A method and system for determining fatigue crack defects of a floating wind power platform

Abstract

The present invention discloses a method and system for determining fatigue crack defects of a floating wind power platform, the method comprising: reconstructing details of a first image set of the floating wind power platform to obtain a second image set; obtaining texture features by converting the first image set into a grayscale image set to obtain a comprehensive feature set; identifying the comprehensive feature set to obtain an actual risk value set of a pixel abnormal point set; and determining fatigue crack defects in corresponding areas of the floating wind power platform according to the numerical values ​​of the risk values ​​in the actual risk value set. The present invention proposes a method and system for determining fatigue crack defects of a floating wind power platform, which improves image quality through detailed reconstruction, and improves the accuracy and reliability of subsequent identification through the extraction of grayscale images and texture features; using a convolutional neural network to identify and analyze the comprehensive feature set can realize accurate identification and risk assessment of defects such as cracks, and can solve the problem of difficulty in accurately identifying fatigue crack defects of floating wind power platforms.

Description

Technical Field

[0001] The present invention relates to the technical field of offshore wind power, and in particular to a fatigue crack defect determination method and system for a floating wind power platform. Background Art

[0002] Traditional offshore wind power projects are usually limited by water depth and geological conditions, while floating wind power platforms can operate in deeper waters. Fatigue cracks on floating wind power platforms are gradually formed when the structure is subjected to long-term alternating loads. By regularly detecting and evaluating fatigue cracks, measures can be taken in advance to repair or replace damaged parts to prevent sudden failures. Existing methods mainly use ultrasonic testing and stress analysis to detect and evaluate defects such as cracks and corrosion on floating wind power platforms.

[0003] However, the propagation and reflection characteristics of ultrasound in materials are affected by the properties and shape of the materials, which may lead to reduced sensitivity and accuracy of ultrasonic testing in some cases. Ultrasonic testing requires professional operating skills and experience, and non-professional operations cannot ensure the accuracy and reliability of the test results. During operation, floating wind power platforms are subject to a variety of environmental loads such as wind loads, wave loads, and flow loads. The complexity and uncertainty of these loads make stress analysis more difficult, resulting in reduced accuracy in identifying fatigue crack defects on floating wind power platforms, making it difficult to improve the safety of floating wind power platforms. Summary of the invention

[0004] The present invention provides a method and system for determining fatigue crack defects of a floating wind power platform, so as to solve the problem that it is difficult to accurately identify fatigue crack defects of a floating wind power platform.

[0005] Collecting images of the floating wind power platform to obtain a first image set;

[0006] Reconstructing the details of the first image set by a detail enhancement model to obtain a second image set; wherein the detail enhancement model is established based on a generator and a discriminator, and the generation loss function of the generator is established based on the adversarial loss and content loss of the generator and the discriminator in the game process;

[0007] Marking a pixel abnormal point set in the second image set based on a dilation operation, and acquiring texture features of the second image set by converting the second image set into a grayscale image set to obtain a comprehensive feature set;

[0008] Performing risk identification on the comprehensive feature set through a convolutional neural network to obtain an actual risk value set of the pixel abnormal point set;

[0009] The fatigue crack defect in the corresponding area of ​​the floating wind power platform is determined according to the numerical value of the risk value in the actual risk value set to obtain a defect identification result.

[0010] In the present invention, the generation loss function of the generator combines the adversarial loss and the content loss, which can ensure that the second image set can retain the important information of the first image set while maintaining authenticity; and in the game process between the generator and the discriminator, the adversarial loss can encourage the generator to generate images that can deceive the discriminator, that is, the details of the real image are as close as possible. This mechanism prompts the generator to continuously optimize its ability to generate details, making fine cracks that may be difficult to detect due to factors such as noise, blur or insufficient light more obvious. Strengthening the boundaries and details of the image through dilation operations helps to highlight small changes in the image. By marking pixel abnormal point sets, areas where cracks may exist can be quickly located, providing a precise target for subsequent risk assessment. Converting images into grayscale image sets can simplify the image processing process, thereby removing the interference of color information on crack detection, making texture features more prominent, and using the extracted texture features as part of the comprehensive feature set can provide a rich source of information for convolutional neural network recognition.

[0011] Compared with the prior art, the present invention improves image quality through detail reconstruction, and the extraction of grayscale images and texture features improves the accuracy and reliability of subsequent recognition; by using a convolutional neural network to identify and analyze the comprehensive feature set, the actual risk value set of each pixel abnormal point set can be obtained, and accurate recognition and risk assessment of defects such as cracks can be achieved, thereby determining the fatigue crack defects in the corresponding area of ​​the floating wind power platform, thereby solving the problem of difficulty in accurately identifying fatigue crack defects in the floating wind power platform.

[0012] As a preferred solution, the generation loss function of the generator is established based on the adversarial loss and content loss of the generator and the discriminator during the game, specifically:

[0013] Controlling the discriminator to score the real image and the generated image generated by the generator, and establishing the adversarial loss of the generator according to the scores obtained by the scoring;

[0014] Establishing a content loss of the generator based on a difference value between the generated image and a preset target image at a feature level;

[0015] Extracting features of the real image and the generated image through the discriminator to obtain a feature atlas of the generator;

[0016] Establishing a feature matching loss of the generator based on the image differences in the feature atlas;

[0017] Establishing a perceptual loss of the generator based on a similarity evaluation result between the generated image and the target image on preset high-level semantic features;

[0018] The generation loss function of the generator is established according to the adversarial loss, the content loss, the feature matching loss and the perceptual loss.

[0019] In this preferred embodiment, the adversarial loss encourages the generator to generate images that can deceive the discriminator, thereby making the generated images visually closer to the real images. The content loss can ensure that the generated image is consistent with the target image at the content level, which helps to retain important information of the original image in image conversion or restoration tasks. The feature matching loss further guides the generator to learn more refined image features by comparing the differences between the real image and the generated image at the feature level, which helps to improve the detail richness and overall quality of the generated image. The perceptual loss is based on the preset high-level semantic feature similarity evaluation results, which helps the generator generate images that are more perceptually natural and realistic.

[0020] As a preferred solution, the anti-loss is specifically:

[0021]

[0022] in, is the image input to the generator, is the prior distribution, The mathematical symbol for expected value, is the generated image generated by the generator, is the probability that the discriminator judges that the generated image is the real image.

[0023] As a preferred solution, a pixel abnormal point set in the second image set is marked based on a dilation operation, and texture features of the second image set are obtained by converting the second image set into a grayscale image set to obtain a comprehensive feature set, specifically:

[0024] On the basis of performing image enhancement processing on the second image set, identifying image abnormal points by controlling the structure element to slide on the second image set, thereby obtaining the pixel abnormal point set;

[0025] Converting the second image set into grayscale images to obtain a grayscale image set;

[0026] Establishing a co-occurrence probability matrix according to the gray level co-occurrence frequency between pixels in the gray image set;

[0027] A comprehensive feature set about the pixel abnormal point set and texture features is obtained from the co-occurrence probability matrix.

