Wind turbine blade surface damage identification method and system based on multi-module collaboration

Abstract

The present invention provides a method and system for identifying surface damage of wind turbine blades based on multi-module collaboration, which relates to the technical field of automatic damage identification. The method includes: using a multi-module collaborative damage identification model obtained based on a YOLOv8 network combined with multiple auxiliary processing modules to identify the damage location and damage type in the surface image of the wind turbine blade; the auxiliary processing module in the multi-module collaborative damage identification model can be any one of an image augmentation module, an image clarity detection module, and a thermal map generation module. Compared with the existing traditional damage detection methods that rely on manpower, it can automatically and accurately complete the damage identification of wind turbine blades, avoid the safety risks of field operations under complex environmental conditions, and avoid the large errors caused by the human measurement results being easily affected by subjective factors, which is suitable for crack damage detection under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic damage identification, and in particular to a method and system for identifying surface damage of a wind turbine blade based on multi-module collaboration. Background Art

[0002] The development of clean energy has been included in the plan. As a representative technology of clean energy, wind power generation converts wind power in nature into electrical energy, which has a good cleaning effect. However, since wind turbine blades need to be exposed to the external environment for a long time and bear the influence of external loads, cracks and shedding of protective films are prone to damage on the surface. In order to ensure that the wind turbine generator is in good operating condition after being put into operation, regular inspections are required by operation and maintenance personnel, which consumes huge human and material resources. Therefore, accurate and rapid positioning of damage can help shorten the standby time of wind turbine blades, reduce the cost of maintenance manpower and material resources, and ensure the continuity of power supply of wind power generation systems.

[0003] Traditional damage detection methods rely on manpower. Under complex environmental conditions, field operations pose safety risks. In addition, manual measurement results are easily affected by subjective factors and produce large errors. There are also disadvantages such as low detection accuracy, susceptibility to interference, and poor robustness. They are not suitable for crack damage detection under complex working conditions. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for determining surface damage of wind turbine blades based on multi-module collaboration, thereby achieving automated and accurate determination of surface damage of wind turbine blades.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] On the one hand, the present invention provides a method for determining surface damage of a wind turbine blade based on multi-module collaboration, comprising the following steps:

[0007] Acquire wind turbine blade surface images.

[0008] A multi-module collaborative damage discrimination model is used to identify the damage location and damage type in the surface image of the wind blade; the multi-module collaborative damage discrimination model is a model obtained by combining a YOLOv8 network with multiple auxiliary processing modules; the auxiliary processing module is any one of the image augmentation module, image clarity detection module and thermal map generation module.

[0009] Optionally, the YOLOv8 network includes a skeleton network, a neck network and a head network; the skeleton network is used to extract fused feature maps of different scales of the wind blade surface image; the neck network is used to generate fused feature maps to be identified of different scales based on the fused feature maps of different scales; the head network is used to determine the damage position and damage type in the wind blade surface image based on the damage position and damage type in each fused feature map to be identified.

[0010] Optionally, the skeleton network includes a first convolutional layer, a second convolutional layer, a first C2F layer, a third convolutional layer, a second C2F layer, a fourth convolutional layer, a third C2F layer, a fifth convolutional layer, a fourth C2F layer, a sixth convolutional layer, a fifth C2F layer and an SPPF spatial pyramid pooling module connected in sequence; the third C2F layer is used to output a first fused feature map, the fourth C2F layer is used to output a second fused feature map, and the SPPF spatial pyramid pooling module is used to output a third fused feature map; the SPPF spatial pyramid pooling module includes a seventh convolutional layer, a parallel maximum pooling module, a first connection layer and an eighth convolutional layer connected in sequence; the parallel maximum pooling module includes a first maximum pooling layer, a second maximum pooling layer and a third maximum pooling layer connected in parallel.

[0011] Optionally, an image augmentation module is added between the second convolutional layer and the first C2F layer, and the image augmentation module is used to generate a number of virtual convolution information according to the convolution information output by the second convolutional layer and random Gaussian noise, and replace the convolution information output by the second convolutional layer with the number of virtual convolution information as the input of the first C2F layer.

