Fan blade defect detection method and device, computer equipment, medium and product

Abstract

The application relates to a fan blade defect detection method and device, computer equipment, a medium and a product. The method comprises the following steps: acquiring a thermal imaging initial image of a fan blade; the thermal imaging initial image comprises initial channel images corresponding to a brightness channel and a chroma channel respectively; performing contrast enhancement processing on the initial channel image under the brightness channel to obtain an enhanced channel image of the fan blade under the brightness channel; obtaining a thermal imaging enhanced image of the fan blade by fusing the enhanced channel image and the initial channel image under the chroma channel; using a feature extraction network containing a multi-dimensional attention component to dynamically fuse attention information corresponding to each dimension of the thermal imaging enhanced image to obtain image features of the thermal imaging enhanced image; and determining a fan blade defect detection result based on the image features. The method can improve the accuracy of fan blade defect detection.

Description

Technical Field

[0001] This application relates to the field of power equipment testing technology, and in particular to a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for detecting defects in wind turbine blades. Background Technology

[0002] With the accelerating global energy transition and the demand for sustainable development, wind power generation has become one of the core pillars of the clean energy system with the greatest potential for large-scale development and strategic value due to its abundant resources, mature technology, environmental friendliness, and significantly improved economic efficiency.

[0003] As the most critical wind energy capture component of wind turbines, wind turbine blades are usually subjected to the effects of wind, sand, salt spray and alternating loads during service, which can lead to cracking and corrosion on the blade surface, as well as delamination and structural damage inside. In order to improve the safety and stability of wind farm operation and reduce the probability of accidents, it is necessary to regularly inspect the wind turbine blades for defects.

[0004] Currently, the commonly used methods for inspecting wind turbine blades are manual inspection and drone visual inspection. However, manual inspection has problems such as high dependence on human labor, high operational risks, and low detection accuracy. Drone inspection also has a significant risk of missing small cracks and internal damage. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for detecting wind turbine blade defects that can improve the accuracy of detection, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a method for detecting defects in wind turbine blades, including:

[0007] Acquire an initial thermal imaging image of the wind turbine blades; the initial thermal imaging image includes initial channel images corresponding to the luminance channel and the chrominance channel, respectively;

[0008] The initial channel image under the brightness channel is subjected to contrast enhancement processing to obtain the enhanced channel image of the wind turbine blade under the brightness channel.

[0009] By fusing the enhanced channel image and the initial channel image under the chroma channel, a thermal imaging enhanced image of the wind turbine blade is obtained;

[0010] Using a feature extraction network containing multi-dimensional attention components, the attention information corresponding to the thermal imaging enhancement image in each dimension is dynamically fused to obtain the image features of the thermal imaging enhancement image.

[0011] Based on the image features, the defect detection results of the wind turbine blades are determined.

[0012] Secondly, this application also provides a wind turbine blade defect detection device, comprising:

[0013] An initial image acquisition module is used to acquire an initial thermal imaging image of the wind turbine blades; the initial thermal imaging image includes initial channel images corresponding to the luminance channel and the chrominance channel, respectively.

[0014] A contrast enhancement module is used to perform contrast enhancement processing on the initial channel image under the brightness channel to obtain an enhanced channel image of the wind turbine blade under the brightness channel.

[0015] The image fusion module is used to obtain the thermal imaging enhancement image of the wind turbine blade by fusing the enhanced channel image and the initial channel image under the chroma channel;

[0016] The feature extraction module is used to dynamically fuse the attention information corresponding to each dimension of the thermal imaging enhancement image using a feature extraction network containing a multi-dimensional attention component, so as to obtain the image features of the thermal imaging enhancement image.

[0017] The defect detection module is used to determine the defect detection result of the wind turbine blade based on the image features.

[0018] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0019] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the method described above.

[0020] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of the method described above.

[0021] The aforementioned wind turbine blade defect detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product can acquire an initial thermal imaging image of the wind turbine blade. This initial thermal imaging image includes initial channel images corresponding to the luminance and chrominance channels, respectively. Subsequently, the initial channel image in the luminance channel undergoes contrast enhancement processing to obtain an enhanced channel image of the wind turbine blade in the luminance channel. The enhanced channel image and the initial channel images in the chrominance channel are then fused to obtain an enhanced thermal imaging image of the wind turbine blade. A feature extraction network containing a multi-dimensional attention component dynamically fuses the attention information corresponding to each dimension of the enhanced thermal imaging image to obtain image features. Finally, the defect detection result of the wind turbine blade is determined based on these image features. On the one hand, this method utilizes the pseudo-color characteristics of thermal imaging to optimize contrast in the image enhancement part, making the defect area more clearly visible, effectively reducing the impact of environmental noise, and improving the identifiability of defects. On the other hand, by dynamically fusing the attention information of the enhanced image in each dimension through a multi-dimensional attention component, it can adaptively enhance defect features of different sizes and locations, effectively improving the accuracy of defect localization and classification, especially improving the detection accuracy of fine cracks and internal damage. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a diagram illustrating the application environment of a wind turbine blade defect detection method in one embodiment.

[0024] Figure 2 This is a flowchart illustrating a wind turbine blade defect detection method in one embodiment;

[0025] Figure 3 This is a schematic diagram of the process for acquiring an initial thermal imaging image of a wind turbine blade in one embodiment;

[0026] Figure 4 This is a flowchart illustrating the process of performing image enhancement processing on the original channel image under each color channel to obtain the enhanced image of the wind turbine blade under the color channel in one embodiment.

[0027] Figure 5 This is a flowchart illustrating a wind turbine blade defect detection method in another embodiment;

[0028] Figure 6This is a schematic diagram illustrating the specific enhancement effect of an image enhancement operation in one embodiment;

[0029] Figure 7 This is a schematic diagram of the structure of a dynamic detection head in one embodiment;

[0030] Figure 8 This is a schematic diagram comparing the detection performance of the improved YOLOv8 in one embodiment;

[0031] Figure 9 This is a structural block diagram of a wind turbine blade defect detection device in one embodiment;

[0032] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] The wind turbine blade defect detection method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, the blade defect detection platform 102 communicates with the server 104 via a network. A data storage system can store the data that the blade defect detection platform 102 needs to process. The data storage system can be integrated onto the blade defect detection platform 102 or placed on the cloud or other network servers. The blade defect detection platform 102 can obtain initial thermal imaging images of the wind turbine blades from the server 104. These initial thermal imaging images include initial channel images corresponding to the luminance and chrominance channels, respectively. Contrast enhancement processing is performed on the initial channel image in the luminance channel to obtain an enhanced channel image of the wind turbine blade in the luminance channel. By fusing the enhanced channel image and the initial channel image in the chrominance channel, an enhanced thermal imaging image of the wind turbine blade is obtained. Subsequently, the blade defect detection platform 102 can use a feature extraction network containing multi-dimensional attention components to dynamically fuse the attention information corresponding to each dimension of the enhanced thermal imaging image to obtain image features of the enhanced thermal imaging image. Based on these image features, the defect detection result of the wind turbine blade is determined. The blade defect detection platform 102 can be integrated into a terminal or server. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be management terminals used by power management personnel, such as drones and inspection equipment. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc., used by power management personnel. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0035] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting defects in wind turbine blades is provided, which is then applied to... Figure 1 The following steps are used as an example to illustrate the blade defect detection platform 102. Among them:

[0036] S202, acquire the initial thermal imaging image of the wind turbine blades. The initial thermal imaging image includes the initial channel images corresponding to the luminance channel and the chrominance channel, respectively.