[0028] This preferred solution converts the color information of the image into grayscale, which helps to reduce the impact of color noise on abnormality detection. By establishing a co-occurrence probability matrix of the grayscale co-occurrence frequency between pixels in the grayscale image set, the texture features in the image can be captured, which helps to distinguish normal areas from abnormal areas.

[0029] As a preferred solution, the image abnormal points are identified by controlling the structural element to slide on the second image set to obtain the pixel abnormal point set, specifically:

[0030] determining the structural element according to the size and shape of the abnormal image region in the second image set;

[0031] An extended image set is obtained by extending the boundary of the primary abnormal region in the second image set; wherein the primary abnormal region is obtained by performing preliminary abnormality detection on the second image set by a machine learning model;

[0032] The structural element is controlled to slide on the image of the extended image set. If there are abnormal points in the area covered by the structural element, the pixel points at the regional position corresponding to the abnormal points are marked as abnormal, so as to obtain the pixel abnormal point set.

[0033] This preferred solution determines the structural elements according to the actual size and shape of the abnormal area, which can make the customized structural elements better adapt to different abnormal areas and improve the accuracy of abnormality detection. The primary abnormal area is obtained through preliminary detection by the machine learning model. Although it can roughly locate the abnormal position, there may be false positives or false negatives. Therefore, by expanding the boundaries of these areas and performing further sliding detection on the expanded image set, more potential abnormal points can be captured, while reducing the anomalies missed due to boundary restrictions.

[0034] As a preferred solution, a co-occurrence probability matrix is ​​established according to the gray level co-occurrence frequency between pixels in the gray image set, specifically:

[0035] Based on the grayscale image set, the grayscale level co-occurrence frequency between each pixel and neighboring pixels at a specified distance and in a specified direction is counted to obtain a co-occurrence frequency set;

[0036] Establishing a gray-level co-occurrence matrix according to the co-occurrence frequency set;

[0037] The gray-level co-occurrence matrix is ​​normalized to obtain the co-occurrence probability matrix.

[0038] In this preferred solution, the gray level co-occurrence matrix can capture the distribution law of the gray value of pixels in the image and their spatial interdependence by counting the gray level co-occurrence frequency between each pixel and the neighboring pixels at a specified distance and in a specified direction. Through the normalization of the gray level co-occurrence matrix, the co-occurrence probability matrix obtained converts the frequency of pixel gray level co-occurrence into a probability form, making the description of texture features more objective and accurate.

[0039] As a preferred solution, the image enhancement processing is specifically as follows:

[0040] Establishing a cumulative distribution function according to the number of pixels corresponding to each gray level in the second image set;

[0041] Normalizing the cumulative distribution function to obtain a mapping function;

[0042] Based on the mapping function, the grayscale of each image in the second image set is mapped into a target grayscale range to obtain the second image set after image enhancement.

[0043] This preferred solution can effectively expand the contrast of the image by mapping the grayscale to a wider or specific target range; contrast enhancement can make the details in the image clearer, thereby improving the visual effect and readability of the image.

[0044] As a preferred solution, the detail enhancement model is established based on the generator and the discriminator, specifically:

[0045] Enhance the details of the image sample by the generator to obtain a generated image;

[0046] Calculating the possibility that the generated image is a real image through the discriminator to obtain a true probability value of the image;

[0047] Based on the true probability value of the image, the generator is optimized by back-propagating the loss value of the generator along the network to each layer of the generator to obtain the optimized generator;

[0048] The detail enhancement model is established according to the optimized generator and the discriminator.

[0049] In this preferred solution, the discriminator is responsible for evaluating the authenticity of the generated image, while the generator strives to generate images that can deceive the discriminator. This adversarial mechanism prompts the generator to continuously improve its generation ability to generate more realistic image samples, thereby ensuring the practicality of the detail enhancement model.

[0050] As a preferred solution, the first image set is reconstructed in detail by a detail enhancement model to obtain a second image set, specifically:

[0051] Extracting features of the first image set through the convolutional layer of the generator in the detail enhancement model to obtain shallow features;

[0052] By deepening the network layer of the detail enhancement model, the shallow features are combined and extracted to obtain deep features of the first image set;

[0053] Instructing the attention mechanism to dynamically adjust the weights of the shallow features and the deep features according to the preset important feature areas;

[0054] The shallow features and the deep features of the first image set after weight adjustment are fused to obtain the second image set.

[0055] This preferred solution combines the shallow features and deep features of the image. Since the shallow features contain more detail information and edge information, while the deep features are rich in higher-level semantic information, through multi-scale feature fusion, feature information at different levels can be fully utilized to generate a second image set containing richer information.

[0056] The present invention also provides a fatigue crack defect determination system for a floating wind power platform, comprising an image module, a reconstruction module, a feature module, an identification module and a judgment module;

[0057] The image module is used to collect images of the floating wind power platform to obtain a first image set;

[0058] The reconstruction module is used to reconstruct the details of the first image set through a detail enhancement model to obtain a second image set; wherein the detail enhancement model is established based on the generator and the discriminator, and the generation loss function of the generator is established based on the confrontation loss and content loss of the generator and the discriminator in the game process;

[0059] The feature module is used to mark a pixel abnormal point set in the second image set based on a dilation operation, and obtain texture features of the second image set by converting the second image set into a grayscale image set to obtain a comprehensive feature set;

[0060] The identification module is used to perform risk identification on the comprehensive feature set through a convolutional neural network to obtain an actual risk value set of the pixel abnormal point set;

[0061] The judgment module is used to determine the fatigue crack defect in the corresponding area of ​​the floating wind power platform according to the numerical value of the risk value in the actual risk value set, and obtain a defect identification result. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1It is a flow chart of a method for determining fatigue crack defects of a floating wind power platform provided in an embodiment of the present application;

[0063] Figure 2 It is a structural schematic diagram of a fatigue crack defect determination system for a floating wind power platform provided in an embodiment of the present application. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0065] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of this application, unless otherwise specified, "several" means two or more.

[0066] Embodiment 1:

[0067] See also Figure 1 The embodiment of the present application provides a method for determining fatigue crack defects of a floating wind power platform, including S1 to S5, and the specific implementation steps are as follows:

[0068] S1. Collect images of a floating wind power platform to obtain a first image set.