[0012] Optionally, an image clarity detection module is added between the skeleton network and the neck network, and the image clarity detection module includes a first image clarity detection submodule connected to the output end of the third C2F layer, a second image clarity detection submodule connected to the output end of the fourth C2F layer, and a third image clarity detection submodule connected to the output end of the SPPF spatial pyramid pooling module; the first image clarity detection submodule is used to perform superpixel reconstruction on the first fusion feature map to obtain a first reconstructed fusion feature map, and output the first reconstructed fusion feature map with a gradient greater than a first threshold; the second image clarity detection submodule is used to perform superpixel reconstruction on the second fusion feature map to obtain a second reconstructed fusion feature map, and output the second reconstructed fusion feature map with a gradient greater than the second threshold; the third image clarity detection submodule is used to perform superpixel reconstruction on the third fusion feature map to obtain a third reconstructed fusion feature map, and output the third reconstructed fusion feature map with a gradient greater than the third threshold.

[0013] Optionally, the neck network includes a second connection layer, a sixth C2F layer, a first upsampling layer, a third connection layer, a seventh C2F layer, a ninth convolutional layer, a fourth connection layer, an eighth C2F layer, a tenth convolutional layer, a fifth connection layer and a ninth C2F layer connected in sequence; the second connection layer is connected to the output end of the second image clarity detection submodule, the second connection layer is connected to the output end of the third image clarity detection submodule through the second upsampling layer, the third connection layer is connected to the output end of the first image clarity detection submodule, the fourth connection layer is connected to the output end of the sixth C2F layer, the fifth connection layer is connected to the output end of the third image clarity detection submodule, and the fifth connection layer is connected to the output end of the tenth convolutional layer.

[0014] Optionally, a heat map generation module is added between the tenth convolutional layer and the fifth connection layer; the heat map generation module is used to convert the feature map output by the tenth convolutional layer into a heat map form.

[0015] Optionally, the heat map generation module includes a global average pooling layer, a fully connected layer and a softmax activation function layer connected in sequence.

[0016] Optionally, the head network includes a first detection head connected to the output end of the seventh C2F layer, a second detection head connected to the output end of the eighth C2F layer, and a third detection head connected to the output end of the ninth C2F layer; the first detection head is used to identify the damage location and damage type in the first to-be-identified fused feature map output by the seventh C2F layer; the second detection head is used to identify the damage location and damage type in the second to-be-identified fused feature map output by the eighth C2F layer; the third detection head is used to identify the damage location and damage type in the third to-be-identified fused feature map output by the ninth C2F layer.

[0017] On the other hand, corresponding to the aforementioned method for determining surface damage of wind blades based on multi-module collaboration, the present invention also provides a system for determining surface damage of wind blades based on multi-module collaboration. When the system for determining surface damage of wind blades based on multi-module collaboration is run by a computer, it executes the method for determining surface damage of wind blades based on multi-module collaboration as described above.

[0018] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0019] The present invention provides a method and system for identifying surface damage of wind turbine blades based on multi-module collaboration, the method comprising: using a multi-module collaborative damage identification model obtained based on a YOLOv8 network combined with multiple auxiliary processing modules to identify the damage location and damage type in the surface image of the wind turbine blade; the auxiliary processing module in the multi-module collaborative damage identification model can be any one of an image augmentation module, an image clarity detection module, and a thermal map generation module. Compared with the existing traditional damage detection methods that rely on manpower, it can automatically and accurately complete the damage identification of wind turbine blades, avoiding the safety risks of field operations under complex environmental conditions, and avoiding the large errors caused by the human measurement results being easily affected by subjective factors, and is suitable for crack damage detection under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0021] Figure 1 A flow chart of a method for determining surface damage of a wind turbine blade based on multi-module collaboration provided in Example 1 of the present invention;

