A method for classifying and identifying surface damage of wind turbine blades based on width learning

Abstract

The present invention relates to the field of computer vision technology, and specifically discloses a method for classifying and identifying surface damage of a fan blade based on width learning, comprising the following steps: collecting fan blade images; preprocessing the collected fan blade images, and denoising the fan blade images by a non-local mean filtering algorithm; extracting feature vectors from the denoised fan blade images by a width learning method; inputting the feature vectors of the fan blade images into a classifier to generate a fan blade damage identification model; denoising the fan blade images to be identified by a non-local mean filtering algorithm, and then inputting the denoised fan blade images into the fan blade damage identification model, and outputting the damage identification results of the fan blade images. The present invention has the advantages of horizontal expansion and incremental learning, and has a fast calculation speed. When the model accuracy is not enough, only the enhanced nodes need to be expanded horizontally. When new data is added, only the new data needs to be trained. The operation is simple and the calculation is fast.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision. Background Art

[0002] Scholars and commercial organizations at home and abroad have begun to use computer vision methods to identify blade damage. The main technologies include image processing, machine learning and deep learning. Patent CN112233091A discloses a method for detecting and locating damage in wind turbine blade images, which is characterized by: using an artificially marked wind turbine blade image database, training a cascade strong classifier according to the classic AdaBoostHaar-like algorithm, and building a deep convolutional neural network damage recognition model, traversing the image through a sliding window method, and identifying it through the trained cascade classifier to determine whether the image contains a damaged area, scaling the detected damaged target area to a uniform size and inputting it into the damage recognition model to determine which type of damage the area belongs to; Patent CN107144569A discloses a wind turbine blade fault diagnosis method based on an SVM classification algorithm with selective search segmentation and a deep convolutional neural network as a feature extractor. First, it selectively searches and segments the image to be detected to obtain a candidate area; then, the deep convolutional neural network is trained through the ImageNet image set, and the network structure except the output layer is extracted as a feature extractor; finally, the extracted features of the wind turbine blade image are used to train the SVM classifier to complete the fault diagnosis task; Patent CN1 10261394A proposes an online real-time diagnosis system for wind turbine blade damage, which uses a quad-rotor drone to capture wind turbine blade images in real time and transmits them to a computer system in real time; dynamically constructs a cloud database to store an image library with a useful VGG-19net image classification method, and the computer system is used to obtain an improved VGG-19net image classification method through image library training, and uses the improved VGG-19net image classification method to classify wind turbine blade images received from a quad-rotor drone to obtain wind turbine blade damage diagnosis and damage grade classification results; Patent CN109741320A discloses a wind turbine blade fault detection method based on aerial images, which collects wind turbine blade images through drones, constructs a three-dimensional visualization model of wind turbine blades based on enhanced images, uses a differential method to subtract the original defective image from the reconstructed image to obtain the residual information of its main component and performs an inverse transformation, uses Alexnet to classify defective blade images into categories such as peeling, cracks, and trachoma, and uses a deep network classification model to classify them into different defect grades according to the degree of damage.

[0003] For the damage recognition algorithm of wind turbine blade images, the deep neural network uses the collected image data as training data to determine the structural parameters of the deep neural network. The method of solidifying the network structural parameters cannot show that the network structure performance has achieved the optimal effect. Although the damage detection method based on deep neural network has a good performance in the classification and recognition of wind turbine blade damage images, it requires a large amount of data to improve the accuracy of damage classification and recognition. At the same time, due to the large number of parameters involved, the complex structure, the large time consumption, and the inability to update the newly generated image data in real time, there are certain limitations. At present, there are few pre-processing operations on the original images of wind turbine blades in the damage detection technology based on deep neural network. In image analysis, the quality of the image directly affects the design and accuracy of the damage recognition algorithm. Therefore, pre-processing is still required before image analysis. Among them, the commonly used mean filter denoising method cannot protect the image details well. While denoising the image, it also destroys the image details and loses the image feature information. There is a tendency to blur the image, and it cannot remove the noise points well. It is basically powerless against salt and pepper noise. In view of the above-mentioned problems, it is necessary to study and design a new type of wind blade surface damage classification and identification method based on width learning to overcome the problems existing in the existing wind blade surface damage classification and identification methods. Summary of the invention