[0069] Step S1 of the embodiment of the present application is specifically as follows:

[0070] The drone is controlled to collect images of the floating wind power platform from different angles and distances according to a preset drone flight path to obtain an initial first image set; wherein the drone is preferably a drone with strong stability, load capacity, wind resistance and endurance, and suitable for offshore operations; and the drone flight path is determined according to the location, size and shape of the floating wind power platform.

[0071] A filter is used to remove noise from images in the initial first image set to obtain a first image set.

[0072] S2. Reconstruct the details of the first image set through a detail enhancement model to obtain a second image set; wherein the detail enhancement model is established based on the generator and the discriminator, and the generation loss function of the generator is established based on the adversarial loss and content loss of the generator and the discriminator in the game process.

[0073] Step S2 of the embodiment of the present application includes S2.1 to S2.2; specifically, it is:

[0074] S2.1, enhancing details of a real image of a preset floating wind power platform by a generator to obtain a generated image;

[0075] The discriminator calculates the possibility that the generated image is a real image and obtains the real probability value of the image; when the input to the discriminator is a real image, the discriminator should output a probability value close to 1, indicating that it believes it is real; when the input is an image generated by the generator, the discriminator should output a probability value close to 0, indicating that it believes it is forged; the discriminator is a convolutional neural network (CNN), the generator is a deconvolutional neural network (Deconvolutional Neural Network), and the discriminant loss function of the discriminator is a cross entropy loss function containing a gradient penalty term;

[0076] Based on the true probability value of the image, the generator is optimized by back-propagating the generator's loss value along the network to each layer of the generator; and during the training process, the generator is controlled to continuously adjust its own parameters to reduce the value of the generated loss function, thereby generating more realistic samples;

[0077] When the generated image is input into the discriminator, the real image corresponding to the generated image is also input into the discriminator, so as to calculate the loss value corresponding to the generated image; the loss value is back-propagated along the network to each layer of the discriminator, and the gradient of each parameter is calculated; the parameters of the discriminator are updated according to the gradient to minimize the discriminant loss function, thereby optimizing the discriminator;

[0078] Traverse the game process of the generator and the discriminator until the preset number of times is reached, or the generation loss function and the discriminant loss function no longer change significantly, and obtain the optimized generator and discriminator;

[0079] A detail enhancement model is established based on the optimized generator and discriminator.

[0080] Among them, the generation loss function of the generator is established as follows:

[0081] Control the discriminator to score the real images and the generated images generated by the generator, and establish the adversarial loss of the generator according to the scores obtained by the scoring; the adversarial loss is used to encourage the generated images to be as close as possible to the real images in distribution to deceive the discriminator;

[0082] Based on the difference between the generated image and the preset target image at the feature level, the content loss of the generator is established; wherein the content loss is used to ensure that the generated image is similar to the target image in content, and the target image is used to guide the generator to generate a reference image that is similar to the real image in visual features, semantic information or structure;

[0083] The discriminator extracts the features of the real image and the generated image to obtain the feature atlas of the generator;

[0084] According to the image differences in the feature atlas, the feature matching loss of the generator is established; wherein the feature matching loss is used to encourage the generated image to be similar to the real image in the intermediate layer features of the discriminator;

[0085] Based on the similarity evaluation results of the generated image and the target image on the preset high-level semantic features, the perceptual loss of the generator is established; wherein the perceptual loss is used to restore the high-frequency details of the image;

[0086] The generative loss function of the generator is established based on adversarial loss, content loss, feature matching loss and perceptual loss; among them, the weight set of the generative loss function is based on Bayesian optimization, which guides the future search direction according to past evaluation results, so as to find the optimal weight combination within a preset small number of evaluation times.

[0087] Among them, the adversarial loss is:

[0088]

[0089] The content loss is:

[0090]

[0091] The feature matching loss is:

[0092]

[0093] The perceptual loss is:

[0094]

[0095] The generated loss function is:

[0096]

[0097] in, is the image input to the generator, is the prior distribution, The mathematical symbol for expected value, is the generated image generated by the generator, is the probability that the discriminator judges that the generated image is a real image.

[0098] represents the expectation of x and z, is a real image;

[0099] Represents a real image In the discriminator Features on the layer, Represents the generated image In the discriminator Features on the layer, It is The dimension of the layer features, is the number of discriminator layers used to calculate feature matching loss;

[0100] Represents a real image In the pre-trained network The feature map on the layer, Represents the generated image In the pre-trained network The feature map on the layer, is the number of elements in the feature map.

[0101] and It is the weight of each loss item in the weight set, which is used to adjust the contribution of adversarial loss, content loss, feature matching loss and perceptual loss to the generation loss function.

[0102] In this embodiment S2.1, the discriminator is used to evaluate the authenticity of the generated image, and the generator is used to try to generate images that can deceive the discriminator. This adversarial mechanism prompts the generator to continuously improve its generation ability to generate more realistic image samples, thereby ensuring the practicality of the detail enhancement model;

[0103] In addition, adversarial loss encourages the generator to generate images that can deceive the discriminator, thereby making the generated images visually closer to real images. Content loss can ensure that the generated image is consistent with the target image at the content level, which helps to retain important information of the original image in image conversion or restoration tasks. Feature matching loss further guides the generator to learn more refined image features by comparing the differences between the real image and the generated image at the feature level, which helps to improve the detail richness and overall quality of the generated image. Perceptual loss is based on the preset high-level semantic feature similarity evaluation results, which helps the generator generate more perceptually natural and realistic images.

[0104] S2.2, extracting features of the first image set through the convolutional layer of the generator in the detail enhancement model to obtain shallow features;

[0105] By deepening the network layer of the detail enhancement model, shallow features are combined and extracted to obtain deep features of the first image set;

[0106] The attention mechanism is guided to dynamically adjust the weights of shallow features and deep features according to the preset important feature areas; the important feature areas correspond to the key elements required by the current task, namely: defect elements such as welding cracks, material fatigue, corrosion fatigue and structural cracks;

[0107] The shallow features and deep features of the first image set after weight adjustment are fused to obtain the second image set.

[0108] Among them, the attention mechanism is established as follows:

[0109] According to the current task requirements, through data analysis, the areas where defects such as welding cracks, material fatigue, corrosion fatigue and structural cracks are located in the input image are marked as key areas;

[0110] Establish a spatial attention module based on the key area, so that the spatial attention module focuses on the spatial position of the input image and enhances the features of the key area by generating a spatial weight distribution map;

[0111] A dynamic adjustment mechanism is introduced into the attention module, so that the first function is taken as the optimization target, and the weights of shallow features and deep features are dynamically adjusted according to the important feature areas to obtain the final attention mechanism.

