A method, device, equipment and medium for detecting defects in the inner cavity of a fan blade

Abstract

The present application discloses a method, device, equipment and medium for detecting defects in the inner cavity of a fan blade, which relates to the field of computer vision and includes: constructing a target data set based on the original defect image, the target defect image and the defect-free image, and dividing the target data set into a training set, a validation set and a test set; inputting the training set into a preset defect detection model, obtaining a feature map based on the backbone network of the model, and obtaining a first output result corresponding to the feature map based on the feature fusion layer of the model; obtaining a second output result corresponding to the first output result according to the noise injection layer of the model, obtaining a target detection result corresponding to the second output result based on the detection head of the model, and determining the trained defect detection model based on the target detection result; optimizing the defect detection model by using the validation set to obtain a target defect detection model, and performing defect detection on the target defect detection model by using the test set, so as to detect the defects in the inner cavity of the blade by using the target defect detection model that meets the expected standard.

Description

Technical Field

[0001] This application relates to the field of computer vision, and particularly to a method, device, equipment and medium for detecting defects in the inner cavity of a fan blade. Background Art

[0002] The existing defect detection technologies for the inner cavity of wind turbine blades have the following disadvantages: 1. Problem of data scarcity: Defects in the inner cavity of fan blades, such as cracks and delaminations, have high actual acquisition costs and small sample sizes, resulting in limited improvement in algorithm performance. 2. Limitations of traditional data augmentation methods in generating images: Traditional methods, such as rotation, flipping, and cropping, only perform affine transformations on the original image, and the generated images are highly correlated with the original data, unable to generate new perspectives or complex texture defects, making it difficult for the model to learn the non-linear distribution characteristics in the real scene. 3. Insufficient training strategies: Traditional hybrid training methods do not distinguish the differences between synthetic data and real data, resulting in difficulty for the model to balance the feature distributions of the two types of data. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method, device, equipment and medium for detecting defects in the inner cavity of a fan blade, which can solve the problem of insufficient generalization ability of the detection model caused by the scarcity of defect samples in the inner cavity of the blade, and improve the robustness of the model to synthetic data noise and the adaptability to the real scene. The specific solutions are as follows:

[0004] In the first aspect, this application provides a method for detecting defects in the inner cavity of a fan blade, including:

[0005] Using the original defect image and obtaining a new target defect image according to a preset generative adversarial network, constructing a target data set based on the original defect image, the target defect image and the defect-free image, and dividing the target data set into a training set, a validation set and a test set according to a preset ratio;

[0006] Inputting the training set into a preset defect detection model, obtaining a feature map based on the backbone network of the preset defect detection model, and obtaining a first output result corresponding to the feature map based on the feature fusion layer of the preset defect detection model; the preset defect detection model includes a backbone network, a feature fusion layer, a noise injection layer and a detection head;

[0007] Obtaining a second output result with noise added corresponding to the first output result according to the noise injection layer of the preset defect detection model, and obtaining a target detection result corresponding to the second output result based on the detection head of the preset defect detection model, so as to determine a trained defect detection model based on the target detection result;

[0008] Optimize the trained defect detection model using the validation set to obtain a target defect detection model, perform defect detection on the images in the test set based on the target defect detection model, and determine whether the target defect detection model meets the expected standard, so as to use the target defect detection model that meets the expected standard to detect the defects in the inner cavity of the blade.

[0009] Optionally, the obtaining of the new target defect image using the original defect image and according to a preset generative adversarial network includes:

[0010] Obtain the preset defect parameters input by the user side; the preset defect parameters include size parameters, direction parameters, and density parameters;

[0011] Input the original defect image and the preset defect parameters into the preset generative adversarial network to obtain a new target defect image.

[0012] Optionally, the obtaining of the feature map based on the backbone network of the preset defect detection model includes:

[0013] Use the backbone network of the preset defect detection model to perform feature extraction operations on the images in the training set to obtain low-level semantic features and high-level semantic features.

[0014] Optionally, the obtaining of the first output result corresponding to the feature map based on the feature fusion layer of the preset defect detection model includes:

[0015] Receive the low-level semantic feature and the high-level semantic feature output by the backbone network through the feature fusion layer of the preset defect detection model;

[0016] Use the feature fusion layer to fuse the low-level semantic feature and the high-level semantic feature to obtain a composite feature map.

[0017] Optionally, the obtaining of the second output result with added noise corresponding to the first output result according to the noise injection layer of the preset defect detection model includes:

[0018] Receive the composite feature map output by the feature fusion layer through the noise injection layer of the preset defect detection model;

[0019] Use the noise injection layer to inject preset noise into the composite feature map and adjust the noise intensity of the preset noise to obtain a noisy feature map; the preset noise includes Gaussian noise and / or salt-and-pepper noise.