[0037] The initial thermal image is a two-dimensional matrix of data generated after the wind turbine blades are detected using an infrared imaging device. The value of each pixel is not the actual color, but represents the temperature or infrared radiation intensity of the corresponding point on the object's surface. It is usually observed by the human eye in the form of a pseudo-color image.

[0038] Luminance and chroma channels are two mathematical components used to decompose the information in a color image. Originating from color models in color science, the purpose of distinguishing between luminance and chroma channels is to separate the light and dark information perceived by the human eye from the pure color information, allowing for independent processing. The luminance channel is essentially the grayscale intensity component of the image, containing information about the image's brightness, contrast, and structural details, and is highly correlated with the human eye's sensitivity to changes in lighting. The chroma channel, on the other hand, is a component describing color attributes, defining the properties of the color itself, and is independent of the color's brightness.

[0039] For example, the blade defect detection platform can acquire an initial thermal imaging image of the wind turbine blade, wherein the initial thermal imaging image includes initial channel images corresponding to the luminance channel and the chrominance channel, respectively.

[0040] In one embodiment, the blade defect detection platform can directly obtain the initial thermal imaging image of the wind turbine blade from the server.

[0041] In one embodiment, the blade defect detection platform can acquire raw data generated by a thermal imaging device after detecting wind turbine blades, process the raw data, and obtain an initial thermal imaging image of the wind turbine blades. Data processing can refer to decomposing or calculating the luminance component matrix and chromaticity component matrix of the raw data under a specific color model through mathematical transformations.

[0042] S204, perform contrast enhancement processing on the initial channel image under the luminance channel to obtain the enhanced channel image of the wind turbine blade under the luminance channel.

[0043] Among them, the initial channel image under the brightness channel refers to the grayscale image of the wind turbine blade. Each element in the image represents the original brightness of the corresponding point on the surface of the wind turbine blade in thermal radiation. This brightness value is directly related to the infrared radiation intensity of the corresponding point captured by the thermal imaging device.

[0044] Contrast enhancement processing refers to the nonlinear mapping process of pixel values ​​in the initial channel image. It is mainly used to make global or local adjustments to the pixel values ​​of the initial channel image under the luminance channel in order to enhance the dynamic range of the image grayscale level or the difference between light and dark in local areas, thereby improving the readability of the image.

[0045] For example, after obtaining the initial thermal imaging image, the blade defect detection platform can perform contrast enhancement processing on the initial channel image in the brightness channel to obtain the enhanced channel image of the wind turbine blade in the brightness channel.

[0046] In one embodiment, the blade defect detection platform is equipped with a contrast enhancement model. After obtaining the initial channel image in the brightness channel, the initial channel image in the brightness channel can be input into the contrast enhancement model for contrast enhancement processing to obtain the enhanced channel image of the wind turbine blade in the brightness channel.

[0047] In one embodiment, the blade defect detection platform can use homomorphic filtering or wavelet transform to enhance the initial channel image in the brightness channel, resulting in an enhanced channel image in the brightness channel. Homomorphic filtering separates the illumination and reflection components through frequency domain processing, effectively suppressing uneven illumination, while wavelet transform enhances defect details through multi-resolution analysis.

[0048] In one embodiment, the blade defect detection platform can perform contrast adaptive histogram equalization on the initial channel image in the brightness channel to obtain an enhanced channel image of the wind turbine blade in the brightness channel.

[0049] For example, the blade defect detection platform can divide the initial channel image in the brightness channel into blocks according to a preset block size to obtain image blocks. For each image block, the brightness distribution of all pixels within the block is statistically analyzed to obtain a histogram. To prevent noise from being excessively amplified in uniform areas, a cropping limit can be preset to specify the maximum number of pixels allowed at any gray level in the histogram. The platform iterates through each gray level of the histogram according to the cropping limit, cropping any gray levels exceeding the limit. The total number of cropped pixels is collected, and the cropped pixels are evenly redistributed to all gray levels. This process is repeated until the count for each gray level does not exceed the cropping limit, resulting in the target histogram. Using a standard histogram equalization formula, the cumulative distribution function of the corresponding image block is calculated based on the target histogram. Bilinear interpolation is then performed on the initial channel image based on each cumulative distribution function to obtain the enhanced channel image. In other words, the new brightness value of each pixel in the initial channel image is determined by the cumulative distribution function of the centers of its four nearest neighboring image blocks. By fusing the transformations of adjacent blocks through bilinear interpolation, abrupt transitions can be effectively avoided.

[0050] In one embodiment, the formula for contrast-limited adaptive histogram equalization of the luminance channel Y(x,y) can be as follows:

[0051] .

[0052] Among them, Y enhanced For incremental channel images, CLAHE() represents an adaptive histogram equalization operation on the brightness channel Y(x,y). c∈[1.0, 3.0] is used to limit contrast amplification, I s Used to define the size of a local region.

[0053] S206. By fusing the enhanced channel image and the initial channel image under the chroma channel, a thermal imaging enhanced image of the wind turbine blade is obtained.

[0054] The initial channel image in the chroma channel is pure color information extracted from the original thermal image, defining only the color attributes of each pixel, such as reddish or bluish tint. The initial channel image in the chroma channel preserves the original pseudo-color mapping relationship of the temperature value assignment, providing color context.

[0055] The process of fusing the enhanced channel image and the initial channel image under the chroma channel refers to the process of recombinating the separately processed luminance component image and chroma component image according to the color model rules. The thermal imaging enhanced image of the wind turbine blade is a color thermal imaging image generated by recombination, which can be obtained by performing an inverse color space transformation on the enhanced channel image and the initial channel image under the chroma channel.