[0004] In order to solve the above-mentioned problems existing in the existing methods for classifying and identifying surface damage of wind turbine blades, the present invention provides a method for classifying and identifying surface damage of wind turbine blades based on width learning.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a method for classifying and identifying surface damage of wind turbine blades based on width learning, comprising the following steps:

[0006] S1. Collecting images of wind turbine blades;

[0007] S2. Preprocessing the collected fan blade image, and denoising the fan blade image by using a non-local mean filtering algorithm;

[0008] S3. Extracting feature vectors from the denoised fan blade image by using a width learning method;

[0009] S4. Inputting the feature vector of the fan blade image into the classifier to generate a fan blade damage recognition model;

[0010] S5. After performing denoising processing on the fan blade image to be identified using a non-local mean filtering algorithm, the image is input into the fan blade damage identification model, and a damage identification result of the fan blade image is output.

[0011] Preferably, in step S1, a video of wind blades is acquired by drone aerial photography, an image of the wind blades in a partial frame of the wind blade video is extracted, the extracted image is detected and the area where the wind blades are located is located.

[0012] Preferably, the step S2 comprises:

[0013] S201. Adjust the fan blade image to m rows * n columns, take point P (x, y) in the image as the point to be filtered, take point P as the center, take ksize rows * ksize columns as the neighborhood block, take ssize rows * ssize columns as the search window, and expand the edge of the fan blade image;

[0014] S202. Clip half_ssize rows from the upper edge, half_ssize rows from the lower edge, half_ssize columns from the left edge, and half_ssize columns from the right edge of the fan blade image after edge expansion to obtain image A;

[0015] S203. offset the position of the current calculation point and the point to be filtered in the search window by (x_offset, y_offset) to obtain image B with the same rows and columns as image A;

[0016] S204. Calculate the pixel difference C of the same coordinate points, pixel difference C = AB;

[0017] S205. Calculate the square of the pixel difference of each pair of points with the same coordinates, the square graph of C = C*C;

[0018] S206. Calculate the integral graph of the position shift of the square graph D by a block parallel method;

[0019] S207. After the integral image is obtained, the neighborhood block of the integral image is obtained according to the corresponding position of the point to be filtered on the integral image, and the mean square error similarity between the neighborhood block and the neighborhood block of the integral image is calculated;

[0020] S208. Calculate weights by mean square error similarity;

[0021] S209. Calculate the filter value and output the denoised fan blade image.

[0022] Preferably, in step S201, the upper edge of the wind blade image is expanded by half_ssize+half_ksize rows, the lower edge is expanded by half_ssize+half_ksize rows, the left edge is expanded by half_ssize+half_ksize columns, and the right edge is expanded by half_ssize+half_ksize columns.

[0023] Preferably, the step S206 includes:

[0024] S2061. Divide each row of data of the wind turbine blade image of m rows and n columns into data blocks of 4 data each, start m*(n / 4) threads, each thread calculates the prefix sum of 4 data in a small block, and all threads calculate in parallel;

[0025] S2062. Start m threads, each thread corresponds to a row of data, and calculate the prefix sum of the last data in all small blocks in the row;

[0026] S2063. Start m*(n / 4) threads, each thread corresponds to a small block, and each row starts from the second small block. Except for the last number, add the number in each small block to the last number of its previous small block to get the prefix sum of all rows;

[0027] S2064. Divide each column of data into blocks of 4 data, start (m / 4)*n threads, each thread calculates the prefix sum of 4 data in a small block, and all threads calculate in parallel;

[0028] S2065. Start n threads, each thread corresponds to a column of data, and calculate the prefix sum of the last data in all small blocks in the column;

[0029] S2066. Start (m / 4)*n threads, each thread corresponds to a small block, and each column starts from the second small block. Except for the last number, the number in each small block is added to the last number of its previous small block to obtain the integral graph D of the original image.

[0030] Preferably, in step S207, the coordinates of point P in the integral graph are P(x+half_ksize, y+half_ksize), and a neighborhood block of the integral graph centered on point P is obtained in the integral graph, and the coordinates of the four vertices of the neighborhood block of the integral graph are P1(x, y), P2(x+half_ksize, y-half_ksize), P3(x-half_ksize, y+half_ksize), and P4(x+half_ksize, y+half_ksize); according to the pixel values ​​IP1, IP2, IP3, and IP4 of points P1, P2, P3, and P4 on the integral graph, the calculation formula of the integral graph is:

[0031] IP=IP4-IP2-IP3+IP1

[0032] The mean square error similarity between the neighborhood block and the integral graph neighborhood block is as follows:

[0033]

[0034] Where: MSE is the mean square error similarity.