[0020] Optionally, the obtaining of the target detection result corresponding to the second output result based on the detection head of the preset defect detection model includes:

[0021] Receive, by a detection head of the preset defect detection model, the noisy feature map output by the noise injection layer;

[0022] Use the detection head to detect the noisy feature map to obtain a target detection result; the target detection result includes a defect position, a defect category, and a confidence level;

[0023] Correspondingly, determining the trained defect detection model based on the target detection result includes:

[0024] Compare the target detection result with the target defect position, the target defect category, and the target confidence level in the original defect image to determine an error result, and use the error result to adjust the parameters of the preset defect detection model to determine the trained defect detection model.

[0025] Optionally, the process of training the preset defect detection model based on the training set is a process of training based on a three-stage model training method; wherein, the training process corresponding to the three-stage model training method includes:

[0026] Extract, from the training set, images with defect sizes within a first preset size range and clear backgrounds as simple samples, and train the preset defect detection model based on the simple samples to perform the first-stage model training;

[0027] Extract, from the training set, images with defect sizes within a second preset size range and with defect blur and / or defect occlusion as difficult samples, and train the preset defect detection model based on the difficult samples to perform the second-stage model training;

[0028] Extract the original defect images from the training set as real samples, and train the preset defect detection model based on the real samples to perform the third-stage model training.

[0029] In a second aspect, the present application provides a device for detecting defects in the inner cavity of a fan blade, including:

[0030] A data set acquisition module, configured to use the original defect images and obtain new target defect images according to a preset generative adversarial network, construct a target data set based on the original defect images, the target defect images, and defect-free images, and divide the target data set into a training set, a validation set, and a test set according to a preset ratio;

[0031] The first result acquisition module is configured to input the training set into a preset defect detection model, obtain a feature map based on the backbone network of the preset defect detection model, and obtain a first output result corresponding to the feature map based on the feature fusion layer of the preset defect detection model; the preset defect detection model includes a backbone network, a feature fusion layer, a noise injection layer, and a detection head.

[0032] The second result acquisition module is configured to obtain a second output result with added noise corresponding to the first output result according to the noise injection layer of the preset defect detection model, and obtain a target detection result corresponding to the second output result based on the detection head of the preset defect detection model, so as to determine a trained defect detection model based on the target detection result.

[0033] The defect detection module is configured to optimize the trained defect detection model by using the validation set to obtain a target defect detection model, perform defect detection on the images in the test set based on the target defect detection model, and determine whether the target defect detection model meets the expected standard, so as to use the target defect detection model that meets the expected standard to detect the defects in the blade inner cavity.

[0034] In a third aspect, the present application provides an electronic device, including:

[0035] A memory for storing a computer program;

[0036] A processor for executing the computer program to implement the foregoing method for detecting defects in the inner cavity of a fan blade.

[0037] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the foregoing method for detecting defects in the inner cavity of a fan blade.

[0038] In this application, a new target defect image is obtained by using the original defect image and based on a preset generative adversarial network. A target data set is constituted based on the original defect image, the target defect image, and the defect-free image, and the target data set is divided into a training set, a validation set, and a test set according to a preset ratio; the training set is input into a preset defect detection model, a feature map is obtained based on the backbone network of the preset defect detection model, and a first output result corresponding to the feature map is obtained based on the feature fusion layer of the preset defect detection model; the preset defect detection model includes a backbone network, a feature fusion layer, a noise injection layer, and a detection head; a second output result after adding noise corresponding to the first output result is obtained according to the noise injection layer of the preset defect detection model, and a target detection result corresponding to the second output result is obtained based on the detection head of the preset defect detection model, so as to determine a trained defect detection model based on the target detection result; the trained defect detection model is optimized by using the validation set to obtain a target defect detection model, the images in the test set are defect-detected based on the target defect detection model, and it is judged whether the target defect detection model meets the expected standard, so as to use the target defect detection model that meets the expected standard to detect the defects in the inner cavity of the blade. As can be seen from the above, this application creates a brand-new target defect image that is similar to the original data distribution but has different perspectives, textures, etc. based on a preset generative adversarial network. At the same time, the original defect image, the target defect image, and the defect-free image are mixed and divided into a training set, a validation set, and a test set, greatly expanding the effective sample size, alleviating the problem of data scarcity, and providing richer learning materials for model training; compared with traditional affine transformation, the preset generative adversarial network can generate defect features with non-linear transformation, such as simulating defects with different lighting changes and material wear degrees. In this way, the potential distribution of real defect images can be learned by using the preset generative adversarial network to generate synthetic samples with brand-new perspectives and complex textures, significantly increasing data diversity and enabling the model to learn more comprehensive defect representations. At the same time, a noise injection layer is added to the preset defect detection model to randomly inject Gaussian noise, salt-and-pepper noise, etc. into the feature map output by the backbone network to simulate interference factors such as sensor noise and uneven lighting in the real detection scenario, further improving the robustness of the model to complex texture defects; the feature fusion layer of the preset defect detection model is used to encode the real data features and the synthetic data features respectively, and the weights of the two types of data are automatically learned through the attention mechanism to suppress the "pseudo-features" in the synthetic data and enhance the key features of the real data, enabling the model to effectively balance the feature distributions of the synthetic data and the real data, using the diversity of the synthetic data to improve the generalization ability and avoiding being misled by its limitations, and finally achieving more reliable defect recognition in the real detection scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for description in the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0040] Figure 1 It is a flowchart of a method for detecting defects in the inner cavity of a wind turbine blade disclosed in the present application;