[0112] Among them, the first function is:

[0113]

[0114] For parameters have:

[0115]

[0116] For parameters have:

[0117]

[0118] in, The goal of the optimization process is to find a loss function The smallest parameter or variable, It refers to the real and error-free feature data in shallow features and deep features. are the preset network parameters. is the weight distribution graph; is the fusion function The parameters of the fusion function The weight is given by the spatial weight distribution map Dynamic decision making; is the result of dynamically adjusting feature fusion. and They are features from the shallow and deep layers of the network respectively; is the input image, is a labeled graph; is the function of the attention mechanism, which is used to , Marking diagram and network parameters , output spatial weight distribution map .

[0119] This embodiment S2.2 combines the shallow features and deep features of the image. Since the shallow features contain more detail information and edge information, while the deep features are rich in higher-level semantic information, multi-scale feature fusion can make full use of feature information at different levels to generate a second image set containing richer information.

[0120] S3. Marking a pixel abnormal point set in the second image set based on a dilation operation, and acquiring texture features of the second image set by converting the second image set into a grayscale image set to obtain a comprehensive feature set.

[0121] Step S3 of the embodiment of the present application includes S3.1 to S3.3; specifically, it is:

[0122] S3.1, establishing a cumulative distribution function according to the number of pixels corresponding to each gray level in the second image set;

[0123] Normalize the cumulative distribution function to between 0 and 1 to obtain the mapping function;

[0124] Based on the mapping function, the grayscale of each image in the second image set is mapped to the target grayscale range to obtain the second image set after image enhancement; wherein, the target grayscale range should avoid the problem of overexposure (loss of information in the highlight area) or underexposure (loss of details in the dark area) of the image. During the mapping process, it is necessary to ensure that the grayscale value after mapping does not exceed the range of 0-255; and while enhancing the contrast, other important details in the image should be maintained as much as possible to ensure that the target grayscale range can include and highlight defects such as welding cracks, material fatigue, corrosion fatigue and structural cracks;

[0125] The mapping function is:

[0126]

[0127] For parameters have:

[0128]

[0129] in, is the gray level (from 0 to L-1), Is grayscale The number of pixels, is the cumulative distribution function, is the total number of pixels in the second image set.

[0130] This embodiment S3.1 can effectively expand the contrast of the image by mapping the grayscale to a wider or specific target range; contrast enhancement can make the details in the image clearer, thereby improving the visual effect and readability of the image.

[0131] S3.2, determining a structural element according to the size and shape of the abnormal region in the second image set; if the abnormal region is elongated and long, a rectangular or elliptical structural element may be selected; if the abnormal region is large and irregular in shape, a larger square or circular structural element may be selected;

[0132] An extended image set is obtained by extending the boundary of the primary abnormal region in the second image set; wherein the primary abnormal region is obtained by performing preliminary abnormality detection on the second image set by a machine learning model;

[0133] The structure element is controlled to slide on the image in the extended image set. For each position, it is checked whether the area covered by the structure element contains abnormal points. If there are abnormal points in the area covered by the structure element, the pixel points at the regional position corresponding to the abnormal points are marked as abnormal to obtain a pixel abnormal point set.

[0134] In this embodiment S3.2, the structural element is determined according to the actual size and shape of the abnormal area, so that the customized structural element can better adapt to different abnormal areas and improve the accuracy of abnormality detection. The primary abnormal area is obtained by preliminary detection through the machine learning model. Although it can roughly locate the abnormal position, there may be false positives or false negatives. Therefore, by expanding the boundaries of these areas and performing further sliding detection on the expanded image set, more potential abnormal points can be captured, while reducing the abnormalities missed due to boundary restrictions.

[0135] S3.3, converting the second image set into grayscale images to obtain a grayscale image set;

[0136] Based on the grayscale image set, at the specified distance d and the specified direction θ, the grayscale level co-occurrence frequency between each pixel and the neighboring pixels is counted to obtain the co-occurrence frequency set;

[0137] A gray-level co-occurrence matrix is ​​established based on the co-occurrence frequency set; each element (i, j) in the gray-level co-occurrence matrix represents the number of occurrences of a pixel with gray-level value i and a pixel with gray-level value j adjacent to it at a specified distance d and a specified direction θ;

[0138] The gray-level co-occurrence matrix is ​​normalized by dividing the value of each element in the gray-level co-occurrence matrix by the sum of all elements in the matrix to obtain the co-occurrence probability matrix;

[0139] A comprehensive feature set about pixel anomaly point set and texture features is obtained from the co-occurrence probability matrix.

[0140] In this embodiment S3.3, the color information of the image is converted into grayscale, which helps to reduce the influence of color noise on abnormality detection. By establishing a co-occurrence probability matrix of the grayscale level co-occurrence frequency between pixels in the grayscale image set, the texture features in the image can be captured, which helps to distinguish normal areas from abnormal areas;

[0141] Moreover, the gray-level co-occurrence matrix can capture the distribution law of pixel gray values ​​in the image and their spatial interdependence by counting the gray-level co-occurrence frequency between each pixel and the neighboring pixels at a specified distance and in a specified direction. Through the normalization of the gray-level co-occurrence matrix, the co-occurrence probability matrix obtained converts the frequency of pixel gray-level co-occurrence into a probability form, making the description of texture features more objective and accurate.

[0142] S4. Perform risk identification on the comprehensive feature set through a convolutional neural network to obtain the actual risk value set of the pixel abnormal point set.

[0143] Step S4 of the embodiment of the present application is specifically as follows:

[0144] The risk identification of the comprehensive feature set is performed through the preset convolutional neural network to obtain the actual risk value set of the pixel abnormal point set;

[0145] Among them, the convolutional neural network is trained with the purpose of accurately distinguishing different types of defects using a floating wind power platform image set containing defective samples and normal samples.

[0146] S5. Determine the fatigue crack defect in the corresponding area of ​​the floating wind power platform according to the numerical value of the risk value in the actual risk value concentration, and obtain a defect identification result.

[0147] Step S5 of the embodiment of the present application includes S5.1 to S5.3; specifically, it is:

[0148] S5.1. Determine the fatigue crack defects in the corresponding area of ​​the floating wind power platform according to the numerical value of the risk value in the actual risk value concentration, and obtain defect identification results of defects such as welding cracks, material fatigue, corrosion fatigue and structural cracks.

[0149] S5.2. Use a finite element analysis model to simulate stress and fatigue damage of various parts of the floating wind power platform under different working conditions, thereby estimating the risk value of each area and obtaining a first risk value set;

[0150] Based on the operation records, fault history, and historical monitoring data of vibration, stress, and temperature of the floating wind power platform, the risk of each area is evaluated through a machine learning algorithm to obtain a second risk value set;

[0151] A risk value assessment set is established according to the first risk value set and the second risk value set; wherein the risk levels in the risk value assessment set include low risk, medium-low risk, medium risk, medium-high risk and high risk.