[0056] For example, after obtaining the fused enhancement channel image, the blade defect detection platform can fuse the enhancement channel image and the initial channel image under the chroma channel to obtain a thermal imaging enhancement image of the wind turbine blade. The color distribution of the thermal imaging enhancement image is similar to that of the original thermal image. Figure 1 However, the temperature contrast in the image is higher, and the thermal anomaly areas are more prominent in both color and brightness.

[0057] S208 uses a feature extraction network containing multi-dimensional attention components to dynamically fuse the attention information corresponding to each dimension of the thermal imaging enhancement image to obtain the image features of the thermal imaging enhancement image.

[0058] Among them, the feature extraction network containing multidimensional attention components is a deep learning network for image feature extraction. Unlike ordinary deep learning networks, this feature extraction network contains multidimensional attention components.

[0059] The multidimensional attention component is a substructure within the feature extraction network. It calculates the importance distribution of input features across multiple dimensions in parallel, generating an attention weight map. These attention weights are then fed back to the original features, enabling adaptive reconstruction and enhancement of the feature representation. By setting the multidimensional attention component, the feature extraction network can selectively focus on specific aspects. The specific dimensions of focus for the multidimensional attention component can be determined by the user based on their actual needs, and may include, but are not limited to, channel, spatial, scale, temporal, and frequency domain dimensions. The attention information corresponding to each dimension is calculated by the multidimensional attention component across different dimensions, representing the distribution of information importance in that dimension.

[0060] For example, the blade defect detection platform can use a feature extraction network containing multi-dimensional attention components to determine the attention information corresponding to each dimension of the thermal imaging enhancement image. Then, it dynamically fuses the attention information corresponding to each dimension of the thermal imaging to calibrate and modulate the thermal imaging enhancement image, enhance important information, suppress secondary information, and finally generate image features with higher information density and stronger dimensional correlation.

[0061] In one embodiment, the attention information includes an attention weight map. After the feature extraction network receives the thermal imaging enhancement image, it can generate an intermediate feature map through some network layers. The intermediate feature map will be simultaneously fed into the multi-dimensional attention component to output the attention weight map of each dimension.

[0062] In one embodiment, dynamically fusing the attention information corresponding to each dimension of thermal imaging refers to applying the attention weights of different dimensions to the intermediate feature map. The dynamic nature of the fusion is reflected in the fact that for different input images, due to the different content of their feature maps, the network will generate completely different attention weight maps, thereby achieving adaptive feature enhancement and suppression related to image content.

[0063] S210, based on image features, determines the defect detection results of wind turbine blades.

[0064] Among them, the defect detection result is the judgment result made after detecting defects in the wind turbine blades based on image features. It is used to reflect information such as whether there are defects in the wind turbine blades, the type of defects, their location, and their extent.

[0065] For example, the blade defect detection platform can perform defect detection on wind turbine blades based on the image features of the obtained thermal imaging enhancement image, and determine the defect detection result of the wind turbine blades.

[0066] In one embodiment, the blade defect detection platform can use the detection network to which the feature extraction network belongs to perform defect detection on the wind turbine blades based on the image features of the thermal imaging enhanced image, and determine the defect detection result of the wind turbine blades.

[0067] In one embodiment, the defect detection network can be a pre-trained detection network with Faster R-CNN or YOLOv5 as the detection core. Faster R-CNN achieves high-precision detection through a region proposal network, while YOLOv5 can balance detection speed and accuracy.

[0068] In one embodiment, the defect detection network can be a YOLOv8 model incorporating a feature extraction network with multidimensional attention components. In this case, the feature extraction network with multidimensional attention components replaces the original detection head in YOLOv8. By deploying attention mechanisms on different dimensions of the features, the representational power of the object detection head is significantly improved without increasing computational overhead.

[0069] In one embodiment, the loss function in the defect detection network can be a generalized cross-union loss function or a distance cross-union loss function, which can be used to evaluate the similarity between the predicted box and the ground truth box, aiming to overcome the limitations of the traditional cross-union loss function and improve the model training effect.

[0070] In one embodiment, the loss function in the defect detection network can be an improved cross-union ratio (CURRR) loss function, which can optimize the performance of bounding box regression in object detection tasks. The improved CURRR loss function introduces a dynamic non-monotonic focusing mechanism, dynamically adjusting the weights of the loss function based on the quality of the samples, particularly for medium-quality samples. This reduces the interference of low-quality samples on model training and improves detection accuracy. The core idea is to evaluate the quality of samples by assessing the deviation of anchor boxes from other samples in the feature space, and then dynamically adjust the weights of each component of the loss function based on this assessment. This allows the model to focus more on samples of moderate quality that play a crucial role in improving model performance.

[0071] In one embodiment, to enhance the generalization ability of the defect detection network, a YOLOv8 model optimized by iteratively training image data of different leaves can be pre-trained to adapt to diverse leaf types and shooting conditions.

[0072] In the aforementioned wind turbine blade defect detection method, an initial thermal imaging image of the wind turbine blade can be acquired. This initial thermal imaging image includes initial channel images corresponding to the luminance and chrominance channels, respectively. Then, contrast enhancement processing is performed on the initial channel image in the luminance channel to obtain an enhanced channel image of the wind turbine blade in the luminance channel. The enhanced channel image and the initial channel image in the chrominance channel are fused to obtain an enhanced thermal imaging image of the wind turbine blade. A feature extraction network containing a multi-dimensional attention component is used to dynamically fuse the attention information corresponding to each dimension of the enhanced thermal imaging image to obtain image features. Finally, the defect detection result of the wind turbine blade is determined based on these image features. This method, on the one hand, utilizes the pseudo-color characteristics of thermal imaging to optimize contrast in the image enhancement part, making the defect area more clearly visible, effectively reducing the influence of environmental noise, and improving the identifiability of defects. On the other hand, by dynamically fusing the attention information of the enhanced image in each dimension through a multi-dimensional attention component, it can adaptively enhance defect features of different sizes and locations, effectively improving the accuracy of defect localization and classification, especially improving the detection accuracy of fine cracks and internal damage.

[0073] In one exemplary embodiment, such as Figure 3 As shown, S202, acquiring the initial thermal imaging image of the wind turbine blades includes:

[0074] S302, acquire the original thermal imaging image of the wind turbine blades in RGB space.

[0075] Among them, the RGB color space is an additive color model, which mainly defines colors by specifying the brightness or intensity of the three primary color channels: red, green, and blue.

[0076] The original thermal image of the wind turbine blades in RGB space refers to the image information directly output by the infrared imaging device after infrared detection of the wind turbine blades. It is a color image encoded using the standard red-green-blue model. Each pixel in the original thermal image contains the intensity values ​​of three channels: R, G, and B. R represents the intensity value of the red channel, G represents the intensity value of the green channel, and B represents the intensity value of the blue channel. This triplet identifies the position of the temperature corresponding to the pixel on the preset color bar.