[0022] Figure 2 A schematic diagram of the structure of a YOLOv8 network combining an image augmentation module, an image clarity detection module and a heat map generation module in a wind turbine blade surface damage identification method based on multi-module collaboration provided in Example 1 of the present invention;

[0023] Figure 3 A schematic diagram of adversarial training of an image augmentation module in a wind turbine blade surface damage identification method based on multi-module collaboration provided in Example 1 of the present invention;

[0024] Figure 4 A schematic diagram of any image clarity detection submodule in the wind turbine blade surface damage identification method based on multi-module collaboration provided in Example 1 of the present invention;

[0025] Figure 5 A schematic diagram of a thermal map generation module in a wind turbine blade surface damage identification method based on multi-module collaboration provided in Example 1 of the present invention;

[0026] Figure 6 A schematic diagram of the structure of a wind turbine blade surface damage identification system based on multi-module collaboration provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The purpose of the present invention is to provide a method and system for determining surface damage of wind turbine blades based on multi-module collaboration, thereby achieving automated and accurate determination of surface damage of wind turbine blades.

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] Embodiment 1:

[0031] This embodiment provides a method for determining surface damage of a wind turbine blade based on multi-module collaboration. Figure 1 As shown in the flowchart, the method provided in this embodiment includes the following steps:

[0032] A1. Obtain the surface image of the wind turbine blade. In this embodiment, a fully autonomous multi-rotor drone is used to cruise over the wind turbine blade to identify abnormal parts, and a camera is used to collect data on the damaged parts. The damage identification task is mainly calculated by the edge node. The input end transmits the roughly collected data information to the GPS navigation module to confirm the position information, and uses the VL53L0X laser ranging module to accurately measure the position. The measurement mode is pre-set, for example, timed continuous measurement is selected. After the VL53L0X laser ranging module is completed after a measurement, the system compares the measurement data and restores the standby state. If the ranging value does not reach the expected setting, the next ranging will be restarted after the measurement time exceeds the defined delay time until the ranging value is within the optimal threshold range. The measurement process is terminated, the ranging cycle ends, and the coordinate information is generated for use in the next link.

[0033] A2. Use a multi-module collaborative damage discrimination model to identify the damage location and damage type in the surface image of the wind blade. The multi-module collaborative damage discrimination model is a model based on the YOLOv8 network combined with multiple auxiliary processing modules; the auxiliary processing module is any one of the image augmentation module, the image clarity detection module and the thermal map generation module. Among them, the YOLOv8 network includes a skeleton network, a neck network and a head network; the skeleton network is used to extract fused feature maps of different scales of the surface image of the wind blade; the neck network is used to generate fused feature maps of different scales to be identified based on fused feature maps of different scales; the head network is used to determine the damage location and damage type in the surface image of the wind blade based on the damage location and damage type in each fused feature map to be identified.

[0034] Specifically, if Figure 2 The structural diagram of the YOLOv8 network combined with the image augmentation module, the image clarity detection module and the heat map generation module is shown, the skeleton network in the YOLOv8 network includes the first convolutional layer, the second convolutional layer, the first C2F layer, the third convolutional layer, the second C2F layer, the fourth convolutional layer, the third C2F layer, the fifth convolutional layer, the fourth C2F layer, the sixth convolutional layer, the fifth C2F layer and the SPPF spatial pyramid pooling module connected in sequence; the third C2F layer is used to output the first fused feature map, the fourth C2F layer is used to output the second fused feature map, and the SPPF spatial pyramid pooling module is used to output the third fused feature map; the SPPF spatial pyramid pooling module includes the seventh convolutional layer, the parallel maximum pooling module, the first connection layer and the eighth convolutional layer connected in sequence; the parallel maximum pooling module includes the first maximum pooling layer, the second maximum pooling layer and the third maximum pooling layer connected in parallel. The wind turbine blade surface image obtained in step A1 is input from the input end of the first convolutional layer. Based on the first and second convolutional layers connected in sequence of the YOLOv8 skeleton network, the image is subjected to two convolution CBS processes. Each convolution is divided into three steps. The data first undergoes a two-dimensional 3×3 convolution. Secondly, to prevent data overfitting, BatchNorm2d is used to perform batch normalization on the input array. Finally, the SiLU activation function nonlinear factor is referenced to comprehensively map the image information.

