Non-supervision defect detection method and equipment for inner cavity of wind power generation blade and medium

Abstract

The invention relates to the field of defect detection, and discloses a wind power generation blade inner cavity unsupervised defect detection method and device and a medium, and the method comprises the steps: collecting a wind power generation blade inner cavity image through a camera carried by an unmanned aerial vehicle, and constructing a wind power generation blade inner cavity defect detection basic data set; performing data enhancement on the basic data set to obtain an enhanced data set; segmenting the enhanced data set to obtain a training set and a test set; constructing a wind power generation blade inner cavity unsupervised defect detection model, completing detection model training by utilizing the training set and introducing an early stop mechanism, and testing through the test set; the detection model passing the test is deployed in a portable computing device, and wind power generation blade inner cavity defect detection is achieved; the method can achieve the detection of the internal defects of the wind power generation blade, and is high in detection speed and precision.

Description

Technical Field

[0001] The present invention relates to the field of defect detection, and particularly to an unsupervised defect detection method for the inner cavity of a wind turbine blade. Background Art

[0002] Currently, the detection of wind turbine blades based on artificial intelligence neural network algorithms mainly uses object detection algorithms (such as YOLO, Faster R CNN, Mask R CNN, etc.). A large number of surface defect images of wind turbine blades are collected by cameras, and the defect areas and defect types are manually labeled in sequence using image annotation software. A supervised image dataset is built and divided into non-overlapping training sets and test sets. The algorithms used learn the defect characteristics supervised based on the training set to obtain the required defect object detection model and deploy it to infer real-time images.

[0003] Currently, the detection parts mainly focus on the outer surface of wind turbine blades. Videos or images of the outer surface of wind turbine blades are collected based on devices such as drones or fixed cameras to achieve real-time defect detection of the outer surface.

[0004] The defects of the above object detection algorithms are that the supervised method relies on manual annotation, which is time-consuming and costly. At the same time, the cost of obtaining a large number of images for training is too high. Usually, the scenario belongs to a small sample dataset, and it is difficult to train a good detection model.

[0005] The defect of the above detection parts mainly focusing on the outer surface of wind turbine blades is that it is difficult to detect the defects existing in the inner cavity of wind turbine blades. In fact, these defects are also extremely risky and urgently need an effective algorithm for the inner cavity of wind turbine blades. Summary of the Invention

[0006] The purpose of the present invention is to propose an unsupervised defect detection method for the inner cavity of a wind turbine blade to solve the technical problem that the current defect detection method cannot effectively detect the internal defects of wind turbine blades.

[0007] Specifically, an unsupervised defect detection method for the inner cavity of a wind turbine blade provided by the present invention includes the following steps:

[0008] S1. Use a camera carried by a drone to collect images of the inner cavity of a wind turbine blade and construct a basic dataset for defect detection of the inner cavity of a wind turbine blade;

[0009] S2. Perform data augmentation on the basic dataset to obtain an augmented dataset;

[0010] S3. Divide the augmented dataset to obtain a training set and a test set;

[0011] S4. Construct an unsupervised defect detection model for the inner cavity of wind turbine blades, complete the training of the detection model using the training set and introducing an early stopping mechanism, and test it through the test set;

[0012] S5. Deploy the tested detection model to a portable computing device to achieve defect detection for the inner cavity of wind turbine blades.

[0013] A storage medium stores instructions and data for implementing an unsupervised defect detection method for the inner cavity of wind turbine blades.

[0014] An unsupervised defect detection device for the inner cavity of wind turbine blades includes: a processor and the storage medium; the processor loads and executes the instructions and data in the storage medium for implementing an unsupervised defect detection method for the inner cavity of wind turbine blades.

[0015] The beneficial effects provided by the present invention are as follows:

[0016] 1. Unsupervised learning reduces the cost of manual annotation

[0017] Traditional defect detection methods for wind turbine blades usually rely on supervised learning, which requires a large amount of defect image data manually annotated, consuming time, labor and high cost. This solution adopts an unsupervised learning method, only requires normal inner cavity images of blades for training, without manually annotating defect data, significantly reducing the cost of data collection and annotation, and at the same time avoiding the subjective errors that may be introduced by manual annotation.