[0152] S5.3. According to the risk value assessment set, determine the risk level of the area corresponding to each defect in the defect identification result, and obtain low-risk area, medium-low risk area, medium-risk area, medium-high risk area and high risk area;

[0153] Use ultrasonic testing technology to perform more accurate defect detection on high-risk areas or other target areas again to obtain detailed record data of fatigue crack defects; the detailed record data includes information such as the location, size, depth and shape of the defects;

[0154] According to the current mission requirements, high-risk areas or other target areas are repaired based on detailed recorded data to ensure the safety of the floating wind power platform.

[0155] Overall, this embodiment has the following beneficial effects:

[0156] In the present invention, the generative loss function of the generator combines the adversarial loss and the content loss, which can ensure that the second image set can retain the important information of the first image set while maintaining authenticity; and in the game process between the generator and the discriminator, the adversarial loss can encourage the generator to generate images that can deceive the discriminator, that is, the details are as close to the real image as possible. This mechanism prompts the generator to continuously optimize its ability to generate details, making fine cracks that may be difficult to detect due to factors such as noise, blur or insufficient light more obvious. Strengthening the boundaries and details of the image through dilation operations helps to highlight subtle changes in the image. By marking pixel abnormal point sets, areas where cracks may exist can be quickly located, providing a precise target for subsequent risk assessment. Converting images to grayscale image sets can simplify the image processing process, thereby removing the interference of color information on crack detection, making texture features more prominent, and using the extracted texture features as part of the comprehensive feature set can provide a rich source of information for convolutional neural network recognition;

[0157] In addition, the present invention improves image quality through detail reconstruction, and the extraction of grayscale images and texture features improves the accuracy and reliability of subsequent recognition; by using a convolutional neural network to identify and analyze the comprehensive feature set, the actual risk value set of each pixel abnormal point set can be obtained, thereby achieving accurate recognition and risk assessment of defects such as cracks, thereby determining fatigue crack defects in corresponding areas of the floating wind power platform, thereby solving the problem of difficulty in accurately identifying fatigue crack defects of the floating wind power platform.

[0158] Embodiment 2:

[0159] See also Figure 2 , an embodiment of the present application provides a fatigue crack defect determination system for a floating wind power platform, comprising an image module 10, a reconstruction module 20, a feature module 30, an identification module 40 and a judgment module 50;

[0160] The image module 10 is used to collect images of the floating wind power platform to obtain a first image set;

[0161] A reconstruction module 20 is used to reconstruct the details of the first image set by using a detail enhancement model to obtain a second image set; wherein the detail enhancement model is established based on the generator and the discriminator, and the generation loss function of the generator is established based on the confrontation loss and content loss of the generator and the discriminator in the game process;

[0162] A feature module 30 is used to mark a pixel abnormal point set in the second image set based on a dilation operation, and obtain texture features of the second image set by converting the second image set into a grayscale image set to obtain a comprehensive feature set;

[0163] The identification module 40 is used to perform risk identification on the comprehensive feature set through a convolutional neural network to obtain an actual risk value set of the pixel abnormal point set;

[0164] The judgment module 50 is used to determine the fatigue crack defect in the corresponding area of ​​the floating wind power platform according to the numerical value of the risk value in the actual risk value set, and obtain a defect identification result.

[0165] In one embodiment, the image module 10 is specifically:

[0166] The drone is controlled to collect images of the floating wind power platform from different angles and distances according to a preset drone flight path to obtain an initial first image set; wherein the drone is preferably a drone with strong stability, load capacity, wind resistance and endurance, and suitable for offshore operations; and the drone flight path is determined according to the location, size and shape of the floating wind power platform.

[0167] A filter is used to remove noise from images in the initial first image set to obtain a first image set.

[0168] In one embodiment, the reconstruction module 20 includes a model unit and an extraction unit;

[0169] The model unit is used to enhance the details of the real image of the preset floating wind power platform through the generator to obtain a generated image;

[0170] The model unit is also used to calculate the possibility that the generated image is a real image through the discriminator to obtain the real probability value of the image; wherein, when the input to the discriminator is a real image, the discriminator should output a probability value close to 1, indicating that it believes it is real; when the input is an image generated by the generator, the discriminator should output a probability value close to 0, indicating that it believes it is forged; wherein the discriminator is a convolutional neural network (CNN), the generator is a deconvolutional neural network (Deconvolutional Neural Network), and the discriminant loss function of the discriminator is a cross entropy loss function containing a gradient penalty term;

[0171] The model unit is also used to optimize the generator based on the true probability value of the image by back-propagating the generator's loss value along the network to each layer of the generator; and during the training process, the generator is controlled to continuously adjust its own parameters to reduce the value of the generated loss function, thereby generating more realistic samples;

[0172] The model unit is also used to input the real image corresponding to the generated image into the discriminator when the generated image is input into the discriminator, so as to calculate the loss value corresponding to the generated image; the loss value is back-propagated to each layer of the discriminator along the network, and the gradient of each parameter is calculated; the parameters of the discriminator are updated according to the gradient to minimize the discriminant loss function, so as to optimize the discriminator;

[0173] The model unit is also used to traverse the game process of the generator and the discriminator until a preset number of times is reached, or the generation loss function and the discriminant loss function no longer change significantly, thereby obtaining the optimized generator and discriminator;

[0174] The model unit is also used to build a detail enhancement model based on the optimized generator and discriminator.

[0175] Among them, the generation loss function of the generator is established as follows:

[0176] Control the discriminator to score the real images and the generated images generated by the generator, and establish the adversarial loss of the generator according to the scores obtained by the scoring; the adversarial loss is used to encourage the generated images to be as close as possible to the real images in distribution to deceive the discriminator;

[0177] Based on the difference between the generated image and the preset target image at the feature level, the content loss of the generator is established; wherein the content loss is used to ensure that the generated image is similar to the target image in content, and the target image is used to guide the generator to generate a reference image that is similar to the real image in visual features, semantic information or structure;

[0178] The discriminator extracts the features of the real image and the generated image to obtain the feature atlas of the generator;

[0179] According to the image differences in the feature atlas, the feature matching loss of the generator is established; wherein the feature matching loss is used to encourage the generated image to be similar to the real image in the intermediate layer features of the discriminator;

[0180] Based on the similarity evaluation results of the generated image and the target image on the preset high-level semantic features, the perceptual loss of the generator is established; wherein the perceptual loss is used to restore the high-frequency details of the image;

[0181] The generative loss function of the generator is established based on adversarial loss, content loss, feature matching loss and perceptual loss; among them, the weight set of the generative loss function is based on Bayesian optimization, which guides the future search direction according to past evaluation results, so as to find the optimal weight combination within a preset small number of evaluation times.