[0041] Figure 2 It is a schematic diagram of a specific method for detecting defects in the inner cavity of a wind turbine blade disclosed in the present application;

[0042] Figure 3 It is a schematic diagram of the structure of a device for detecting defects in the inner cavity of a wind turbine blade disclosed in the present application;

[0043] Figure 4 It is a schematic diagram of the structure of an electronic device disclosed in the present application. Detailed implementation manners

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of them. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0045] The existing technologies for detecting defects in the inner cavity of wind turbine blades have the following disadvantages: 1. The problem of data scarcity: The actual acquisition cost of defects in the inner cavity of wind turbine blades is high and the sample size is small, resulting in limited improvement in algorithm performance. 2. Limitations in generating images by traditional data augmentation methods: Traditional methods only perform affine transformation on the original image, and the generated images are highly correlated with the original data, and cannot generate new perspectives or complex texture defects, resulting in the model being difficult to learn the non-linear distribution characteristics in the real scene. 3. Insufficient training strategies: Traditional hybrid training methods do not distinguish the differences between synthetic data and real data, resulting in the model being difficult to balance the feature distributions of the two types of data. Therefore, the present application provides a method for detecting defects in the inner cavity of wind turbine blades, which can solve the problem of insufficient generalization ability of the detection model caused by the scarcity of defect samples in the inner cavity of the blades, and improve the robustness of the model to synthetic data noise and the adaptability to the real scene.

[0046] See Figure 1 As shown, the embodiments of the present application disclose a method for detecting defects in the inner cavity of a wind turbine blade, including:

[0047] Step S11: Use the original defect image and obtain a new target defect image according to a preset generative adversarial network. Based on the original defect image, the target defect image, and the defect-free image, construct a target data set, and divide the target data set into a training set, a validation set, and a test set according to a preset ratio.

[0048] In this embodiment, first, the original defect image can be used and a new target defect image can be generated according to a preset generative adversarial network, which specifically includes: first, preset defect parameters input by the user terminal can be obtained; wherein, the preset defect parameters include, but are not limited to, size parameters, direction parameters, and density parameters; then, the original defect image and the preset defect parameters are input into the preset generative adversarial network to obtain a new target defect image.

[0049] It can be understood that traditional data augmentation methods generate new images through simple geometric or pixel transformations. Although they can alleviate the problem of insufficient data, their limitations are significant. Traditional methods, such as rotation, flipping, and cropping, only perform affine transformations on the original image, and the generated images are highly correlated with the original data, unable to generate new perspectives or complex texture defects, such as delamination bubbles in the inner cavity of a fan blade, resulting in the model being difficult to learn the non-linear distribution characteristics in the real scene. In addition, in the detection of the inner cavity of a fan blade, due to the presence of complex backgrounds such as reflective surfaces and porous core materials, the signal-to-noise ratio of the image will decrease. Although traditional filtering algorithms can enhance the contrast of the image and make the details in the image more obvious, they will also amplify the noise, resulting in a decrease in image quality and further affecting the accuracy of defect detection. To address this problem, this embodiment can use a generative adversarial network (GAN, Generative Adversarial Networks) to generate defect samples in the inner cavity of the blade. Among them, GAN undergoes adversarial training between a generator and a discriminator. The generator can learn the characteristics and distribution rules of real defect images, thereby synthesizing images highly similar to real defects. Moreover, by inputting defect parameters, such as size, direction, density, etc., defect morphologies that conform to actual physical laws can be generated, increasing the diversity and authenticity of the generated images. For example, defect images such as cracks of different sizes and directions or delamination bubbles of different densities can be generated, enabling the model to learn richer defect characteristics.

[0050] After obtaining the original defect image and the newly generated target defect image, the original defect image, the target defect image, and the defect-free image can be integrated together to form a complete data set, that is, the target data set, providing a richer data resource for subsequent model training. Among them, the addition of the defect-free image is to enable the model to not only learn the characteristics of defects but also learn the image characteristics under normal conditions, so as to better distinguish between defective and defect-free situations.

[0051] Furthermore, for the newly constructed target dataset, it can be divided into a training set, a validation set, and a test set according to the ratio of 8:1:1. The training set is used to train the model so that it can learn the features and patterns in the dataset. The validation set is used to verify the performance of the model, adjust the model parameters, and prevent overfitting. The test set is used to finally evaluate the performance of the model and judge the effect of the model in actual applications.