[0018] 2. Adapt to the complex inner cavity environment and improve the detection coverage

[0019] Existing technologies mainly focus on the defect detection of the outer surface of wind turbine blades, while the inner cavity defects also have high risks, but are ignored due to the complex environment and difficulty in accessing. This solution can flexibly enter the inner cavity of the blade for image collection by using a drone equipped with a high-definition camera, and combined with image registration technology, ensures a comprehensive detection coverage, effectively solving the problem of inner cavity defect detection.

[0020] 3. High-precision pixel-level defect positioning

[0021] This solution introduces a small-sample anomaly detection method (RegAD) based on image registration, combined with unsupervised fine-tuning of the Resnet18 pre-trained model, which can compare the differences between normal images and test images during model inference and output a defect area mask with pixel-level accuracy. This high-precision defect positioning ability can accurately identify tiny defects, providing a reliable basis for subsequent repair and maintenance.

[0022] 4. Data augmentation improves the generalization ability of the model

[0023] Through geometric data augmentation (such as random rotation, translation, and shearing) and non-geometric data augmentation (such as random flipping, cropping, and noise injection) methods, this solution significantly amplifies the diversity of the training dataset and improves the generalization ability of the model. This enables the model to better adapt to detection tasks under different lighting, angles, and background conditions, and enhances the robustness of detection.

[0024] 5. Early stopping mechanism to optimize training efficiency

[0025] During the model training process, an early stopping mechanism is introduced to automatically stop training when the model performance no longer improves, avoiding ineffective consumption of computing resources and model overfitting problems. This not only reduces the training cost but also ensures the model is deployed in an optimal state, further enhancing the detection efficiency and accuracy.

[0026] 6. Deployment on portable devices for real-time detection

[0027] This solution deploys the trained model to portable computing devices (such as laptops or cloud servers). Combining real-time shooting by drones and wireless transmission technology, it can achieve real-time detection of defects in the inner cavity of wind turbine blades. This deployment method is flexible and efficient, suitable for on-site operations, and significantly improves the timeliness and practicality of detection.

[0028] 7. Defect visualization and recording

[0029] By annotating the defect area mask to the corresponding image and visualizing it, this solution can intuitively present the detection results, facilitating technicians to quickly locate and evaluate defects. At the same time, the defect area is saved in the form of an image mask, providing reliable data support for subsequent analysis and historical records. Brief description of the drawings

[0030] Figure 1 is a schematic diagram of the process of the method of the present invention;

[0031] Figure 2 is a schematic diagram of the structure of the detection model of the present invention;

[0032] Figure 3 is a schematic diagram of the operation of the hardware device in the embodiment of the present invention. Detailed implementation manners

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in conjunction with the drawings.

[0034] Before formally elaborating on the present invention, first a general description of the solution of the present invention is given for easy understanding.

[0035] Please refer to Figure 1, a method for unsupervised defect detection of the inner cavity of a wind turbine blade provided by the present invention includes:

[0036] S1. Use a camera carried by a drone to collect images of the inner cavity of the wind turbine blade, and construct a basic dataset for defect detection of the inner cavity of the wind turbine blade;

[0037] It should be noted that to construct a basic dataset for defect detection of the inner cavity of the wind turbine blade, in the wind turbine blade scenario, use a drone carrying a high-definition camera to take pictures of the inner cavity of the wind turbine blade on-site. The inner cavity pictures include common defect types that may exist.

[0038] Based on the prior knowledge of surface defect detection, manually classify the images of the normal inner cavity surface and the defective inner cavity surface. Obtain two sets of images, namely the images of the normal inner cavity surface and the defective inner cavity surface. Record the defect types and regions existing in the images with pixel-level accuracy. The defect types are recorded in the form of text documents, and the defect regions are recorded in the form of image masks.

[0039] S2. Perform data augmentation on the basic dataset to obtain an augmented dataset;

[0040] It should be noted that in the present invention, geometric data augmentation methods and non-geometric data augmentation methods are used for data augmentation.

[0041] As an embodiment, the geometric data augmentation method focuses on changing the geometric shape of the image. Use the image random rotation method, that is, rotate the image between 0 and 360 degrees; the image random translation method, that is, move the image to provide different views; the image random shearing method, that is, move a part of the image in one direction while the other part moves in the opposite direction.