[0182] Among them, the adversarial loss is:

[0183]

[0184] The content loss is:

[0185]

[0186] The feature matching loss is:

[0187]

[0188] The perceptual loss is:

[0189]

[0190] The generated loss function is:

[0191]

[0192] in, is the image input to the generator, is the prior distribution, The mathematical symbol for expected value, is the generated image generated by the generator, is the probability that the discriminator judges that the generated image is a real image.

[0193] represents the expectation of x and z, is a real image;

[0194] Represents a real image In the discriminator Features on the layer, Represents the generated image In the discriminator Features on the layer, It is The dimension of the layer features, is the number of discriminator layers used to calculate feature matching loss;

[0195] Represents a real image In the pre-trained network The feature map on the layer, Represents the generated image In the pre-trained network The feature map on the layer, is the number of elements in the feature map.

[0196] and It is the weight of each loss item in the weight set, which is used to adjust the contribution of adversarial loss, content loss, feature matching loss and perceptual loss to the generation loss function.

[0197] In the model unit of this embodiment, the discriminator is used to evaluate the authenticity of the generated image, and the generator is used to try to generate images that can deceive the discriminator. This adversarial mechanism prompts the generator to continuously improve its generation ability to generate more realistic image samples, thereby ensuring the practicality of the detail enhancement model;

[0198] In addition, adversarial loss encourages the generator to generate images that can deceive the discriminator, thereby making the generated images visually closer to real images. Content loss can ensure that the generated image is consistent with the target image at the content level, which helps to retain important information of the original image in image conversion or restoration tasks. Feature matching loss further guides the generator to learn more refined image features by comparing the differences between the real image and the generated image at the feature level, which helps to improve the detail richness and overall quality of the generated image. Perceptual loss is based on the preset high-level semantic feature similarity evaluation results, which helps the generator generate more perceptually natural and realistic images.

[0199] An extraction unit, configured to extract features of the first image set through a convolutional layer of a generator in a detail enhancement model to obtain shallow features;

[0200] The extraction unit is further used to combine and extract shallow features by deepening the network layer of the detail enhancement model to obtain deep features of the first image set;

[0201] The extraction unit is also used to guide the attention mechanism to dynamically adjust the weights of shallow features and deep features according to the preset important feature areas; the important feature areas correspond to the key elements required by the current task, namely: defect elements such as welding cracks, material fatigue, corrosion fatigue and structural cracks;

[0202] The extraction unit is also used to fuse the shallow features and deep features of the first image set after weight adjustment to obtain the second image set.

[0203] Among them, the attention mechanism is established as follows:

[0204] According to the current task requirements, through data analysis, the areas where defects such as welding cracks, material fatigue, corrosion fatigue and structural cracks are located in the input image are marked as key areas;

[0205] Establish a spatial attention module based on the key area, so that the spatial attention module focuses on the spatial position of the input image and enhances the features of the key area by generating a spatial weight distribution map;

[0206] A dynamic adjustment mechanism is introduced into the attention module, so that the first function is taken as the optimization target, and the weights of shallow features and deep features are dynamically adjusted according to the important feature areas to obtain the final attention mechanism.

[0207] Among them, the first function is:

[0208]

[0209] For parameters have:

[0210]

[0211] For parameters have:

[0212]

[0213] in, The goal of the optimization process is to find a loss function The smallest parameter or variable, It refers to the real and error-free feature data in shallow features and deep features. are the preset network parameters. is the weight distribution graph; is the fusion function The parameters of the fusion function The weight is given by the spatial weight distribution map Dynamic decision making; is the result of dynamically adjusting feature fusion. and They are features from the shallow and deep layers of the network respectively; is the input image, is a labeled graph; is the function of the attention mechanism, which is used to , Marking diagram and network parameters , output spatial weight distribution map .

[0214] The extraction unit of this embodiment fuses the shallow features and deep features of the image. Since the shallow features contain more detail information and edge information, while the deep features are rich in higher-level semantic information, multi-scale feature fusion can fully utilize feature information at different levels to generate a second image set containing richer information.

[0215] In one embodiment, the feature module 30 includes an enhancement unit, an anomaly unit, and a grayscale unit;

[0216] The enhancement unit is used to establish a cumulative distribution function according to the number of pixels corresponding to each gray level in the second image set;

[0217] The enhancement unit is also used to normalize the cumulative distribution function to between 0 and 1 to obtain the mapping function;

[0218] The enhancement unit is further used to map the grayscale of each image in the second image set to a target grayscale range based on a mapping function to obtain a second image set after image enhancement; wherein the target grayscale range should avoid the problem of overexposure (loss of information in highlight areas) or underexposure (loss of details in dark areas) of the image. During the mapping process, it is necessary to ensure that the grayscale value after mapping does not exceed the range of 0-255; and while enhancing the contrast, other important details in the image should be maintained as much as possible to ensure that the target grayscale range can include and highlight defects such as welding cracks, material fatigue, corrosion fatigue and structural cracks;

[0219] The mapping function is:

[0220]

[0221] For parameters have:

[0222]

[0223] in, is the gray level (from 0 to L-1), Is grayscale The number of pixels, is the cumulative distribution function, is the total number of pixels in the second image set.

[0224] The enhancement unit of this embodiment can effectively expand the contrast of the image by mapping the grayscale to a wider or specific target range; contrast enhancement can make the details in the image clearer, thereby improving the visual effect and readability of the image.

[0225] an abnormal unit, used to determine a structural element according to the size and shape of the abnormal region of the image in the second image set; wherein, if the abnormal region is elongated and long, a rectangular or elliptical structural element may be selected; if the abnormal region is large and irregular in shape, a larger square or circular structural element may be selected;

[0226] An anomaly unit, used to obtain an extended image set by expanding the boundary of a primary anomaly region in the second image set; wherein the primary anomaly region is obtained by performing preliminary anomaly detection on the second image set by a machine learning model;

[0227] The abnormal unit is used to control the sliding of the structural element on the image of the extended image set. For each position, it is checked whether the area covered by the structural element contains abnormal points. If there are abnormal points in the area covered by the structural element, the pixel points at the regional position corresponding to the abnormal points are marked as abnormal, and a pixel abnormal point set is obtained.

[0228] The anomaly unit of this embodiment determines the structural element according to the actual size and shape of the abnormal area, which can make the customized structural element better adapt to different abnormal areas and improve the accuracy of anomaly detection. The primary abnormal area is obtained by preliminary detection through the machine learning model. Although it can roughly locate the abnormal position, there may be false positives or false negatives. Therefore, by expanding the boundaries of these areas and performing further sliding detection on the expanded image set, more potential abnormal points can be captured, while reducing the anomalies missed due to boundary restrictions.