[0077] A raw thermal image comprises raw channel images for each of its multiple color channels. These color channels are the red, green, and blue channels. Each raw channel image is a separate two-dimensional scalar matrix obtained by decomposing the raw thermal image along the channel dimension. Each raw channel image represents the intensity distribution of the corresponding primary color component in the raw thermal image. For example, each pixel value in the red channel raw image represents the intensity of the red component, each pixel value in the green channel raw image represents the intensity of the green component, and each pixel value in the blue channel raw image represents the intensity of the blue component.

[0078] For example, the blade defect detection platform can acquire raw thermal images of wind turbine blades in RGB space.

[0079] In one embodiment, the blade defect detection platform can directly obtain the original thermal imaging image of the wind turbine blade in RGB space from the server.

[0080] In one embodiment, the blade defect detection platform can communicate directly with the infrared imaging device to control the infrared imaging device to perform infrared monitoring and scanning of the wind turbine blades, thereby obtaining the original thermal imaging image of the wind turbine blades in RGB space.

[0081] S304, for each color channel, perform image enhancement processing on the original channel image under the color channel to obtain the enhanced image of the wind turbine blade under the color channel.

[0082] Image enhancement processing refers to applying enhancement algorithms to the original channel images for each color channel to simulate the human eye's sensitivity to details at different scales. Understandably, the enhancement algorithms used for different original channel images can be the same or different.

[0083] For example, the blade defect detection platform can perform image enhancement processing on the original channel image under each color channel to obtain an enhanced image of the wind turbine blade under the color channel. By independently enhancing each RGB channel, information can be extracted and optimized from temperature ranges with different color codes. For example, enhancing the red channel can make high-temperature defects more prominent, while enhancing the blue channel can make low-temperature defects or background contrast more obvious.

[0084] In one embodiment, image enhancement operations may include contrast stretching, histogram equalization, filtering and denoising, etc.

[0085] S306 combines the enhanced images of each color channel to obtain the RGB enhanced image of the wind turbine blade in the RGB space.

[0086] Here, "combination" refers to the operation of merging the enhanced images of each color channel obtained separately according to the agreed channel order to form an RGB image array containing complete color information.

[0087] The RGB enhanced image of wind turbine blades in RGB space refers to the standard RGB three-channel color image produced by superimposing the enhanced images of each color channel in the channel dimension.

[0088] For example, after obtaining the enhanced images of each color channel, the blade defect detection platform can superimpose the enhanced images on the channel dimension according to the agreed channel order to obtain the RGB enhanced image of the wind turbine blade in the RGB space.

[0089] S308 converts the RGB enhanced image to YUV space to obtain the initial thermal imaging image of the wind turbine blades.

[0090] The YUV color space is a color coding system that separates color information into luminance and chrominance components. This aligns with the human visual system's greater sensitivity to changes in brightness than to changes in color, making it a widely used color model in digital image and video processing. The Y component in the YUV space represents the image's luminance information, containing all details and contours, and is a single-channel grayscale image. The U and V components together represent the image's color information, describing the difference between blue and luminance, and the difference between red and luminance, respectively, and are two single-channel images.

[0091] For example, after obtaining the RGB enhanced image of the wind turbine blade in the RGB space, the blade defect detection platform can use a preset conversion formula to linearly transform the color value of each pixel from the RGB space to the YUV space, thereby generating a three-channel image represented in the YUV color space, and thus obtaining the initial thermal imaging image of the wind turbine blade.

[0092] In one embodiment, the blade defect detection platform can calculate the YUV value corresponding to the RGB value of each pixel in the RGB enhanced image, thereby producing a new three-channel image represented in the YUV color space.

[0093] In the above embodiments, by enhancing the original channel image under each color channel and then combining the enhanced images under each color channel, the RGB enhanced image of the wind turbine blade is obtained. This optimizes the contrast and detail of each color component, reduces the impact of noise, and converts the RGB enhanced image to YUV space, which facilitates more refined processing of the brightness channel separately and reduces the risk of color distortion.

[0094] In one embodiment, such as Figure 4As shown in step S304, for each color channel, image enhancement processing is performed on the original channel image under the color channel to obtain the enhanced image of the wind turbine blade under the color channel, including:

[0095] S402 uses multiple Gaussian filters of different scales for each color channel to perform image enhancement processing on the original channel image under the color channel, and obtains the sub-image of the wind turbine blade corresponding to each scale under the color channel.

[0096] The Gaussian filter is a linear smoothing filter based on the Gaussian function, primarily used to suppress noise, smooth images, and preserve edge and detail features as much as possible. Its core principle is to blur the target pixel and its neighborhood pixels using a weighted average, with the weights determined by the Gaussian function. Gaussian filters of different sizes have different scale parameters, i.e., standard deviation. The standard deviation defines the width of the Gaussian function, determining the range of its smoothing intensity. The larger the scale parameter of each filter, the wider its spatial range of action, the stronger its smoothing ability, and the larger the image structure it can preserve.

[0097] Image enhancement processing refers to the process of performing two-dimensional convolution operations on the original channel image in parallel with multiple Gaussian kernels of different sizes, with each convolution generating a smoothed sub-image of the original channel image at the corresponding scale.

[0098] Each sub-image at each scale is an output image obtained by performing multi-scale linear decomposition on the original channel image. The original channel image is convolved with each Gaussian filter of a different size to obtain the sub-image at each scale. Each sub-image filters out more refined spatial rating information than the current filter scale, while retaining structural information at that scale and coarser scales.

[0099] For example, the blade defect detection platform can use multiple Gaussian filters of different sizes for each color channel to enhance the original channel image of the color channel, thereby obtaining the sub-images of the wind turbine blades at each scale under the color channel.

[0100] S404, combine the sub-images to obtain the original enhanced image of the wind turbine blades in the color channel.

[0101] Among them, the original enhanced image of the wind turbine blade in the color channel refers to the single image output by taking multiple sub-images of different scales in the same color channel as input, calculating them through preset fusion rules, and finally outputting an enhanced version of the color channel data after multi-scale fusion.

[0102] For example, after obtaining sub-images at each scale, the blade defect detection platform can combine the sub-images to obtain the original enhanced image of the wind turbine blade in the color channel.

[0103] In one embodiment, the blade defect detection platform is equipped with a combined model. After obtaining the sub-images at each scale, for each scale, the sub-images at that scale can be input into the combined model to obtain the original enhanced image of the wind turbine blade in the color channel.

[0104] In one embodiment, each color channel is configured with its own preset combination weight. After obtaining each sub-image under the color channel, the sub-images can be combined according to the preset combination weight to obtain the original enhanced image under that color channel.

