Generative adversarial network-based fan blade image generation and defect identification method

Abstract

The invention relates to the technical field of wind turbine inspection, in particular to a wind turbine blade image generation and defect identification method based on a generative adversarial network. According to the technical scheme, the fan blade image generation and defect identification method based on the generative adversarial network comprises a defect image generation method; in an image generation part, a defective fan blade is taken as a research object, an improved CycleGAN model is adopted to generate defect data, and a dense convolution block is used for replacing a residual block in a generator to prevent a gradient disappearance phenomenon of the generated model, so that the network can be designed deeper, the training process is more stable, and the convergence speed is higher; a coordinate attention mechanism is added behind a convolutional layer and a dense convolutional block of the generator, so that the generator can more accurately position the defect, focus errors are avoided, and the generated fan blade defect sample is clearer.

Description

Technical Field

[0001] The present invention relates to the technical field of fan inspection, and particularly to a method for generating and defect-identifying fan blade images based on a generative adversarial network. Background Art

[0002] As the core component of a wind turbine that converts wind energy into mechanical energy, the blade accounts for 15% - 25% of the total cost of the fan. However, the blade is very large in size, and it is inconvenient to detect its own defects. Especially when operating under heavy load for a long time in a harsh natural environment, defects are extremely likely to occur. The defects of the blade directly lead to a 5% - 30% reduction in power generation efficiency. If the faults of the fan blade cannot be detected and processed in time, it will bring huge economic losses to the wind farm, and even cause faults in other components, which is a huge hidden danger to the safe operation of the fan. The current methods have problems such as low accuracy and high missed detection rate. It is still necessary for staff to manually interpret and mark the missed and misdetected images for recheck. The long-term and high-intensity repetitive detection operations of analysts are likely to cause visual fatigue, resulting in the inability to guarantee the recognition accuracy. At the same time, training and testing these intelligent methods require a large number of defect samples, which cannot be satisfied in reality. Therefore, in view of the above problems, research is carried out in the fields of image generation and defect recognition to generate fan blade defect images and improve the accuracy of defect recognition algorithms, providing certain guidance for the surface inspection of fans. In view of the fact that the research and development of blade surface defect detection is urgently needed for the wind power industry, the present invention is proposed. Summary of the Invention

[0003] In order to overcome the problems of low accuracy and high missed detection rate in the current methods, it is still necessary for staff to manually interpret and mark the missed and misdetected images for recheck. The long-term and high-intensity repetitive detection operations of analysts are likely to cause visual fatigue, resulting in the inability to guarantee the recognition accuracy. At the same time, training and testing these intelligent methods require a large number of defect samples, which cannot be satisfied in reality.

[0004] The technical solution of the present invention is: a method for generating and defect-identifying fan blade images based on a generative adversarial network, including the following steps: S101: First, clarify the purpose and requirements of image collection, and preprocess the images; S102: Construct a basic model and configure the parameters of the model; S103: Introduce a coordinate attention mechanism and adjust the parameters of the coordinate attention mechanism; S104: Replace the residual block with a dense convolutional block and adjust the parameters of the dense convolutional block; S105: Train the CycleGAN model, and evaluate its performance and adjust the parameters; S106: Generate incomplete images of wind turbine blades; S107: Construct a GANomaly model and configure its parameters; S108: Improve the GANomaly model by replacing the discriminator; S109: Train and test the GANomaly model, and evaluate its performance and adjust the parameters.

[0005] Preferably, when performing data preprocessing, the following steps are included: S201: Use drone inspections, on-site manual shootings, or access to public datasets as the image source; S202: Collect data related to the collection time period, location, and weather conditions, collect images according to the plan, and record the shooting conditions and blade status of each image; S203: Determine the data source according to the task requirements, formulate a detailed collection plan including various factors that may affect image quality, execute data collection, and record relevant information; S204: Remove blurred, duplicate, or irrelevant images, classify and annotate the images, and adjust the image size, resolution, and format to meet the model input requirements; S205: Perform data augmentation on the images, including rotation, flipping, and cropping, to increase data diversity.