[0052] Step S12: Input the training set into a preset defect detection model, obtain a feature map based on the backbone network of the preset defect detection model, and obtain a first output result corresponding to the feature map based on the feature fusion layer of the preset defect detection model; the preset defect detection model includes a backbone network, a feature fusion layer, a noise injection layer, and a detection head.

[0053] In this embodiment, first, the training set in step S11 can be input into a preset defect detection model, such as the YOLOv11 model. Then, the backbone network of the preset defect detection model can be used to perform feature extraction operations on the images in the training set to obtain low-level semantic features and high-level semantic features. Next, the low-level semantic features and high-level semantic features output by the backbone network can be received through the feature fusion layer of the preset defect detection model. Finally, the low-level semantic features and high-level semantic features can be fused by the feature fusion layer to obtain a composite feature map.

[0054] It can be understood that when the training set images are input into the model body, they first pass through the backbone network. The main function of the backbone network is to extract features from the input images, gradually extracting from the low-level features of the images, such as simple features like edges and colors, to high-level semantic features, such as more abstract features like the shape and category of objects. At the same time, it constructs the basis of a multi-scale feature pyramid, and features at different scales can capture information about objects of different sizes in the image, providing rich feature representations for subsequent feature fusion and object detection; the output result of the backbone network will be input into the feature fusion layer. The role of the feature fusion layer is to fuse the multi-scale features extracted by the backbone network. By different fusion methods, such as addition, splicing, etc., it integrates the feature information at different scales, enhancing the model's adaptability to small targets and complex backgrounds because small targets may be obvious only in the feature maps of certain specific scales, enabling the model to better handle various different image scenarios.

[0055] Step S13: Obtain a second output result with added noise corresponding to the first output result according to the noise injection layer of the preset defect detection model, and obtain a target detection result corresponding to the second output result based on the detection head of the preset defect detection model, so as to determine the trained defect detection model based on the target detection result.

[0056] In this embodiment, first, the composite feature map output by the feature fusion layer can be received through the noise injection layer of the preset defect detection model; then, the preset noise is injected into the composite feature map by the noise injection layer, and the noise intensity of the preset noise is adjusted to obtain a noisy feature map; wherein, the preset noise can include Gaussian noise and / or salt-and-pepper noise. Further, the noisy feature map output by the noise injection layer can be received through the detection head of the preset defect detection model; then, the detection head is used to detect the noisy feature map to obtain a target detection result; wherein, the target detection result can include the defect position, defect category, confidence level, etc. Finally, the target detection result can be compared with the target defect position, target defect category, and target confidence level in the original defect image to determine an error result, and the parameters of the preset defect detection model are adjusted using the error result to determine the trained defect detection model.

[0057] It should be noted that the images generated by using the generative adversarial network may have the situation that the generator is not perfect, which will cause some unrealistic textures or artifacts to appear in the images, such as blurred edges, making the boundaries of the objects in the images look unclear, or there are unnatural color transitions, that is, the change of colors does not conform to the real situation. These problems will affect the accurate understanding and judgment of the model on the images. On the other hand, during the process of generating images, the generator may inadvertently introduce high-frequency noise, which does not exist in real data, and during the learning process of the model, these noises may be misinterpreted as features of the images, resulting in deviations in the judgment of the model, such as misjudging the noise as part of the defect. To solve the above problems, in this embodiment, a noise injection layer can be added between the feature fusion layer and the detection head of the preset defect detection model. By introducing noise during the training process, the model is forced to learn how to accurately identify defects under the interference of noise, thereby improving the model's ability to identify defects and anti-interference ability. At the same time, since adversarial perturbations may cause the model to overfit to high-frequency noise or artifacts, adding a noise injection layer can enable the model to adapt to these interferences during training, avoid the model relying too much on the features of noise or artifacts, and thus improve the model's resistance to adversarial attacks and make the model more robust.

[0058] For example, in the noise injection layer, Gaussian noise or salt-and-pepper noise can be added. At the same time, the noise amplitude can be controlled by setting the noise intensity. Since the detection head is responsible for the final feature decoding and prediction, injecting noise before the detection head can directly affect the decision boundary of the model, that is, the criterion based on which the model determines whether there is a target in the image and the category and position of the target, so that the model can adapt to the existence of noise during training and improve its robustness in practical applications.

[0059] Further, the output result of the noise injection layer is input into the detection head. The detection head is used to detect the noise-added feature map, and through a series of calculations and judgments, the target detection result is obtained. The target detection result includes information such as the defect position (determining the specific coordinate position of the defect in the image), the defect category (judging the type of the defect, such as crack, delamination, etc.), and the confidence level (representing the evaluation of the reliability of the model's detection result). These results are the judgments of the model on whether there are defects in the input image and the defect-related attributes. The target detection result is compared with the target defect position, the target defect category, and the target confidence level in the original defect image. Among them, the original defect image is a part of the training data, which contains known real defect information. By comparing the detection result with the real information, the detection error of the model, that is, the error result, can be determined. According to the magnitude and direction of the error, corresponding optimization algorithms, such as stochastic gradient descent, can be used to update the parameters of the model, so that the model can reduce the error and improve the detection accuracy in subsequent training and detection. By continuously repeating this comparison and adjustment process, the performance of the model is gradually optimized, and finally the trained defect detection model is determined, so that it can more accurately detect the defects in the image.