[0105] In one embodiment, the blade defect detection platform is equipped with an enhancement formula that can directly perform image enhancement processing on the original channel image and combined processing of each sub-image to obtain the final original enhanced image. The formula is as follows:

[0106] .

[0107] Where I(x, y) refers to the input image, which in this embodiment is the original channel image, and S(x, y) is the output image, which in this embodiment is the original enhanced image. G i (x, y) is the Gaussian kernel at the i-th scale, w i Let N be the weight of the i-th scale, and N be the number of scales.

[0108] S406 performs a color restoration operation on the original enhanced image to obtain an enhanced image of the wind turbine blades in the color channel.

[0109] Color restoration refers to applying a color correction transformation to an image that has undergone nonlinear enhancement to compensate for color distortion that may have been introduced during the enhancement process. This helps restore the naturalness and consistency of image colors. Color restoration operations may include, but are not limited to, linear mapping based on the statistics of the original image, histogram matching, color gain adjustment, and the calculation of color restoration factors in a multi-scale framework.

[0110] Among them, the enhanced image of the wind turbine blade in the color channel refers to the final single-channel image obtained after color restoration and correction of the original enhanced image. It not only retains the details and contrast improvement brought by multi-scale enhancement, but also makes reasonable adjustments to the color performance, providing high-quality image components for subsequent cross-channel synthesis.

[0111] For example, the blade defect detection platform can perform color restoration on the original enhanced image to obtain an enhanced image of the wind turbine blade in the color channel.

[0112] In one embodiment, a color restoration factor for each pixel can be calculated and applied to the original enhanced image. The color restoration factor can be constructed based on the relative intensities of each channel of the original image to preserve color proportions.

[0113] In one embodiment, the blade defect detection platform is equipped with a color restoration function, which calculates an enhanced image of the wind turbine blade in the color channel. The formula for the color restoration function is shown below:

[0114] .

[0115] Where S(x, y) is the output image, which in this embodiment is an enhanced image of the wind turbine blades in the color channel. chorm (x, y) represents the chromaticity components of the image, and α is the weighting coefficient.

[0116] In the above embodiments, by using multiple Gaussian filters of different sizes to perform image enhancement processing on the original channel images under the color channels, comprehensive information capture from local details to global structure can be achieved. Secondly, the sub-images obtained after enhancement processing are combined to obtain the original enhanced image under the color channels, which can enhance details while optimizing overall contrast. Finally, color restoration is performed on the original enhanced image to obtain the final enhanced image under the color channels, making the image color representation more natural and consistent, and possessing higher feature discrimination ability.

[0117] In one embodiment, S202, acquiring the original thermal imaging image of the wind turbine blade in RGB space may include: using non-destructive testing technology to perform defect detection on the wind turbine blade, determining the defect area of ​​the wind turbine blade, and acquiring the original thermal imaging image of the defect area in RGB space.

[0118] Non-destructive testing (NDT) technology refers to detection techniques that indirectly infer or assess the internal integrity of a material by detecting abnormal signals generated by discontinuities in the internal structure when a physical field propagates through the material, without damaging the structural integrity or functionality of the wind turbine blades. For example, NDT techniques may rely on acoustic emission or ultrasonic waves.

[0119] In one embodiment, acoustic emission detection is a detection technology that uses the transient elastic wave signals actively released by defects in a material during stress or temperature changes to achieve real-time location and severity assessment of defects. Designers can deploy multiple high-sensitivity acoustic emission sensor arrays at key locations on wind turbine blades. When the blade is under load, if there are active defects inside, they will release weak stress waves. By detecting the stress wave signals, the defects can be located and their severity assessed in real time.

[0120] In one embodiment, ultrasonic testing involves emitting high-frequency ultrasonic pulses into the material and receiving the echoes reflected from the defect interface or the bottom surface of the material. By analyzing the arrival time, amplitude, and waveform characteristics of the echoes, the location, size, and nature of the defects can be detected. Designers can use an ultrasonic probe to couple ultrasonic waves onto the blade surface. The sound waves can propagate in the composite material and are emitted when they encounter defects or the bottom surface. The probe can receive the echoes, scan and record the echo information of the entire area to generate a scanned image. The blade defect detection platform can then use this scanned image to detect the location, size, and nature of the defects.

[0121] Among them, the defect area refers to the area where defects are initially detected in the wind turbine blades through non-destructive testing.

[0122] For example, when it is necessary to perform defect detection on wind turbine blades, the blade defect detection platform can communicate with non-destructive testing equipment to control the non-destructive testing equipment to perform non-destructive testing on the wind turbine blades and obtain non-destructive testing data of the wind turbine blades. The blade defect detection platform can determine the defect area of ​​the wind turbine blades based on the non-destructive testing data, and then control the infrared imaging equipment to perform thermal imaging scanning on the defect area of ​​the wind turbine blades to obtain the original thermal imaging image of the defect area in RGB space.

[0123] In the above embodiments, by first performing non-destructive testing on the wind turbine blades to determine the defect area, and then acquiring the original thermal imaging image of the defect area in RGB space, the defect location of the wind turbine blades can be coarsely located before precise defect detection based on the original thermal imaging image. Then, the original thermal imaging image can be acquired based on the coarse location result, which can effectively improve the detection efficiency and targeting of wind turbine blade defects, and improve the accuracy and reliability of defect detection.

[0124] In one embodiment, the multidimensional attention component may include a cascaded scale-aware attention component, a spatial-aware attention component, and a task-aware attention component. S206, using a feature extraction network containing the multidimensional attention component, the attention information corresponding to each dimension of the thermal imaging enhancement image is dynamically fused to obtain the image features of the thermal imaging enhancement image. This includes: using a feature extraction network containing the multidimensional attention component to dynamically fuse the attention information corresponding to scale, space, and task-aware aspects of the thermal imaging enhancement image to obtain the image features of the thermal imaging enhancement image.

[0125] The scale-aware attention component is a network module that dynamically assigns weights based on the importance of the input feature map at different scales. It calculates a set of weights by analyzing the contribution of features at different scales to the current task, and then uses these weights to perform weighted fusion or selection of multi-scale features in the channel or spatial dimensions. The scale-aware attention component is deployed only at the hierarchical level, learning the relative importance of different semantic levels to appropriately enhance features according to the scale of the object.

[0126] The spatially aware attention component is a network module that dynamically assigns weights based on the importance of the input feature map at different two-dimensional spatial locations. It analyzes each location of the feature map in both the height and width dimensions, generating a two-dimensional attention weight matrix. Each value in the matrix represents the importance score of the corresponding empty Agin location for the current task, thereby enhancing the feature response of key regions and suppressing background or irrelevant regions. The spatially aware attention component is deployed in the spatial dimension (height × width) to learn coherent and discriminative representations of spatial locations.