[0006] Preferably, when constructing the CycleGAN basic model, the following steps are included: S301: Deeply understand the basic principles and architecture of CycleGAN, including the generator, discriminator, and loss function; S302: Determine whether the CycleGAN model is applicable according to the task requirements for generating wind turbine blade images, and consider possible improvements; S303: Select appropriate generator and discriminator structures according to the task requirements, and select a suitable loss function to optimize the model training process; S304: Confirm the training parameters, including the learning rate, optimizer, number of training epochs, and batch size, to ensure that the model can be effectively trained.

[0007] Preferably, when introducing the coordinate attention mechanism, the following steps are included: S401: Deeply understand the basic principles of the attention mechanism and its applications in different fields, and design a suitable coordinate attention module according to the characteristics of wind turbine blade images to capture the spatial position information in the images; S402: Determine the integration position of the coordinate attention module according to the structure of the generator, including the encoder, transformer, or decoder; S403: Incorporate the coordinate attention module into the generator and adjust the relevant parameters and structures to ensure the stability and effectiveness of the model.

[0008] Preferably, when replacing the residual block with a dense convolutional block, the following steps are included: S501: Consult relevant papers and materials on the residual block to understand its structure and principle, and analyze the application and advantages of the residual block in deep learning; S502: Analyze the advantages of the dense convolutional block in feature extraction and image generation. In the converter of CycleGAN, replace the residual block with a dense convolutional block and ensure that the structure of the replaced converter is stable and effective; S503: Adjust the relevant parameters according to the characteristics of the dense convolutional block and the task requirements, including the convolutional kernel size, stride, and number of channels; S504: Verify the impact of the adjusted parameters on the model performance through experiments and select the optimal parameter combination.

[0009] Preferably, when training and optimizing the model, the following steps are included: S601: Divide the preprocessed fan blade image dataset into a training set, a validation set, and a test set to ensure the independence and representativeness of each dataset; S602: Perform data augmentation on the training set, including rotation, flipping, or cropping. Use a data augmentation library or tool to batch process the training set images to improve data diversity; S603: Initialize the parameters of the CycleGAN model according to the configured parameters to ensure reasonable parameter initialization and avoid the problems of gradient disappearance or explosion during model training; S604: Use the training set to train the CycleGAN model, monitor the changes in the loss function and metrics, and adjust the training strategy in a timely manner; S605: Adjust the learning rate, optimizer, or loss function parameters according to the training process and results, and save the model weights and logs during the training process for subsequent analysis and comparison.

[0010] Preferably, when performing performance evaluation, the following steps are included: S701: Select suitable evaluation metrics according to the task requirements, including accuracy, recall, and F1 score, to ensure the rationality and effectiveness of the evaluation metrics; S702: Use the validation set to perform performance evaluation on the trained CycleGAN model, record the evaluation results, including the numerical values of each metric and the image generation effect; S703: Analyze the advantages and disadvantages of the model according to the evaluation metrics and results, and identify the problems existing in the model during the image generation process; S704: Adjust the model parameters or structure according to the analysis results, and verify the performance of the optimized model through experiments to ensure the effectiveness of the improvement.

[0011] Preferably, when generating defective images, the following steps are included: S801: Select one or more normal blade images from the dataset as input, ensuring that the input image quality is good and meets the model input requirements; S802: Preprocess the input image, including resizing and normalization, to ensure that the preprocessed image meets the model input requirements and is convenient for the model to process; S803: Input the preprocessed normal blade image into the trained CycleGAN model to generate the corresponding defective blade image, and save the generation result; S804: Screen the generated defective images, remove the images with poor quality or non-compliance, and ensure that the screened defective images have good quality and can be used for subsequent analysis and applications.