[0035] Preferably, in step S208, the formula for calculating the mean square error similarity weight is:

[0036]

[0037] Where: sum is the sum of Gaussian coefficients, MES(P, P') is the neighborhood block similarity between point P and any point P' in the search window, and h is the filter coefficient that controls the degree of filter smoothness.

[0038] Preferably, in step S209, the filter value of point P is obtained by weighted average of the pixel values ​​of all points in the search window, and the formula is:

[0039] NLmeans(P)=∑W(P,P')*I(P')

[0040] Where: W(P,P') is the Gaussian weight of point P and any point P' in the search window, and I(P') is the integral graph of any point P' in the search window.

[0041] Preferably, in step S3, extracting the feature vector of the fan blade image by the width learning method includes: connecting the initial fan blade image data through a linear mapping function and inputting a weight matrix transformation to obtain a mapping node, and the formula is:

[0042]

[0043] Where: X is the input data for model training, X∈R a×b , R is a set of real numbers, a is the number, b is the dimension, is the activation function of the feature node, is the weight matrix of the i-th feature node, For the input layer to map feature layer and The corresponding bias matrix;

[0044] The mapping node group obtained by n transformations is:

[0045] Z n =[Z1,Z2,Z3,......,Z n ]

[0046] The enhancement node is obtained by transforming the mapping node through nonlinear mapping and activation function. The formula is:

[0047]

[0048] Where: j is the activation function of the enhanced node, is the random connection weight matrix that maps the feature nodes to the enhanced node layer of the jth group, is the j-th group bias matrix;

[0049] The enhanced node group obtained by m transformations is:

[0050] H m =[H1,H2,H3,......,H m ]

[0051] The mapping node and the enhancement node are used together as the input of the width learning neural network to obtain the feature vector of the fan blade image. The formula is:

[0052] Y=[Z n |H m ]W m

[0053] Where: Y is the feature vector of the input fan blade image, [Z n |H m ] matrix is ​​the actual input of the width learning network, W m is the weight matrix connecting the system input to the output layer;

[0054] The characteristic vector of the fan blade image is solved by solving the ridge regression pseudo-inverse method. The formula is:

[0055] W m =(λI+A T A) -1 A T Y

[0056] Where A=[Z n |H m ], λ is the regularization coefficient, I is the unit matrix, and T is the symbol for matrix transposition in mathematical operations.

[0057] Preferably, in step S5, the damage identification results corresponding to the data to be identified include four types: normal blades, surface cracks, surface damage and corrosion.

[0058] The invention discloses a method for classifying and identifying surface damage of a fan blade based on width learning. In the preprocessing stage, an optimization algorithm based on non-local mean filtering is used to process the damage image of the fan blade, which makes full use of the redundant information in the image. While effectively removing noise, the detail features of the damage image of the fan blade can be retained to the greatest extent. At the same time, a general parallel computing architecture is used to optimize the non-local mean filtering algorithm based on the integral graph. When the weighted sum of pixels is calculated, the general parallel computing architecture is used to parallelly calculate the integral graph to accelerate the calculation, so that the calculation speed of the preprocessing is greatly improved. At the same time, a single hidden layer does not have a complex coupling relationship between layers, has a simple structure, has the advantages of horizontal expansion and incremental learning, and has a fast calculation speed. When the model accuracy is not enough, it does not require a large amount of data to retrain the model, nor does it increase the number of layers and adjust the calculation amount of parameters like a deep network. It does not need to use gradient descent to update the weights. It only needs to horizontally expand the enhanced nodes or feature nodes to adjust the model, so as to achieve the improvement of the model performance. When new data is added, the model does not need to retrain all data, but only needs to train the new data. The operation is simple and the calculation is fast. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of a method for classifying and identifying surface damage of a wind turbine blade based on width learning according to an embodiment of the present invention;

[0060] Figure 2 is a preprocessing flow chart based on a non-local mean filtering algorithm according to an embodiment of the present invention;

[0061] Figure 3 This is an integral graph calculation process based on the block parallel idea in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0063] A method for classifying and identifying surface damage of a fan blade based on width learning of the present invention is as follows Figure 1 As shown, the following steps are included:

[0064] S1. Collecting images of wind turbine blades;

[0065] The wind blades can be photographed by a drone to obtain a video of the wind blades, extract images of the wind blades in some frames of the video of the wind blades, detect the extracted images and locate the area where the wind blades are located.