[0127] The task-aware attention component is a network module that can dynamically adjust feature representations according to the specific needs of downstream tasks. Task-aware attention components are typically deployed on channels and can guide different feature channels to favor different tasks based on the responses of different convolutional kernels to objects, such as classification, bounding box regression, and center point learning.

[0128] For example, the blade defect detection platform can use a feature extraction network containing multi-dimensional attention components to dynamically fuse the attention information corresponding to the thermal imaging enhancement image under scale, space and task perception respectively, to obtain the image features of the thermal imaging enhancement image.

[0129] In one embodiment, after the thermal imaging enhancement image is input into the feature extraction network, the feature extraction network first extracts multi-scale features. The scale-aware attention component dynamically calculates the weights of each scale feature based on the image content and fuses them to obtain a scale-optimized feature map. Subsequently, the scale-optimized feature map is fed into the spatial awareness attention component, which calculates a spatial weight map and multiplies it with the scale-optimized feature map to highlight important spatial locations and suppress secondary spatial locations, resulting in a spatially focused feature map. The spatially focused feature map is then fed into the task-aware attention component, which generates task-specific attention weights based on the specific task the network needs to perform, and finally modulates the spatially focused feature map to obtain the image features of the thermal imaging enhancement image.

[0130] In the above embodiments, the cascaded multidimensional attention mechanism can achieve adaptive selection at scale, intelligent focusing in space, and precise modulation in task, enabling the feature extraction network to extract highly pure and discriminative image features from complex thermal imaging enhanced images, thereby improving the accuracy and computational efficiency of wind turbine blade defect detection.

[0131] In one embodiment, the defect detection result is based on a defect detection network. The wind turbine blade defect detection method may further include: determining incremental training samples containing an initial thermal imaging image if the defect detection result is incorrect; and incrementally training the defect detection network using the incremental training samples to obtain an updated defect detection network.

[0132] Among them, an error in the defect detection result can refer to the fact that when a defect detection network is used to detect defects in the input image features, the output judgment is incorrect. For example, it could be a missed defect, a false detection, a misclassification, or an inaccurate location of the defect.

[0133] In one embodiment, the determination of the defect detection result can be made by the testing personnel themselves by comparing the output defect detection result with the actual operation and maintenance result.

[0134] Incremental training samples refer to the training data set containing the initial thermal imaging images that caused the model to make detection errors. The increment emphasizes that the training samples supplement the original training set, rather than replace it.

[0135] The updated defect detection network refers to the new detection network whose parameters have been optimized and adjusted after incremental training with incremental training samples.

[0136] For example, the blade defect detection platform can feed back the defect detection results to the inspectors. The inspectors can then perform maintenance checks on the wind turbine blades based on the defect detection results and provide feedback on the accuracy of the defect detection results. If the blade defect detection platform determines that the defect detection results are incorrect based on the information provided by the inspectors, it can determine incremental training samples containing the initial thermal imaging images. The incremental training samples are then used to incrementally train the defect detection network to obtain an updated defect detection network. Subsequently, the updated defect detection network can be used to detect defects in the wind turbine blades, thereby improving the accuracy of defect detection.

[0137] In one embodiment, the current defect detection network is iteratively trained using incremental training samples containing initial thermal imaging images. The loss function in the defect detection network penalizes the model's errors on the incremental training samples. Once the training converges, a defect detection network containing new network parameters is obtained.

[0138] In the above embodiments, during the actual detection process, diverse data can be continuously injected based on the actual defect detection results for model iterative training, enabling the defect detection network to learn from errors and continuously evolve in actual deployment, effectively enhancing the long-term reliability of the defect detection network and its adaptability to different working conditions.

[0139] In one embodiment, such as Figure 5 As shown, a method for detecting defects in wind turbine blades is provided, which specifically includes the following steps:

[0140] S501, acquire the original thermal imaging image of the wind turbine blades in RGB space; the original thermal imaging image includes the original channel images of each of the multiple color channels.

[0141] The raw thermal image is essentially composed of illumination and reflection components. It also has multiple color channels: red, green, and blue.

[0142] S502 uses multiple Gaussian filters of different scales to perform image enhancement processing on the original channel image under each color channel, thereby obtaining the sub-images of the wind turbine blades at each scale under each color channel.

[0143] S503 combines the sub-images to obtain the original enhanced image of the wind turbine blades in the color channel.

[0144] The process of obtaining the original enhanced image of the wind turbine blades in the color channel can essentially be considered as performing convolution and logarithmic transformation on the original image using multiple Gaussian filters of different scales on three separate color channels to simulate the human eye's sensitivity to details at different scales.

[0145] S504 performs a color restoration operation on the original enhanced image to obtain an enhanced image of the wind turbine blades in the color channel.

[0146] The purpose of color restoration is to preserve or restore color information that may have been lost during image processing.

[0147] S505 combines the enhanced images of each color channel to obtain the RGB enhanced image of the wind turbine blade in the RGB space.

[0148] S506 converts the RGB enhanced image to YUV space to obtain the initial thermal imaging image of the wind turbine blades.

[0149] S507 performs contrast enhancement processing on the initial channel image in the luminance channel to obtain the enhanced channel image of the wind turbine blades in the luminance channel.

[0150] For example, the contrast enhancement process can be a contrast-limited adaptive histogram equalization process applied to the luminance channel Y(x,y).

[0151] S508 obtains a thermal imaging enhancement image of the wind turbine blades by fusing the enhanced channel image and the initial channel image under the chroma channel.

[0152] In one embodiment, the specific enhancement effect achieved through the above image enhancement operation can be as follows: Figure 6 As shown. Figure 6 Specifically, the image includes the original thermal imaging image of the wind turbine blade, the RGB enhanced image, and the thermal imaging enhanced image. It can be seen that the RGB enhanced image enhances details while retaining the naturalness of the colors. The thermal imaging enhanced image further optimizes the brightness and contrast to address the false-color characteristics of thermal imaging. The defects at the bottom of the image, which are internal defects of the blade, can be clearly seen, making it more suitable for industrial inspection scenarios.

[0153] S509 uses a feature extraction network containing multi-dimensional attention components to dynamically fuse the attention information corresponding to each dimension of the thermal imaging enhancement image to obtain the image features of the thermal imaging enhancement image.