[0012] Preferably, when improving the GANomaly model, the following steps are included; S901: Add label smoothing technology after the reconstructed image of the GANomaly model, and reasonably design the parameters and strategies of the label smoothing technology to ensure that it can be seamlessly docked with the GANomaly model and play a role; S902: Monitor the performance changes of the model during the training process to evaluate the effectiveness of the label smoothing technology; S903: Replace the original discriminator of the GANomaly model with a PatchGAN discriminator; S904: Ensure the correctness and integrity of the replacement process, avoid introducing new errors or vulnerabilities, and at the same time, adjust the parameters of the PatchGAN discriminator to adapt to the task requirements and hardware conditions.

[0013] Preferably, when training and testing the GANomaly model, the following steps are included: S1001: Use the normal fan blade image dataset as the training set to train the GANomaly model, and at the same time prepare a small number of real defective blade images as the test set to evaluate the performance of the model; S1002: Ensure the rationality and effectiveness of the dataset division, avoid data leakage or repeated use, and at the same time consider the diversity and representativeness of the dataset to improve the generalization ability of the model; S1003: Use the training set to train the GANomaly model. During the training process, monitor the changes in the loss function, adjust parameters such as the learning rate to optimize the training effect, and record the logs and model weights during the training process; S1004: Use the test set to evaluate the performance of the trained GANomaly model. The evaluation metrics include recognition accuracy, recall rate, and F1 score. Adjust the model parameters according to the evaluation results and performance metrics to optimize its performance.

[0014] Advantages of the present invention: In the image generation part, taking the defective fan blade as the research object, the improved CycleGAN model is used to generate defective data. By replacing the residual block with a dense convolutional block in the generator, the phenomenon of gradient disappearance in the generation model is prevented, enabling the network to be designed deeper, with a more stable training process and a faster convergence speed. By adding a coordinate attention mechanism after the convolutional layer and the dense convolutional block of the generator, the generator can more accurately locate the defective part, avoid incorrect focus points, and generate clearer defective fan blade samples. Finally, using the improved CycleGAN model to transfer common defective fan blade data to normal fan blade data, a large number of high-quality defective fan blade samples are generated, solving the problem of insufficient samples in the research of fan blade defect recognition, effectively expanding the defective data set, and meeting the requirements of fan blade defect recognition. In the defect recognition part, the present invention uses the GANomaly algorithm for defect recognition, trains with normal fan blade samples, and the generator model learns the data distribution of normal samples. When a defective fan blade is input, the discriminator will determine that the data is abnormal. The present invention improves the GANomaly algorithm. By adding label smoothing technology after reconstructing the image, overfitting of the model is prevented. By using a PatchGAN discriminator to replace the original discriminator, the model pays more attention to the defect details of the image. Description of the Drawings

[0015] Figure 1 Shown is a schematic framework flow diagram of a method for generating and recognizing fan blade images based on a generative adversarial network according to the present invention; Figure 2 Shown is a schematic data preprocessing flow diagram of a method for generating and recognizing fan blade images based on a generative adversarial network according to the present invention; Figure 3 Shown is a schematic defective image generation flow diagram of a method for generating and recognizing fan blade images based on a generative adversarial network according to the present invention; Figure 4 Shown is a technical roadmap of a method for generating and recognizing fan blade images based on a generative adversarial network according to the present invention; Figure 5 Shown is a structural diagram of an improved generator of the CycleGAN algorithm of a method for generating and recognizing fan blade images based on a generative adversarial network according to the present invention. Detailed Embodiments

[0016] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0017] Please refer to Figures 1 - 5 , the present invention provides an embodiment: a method for generating and defect identifying of fan blade images based on a generative adversarial network, including the following steps: S101: First, clarify the purpose and requirements of image collection, and preprocess the images; S102: Build a basic model and configure the parameters of the model; S103: Introduce a coordinate attention mechanism and adjust the parameters of the coordinate attention mechanism; S104: Replace the residual block with a dense convolutional block and adjust the parameters of the dense convolutional block; S105: Train the CycleGAN model, and evaluate and adjust its performance and parameters; S106: Generate incomplete fan blade images; S107: Build a GANomaly model and configure the model parameters; S108: Improve the GANomaly model and replace the discriminator; S109: Train and test the GANomaly model, and evaluate and adjust its performance and parameters.