[0060] Step S14: Optimize the trained defect detection model using the validation set to obtain a target defect detection model, and perform defect detection on the images in the test set based on the target defect detection model, and determine whether the target defect detection model meets the expected standard, so as to use the target defect detection model that meets the expected standard to detect the defects in the blade inner cavity.

[0061] In this embodiment, the preset defect detection model can be trained based on the three-stage model training method and using the training set; among them, the training process corresponding to the three-stage model training method can include: extracting images with defect sizes within the first preset size range and clear backgrounds from the training set as simple samples, and training the preset defect detection model based on the simple samples to perform the first-stage model training; extracting images with defect sizes within the second preset size range and having defect blur and / or defect occlusion from the training set as difficult samples, and training the preset defect detection model based on the difficult samples to perform the second-stage model training; extracting the original defect images from the training set as real samples, and training the preset defect detection model based on the real samples to perform the third-stage model training.

[0062] It is understandable that real data with larger defect sizes and simple backgrounds, as well as data synthesized through a generative adversarial network (GAN), are first selected to train the model. These data are characterized by obvious defects and less background interference, which is convenient for the model to learn the basic defect morphology and enables the model to establish a preliminary understanding of defects. For example, in the detection of the inner cavity of a blade, large cracks have obvious edge features, and the model can quickly establish sensitivity to the defect location and prior knowledge of the defect shape by learning these features, laying a good foundation for subsequent learning. After the model has a certain basic recognition ability through simple sample training, synthetic data with smaller defect sizes and cases of blur or occlusion are gradually added. These data simulate the complex situations that may occur in real industrial scenarios, enabling the model to focus on learning how to resist these interference factors on the basis of having mastered the core features, further improving the model's anti-interference ability and the ability to handle complex situations. Finally, complex real defect data is introduced to fine-tune the model. The real data contains various complex situations that may be encountered in actual detection and has a certain distribution difference from the previous synthetic data. By using real data for fine-tuning, the model can adjust its decision boundary to bridge the distribution difference between the synthetic data and the real data. This fine-tuning method can significantly improve the training efficiency of the model, enabling the model to better adapt to the actual defect detection task and further improving the model's performance.

[0063] In this way, the three-stage model training method in this embodiment improves the performance and generalization ability of the model in the task of detecting defects in the inner cavity of the blade by reasonably arranging the difficulty and type of training samples, enabling the model to gradually learn and improve.

[0064] Experimental results:

[0065] On a test set of 100 inner cavity pictures of blades (50 with defects and 50 without defects), a comparative experiment was conducted between the model proposed in this embodiment and the original YOLOv11 model. The recall rate of the original YOLOv11 model on the test set was 74%, and the precision rate was 73%. The recall rate of the new model proposed in this embodiment on the test set was 82%, and the precision rate was 84%.

[0066] The results show that the recall rate and precision rate of the new model proposed in this embodiment on the test set both far exceed the baseline model YOLOv11. In summary, the method for detecting defects in the inner cavity of the fan blade proposed in this embodiment has greatly improved the detection ability and generalization ability of the original detection model.

[0067] As can be seen from the above, and referring to Figure 2As shown, in this embodiment, first, new target defect samples are obtained using the original defect samples and based on the GNA network. A target data set is constructed based on the original defect samples, target defect samples, and defect-free samples, and the target data set is divided into a training set, a validation set, and a test set according to a preset ratio. Then, the training set is input into a preset defect detection model, a feature map is obtained based on the backbone network of the preset defect detection model, and a first output result corresponding to the feature map is obtained based on the feature fusion layer of the preset defect detection model. A second output result with added noise corresponding to the first output result is obtained according to the noise injection layer of the preset defect detection model, and a target detection result corresponding to the second output result is obtained based on the detection head of the preset defect detection model, so as to determine the trained defect detection model based on the target detection result. Finally, the trained defect detection model is optimized using the validation set to obtain a target defect detection model, and the images in the test set are defect-detected using the target defect detection model to determine whether the target defect detection model meets the expected standard, so as to use the target defect detection model that meets the expected standard to detect the defects in the inner cavity of the blade. In this way, based on the preset generative adversarial network, brand-new target defect images similar to the original data distribution but with different perspectives, textures, etc. are created. At the same time, the original defect images, target defect images, and defect-free images are mixed and divided into a training set, a validation set, and a test set, greatly expanding the effective sample size, alleviating the problem of data scarcity, and providing richer learning materials for model training; compared with traditional affine transformation, the preset generative adversarial network can generate defect features with non-linear transformation, such as simulating defects with different lighting changes and material wear degrees. Therefore, the preset generative adversarial network can be used to learn the latent distribution of real defect images to generate synthetic samples with brand-new perspectives and complex textures, significantly increasing data diversity and enabling the model to learn more comprehensive defect representations. At the same time, a noise injection layer is added to the preset defect detection model to randomly inject Gaussian noise, salt-and-pepper noise, etc. into the feature map output by the backbone network, simulating interference factors such as sensor noise and uneven lighting in the real detection scenario, further improving the robustness of the model to complex texture defects; the feature fusion layer of the preset defect detection model is used to encode the real data features and synthetic data features respectively, and the weights of the two types of data are automatically learned through the attention mechanism, suppressing the "pseudo-features" in the synthetic data and enhancing the key features of the real data, enabling the model to effectively balance the feature distributions of the synthetic data and the real data, using the diversity of the synthetic data to improve the generalization ability and avoiding being misled by its limitations, and finally achieving more reliable defect recognition in the real detection scenario.