[0154] In one implementation, the defect detection network is based on YOLOv8, with improvements in two areas: head replacement and loss function optimization. Head replacement involves replacing the original YOLOv8 model's detection head with a dynamic head. This dynamic head mechanism integrates scale-aware, spatial-aware, and task-aware attention mechanisms within a unified framework. By deploying attention mechanisms on different dimensions of the features, the dynamic head significantly improves the representational power of the object detection head without increasing computational overhead. Its core lies in integrating scale, spatial, and task-aware attention, utilizing self-attention across the three dimensions of the input features. The dynamic detection head structure is as follows: Figure 7 As shown.

[0155] Loss function optimization refers to replacing the traditional Cross-Union Ratio (CUNR) loss function in the original model with an improved CUNR-based loss function, aiming to optimize the performance of bounding box regression in object detection tasks. The improved CUNR-based loss function introduces a dynamic non-monotonic focusing mechanism, dynamically adjusting the weights of the loss function based on sample quality, particularly for medium-quality samples. This reduces the interference of low-quality samples on model training and improves detection accuracy. The core idea of ​​the improved CUNR-based loss function is to evaluate sample quality by assessing the degree of deviation of anchor boxes from other samples in the feature space, and then dynamically adjust the weights of each component of the loss function based on this assessment. This allows the model to focus more on samples of moderate quality that play a crucial role in improving model performance.

[0156] In one embodiment, the detection performance of conventional YOLOv8 is compared with that of the improved YOLOv8 in this embodiment. Figure 8 As shown, it can be seen that whether YOLOv8 is used alone for detection or YOLOv8 is used in conjunction with image enhancement processing, the improved YOLOv8 has a stronger defect detection effect, especially in the detection of small cracks and internal damage.

[0157] S510 determines the defect detection results of wind turbine blades based on image features.

[0158] S511, in the event of an error in the defect detection result, determine the incremental training samples that include the initial thermal imaging image.

[0159] S512 uses incremental training samples to incrementally train the defect detection network, resulting in an updated defect detection network.

[0160] The aforementioned wind turbine blade defects and detection methods can specifically include the following beneficial effects:

[0161] First, it can significantly improve detection accuracy. In existing technologies, traditional image processing methods (such as Canny edge detection) are easily affected by lighting, shadows, and surface reflections, leading to missed detections of minute cracks. The image enhancement method in this embodiment decomposes the image using Retinex theory and optimizes brightness and contrast for the pseudo-color characteristics of thermal imaging, making the defect area more clearly visible. This reduces the impact of environmental noise, improves the identifiability of defects, and thus significantly reduces the false negative and false positive rates. Meanwhile, existing deep learning models (such as Mask R-CNN) suffer from blurred boundary segmentation in complex backgrounds and are sensitive to low-contrast defects. In this embodiment, the scale-aware, spatial-aware, and task-aware attention mechanisms of the dynamic detection head adaptively enhance defect features at different scales and locations. At the same time, it dynamically adjusts sample weights based on an improved loss function of cross-union ratio, focusing on medium-quality samples and optimizing bounding box regression. This improves the accuracy of defect localization and classification, especially in the detection of small cracks and internal damage.

[0162] Secondly, it effectively enhances the generalization ability of the defect detection method. Existing models often overfit due to limited training data and have poor adaptability across different models and scenarios. This embodiment iteratively injects data from diverse blade types and shooting conditions, enabling the model to continuously learn defect characteristics under different environments, reducing the model's dependence on specific datasets. This improves the system's stability under varying inspection conditions (such as different weather and angles) and makes it suitable for various wind turbine blade models.

[0163] Third, it optimizes detection real-time performance and computational efficiency. Existing deep learning methods, such as Mask R-CNN, have high computational complexity when processing high-resolution images, making them difficult to meet the real-time detection requirements of UAVs. This embodiment accelerates inference while maintaining accuracy by leveraging the lightweight design of YOLOv8 and the fact that the dynamic head has no additional computational overhead. This makes the system more suitable for embedded devices or online inspection scenarios, achieving efficient real-time detection.

[0164] Fourth, it reduces the reliance on labeled data. Existing deep learning models require a large amount of pixel-level labeled data, which is costly. In this embodiment, image quality is improved through enhancement algorithms, making defect features more obvious and reducing the model's sensitivity to the accuracy and quantity of annotations. Combined with iterative training, the system can achieve good performance with only a small amount of labeled data, reducing the burden of manual annotation.

[0165] Fifth, it enables effective detection of internal defects in wind turbine blades. Thermal imaging technology reflects the internal condition of an object by capturing the temperature distribution on its surface. When defects (such as delamination or bubbles) exist inside the wind turbine blade, they alter the heat conduction characteristics of that area, resulting in abnormal hot or cold spots in thermal imaging. Therefore, the thermal imaging image itself contains information about the internal structure. This embodiment combines thermal imaging with an improved deep learning model to achieve non-contact, rapid detection of internal defects in wind turbine blades. This represents a fundamental difference and advantage over existing mainstream visual inspection methods, filling the gap in efficient and visual diagnosis of the internal health status of wind turbine blades. It enables earlier detection of potential safety hazards and avoids structural damage to the blade caused by the spread of internal damage.

[0166] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0167] Based on the same inventive concept, this application also provides a wind turbine blade defect detection device for implementing the wind turbine blade defect detection method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the wind turbine blade defect detection device provided below can be found in the limitations of the wind turbine blade defect detection method described above, and will not be repeated here.

[0168] In one exemplary embodiment, such as Figure 9 As shown, a wind turbine blade defect detection device 900 is provided, comprising: an initial image acquisition module 901, a contrast enhancement module 902, an image fusion module 903, a feature extraction module 904, and a defect detection module 905, wherein:

[0169] The initial image acquisition module 901 is used to acquire the initial thermal imaging image of the wind turbine blades; the initial thermal imaging image includes the initial channel images corresponding to the luminance channel and the chrominance channel, respectively.

[0170] The contrast enhancement module 902 is used to perform contrast enhancement processing on the initial channel image under the brightness channel to obtain the enhanced channel image of the wind turbine blade under the brightness channel.

[0171] The image fusion module 903 is used to obtain a thermal imaging enhancement image of the wind turbine blade by fusing the enhanced channel image and the initial channel image under the chroma channel.

[0172] The feature extraction module 904 is used to dynamically fuse the attention information corresponding to each dimension of the thermal imaging enhancement image using a feature extraction network containing a multi-dimensional attention component, so as to obtain the image features of the thermal imaging enhancement image.

[0173] The defect detection module 905 is used to determine the defect detection results of the wind turbine blades based on image features.