[0018] Preferably, when performing data preprocessing, it includes the following steps: S201: Use drone patrol shooting, on-site manual shooting or accessing public data sets as the image source; S202: Collect data related to the collection time period, location and weather conditions, collect images according to the plan, and record the shooting conditions and blade status of each image; S203: Determine the data source according to the task requirements, formulate a detailed collection plan, including various factors that may affect the image quality, execute data collection, and record relevant information; S204: Remove blurred, duplicate or irrelevant images, classify and label the images, and adjust the image size, resolution and format to meet the model input requirements; S205: Perform data augmentation processing on the images, including rotation, flipping and cropping, to increase the diversity of the data.

[0019] Preferably, when building the CycleGAN basic model, it includes the following steps: S301: Deeply understand the basic principles and architectures of CycleGAN, including the generator, discriminator and loss function; S302: Determine whether the CycleGAN model is applicable according to the task requirements generated from the fan blade images, and consider possible improvements; S303: Select appropriate generator and discriminator structures according to the task requirements, and select suitable loss functions to optimize the model training process; S304: Confirm the training parameters, including the learning rate, optimizer, number of training epochs, and batch size, to ensure that the model can be effectively trained.

[0020] Preferably, when introducing the coordinate attention mechanism, the following steps are included: S401: Deeply understand the basic principles of the attention mechanism and its applications in different fields, and design a suitable coordinate attention module according to the characteristics of the fan blade images to capture the spatial position information in the images; S402: Determine the integration position of the coordinate attention module according to the structure of the generator, including the encoder, transformer, or decoder; S403: Incorporate the coordinate attention module into the generator, and adjust the relevant parameters and structures to ensure the stability and effectiveness of the model.

[0021] Preferably, when replacing the residual block with a dense convolutional block, the following steps are included: S501: Consult relevant papers and materials on the residual block, understand its structure and principle, and analyze the applications and advantages of the residual block in deep learning; S502: Analyze the advantages of the dense convolutional block in feature extraction and image generation. In the transformer of CycleGAN, replace the residual block with a dense convolutional block, and ensure that the structure of the replaced transformer is stable and effective; S503: Adjust the relevant parameters according to the characteristics of the dense convolutional block and the task requirements, including the convolutional kernel size, stride, and number of channels; S504: Verify the impact of the adjusted parameters on the model performance through experiments, and select the optimal parameter combination.

[0022] Preferably, when performing model training and optimization, the following steps are included: S601: Divide the preprocessed fan blade image dataset into a training set, a validation set, and a test set to ensure the independence and representativeness of each dataset; S602: Perform data augmentation on the training set, including rotation, flipping, or cropping. Use a data augmentation library or tool to batch process the training set images to improve data diversity; S603: Initialize the parameters of the CycleGAN model according to the configured parameters to ensure reasonable parameter initialization and avoid the problems of gradient disappearance or explosion during the model training process; S604: Use the training set to train the CycleGAN model, monitor the changes in the loss function and metrics, and adjust the training strategy in a timely manner; S605: According to the training process and results, adjust the learning rate, optimizer, or loss function parameters, and save the model weights and logs during the training process for subsequent analysis and comparison.

[0023] Preferably, when performing performance evaluation, the following steps are included: S701: Select suitable evaluation metrics according to the task requirements, including accuracy, recall, and F1 score, to ensure the rationality and effectiveness of the evaluation metrics; S702: Use the validation set to perform performance evaluation on the trained CycleGAN model, record the evaluation results, including the numerical values of various metrics and the image generation effect; S703: Analyze the advantages and disadvantages of the model according to the evaluation metrics and results, and identify the problems existing in the model during the image generation process; S704: Adjust the model parameters or structure according to the analysis results, and verify the performance of the optimized model through experiments to ensure the effectiveness of the improvement.