[0068] See Figure 3 As shown, an embodiment of the present application also discloses a device, including:

[0069] The dataset acquisition module 11 is used to utilize the original defect images and obtain new target defect images according to a preset generative adversarial network, construct a target dataset based on the original defect images, the target defect images, and defect-free images, and divide the target dataset into a training set, a validation set, and a test set according to a preset ratio;

[0070] The first result acquisition module 12 is used to input the training set into a preset defect detection model, obtain a feature map based on the backbone network of the preset defect detection model, and obtain a first output result corresponding to the feature map based on the feature fusion layer of the preset defect detection model; the preset defect detection model includes a backbone network, a feature fusion layer, a noise injection layer, and a detection head;

[0071] The second result acquisition module 13 is used to obtain a second output result with noise added corresponding to the first output result according to the noise injection layer of the preset defect detection model, and obtain a target detection result corresponding to the second output result based on the detection head of the preset defect detection model, so as to determine a trained defect detection model based on the target detection result;

[0072] The defect detection module 14 is used to optimize the trained defect detection model by using the validation set to obtain a target defect detection model, perform defect detection on the images in the test set based on the target defect detection model, and determine whether the target defect detection model meets the expected standard, so as to use the target defect detection model that meets the expected standard to detect the defects in the inner cavity of the blade.

[0073] As can be seen from the above, based on the preset generative adversarial network, this application creates a brand-new target defect image that is similar to the original data distribution but different in perspective, texture, etc. At the same time, the original defect image, the target defect image, and the defect-free image are mixed and divided into a training set, a validation set, and a test set, significantly expanding the effective sample size, alleviating the problem of data scarcity, and providing richer learning materials for model training. Compared with traditional affine transformation, the preset generative adversarial network can generate defect features with non-linear transformation, such as simulating defects with different lighting changes and material wear degrees. In this way, the potential distribution of real defect images can be learned using the preset generative adversarial network to generate synthetic samples with brand-new perspectives and complex textures, significantly increasing data diversity and enabling the model to learn more comprehensive defect representations. At the same time, a noise injection layer is added to the preset defect detection model to randomly inject Gaussian noise, salt-and-pepper noise, etc. into the feature map output by the backbone network, simulating interference factors such as sensor noise and uneven lighting in real detection scenarios, further improving the robustness of the model to defects with complex textures. The feature fusion layer of the preset defect detection model is used to encode the real data features and the synthetic data features respectively, and the weights of the two types of data are automatically learned through the attention mechanism, suppressing the "pseudo-features" in the synthetic data and enhancing the key features of the real data, enabling the model to effectively balance the feature distributions of the synthetic data and the real data, using the diversity of the synthetic data to improve the generalization ability and avoiding being misled by its limitations, and finally achieving more reliable defect recognition in real detection scenarios.

[0074] In some specific embodiments, the dataset acquisition module 11 includes:

[0075] A parameter acquisition unit for acquiring preset defect parameters input by the user terminal; the preset defect parameters include size parameters, direction parameters, and density parameters;

[0076] An image acquisition unit for inputting the original defect image and the preset defect parameters into the preset generative adversarial network to obtain a new target defect image.

[0077] In some specific embodiments, the first result acquisition module 12 includes:

[0078] A feature extraction unit for performing feature extraction operations on the images in the training set using the backbone network of the preset defect detection model to obtain low-level semantic features and high-level semantic features.

[0079] In some specific embodiments, the first result acquisition module 12 includes:

[0080] A feature receiving unit for receiving the low-level semantic features and the high-level semantic features output by the backbone network through the feature fusion layer of the preset defect detection model;

[0081] The first feature map determination unit is configured to fuse the underlying semantic features and the high-level semantic features by using the feature fusion layer to obtain a composite feature map.

[0082] In some specific embodiments, the second result acquisition module 13 includes:

[0083] The first feature map reception unit is configured to receive the composite feature map output by the feature fusion layer through the noise injection layer of the preset defect detection model;

[0084] The second feature map determination unit is configured to inject preset noise into the composite feature map by using the noise injection layer and adjust the noise intensity of the preset noise to obtain a noisy feature map; the preset noise includes Gaussian noise and / or salt-and-pepper noise.