[0174] In one embodiment, the initial image acquisition module 901 is used to: acquire a raw thermal imaging image of the wind turbine blade in RGB space; the raw thermal imaging image includes raw channel images of each of multiple color channels; for each color channel, perform image enhancement processing on the raw channel images of the color channel to obtain an enhanced image of the wind turbine blade in the color channel; combine the enhanced images of each color channel to obtain an RGB enhanced image of the wind turbine blade in RGB space; and convert the RGB enhanced image to YUV space to obtain the initial thermal imaging image of the wind turbine blade.

[0175] In one embodiment, the image fusion module 903 is used to: for each color channel, use multiple Gaussian filters of different scales to perform image enhancement processing on the original channel image under the color channel to obtain the sub-images of the wind turbine blades corresponding to each scale under the color channel; combine the sub-images to obtain the original enhanced image of the wind turbine blades under the color channel; and perform color restoration operation on the original enhanced image to obtain the enhanced image of the wind turbine blades under the color channel.

[0176] In one embodiment, the initial image acquisition module 901 is used to perform defect detection on the wind turbine blades using non-destructive testing technology, determine the defect area of ​​the wind turbine blades, and acquire the original thermal imaging image of the defect area in RGB space.

[0177] In one embodiment, the multidimensional attention component includes a cascaded scale-aware attention component, a spatial-aware attention component, and a task-aware attention component. The feature extraction module 904 is used to: dynamically fuse the attention information corresponding to the thermal imaging enhancement image under scale, spatial, and task-aware conditions using a feature extraction network containing the multidimensional attention component, to obtain the image features of the thermal imaging enhancement image.

[0178] In one embodiment, the defect detection result is realized based on a defect detection network, and the wind turbine blade defect detection device 900 further includes:

[0179] The incremental training sample determination module is used to determine incremental training samples containing the initial thermal imaging image when the defect detection results are incorrect.

[0180] The incremental training module is used to incrementally train the defect detection network using incremental training samples to obtain an updated defect detection network.

[0181] Each module in the aforementioned wind turbine blade defect detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0182] In one exemplary embodiment, a computer device is provided, which may be a server integrating a blade defect detection platform, and its internal structure diagram may be as follows. Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores relevant data for wind turbine blade defect detection methods. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a wind turbine blade defect detection method.

[0183] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0184] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0185] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0186] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0187] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0188] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0189] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0190] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting defects in wind turbine blades, characterized in that, The method includes: Acquire raw thermal imaging images of wind turbine blades in RGB space; the raw thermal imaging images include raw channel images of each of multiple color channels. For each of the color channels, image enhancement processing is performed on the original channel image under the color channel to obtain the enhanced image of the wind turbine blade under the color channel; By combining the enhanced images of each of the aforementioned color channels, an RGB enhanced image of the wind turbine blade in the RGB space is obtained; The RGB enhanced image is converted to YUV space to obtain the initial thermal imaging image of the wind turbine blade; the initial thermal imaging image includes initial channel images corresponding to the luminance channel and the chrominance channel, respectively. The initial channel image under the brightness channel is subjected to contrast enhancement processing to obtain the enhanced channel image of the wind turbine blade under the brightness channel. By fusing the enhanced channel image and the initial channel image under the chroma channel, a thermal imaging enhanced image of the wind turbine blade is obtained; A feature extraction network containing multidimensional attention components is used to dynamically fuse the attention information corresponding to the thermal imaging enhancement image under scale, space and task perception respectively to obtain the image features of the thermal imaging enhancement image; the multidimensional attention components include cascaded scale perception attention components, space perception attention components and task perception attention components. Based on the image features, the defect detection results of the wind turbine blades are determined.

2. The method according to claim 1, characterized in that, The step of performing image enhancement processing on the original channel image under each color channel to obtain the enhanced image of the wind turbine blade under the color channel includes: For each color channel, multiple Gaussian filters of different scales are used to perform image enhancement processing on the original channel image under the color channel to obtain the sub-image of the wind turbine blade corresponding to each scale under the color channel; By combining the sub-images, the original enhanced image of the wind turbine blade under the color channel is obtained; A color restoration operation is performed on the original enhanced image to obtain an enhanced image of the wind turbine blade in the color channel.

3. The method according to claim 1, characterized in that, The acquisition of the raw thermal imaging image of the wind turbine blades in RGB space includes: Non-destructive testing technology is used to detect defects in the wind turbine blades and determine the defect areas of the wind turbine blades. Obtain the original thermal imaging image of the defective region in RGB space.

4. The method according to claim 1, characterized in that, The defect detection results are achieved based on a defect detection network; the method further includes: In the event that the defect detection result is incorrect, an incremental training sample containing the initial thermal imaging image is determined. The defect detection network is incrementally trained using the incremental training samples to obtain an updated defect detection network.

5. A device for detecting defects in wind turbine blades, characterized in that, The device includes: An initial image acquisition module is used to acquire a raw thermal imaging image of the wind turbine blade in RGB space; the raw thermal imaging image includes raw channel images of each of multiple color channels; for each color channel, image enhancement processing is performed on the raw channel image of that color channel to obtain an enhanced image of the wind turbine blade in that color channel; the enhanced images of each color channel are combined to obtain an RGB enhanced image of the wind turbine blade in RGB space; the RGB enhanced image is converted to YUV space to obtain the initial thermal imaging image of the wind turbine blade; the initial thermal imaging image includes initial channel images corresponding to the luminance channel and chrominance channel respectively; A contrast enhancement module is used to perform contrast enhancement processing on the initial channel image under the brightness channel to obtain an enhanced channel image of the wind turbine blade under the brightness channel. The image fusion module is used to obtain the thermal imaging enhancement image of the wind turbine blade by fusing the enhanced channel image and the initial channel image under the chroma channel; The feature extraction module is used to dynamically fuse the attention information corresponding to the thermal imaging enhancement image under scale, space and task perception respectively using a feature extraction network containing multi-dimensional attention components to obtain the image features of the thermal imaging enhancement image; the multi-dimensional attention components include cascaded scale perception attention components, space perception attention components and task perception attention components. The defect detection module is used to determine the defect detection result of the wind turbine blade based on the image features.

6. The apparatus according to claim 5, characterized in that, The image fusion module is used for: For each color channel, multiple Gaussian filters of different scales are used to perform image enhancement processing on the original channel image under the color channel to obtain the sub-image of the wind turbine blade corresponding to each scale under the color channel; the sub-images are combined to obtain the original enhanced image of the wind turbine blade under the color channel; color restoration is performed on the original enhanced image to obtain the enhanced image of the wind turbine blade under the color channel.

7. The apparatus according to claim 5, characterized in that, The initial image acquisition module is used for: Non-destructive testing technology is used to detect defects in the wind turbine blades, and the defect areas of the wind turbine blades are determined; the original thermal imaging image of the defect areas in RGB space is obtained.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images