[0024] Preferably, when generating defect images, the following steps are included: S801: Select one or more normal leaf images from the dataset as input, ensuring that the input image quality is good and meets the model input requirements; S802: Preprocess the input image, including resizing and normalization, to ensure that the preprocessed image meets the model input requirements and is convenient for the model to process; S803: Input the preprocessed normal leaf image into the trained CycleGAN model to generate the corresponding defective leaf image, and save the generation result; S804: Screen the generated defect images, remove the images with poor quality or non-compliance, and ensure that the screened defect images have good quality and can be used for subsequent analysis and applications.

[0025] Preferably, when improving the GANomaly model, the following steps are included; S901: Add label smoothing technology after the reconstructed image of the GANomaly model, and reasonably design the parameters and strategies of the label smoothing technology to ensure that it can be seamlessly docked with the GANomaly model and play a role; S902: Monitor the performance changes of the model during the training process to evaluate the effectiveness of the label smoothing technology; S903: Replace the original discriminator of the GANomaly model with a PatchGAN discriminator; S904: Ensure the correctness and integrity of the replacement process, avoid introducing new errors or vulnerabilities. Meanwhile, adjust the parameters of the PatchGAN discriminator to adapt to the task requirements and hardware conditions.

[0026] Preferably, when training and testing the GANomaly model, the following steps are included: S1001: Use the normal wind turbine blade image dataset as the training set to train the GANomaly model. Meanwhile, prepare a small number of real defective blade images as the test set to evaluate the performance of the model; S1002: Ensure the rationality and effectiveness of the dataset division, avoid data leakage or repeated use, and consider the diversity and representativeness of the dataset to improve the generalization ability of the model; S1003: Use the training set to train the GANomaly model. During the training process, monitor the change of the loss function, adjust parameters such as the learning rate to optimize the training effect, and record the logs and model weights during the training process; S1004: Use the test set to evaluate the performance of the trained GANomaly model. The evaluation metrics include recognition accuracy, recall rate, and F1 score. Adjust the model parameters according to the evaluation results and performance metrics to optimize its performance.

[0027] Example 1 S1101: Use the wind turbine blade images taken by drone inspection, the wind turbine blade images taken manually on-site, and access and download the wind turbine blade images in the public dataset; S1102: Determine that the collection time period is spring and autumn, the locations are multiple wind farms, and the weather conditions are sunny or cloudy. Develop a detailed collection plan, including shooting angles, distances, and lighting conditions, etc.; S1103: Remove blurred, repeated, or irrelevant images, classify the images, such as normal blades, cracked blades, worn blades, etc., and make annotations. Adjust the image size to 256x256 pixels, the resolution is 72DPI, and the format is JPEG to meet the model input requirements; S1104: Perform rotation (±15°), flipping (horizontal / vertical), and cropping (randomly crop to 224x224 pixels) on the images to increase the data diversity; S1105: Select the generator with the ResNet structure and the discriminator with the PatchGAN structure. Select the cycle consistency loss, adversarial loss, and identity loss as the total loss function. Set the learning rate to 0.0002, the optimizer to Adam, the number of training epochs to 100, and the batch size to 16; S1106: Design a coordinate attention module suitable for fan blade images to capture spatial position information in the images and determine that the coordinate attention module is incorporated into the encoder part of the generator; S1107: Incorporate the coordinate attention module into the encoder part of the generator and adjust relevant parameters of the coordinate attention module, such as the size and number of channels of the feature map, to ensure the stability and effectiveness of the model; S1108: In the CycleGAN's transformer, replace the residual block with a dense convolutional block and adjust relevant parameters of the dense convolutional block, such as the convolutional kernel size of 3x3, the stride of 1, and the number of channels gradually increasing according to the size of the feature map; S1109: Divide the preprocessed fan blade image dataset into a training set (80%), a validation set (10%), and a test set (10%). Perform data augmentation on the training set, including rotation, flipping, or cropping, to improve data diversity; S1110: Select accuracy, recall, and F1 score as evaluation metrics, use the validation set to evaluate the performance of the trained CycleGAN model, and record the evaluation results; S1111: Select one or more normal blade images from the dataset as input, and adjust the size and normalize the input images; S1112: Input the preprocessed normal blade images into the trained CycleGAN model to generate corresponding defective blade images, and screen the generated defective images to remove images with poor quality or non-compliance; S1113: Add label smoothing technology after the reconstructed images of the GANomaly model, replace the original discriminator of the GANomaly model with a PatchGAN discriminator, and adjust the parameters of the PatchGAN discriminator to adapt to task requirements and hardware conditions; S1114: Use the normal fan blade image dataset as the training set, prepare a small number of real defective blade images as the test set, and use the training set to train the GANomaly model.