[0229] A grayscale unit, used for converting the second image set into grayscale images to obtain a grayscale image set;

[0230] The grayscale unit is also used to count the grayscale level co-occurrence frequencies between each pixel and its neighboring pixels at a specified distance d and a specified direction θ based on the grayscale image set to obtain a co-occurrence frequency set;

[0231] The gray level unit is also used to establish a gray level co-occurrence matrix based on the co-occurrence frequency set; each element (i, j) in the gray level co-occurrence matrix represents the number of occurrences of a pixel with a gray level value i and a pixel with a gray level value j adjacent to it at a specified distance d and a specified direction θ;

[0232] The gray level unit is also used to normalize the gray level co-occurrence matrix by dividing the value of each element in the gray level co-occurrence matrix by the sum of all elements in the matrix to obtain a co-occurrence probability matrix;

[0233] The grayscale unit is also used to obtain a comprehensive feature set about pixel abnormal point set and texture features from the co-occurrence probability matrix.

[0234] The grayscale unit of this embodiment converts the color information of the image into grayscale, which helps to reduce the impact of color noise on abnormality detection. By establishing a co-occurrence probability matrix of the grayscale level co-occurrence frequency between pixels in the grayscale image set, the texture features in the image can be captured, which helps to distinguish normal areas from abnormal areas;

[0235] Moreover, the gray-level co-occurrence matrix can capture the distribution law of pixel gray values ​​in the image and their spatial interdependence by counting the gray-level co-occurrence frequency between each pixel and the neighboring pixels at a specified distance and in a specified direction. Through the normalization of the gray-level co-occurrence matrix, the co-occurrence probability matrix obtained converts the frequency of pixel gray-level co-occurrence into a probability form, making the description of texture features more objective and accurate.

[0236] In one embodiment, the identification module 40 is specifically:

[0237] The risk identification of the comprehensive feature set is performed through the preset convolutional neural network to obtain the actual risk value set of the pixel abnormal point set;

[0238] Among them, the convolutional neural network is trained with the purpose of accurately distinguishing different types of defects using a floating wind power platform image set containing defective samples and normal samples.

[0239] In one embodiment, the determination module 50 includes an identification unit, a risk unit, and a repair unit;

[0240] Among them, the identification unit is used to determine the fatigue crack defects in the corresponding area of ​​the floating wind power platform according to the numerical value of the risk value in the actual risk value concentration, and obtain defect identification results of defects such as welding cracks, material fatigue, corrosion fatigue and structural cracks.

[0241] A risk unit, used for simulating stress and fatigue damage of various parts of the floating wind power platform under different working conditions using a finite element analysis model, thereby estimating the risk value of each area and obtaining a first risk value set;

[0242] The risk unit is also used to evaluate the risk of each area through a machine learning algorithm based on the operation record, fault history, and historical monitoring data of vibration, stress, and temperature of the floating wind power platform to obtain a second risk value set;

[0243] The risk unit is also used to establish a risk value assessment set based on the first risk value set and the second risk value set; wherein the risk levels in the risk value assessment set include low risk, medium-low risk, medium risk, medium-high risk and high risk.

[0244] The repair unit is used to determine the risk level of the area corresponding to each defect in the defect identification result according to the risk value assessment set, and obtain low-risk area, medium-low risk area, medium-risk area, medium-high risk area and high risk area;

[0245] The repair unit is also used to use ultrasonic testing technology to perform more accurate defect detection on high-risk areas or other target areas again to obtain detailed record data of fatigue crack defects; wherein the detailed record data includes information such as the location, size, depth and shape of the defect;

[0246] The repair unit is also used to repair high-risk areas or other target areas according to the current mission requirements and based on detailed recorded data to ensure the safety of the floating wind power platform.

[0247] Overall, this embodiment has the following beneficial effects:

[0248] In the present invention, the generative loss function of the generator combines the adversarial loss and the content loss, which can ensure that the second image set can retain the important information of the first image set while maintaining authenticity; and in the game process between the generator and the discriminator, the adversarial loss can encourage the generator to generate images that can deceive the discriminator, that is, the details are as close to the real image as possible. This mechanism prompts the generator to continuously optimize its ability to generate details, making fine cracks that may be difficult to detect due to factors such as noise, blur or insufficient light more obvious. Strengthening the boundaries and details of the image through dilation operations helps to highlight subtle changes in the image. By marking pixel abnormal point sets, areas where cracks may exist can be quickly located, providing a precise target for subsequent risk assessment. Converting images to grayscale image sets can simplify the image processing process, thereby removing the interference of color information on crack detection, making texture features more prominent, and using the extracted texture features as part of the comprehensive feature set can provide a rich source of information for convolutional neural network recognition;

[0249] In addition, the present invention improves image quality through detail reconstruction, and the extraction of grayscale images and texture features improves the accuracy and reliability of subsequent recognition; by using a convolutional neural network to identify and analyze the comprehensive feature set, the actual risk value set of each pixel abnormal point set can be obtained, thereby achieving accurate recognition and risk assessment of defects such as cracks, thereby determining fatigue crack defects in corresponding areas of the floating wind power platform, thereby solving the problem of difficulty in accurately identifying fatigue crack defects of the floating wind power platform.

[0250] Embodiment three:

[0251] An embodiment of the present application provides a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the fatigue crack defect determination method of a floating wind power platform;