[0035] In some optional implementations, an image augmentation module is added between the second convolution layer and the first C2F layer. The image augmentation module is used to generate a number of virtual convolution information based on the convolution information output by the second convolution layer and random Gaussian noise, and replace the convolution information output by the second convolution layer with the virtual convolution information as the input of the first C2F layer. The image augmentation module is based on the GAN generative adversarial network, which efficiently combines the generative network (G model) and the discriminative network (D model) in machine learning. The generative network G model is used to generate an image to deceive the discriminative network D model, and the D model tries its best to distinguish the true from the false by the image difference, thereby screening out the most authentic image source information through a zero-sum game. Figure 3 As shown in the schematic diagram, the convolution information output by the second convolutional layer in the previous step randomly generates an n-dimensional vector of mean distribution or random Gaussian noise for sampling, and one of the two types of data is used as the input of the discriminant model. The D model determines the true or false probability of the data, and calculates the loss function according to the probability value, which is fed back to the G model. Finally, the parameters of the G model are updated by updating the D model. The data is enhanced by iterative training, and the processing capabilities of the two models are improved synchronously until the data generated by the G model is sufficient to be indistinguishable from the real one and reaches a certain balance with the ability of the D model. In this mutual confrontation, the global optimum is achieved, and high-quality data is output. During the evaluation process, the two models continue to evolve, and the generative adversarial network conducts image information game, so that the trained image generated data and the real data have the same distribution.

[0036] The output of the image augmentation module is processed alternately by four C2F modules and convolutions in the YOLOv8 architecture skeleton network (the first C2F layer and the third convolution layer, the second C2F layer and the fourth convolution layer, the third C2F layer and the fifth convolution layer, and the fourth C2F layer and the sixth convolution layer). In each C2F layer, the convolution data is processed by channel. Part of the data A remains unchanged, and the other part of the data B is processed by deep convolution. The two are fused to form a data body, and then the processed data B is further deeply convolved and fused with the data body. Each time, only the data of the previous deep convolution is repeatedly processed. The feature fusion is achieved by fully connecting the data of the low-level feature map with the high-level feature map so that they have corresponding adjustments in the number of channels and size. The fused feature map is used as the output of the C2F layer for use by subsequent networks. By fusing feature maps of different levels, the C2F layer improves the image perception ability of the model itself. After the C2F layer and the convolution layer in the skeleton network are alternately connected four times, a C2F layer is finally added as the convolution end to end the feature fusion process of the skeleton network and enter the SPPF module. After convolution of the feature data, three parallel convolution kernels with a maximum pooling of 5*5 are used to ensure that the size of the SPPF output feature map remains unchanged, which is conducive to reducing the computational workload and completing multi-scale feature fusion. This step prepares for the neck network data processing.

[0037] In another optional embodiment, an image clarity detection module is added between the skeleton network and the neck network, and the image clarity detection module includes a first image clarity detection submodule connected to the output end of the third C2F layer, a second image clarity detection submodule connected to the output end of the fourth C2F layer, and a third image clarity detection submodule connected to the output end of the SPPF spatial pyramid pooling module; the first image clarity detection submodule is used to perform superpixel reconstruction on the first fused feature map output by the third C2F layer to obtain a first reconstructed fused feature map, and output the first reconstructed fused feature map with a gradient greater than a first threshold; the second image clarity detection submodule is used to perform superpixel reconstruction on the second fused feature map output by the fourth C2F layer to obtain a second reconstructed fused feature map, and output the second reconstructed fused feature map with a gradient greater than the second threshold; the third image clarity detection submodule is used to perform superpixel reconstruction on the third fused feature map output by the SPPF spatial pyramid pooling module to obtain a third reconstructed fused feature map, and output the third reconstructed fused feature map with a gradient greater than the third threshold.