[0042] As an embodiment, the non-geometric data augmentation method focuses on the visual appearance of the image rather than its geometric shape. Use the image random flipping method, that is, flip the image horizontally or vertically; the image random cropping method, that is, randomly crop a part of the image and enlarge it to the size of the original image; the image random noise method, that is, inject random image noise generated by a computer noise generator into the image.

[0043] S3. Segment the augmented dataset to obtain a training set and a test set;

[0044] It should be noted that in the present invention, based on the dataset segmentation strategy, the data is segmented into a training set and a test set. All the images in the training set are of normal blade inner cavities, and the test set includes images of both normal and defective blade inner cavities.

[0045] The dataset splitting strategy is to classify the images of the normal inner cavity surface and the defective inner cavity surface manually. 90% of the images of the normal inner cavity surface are divided into the training set, and the remaining 10% of the images of the normal inner cavity surface and all the images of the defective inner cavity surface are divided into the test set, and it is ensured that there are no duplicate images between the training set and the test set.

[0046] In addition, in the present invention, the dataset architecture is built in the format of the MVTec anomaly detection dataset, including the basic dataset for detecting the inner cavity defects of the wind turbine blade obtained by the camera carried by the drone; the augmented dataset obtained by data augmentation; the defect types and regions existing in the image are recorded with pixel-level accuracy, and the defect types are recorded in the form of a text document, and the defect regions are recorded in the form of an image mask.

[0047] S4. Construct an unsupervised defect detection model for the inner cavity of the wind turbine blade, use the training set and introduce an early stopping mechanism to complete the training of the detection model, and test it through the test set;

[0048] As an embodiment, in the present invention, the Resnet18 pre-trained model is fine-tuned unsupervised on the training set composed of the images of the normal inner cavity of the blade (the structure reference Figure 2 ). During the model inference process, the model compares the differences between the normal and test images, outputs the defect region of the image, which is a mask with pixel-level accuracy, and annotates it to the image. An early stopping mechanism is set to reduce the training cost and obtain an excellent detection model at the same time.

[0049] It should be noted that in the training process of an introduced small-sample anomaly detection framework RegAD based on image registration, when the model continues to train but the performance deteriorates, the training is stopped in advance and the best model is saved to prevent the consumption of invalid computing resources and the decline of the model performance.

[0050] S5. Deploy the tested detection model to a portable computing device to realize the detection of the inner cavity defects of the wind turbine blade.

[0051] In the present invention, the video of the inner cavity of the wind turbine blade is obtained in real time in the form of being taken by the camera carried by the drone and transmitted to the portable computing device through a wireless network.

[0052] Use the video frame extraction algorithm to intercept the key frames of the video as the images of the inner cavity of the wind turbine blade.

[0053] Input the images of the inner cavity of the wind turbine blade into the trained unsupervised defect detection model for the inner cavity of the wind turbine blade, and output the defect region image mask through the neural network inference.

[0054] Annotate the defect region image mask to the corresponding image and visualize it, and save the defect region in the form of an image mask.

[0055] Please refer to Figure 3 , Figure 3 which is a schematic diagram of the operation of the hardware device according to an embodiment of the present invention. The hardware device specifically includes: an unsupervised defect detection device 401 for the inner cavity of a wind power blade, a processor 402, and a storage medium 403.

[0056] An unsupervised defect detection device 401 for the inner cavity of a wind power blade: The unsupervised defect detection device 401 for the inner cavity of a wind power blade implements the unsupervised defect detection method for the inner cavity of a wind power blade.

[0057] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the unsupervised defect detection method for the inner cavity of a wind power blade.

[0058] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the unsupervised defect detection method for the inner cavity of a wind power blade.

[0059] The beneficial effects of the present invention are as follows:

[0060] 1. Unsupervised learning, reducing the cost of manual annotation

[0061] Traditional defect detection methods for wind power blades usually rely on supervised learning, which requires a large amount of defect image data manually annotated, consuming time, labor and high cost. This solution adopts an unsupervised learning method, only requiring normal inner cavity images of the blade for training, without the need for manual annotation of defect data, significantly reducing the cost of data collection and annotation, and at the same time avoiding the subjective errors that may be introduced by manual annotation.