[0085] In some specific embodiments, the second result acquisition module 13 includes:

[0086] The second feature map reception unit is configured to receive the noisy feature map output by the noise injection layer through the detection head of the preset defect detection model;

[0087] The result determination unit is configured to detect the noisy feature map by using the detection head to obtain a target detection result; the target detection result includes a defect position, a defect category, and a confidence level;

[0088] Correspondingly, the second result acquisition module 13 includes:

[0089] The parameter adjustment unit is configured to compare the target detection result with the target defect position, the target defect category, and the target confidence level in the original defect image to determine an error result, and use the error result to adjust the parameters of the preset defect detection model to determine a trained defect detection model.

[0090] In some specific embodiments, the process of training the preset defect detection model based on the training set is a process of training based on a three-stage model training method; wherein, the training process corresponding to the three-stage model training method includes:

[0091] The first training unit is configured to extract images with defect sizes within a first preset size range and clear backgrounds from the training set as simple samples, and train the preset defect detection model based on the simple samples to perform first-stage model training;

[0092] A second training unit, configured to extract, from a training set, images with defect sizes within a second preset size range and having defect blurring and / or defect occlusion as difficult samples, and train the preset defect detection model based on the difficult samples to perform second-stage model training;

[0093] A third training unit, configured to extract original defect images from a training set as real samples, and train the preset defect detection model based on the real samples to perform third-stage model training.

[0094] Further, an embodiment of the present application also discloses an electronic device. Figure 4 FIG. 20 is a structural diagram of an electronic device shown according to an exemplary embodiment. The content in the figure should not be regarded as any limitation on the scope of use of the present application.

[0095] Figure 4 FIG. 20 is a schematic structural diagram of an electronic device 20 provided by an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. Among them, the memory 22 is used to store a computer program, and the computer program is loaded and executed by the processor 21 to implement the relevant steps in the fan blade inner cavity defect detection method disclosed in any of the foregoing embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0096] In this embodiment, the power supply 23 is used to provide working voltages for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows is any communication protocol applicable to the technical solution of the present application, and no specific limitation is imposed thereon here; the input / output interface 25 is used to obtain external input data or output data to the outside, and its specific interface type can be selected according to specific application requirements, and no specific limitation is made here.

[0097] In addition, the memory 22, as a carrier for resource storage, may be a read-only memory, a random access memory, a magnetic disk, or an optical disc, etc. The resources stored thereon may include an operating system 221, a computer program 222, etc., and the storage method may be temporary storage or permanent storage.

[0098] Among them, the operating system 221 is used to manage and control each hardware device and computer program 222 on the electronic device 20, and it can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program that can be used to complete the fan blade inner cavity defect detection method executed by the electronic device 20 disclosed in any of the foregoing embodiments, the computer program 222 can further include computer programs that can be used to complete other specific tasks.

[0099] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the foregoing disclosed fan blade inner cavity defect detection method. For the specific steps of this method, reference can be made to the corresponding content disclosed in the foregoing embodiments, and details will not be repeated here.

[0100] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts between the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and reference can be made to the description of the method part for related parts.

[0101] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0102] The steps of the method or algorithm described in combination with the embodiments disclosed in this article can be directly implemented by hardware, software modules executed by a processor, or a combination of the two. The software module can be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0103] Finally, it should also be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the process, method, article or device comprising said element.

[0104] The technical solutions provided in this application have been introduced in detail above. Specific examples are used in this text to elaborate on the principles and implementation manners of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A method for detecting defects in the inner cavity of a fan blade, characterized in that, Including: Using the original defect image and obtaining a new target defect image according to a preset generative adversarial network, constructing a target data set based on the original defect image, the target defect image and the defect-free image, and dividing the target data set into a training set, a validation set and a test set according to a preset ratio; Inputting the training set into a preset defect detection model, obtaining a feature map based on the backbone network of the preset defect detection model, and obtaining a first output result corresponding to the feature map based on the feature fusion layer of the preset defect detection model; The preset defect detection model includes a backbone network, a feature fusion layer, a noise injection layer and a detection head; Obtaining a second output result with added noise corresponding to the first output result according to the noise injection layer of the preset defect detection model, and obtaining a target detection result corresponding to the second output result based on the detection head of the preset defect detection model, so as to determine a trained defect detection model based on the target detection result; Optimizing the trained defect detection model by using the validation set to obtain a target defect detection model, performing defect detection on the images in the test set based on the target defect detection model, and determining whether the target defect detection model meets the expected standard, so as to use the target defect detection model that meets the expected standard to detect the defects in the inner cavity of the blade; Among them, the process of training the preset defect detection model based on the training set is a process of training based on a three-stage model training method; among them, the training process corresponding to the three-stage model training method includes: Extracting images with defect sizes within a first preset size range and clear backgrounds from the training set as simple samples, and training the preset defect detection model based on the simple samples to perform the first-stage model training; Extracting images with defect sizes within a second preset size range and having defect blur and / or defect occlusion from the training set as difficult samples, and training the preset defect detection model based on the difficult samples to perform the second-stage model training; Extracting the original defect image from the training set as a real sample, and training the preset defect detection model based on the real sample to perform the third-stage model training.