[0028] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the gist of the present invention.

Claims

1. A method for generating fan blade images and defect recognition based on a generative adversarial network, characterized in that: It includes the following steps: S101: First, clarify the purpose and requirements of image collection, and preprocess the images; S102: Build a basic model and configure the parameters of the model; S103: Introduce the coordinate attention mechanism and adjust the parameters of the coordinate attention mechanism; S104: Replace the residual block with a dense convolutional block and adjust the parameters of the dense convolutional block; S105: Train the CycleGAN model, and evaluate and adjust its performance and parameters; S106: Generate incomplete images of wind turbine blades; S107: Build the GANomaly model and configure the model parameters; S108: Improve the GANomaly model and replace the discriminator; S109: Train and test the GANomaly model, and evaluate and adjust its performance and parameters.

2. The method for generating and defect-identifying fan blade images based on a generative adversarial network according to claim 1, wherein: When performing data preprocessing, it includes the following steps: S201: Use drone patrol shooting, on-site manual shooting, or access to public datasets as the image source; S202: Collect data related to the collection time period, location, and weather conditions, collect images according to the plan, and record the shooting conditions and blade status of each image; S203: Determine the data source according to the task requirements, formulate a detailed collection plan, including various factors that may affect image quality, execute data collection, and record relevant information; S204: Remove blurred, duplicate, or irrelevant images, classify and label the images, and adjust the image size, resolution, and format to meet the model input requirements; S205: Perform data augmentation on the images, including rotation, flipping, and cropping, to increase the diversity of the data.

3. A method for generating and defect identifying of fan blade images based on generative adversarial network according to claim 1, characterized in that: When building the CycleGAN basic model, it includes the following steps: S301: Deeply understand the basic principles and architecture of CycleGAN, including the generator, discriminator, and loss function; S302: Determine whether the CycleGAN model is applicable according to the task requirements of generating wind turbine blade images, and consider possible improvements; S303: Select appropriate generator and discriminator structures according to the task requirements, and select a suitable loss function to optimize the model training process; S304: Confirm the training parameters, including the learning rate, optimizer, number of training epochs, and batch size, to ensure that the model can be effectively trained.

4. A method for generating and defect-identifying fan blade images based on a generative adversarial network according to claim 1, characterized in that: When introducing the coordinate attention mechanism, it includes the following steps: S401: Deeply understand the basic principles of the attention mechanism and its applications in different fields, and design a suitable coordinate attention module according to the characteristics of wind turbine blade images to capture the spatial position information in the images; S402: Determine the integration position of the coordinate attention module according to the structure of the generator, including the encoder, transformer, or decoder; S403: Integrate the coordinate attention module into the generator and adjust the relevant parameters and structures to ensure the stability and effectiveness of the model.

5. A method for generating and defect-identifying a fan blade image based on a generative adversarial network according to claim 1, characterized in that: When replacing the residual block with a dense convolutional block, it includes the following steps: S501: Consult relevant papers and materials on the residual block, understand its structure and principle, and analyze the application and advantages of the residual block in deep learning; S502: Analyze the advantages of dense convolutional blocks in feature extraction and image generation. In the transformer of CycleGAN, replace the residual blocks with dense convolutional blocks and ensure the stability and effectiveness of the replaced transformer structure; S503: Adjust relevant parameters according to the characteristics of dense convolutional blocks and task requirements, including the convolutional kernel size, stride, and number of channels; S504: Verify the impact of the adjusted parameters on the model performance through experiments and select the optimal parameter combination.