[0038] like Figure 4 As shown in the schematic diagram, each image clarity detection submodule contains multiple ResBlock residual blocks. Multiple ResBlock modules are used to merge different feature information. The UpSample upsampling layer connected thereto is used to amplify the image and improve the image resolution. After convolution, a feature map with doubled number of channels is obtained. Image cutting and pixel rearrangement are performed. The two channel feature maps are inserted into each other and the size is doubled, so that the image is finally reconstructed into superpixels, retaining the original contrast information of the image. The Tenegrad gradient function is used to measure the clarity, and the sobel operator is used to calculate the horizontal and vertical gradients. At this time, the clearer the image, the larger the gradient value. By setting a reasonable threshold, images below the threshold will not be selected, while images above the threshold meet the clarity requirements of damage detection and further participate in the model calculation.

[0039] Specifically, the neck network in the YOLOv8 network includes a second connection layer, a sixth C2F layer, a first upsampling layer, a third connection layer, a seventh C2F layer, a ninth convolutional layer, a fourth connection layer, an eighth C2F layer, a tenth convolutional layer, a fifth connection layer and a ninth C2F layer connected in sequence; the second connection layer is connected to the output end of the second image clarity detection submodule, the second connection layer is connected to the output end of the third image clarity detection submodule through the second upsampling layer, the third connection layer is connected to the output end of the first image clarity detection submodule, the fourth connection layer is connected to the output end of the sixth C2F layer, the fifth connection layer is connected to the output end of the third image clarity detection submodule, and the fifth connection layer is connected to the output end of the tenth convolutional layer. The neck structure belongs to the supplementary processing part of the skeleton network. One part is subjected to feature fusion through the skeleton network, and the other part of the feature data is processed by the third C2F layer and the fourth C2F layer and enters the corresponding EDSR image clarity detection submodule in the neck network. The residual idea is used to solve the super-resolution problem. The Mult residual scaling layer is used after the convolution layer and the Relu activation function layer, so that the residual block multiplies the convolution-processed data by an arbitrarily selected decimal before adding, to ensure the stability of the network. Compared with the original module, the BatchNorm layer after removing two convolution layers, the middle ReLU activation function layer remains unchanged. Under the same computing resources, EDSR can stack more layers and extract more features in each layer.

[0040] In some other implementations, a heat map generation module is added between the tenth convolutional layer and the fifth connection layer; the heat map generation module is used to convert the feature map output by the tenth convolutional layer into a heat map form. In this embodiment, the heat map generation module is based on the CAM attention mechanism, such as Figure 5As shown in the schematic diagram, the heat map generation module includes the eleventh convolutional layer, the twelfth convolutional layer, the thirteenth convolutional layer and the fourteenth convolutional layer of different sizes connected in sequence, and then the global average pooling layer, the fully connected layer and the softmax activation function layer connected in sequence. The neck network mainly performs secondary processing on the feature information extracted by the third C2F layer and the fourth C2F layer in the skeleton network to integrate the features of the skeleton network. The C2F operation process is repeated in the neck network for the screened images with higher clarity, and the CAM attention mechanism is introduced in the last convolutional layer. This attention mechanism performs global average pooling on the convolutional data. The method of extracting features by global average pooling can prevent information overfitting, and expand the pooled results in one-dimensional form to form a fully connected layer to comprehensively display the output features. The weights of the fully connected layer are multiplied by the feature map set in different batches and different numbers of iterations, and then the classification results are calculated by the softmax activation function to enhance the network learning ability. At the same time, the weights of all feature maps in the last convolutional layer for the image category are calculated, and the damage features of the image are mapped out in the form of a heat map by weighting the heat map and the original image, so as to realize the visualization of the weighted results of the original image and the heat map, and then the image recognition results are obtained.