[0062] 2. Adapt to the complex inner cavity environment and improve the detection coverage

[0063] Existing technologies mainly focus on the defect detection of the outer surface of wind power blades, while inner cavity defects also have high risks, but are ignored due to the complex environment and difficulty in accessing. This solution can flexibly enter the inner cavity of the blade for image collection by using a drone equipped with a high-definition camera, and combined with image registration technology, ensuring a comprehensive detection coverage and effectively solving the problem of inner cavity defect detection.

[0064] 3. High-precision pixel-level defect localization

[0065] This solution introduces a small-sample anomaly detection method (RegAD) based on image registration, combined with unsupervised fine-tuning of the Resnet18 pre-trained model, which can compare the differences between normal images and test images during model inference and output a defect region mask with pixel-level accuracy. This high-precision defect localization ability can accurately identify tiny defects and provide a reliable basis for subsequent repair and maintenance.

[0066] 4. Data augmentation to enhance the generalization ability of the model

[0067] Through geometric data augmentation (such as random rotation, translation, and shearing) and non-geometric data augmentation (such as random flipping, cropping, and noise injection) methods, this solution significantly amplifies the diversity of the training dataset and enhances the generalization ability of the model. This enables the model to better adapt to detection tasks under different lighting, angles, and background conditions, improving the robustness of the detection.

[0068] 5. Early stopping mechanism to optimize the training efficiency

[0069] During the model training process, an early stopping mechanism is introduced to automatically stop the training when the model performance no longer improves, avoiding the consumption of invalid computing resources and the problem of model overfitting. This not only reduces the training cost but also ensures the deployment of the model in the optimal state, further enhancing the detection efficiency and accuracy.

[0070] 6. Deployment on portable devices for real-time detection

[0071] This solution deploys the trained model to portable computing devices (such as laptops or cloud servers). Combining real-time shooting by drones and wireless transmission technology, it can achieve real-time detection of defects in the inner cavity of wind turbine blades. This deployment method is flexible and efficient, suitable for on-site operations, and significantly improves the timeliness and practicality of the detection.

[0072] 7. Defect visualization and recording

[0073] By annotating the defect area mask to the corresponding image and visualizing it, this solution can intuitively present the detection results, facilitating technicians to quickly locate and evaluate the defects. At the same time, the defect area is saved in the form of an image mask, providing reliable data support for subsequent analysis and historical records.

[0074] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for unsupervised defect detection in the inner cavity of a wind turbine blade, characterized in that: include: S1. Use the camera carried by the drone to collect the inner cavity image of the wind turbine blade and build a basic data set for wind turbine blade inner cavity defect detection; S2. Performing data enhancement on the basic data set to obtain an enhanced data set; S3, split the enhanced data set into a training set and a test set; S4. Construct an unsupervised defect detection model for the inner cavity of a wind turbine blade, use the training set and introduce an early stop mechanism to complete the detection model training, and test it with a test set; S5. Deploy the tested detection model to a portable computing device to realize the defect detection of the inner cavity of the wind turbine blade.

2. The unsupervised defect detection method for the inner cavity of a wind turbine blade according to claim 1, characterized in that: The basic data set in step S1 includes: a normal inner cavity surface image and a defective inner cavity surface image.

3. The unsupervised defect detection method for the inner cavity of a wind turbine blade according to claim 1, characterized in that: The data enhancement in step S2 includes: a geometric data enhancement method and a non-geometric data enhancement method.

4. The unsupervised defect detection method for the inner cavity of a wind turbine blade according to claim 1, characterized in that: When performing segmentation in step S3, the training set and the test set are divided according to a preset ratio, wherein the training set is entirely normal lumen surface images, and the test set includes some normal lumen surface images and all defective lumen surface images.

5. The unsupervised defect detection method for the inner cavity of a wind turbine blade according to claim 1, characterized in that: The wind turbine blade inner cavity unsupervised defect detection model described in step S4 adopts the Resnet18 pre-trained model.

6. The unsupervised defect detection method for the inner cavity of a wind turbine blade according to claim 5, characterized in that: During the model training process in step S4, a small sample anomaly detection framework RegAD based on image registration is also introduced.

7. A storage medium, characterized in that: The storage medium stores instructions and data for implementing the unsupervised defect detection method for the inner cavity of a wind turbine blade as described in any one of claims 1 to 6.

8. An unsupervised defect detection device for the inner cavity of a wind turbine blade, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement the unsupervised defect detection method for the inner cavity of a wind turbine blade as described in any one of claims 1 to 6.

Images