6. A method for generating and defect-identifying fan blade images based on a generative adversarial network according to claim 1, characterized in that: When training and optimizing the model, the following steps are included: S601: Divide the preprocessed wind turbine blade image dataset into a training set, a validation set, and a test set to ensure the independence and representativeness of each dataset; S602: Perform data augmentation on the training set, including rotation, flipping, or cropping. Use a data augmentation library or tool to batch process the training set images to improve data diversity; S603: Initialize the parameters of the CycleGAN model according to the configured parameters to ensure reasonable parameter initialization and avoid the problems of gradient disappearance or explosion during model training; S604: Use the training set to train the CycleGAN model, monitor the changes in the loss function and metrics, and adjust the training strategy in a timely manner; S605: Adjust the learning rate, optimizer, or loss function parameters according to the training process and results, and save the model weights and logs during the training process for subsequent analysis and comparison.

7. A method for generating and defect identifying of fan blade images based on generative adversarial network according to claim 1, characterized in that: When performing performance evaluation, the following steps are included: S701: Select suitable evaluation metrics according to the task requirements, including accuracy, recall, and F1 score, to ensure the rationality and effectiveness of the evaluation metrics; S702: Use the validation set to perform performance evaluation on the trained CycleGAN model, record the evaluation results, including the values of various metrics and the image generation effect; S703: Analyze the advantages and disadvantages of the model according to the evaluation metrics and results, and identify the problems existing in the model during the image generation process; S704: Adjust the model parameters or structure according to the analysis results, and verify the performance of the optimized model through experiments to ensure the effectiveness of the improvement.

8. A method for generating and defect-identifying fan blade images based on a generative adversarial network according to claim 1, characterized in that: When generating defect images, the following steps are included: S801: Select one or more normal blade images from the dataset as input, ensuring that the input image quality is good and meets the model input requirements; S802: Preprocess the input image, including resizing and normalization, to ensure that the preprocessed image meets the model input requirements and is convenient for model processing; S803: Input the preprocessed normal blade image into the trained CycleGAN model to generate the corresponding defective blade image and save the generation result; S804: Screen the generated defect images, remove the images with poor quality or non-compliance, and ensure that the screened defect images have good quality and can be used for subsequent analysis and applications.

9. A method for generating and defect identifying of fan blade images based on generative adversarial network according to claim 1, characterized in that: When improving the GANomaly model, the following steps are included; S901: Add label smoothing technology after the reconstructed image of the GANomaly model. At the same time, reasonably design the parameters and strategies of the label smoothing technology to ensure that it can be seamlessly docked with the GANomaly model and play a role; S902: Monitor the performance changes of the monitoring model during the training process to evaluate the effectiveness of label smoothing techniques; S903: Replace the original discriminator of the GANomaly model with a PatchGAN discriminator; S904: Ensure the correctness and integrity of the replacement process, avoid introducing new errors or vulnerabilities, and at the same time, adjust the parameters of the PatchGAN discriminator to adapt to the task requirements and hardware conditions.

10. A method for generating and defect identifying of fan blade images based on generative adversarial network according to claim 1, characterized in that: When training and testing the GANomaly model, the following steps are included: S1001: Use the normal fan blade image dataset as the training set to train the GANomaly model, and at the same time prepare a small number of real defective blade images as the test set to evaluate the performance of the model; S1002: Ensure the rationality and effectiveness of the dataset division, avoid data leakage or repeated use, and at the same time consider the diversity and representativeness of the dataset to improve the generalization ability of the model; S1003: Use the training set to train the GANomaly model. During the training process, monitor the changes in the loss function, adjust parameters such as the learning rate to optimize the training effect, and at the same time record the logs and model weights during the training process; S1004: Use the test set to evaluate the performance of the trained GANomaly model. The evaluation metrics include recognition accuracy, recall rate, and F1 score. Adjust the model parameters according to the evaluation results and performance metrics to optimize its performance.

Images