[0041] Specifically, the head network in the YOLOv8 network includes a first detection head connected to the output end of the seventh C2F layer, a second detection head connected to the output end of the eighth C2F layer, and a third detection head connected to the output end of the ninth C2F layer; the first detection head is used to identify the damage location and damage type in the first fused feature map to be identified output by the seventh C2F layer; the second detection head is used to identify the damage location and damage type in the second fused feature map to be identified output by the eighth C2F layer; the third detection head is used to identify the damage location and damage type in the third fused feature map to be identified output by the ninth C2F layer. The first detection head, the second detection head, and the third detection head output feature maps of sizes 1*66*80*80, 1*66*40*40, and 1*66*20*20, respectively, to predict damage at different scales. Based on the traditional YOLOv8 network, the decoupled head structure is used to divide the detection results into two parts for convolution. One part is used for target recognition and the location of the damage is identified. The number of channels of the regression head is changed to 4*reg_max (reg_max defaults to 16), while the other part is used for classification and output of the damage type to improve the detection effect. Furthermore, the results output by the head network can be transmitted to the display of the fully autonomous multi-rotor drone control end. The operator can quickly receive high-definition pictures of the damage location at the command center, which is convenient for further formulating repair plans.

[0042] This embodiment provides a method for identifying surface damage of wind turbine blades based on multi-module collaboration, using a multi-module collaborative damage identification model based on the YOLOv8 network combined with multiple auxiliary processing modules to identify the damage location and damage type in the surface image of the wind turbine blade; the auxiliary processing module in the multi-module collaborative damage identification model can be any one of the image augmentation module, image clarity detection module and thermal map generation module. Compared with the existing traditional damage detection methods that rely on manpower, the solution of this embodiment can automatically and accurately complete the damage identification of wind turbine blades, avoiding the safety risks of field operations under complex environmental conditions, and avoiding the large errors caused by the human measurement results that are easily affected by subjective factors, and is suitable for crack damage detection under complex working conditions.

[0043] Embodiment 2:

[0044] In addition, the method of embodiment 1 of the present invention can also be used by Figure 6 The wind turbine blade surface damage identification system based on multi-module collaboration is implemented as shown in the figure. Figure 6 As shown, the wind turbine blade surface damage identification system based on multi-module collaboration may include a wind turbine blade surface image acquisition module, an image augmentation module, an image clarity detection module, a thermal map generation module, a damage type identification module, and a damage location identification module; some of these modules may also have sub-units for realizing their functions, such as the image clarity detection module including a first image clarity detection sub-module, a second image clarity detection sub-module, and a third image clarity detection sub-module. Of course, Figure 6 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different functions. Figure 6 One or at least two components of the system shown.

