio (IUU) between the highest-scoring bounding box and the current bounding box. This is the Gaussian variance hyperparameter, used to control the severity of the penalty.

6. The quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models according to claim 1, characterized in that, The semantic segmentation expert model is based on an encoder-decoder architecture. It uses skip connections to splice and fuse high-level semantic information from the deep layers of the network with precise localization and edge information from the shallow layers of the network, and outputs a pixel-level semantic segmentation mask.

7. The quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models according to claim 1, characterized in that, Defect area The calculation formula is: ； Defect length The calculation formula is: ； in, This represents the total number of defective pixels. This represents the total number of pixels in the skeleton lines. The ground sampling distance is calculated using the drone's camera parameters and flight altitude.

8. A quantitative analysis system for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models, used to implement the quantitative analysis method for offshore wind turbine blade defects as described in any one of claims 1-7, characterized in that, include: The image acquisition module is used to acquire raw image data of offshore wind turbine blades in operation; The image pre-enhancement module is used to perform global noise reduction and motion blur removal on the original image using an image enhancement model to obtain an enhanced image. The image slicing module is used to slice the enhanced image into multiple non-overlapping image blocks; The dynamic routing module is used to classify the defect type of each image patch using a routing model and dynamically distribute the image patch to the corresponding expert model according to the probability vector. The heterogeneous expert processing module includes parallel instance segmentation expert models and semantic segmentation expert models, which are used to handle discrete defects and continuous defects, respectively. The results summary and quantitative analysis module is used to reconstruct the local masks output by each expert model into a global segmentation mask image, and calculate the number of defects, total area and total length, and generate a structured defect detection report.

9. The quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models according to claim 8, characterized in that, The dynamic routing module and the heterogeneous expert processing module communicate asynchronously via message queues, supporting real-time distribution of image blocks and parallel inference of expert models.

10. The quantitative analysis method for offshore wind turbine blade defects based on dynamic routing and heterogeneous expert models according to claim 8, characterized in that, The results summary and quantitative analysis module uses a two-pass scanning method to perform a connected component labeling algorithm and a Zhang-Suen thinning algorithm to perform morphological skeleton extraction. The quantitative analysis results are stored in a database and compared with historical data to generate trend analysis charts.

Images