2. The method for detecting defects in the inner cavity of a fan blade according to claim 1, characterized in that The obtaining of the new target defect image by using the original defect image and according to a preset generative adversarial network includes: Obtaining preset defect parameters input by the user side; the preset defect parameters include size parameters, direction parameters and density parameters; Inputting the original defect image and the preset defect parameters into a preset generative adversarial network to obtain a new target defect image.

3. The method for detecting defects in the inner cavity of a fan blade according to claim 1, wherein, The obtaining of the feature map based on the backbone network of the preset defect detection model includes: Using the backbone network of the preset defect detection model to perform feature extraction operations on the images in the training set to obtain low-level semantic features and high-level semantic features.

4. The method for detecting defects in the inner cavity of a fan blade according to claim 3, characterized in that The obtaining of the first output result corresponding to the feature map based on the feature fusion layer of the preset defect detection model includes: Receiving the low-level semantic features and the high-level semantic features output by the backbone network through the feature fusion layer of the preset defect detection model; The underlying semantic features and the high-level semantic features are fused by using the feature fusion layer to obtain a composite feature map.

5. The method for detecting defects in the inner cavity of a fan blade according to claim 4, characterized in that, Obtaining the second output result with added noise corresponding to the first output result according to the noise injection layer of the preset defect detection model includes: Receiving the composite feature map output by the feature fusion layer through the noise injection layer of the preset defect detection model; Injecting preset noise into the composite feature map by using the noise injection layer, and adjusting the noise intensity of the preset noise to obtain a noisy feature map; the preset noise includes Gaussian noise and / or salt-and-pepper noise.

6. The method for detecting defects in the inner cavity of a fan blade according to claim 5, characterized in that Obtaining the target detection result corresponding to the second output result based on the detection head of the preset defect detection model includes: Receiving the noisy feature map output by the noise injection layer through the detection head of the preset defect detection model; Detecting the noisy feature map by using the detection head to obtain a target detection result; the target detection result includes a defect position, a defect category, and a confidence level; Correspondingly, determining the trained defect detection model based on the target detection result includes: Comparing the target detection result with the target defect position, the target defect category, and the target confidence level in the original defect image to determine an error result, and using the error result to adjust the parameters of the preset defect detection model to determine the trained defect detection model.

7. A defect detection device for the inner cavity of a fan blade, characterized in that, Including: A dataset acquisition module, configured to use the original defect image and obtain a new target defect image according to a preset generative adversarial network, construct a target dataset based on the original defect image, the target defect image, and the defect-free image, and divide the target dataset into a training set, a validation set, and a test set according to a preset ratio; A first result acquisition module, configured to input the training set into a preset defect detection model, obtain a feature map based on the backbone network of the preset defect detection model, and obtain a first output result corresponding to the feature map based on the feature fusion layer of the preset defect detection model; The preset defect detection model includes a backbone network, a feature fusion layer, a noise injection layer, and a detection head; A second result acquisition module, configured to obtain a second output result with added noise corresponding to the first output result according to the noise injection layer of the preset defect detection model, and obtain a target detection result corresponding to the second output result based on the detection head of the preset defect detection model, so as to determine the trained defect detection model based on the target detection result; A defect detection module, configured to optimize the trained defect detection model by using the validation set to obtain a target defect detection model, perform defect detection on the images in the test set based on the target defect detection model, and determine whether the target defect detection model meets the expected standard, so as to use the target defect detection model that meets the expected standard to detect the defects in the inner cavity of the blade; Among them, the process of training the preset defect detection model based on the training set is a process of training based on a three-stage model training method; among them, the relevant functional modules corresponding to the three-stage model training method include: The first training module is used to extract images with defect sizes within the first preset size range and clear backgrounds from the training set as simple samples, and train a preset defect detection model based on the simple samples to perform the first-stage model training; The second training module is used to extract images with defect sizes within the second preset size range and having defect blurring and / or defect occlusion from the training set as difficult samples, and train the preset defect detection model based on the difficult samples to perform the second-stage model training; The third training module is used to extract original defect images from the training set as real samples, and train the preset defect detection model based on the real samples to perform the third-stage model training.

8. An electronic device, characterized in that, Comprising: A memory for storing a computer program; A processor for executing the computer program to implement the method for detecting defects in the inner cavity of a wind turbine blade according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, For storing a computer program, which when executed by a processor implements the method for detecting defects in the inner cavity of a wind turbine blade according to any one of claims 1 to 6.

Images