[0045] Specific examples are used in this article, but the above description is only to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas; technicians in this field should understand that the modules or steps of the present invention can be implemented by a general computer device, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by a computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps in them can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0046] Meanwhile, for those skilled in the art, according to the concept of the present invention, there will be changes in the specific implementation and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for identifying surface damage of wind turbine blades based on multi-module collaboration, characterized in that: include: Acquire the surface image of the wind turbine blade; Using a multi-module collaborative damage discrimination model, identifying the damage location and damage type in the surface image of the wind turbine blade; The multi-module collaborative damage discrimination model is a model obtained by combining a plurality of auxiliary processing modules based on a YOLOv8 network; the auxiliary processing module is any one of an image augmentation module, an image clarity detection module and a thermal map generation module; The YOLOv8 network includes a skeleton network, a neck network and a head network; the skeleton network is used to extract fused feature maps of different scales of the wind turbine blade surface image; the neck network is used to generate fused feature maps to be identified of different scales based on the fused feature maps of different scales; the head network is used to determine the damage position and damage type in the wind turbine blade surface image based on the damage position and damage type in each fused feature map to be identified; The skeleton network includes a first convolutional layer, a second convolutional layer, a first C2F layer, a third convolutional layer, a second C2F layer, a fourth convolutional layer, a third C2F layer, a fifth convolutional layer, a fourth C2F layer, a sixth convolutional layer, a fifth C2F layer and an SPPF spatial pyramid pooling module connected in sequence; the third C2F layer is used to output a first fused feature map, the fourth C2F layer is used to output a second fused feature map, and the SPPF spatial pyramid pooling module is used to output a third fused feature map; the SPPF spatial pyramid pooling module includes a seventh convolutional layer, a parallel maximum pooling module, a first connection layer and an eighth convolutional layer connected in sequence; the parallel maximum pooling module includes a first maximum pooling layer, a second maximum pooling layer and a third maximum pooling layer connected in parallel with each other; The image clarity detection module is added between the skeleton network and the neck network, and the image clarity detection module includes a first image clarity detection submodule connected to the output end of the third C2F layer, a second image clarity detection submodule connected to the output end of the fourth C2F layer, and a third image clarity detection submodule connected to the output end of the SPPF spatial pyramid pooling module; the first image clarity detection submodule is used to perform superpixel reconstruction on the first fusion feature map to obtain a first reconstructed fusion feature map, and output the first reconstructed fusion feature map with a gradient greater than a first threshold; the second image clarity detection submodule is used to perform superpixel reconstruction on the second fusion feature map to obtain a second reconstructed fusion feature map, and output the second reconstructed fusion feature map with a gradient greater than a second threshold; the third image clarity detection submodule is used to perform superpixel reconstruction on the third fusion feature map to obtain a third reconstructed fusion feature map, and output the third reconstructed fusion feature map with a gradient greater than a third threshold; The neck network includes a second connection layer, a sixth C2F layer, a first upsampling layer, a third connection layer, a seventh C2F layer, a ninth convolutional layer, a fourth connection layer, an eighth C2F layer, a tenth convolutional layer, a fifth connection layer and a ninth C2F layer connected in sequence; the second connection layer is connected to the output end of the second image definition detection submodule, the second connection layer is connected to the output end of the third image definition detection submodule through the second upsampling layer, the third connection layer is connected to the output end of the first image definition detection submodule, the fourth connection layer is connected to the output end of the sixth C2F layer, the fifth connection layer is connected to the output end of the third image definition detection submodule, and the fifth connection layer is connected to the output end of the tenth convolutional layer; The head network includes a first detection head connected to the output end of the seventh C2F layer, a second detection head connected to the output end of the eighth C2F layer, and a third detection head connected to the output end of the ninth C2F layer; the first detection head is used to identify the damage position and damage type in the first fused feature map to be identified output by the seventh C2F layer; the second detection head is used to identify the damage position and damage type in the second fused feature map to be identified output by the eighth C2F layer; the third detection head is used to identify the damage position and damage type in the third fused feature map to be identified output by the ninth C2F layer.

2. The method for identifying surface damage of a wind turbine blade based on multi-module collaboration according to claim 1 is characterized in that: The image augmentation module is added between the second convolutional layer and the first C2F layer, and the image augmentation module is used to generate a plurality of virtual convolution information according to the convolution information output by the second convolutional layer and random Gaussian noise, and replace the convolution information output by the second convolutional layer with the plurality of virtual convolution information as the input of the first C2F layer.

3. The method for identifying surface damage of a wind turbine blade based on multi-module collaboration according to claim 2 is characterized in that: A heat map generation module is added between the tenth convolutional layer and the fifth connection layer; the heat map generation module is used to convert the feature map output by the tenth convolutional layer into a heat map form.

4. The method for identifying surface damage of a wind turbine blade based on multi-module collaboration according to claim 3 is characterized in that: The heat map generation module includes a global average pooling layer, a fully connected layer and a softmax activation function layer connected in sequence.

5. A wind turbine blade surface damage identification system based on multi-module collaboration, characterized in that: When the wind turbine blade surface damage identification system based on multi-module collaboration is run by a computer, it executes the wind turbine blade surface damage identification method based on multi-module collaboration as described in any one of claims 1 to 4.

Images