[0252] Wherein, the fatigue crack defect determination method of a floating wind power platform, if implemented in the form of a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0253] The above are preferred embodiments of the present invention. It should be noted that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for determining fatigue crack defects of a floating wind power platform, characterized in that: include: Collecting images of the floating wind power platform to obtain a first image set; Reconstructing the details of the first image set by a detail enhancement model to obtain a second image set; wherein the detail enhancement model is established based on a generator and a discriminator, and the generation loss function of the generator is established based on the adversarial loss and content loss of the generator and the discriminator in the game process; Marking a pixel abnormal point set in the second image set based on a dilation operation, and acquiring texture features of the second image set by converting the second image set into a grayscale image set to obtain a comprehensive feature set; Performing risk identification on the comprehensive feature set through a convolutional neural network to obtain an actual risk value set of the pixel abnormal point set; Determine fatigue crack defects in a corresponding area of ​​the floating wind power platform according to the numerical value of the risk value in the actual risk value set, and obtain a defect identification result; The method of marking the pixel abnormal point set in the second image set based on the dilation operation, and obtaining the texture features of the second image set by converting the second image set into a grayscale image set to obtain a comprehensive feature set is as follows: on the basis of image enhancement processing on the second image set, the image abnormal points are identified by controlling the structure element to slide on the second image set to obtain the pixel abnormal point set; the second image set is converted into a grayscale image to obtain a grayscale image set; a co-occurrence probability matrix is ​​established according to the grayscale level co-occurrence frequency between pixels in the grayscale image set; and a comprehensive feature set about the pixel abnormal point set and texture features is obtained from the co-occurrence probability matrix; The image abnormal points are identified by controlling the structural element to slide on the second image set, and the pixel abnormal point set is obtained, specifically: determining the structural element according to the size and shape of the abnormal image region in the second image set; An extended image set is obtained by extending the boundary of the primary abnormal region in the second image set; wherein the primary abnormal region is obtained by performing preliminary abnormality detection on the second image set by a machine learning model; Controlling the structural element to slide on the image of the extended image set, if there is an abnormal point in the area covered by the structural element, marking the pixel point at the regional position corresponding to the abnormal point as abnormal, to obtain the pixel abnormal point set; The first image set is reconstructed in detail by using a detail enhancement model to obtain a second image set, specifically: Extracting features of the first image set through the convolutional layer of the generator in the detail enhancement model to obtain shallow features; By deepening the network layer of the detail enhancement model, the shallow features are combined and extracted to obtain deep features of the first image set; Instructing the attention mechanism to dynamically adjust the weights of the shallow features and the deep features according to the preset important feature areas; The shallow features and the deep features of the first image set after weight adjustment are fused to obtain the second image set.

2. A method for determining fatigue crack defects of a floating wind power platform according to claim 1, characterized in that: The generation loss function of the generator is established based on the adversarial loss and content loss of the generator and the discriminator during the game, specifically: Controlling the discriminator to score the real image and the generated image generated by the generator, and establishing the adversarial loss of the generator according to the scores obtained by the scoring; Establishing a content loss of the generator based on a difference value between the generated image and a preset target image at a feature level; Extracting features of the real image and the generated image through the discriminator to obtain a feature atlas of the generator; Establishing a feature matching loss of the generator based on the image differences in the feature atlas; Establishing a perceptual loss of the generator based on a similarity evaluation result between the generated image and the target image on preset high-level semantic features; The generation loss function of the generator is established according to the adversarial loss, the content loss, the feature matching loss and the perceptual loss.

3. A method for determining fatigue crack defects of a floating wind power platform according to claim 2, characterized in that: The adversarial loss is specifically: ； in, is the image input to the generator, is the prior distribution, The mathematical symbol for expected value, is the generated image generated by the generator, is the probability that the discriminator judges that the generated image is the real image.

4. A method for determining fatigue crack defects of a floating wind power platform according to claim 1, characterized in that: According to the gray level co-occurrence frequency between pixels in the gray image set, a co-occurrence probability matrix is ​​established, which is specifically: Based on the grayscale image set, the grayscale level co-occurrence frequency between each pixel and neighboring pixels at a specified distance and in a specified direction is counted to obtain a co-occurrence frequency set; Establishing a gray-level co-occurrence matrix according to the co-occurrence frequency set; The gray-level co-occurrence matrix is ​​normalized to obtain the co-occurrence probability matrix.

5. The method for determining fatigue crack defects of a floating wind power platform according to claim 1, characterized in that: The image enhancement processing is specifically: Establishing a cumulative distribution function according to the number of pixels corresponding to each gray level in the second image set; Normalizing the cumulative distribution function to obtain a mapping function; Based on the mapping function, the grayscale of each image in the second image set is mapped into a target grayscale range to obtain the second image set after image enhancement.

6. A method for determining fatigue crack defects of a floating wind power platform according to claim 1, characterized in that: The detail enhancement model is established based on the generator and the discriminator, specifically: Enhance the details of the image sample by the generator to obtain a generated image; Calculating the possibility that the generated image is a real image through the discriminator to obtain a true probability value of the image; Based on the true probability value of the image, the generator is optimized by back-propagating the loss value of the generator along the network to each layer of the generator to obtain the optimized generator; The detail enhancement model is established according to the optimized generator and the discriminator.

7. A fatigue crack defect determination system for a floating wind power platform, characterized in that: It includes image module, reconstruction module, feature module, recognition module and judgment module; The image module is used to collect images of the floating wind power platform to obtain a first image set; The reconstruction module is used to reconstruct the details of the first image set through a detail enhancement model to obtain a second image set; wherein the detail enhancement model is established based on the generator and the discriminator, and the generation loss function of the generator is established based on the confrontation loss and content loss of the generator and the discriminator in the game process; The feature module is used to mark a pixel abnormal point set in the second image set based on a dilation operation, and obtain texture features of the second image set by converting the second image set into a grayscale image set to obtain a comprehensive feature set; The identification module is used to perform risk identification on the comprehensive feature set through a convolutional neural network to obtain an actual risk value set of the pixel abnormal point set; The judgment module is used to determine the fatigue crack defect of the corresponding area in the floating wind power platform according to the numerical value of the risk value in the actual risk value set, and obtain a defect identification result; The method of marking the pixel abnormal point set in the second image set based on the dilation operation, and obtaining the texture features of the second image set by converting the second image set into a grayscale image set to obtain a comprehensive feature set is as follows: on the basis of image enhancement processing on the second image set, the image abnormal points are identified by controlling the structure element to slide on the second image set to obtain the pixel abnormal point set; the second image set is converted into a grayscale image to obtain a grayscale image set; a co-occurrence probability matrix is ​​established according to the grayscale level co-occurrence frequency between pixels in the grayscale image set; and a comprehensive feature set about the pixel abnormal point set and texture features is obtained from the co-occurrence probability matrix; The image abnormal points are identified by controlling the structural element to slide on the second image set, and the pixel abnormal point set is obtained, specifically: determining the structural element according to the size and shape of the abnormal image region in the second image set; An extended image set is obtained by extending the boundary of the primary abnormal region in the second image set; wherein the primary abnormal region is obtained by performing preliminary abnormality detection on the second image set by a machine learning model; Controlling the structural element to slide on the image of the extended image set, if there is an abnormal point in the area covered by the structural element, marking the pixel point at the regional position corresponding to the abnormal point as abnormal, to obtain the pixel abnormal point set; The first image set is reconstructed in detail by using a detail enhancement model to obtain a second image set, specifically: Extracting features of the first image set through the convolutional layer of the generator in the detail enhancement model to obtain shallow features; By deepening the network layer of the detail enhancement model, the shallow features are combined and extracted to obtain deep features of the first image set; Instructing the attention mechanism to dynamically adjust the weights of the shallow features and the deep features according to the preset important feature areas; The shallow features and the deep features of the first image set after weight adjustment are fused to obtain the second image set.

Images