[0066] S2. Preprocessing the collected fan blade image, and denoising the fan blade image by using a non-local mean filtering algorithm;

[0067] like Figure 2 As shown, step S2 includes:

[0068] S201. Adjust the fan blade image to a predetermined size of m rows and n columns, preferably 256×256, take point P (x, y) in the image as the point to be filtered, take point P as the center, take ksize rows and ksize columns as the neighborhood block, take ssize rows and ssize columns as the search window, and expand the edge of the fan blade image;

[0069] Generally speaking, considering the complexity of the algorithm, the rectangular window size ssize*ssize is generally 21x21, and the neighborhood block size ksize*ksize can be 7x7. However, in practice, it is often necessary to select appropriate parameters based on the noise. When the standard deviation σ of the Gaussian noise is larger, in order to improve the robustness of the algorithm, the block area needs to be increased, and the search area also needs to be increased. Since ksize and ssize are both odd numbers, their half values ​​are usually passed in: half_ksize and half_ssize. Therefore, the size of the neighborhood block is (2*half_ksize+1)*(2*half_ksize+1), and the size of the search window is (2*half_ssize+1)*(2*half_ssize+1). Since each point to be filtered in the original image requires a complete search window and multiple neighborhood blocks, the image edge must be expanded before non-local mean filtering. Preferably, the upper edge of the fan blade image is expanded by half_ssize+half_ksize rows, the lower edge is expanded by half_ssize+half_ksize rows, the left edge is expanded by half_ssize+half_ksize columns, and the right edge is expanded by half_ssize+half_ksize columns;

[0070] S202. Clip half_ssize rows from the upper edge, half_ssize rows from the lower edge, half_ssize columns from the left edge, and half_ssize columns from the right edge of the fan blade image after edge expansion to obtain image A;

[0071] S203. offset the position of the current calculation point and the point to be filtered in the search window by (x_offset, y_offset) to obtain image B with the same rows and columns as image A;

[0072] Among them, -half_ssize≤x_offset≤half_ssize, -half_ssize≤y_offset≤half_ssize;

[0073] S204. Calculate the pixel difference C of the same coordinate points, pixel difference C = AB;

[0074] S205. Calculate the square of the pixel difference of each pair of points with the same coordinates, the square of C D = C * C;

[0075] S206. Calculate the integral graph of the position shift of the square graph D by a block parallel method;

[0076] Use the idea of ​​block parallelism to calculate the integral graph of the position offset of the square graph D, such as Figure 3 There are 6 main steps shown:

[0077] S2061. Divide each row of the wind turbine blade image of m rows and n columns into a data block of 4 data, start m*(n / 4) threads, each thread calculates the prefix sum of 4 data in a small block, and all threads calculate in parallel;

[0078] S2062. Start m threads, each thread corresponds to a row of data, and calculate the prefix sum of the last data in all small blocks in the row;

[0079] S2063. Start m*(n / 4) threads, each thread corresponds to a small block, and each row starts from the second small block. Except for the last number, add the number in each small block to the last number of its previous small block to get the prefix sum of all rows;

[0080] S2064. Divide each column of data into blocks of 4 data, start (m / 4)*n threads, each thread calculates the prefix sum of 4 data in a small block, and all threads calculate in parallel;

[0081] S2065. Start n threads, each thread corresponds to a column of data, and calculate the prefix sum of the last data in all small blocks in the column;

[0082] S2066. Start (m / 4)*n threads, each thread corresponds to a small block, and each column starts from the second small block. Except for the last number, add the number in each small block to the last number of its previous small block to obtain the integral graph D of the original image;

[0083] S207. After the integral image is obtained, the neighborhood block of the integral image is obtained according to the corresponding position of the point to be filtered on the integral image, and the mean square error similarity between the neighborhood block and the neighborhood block of the integral image is calculated;

[0084] The rows and columns of the neighborhood blocks are the same, both with m rows and n columns. The mean square error (MSE) calculation formula of two neighborhood blocks is:

[0085]

[0086] Wherein, MSE(A, B) is the mean square error of the two neighborhood blocks of image A and image B, A(i, j) is the pixel value of point (i, j) in the neighborhood block of A, and B(i, j) is the pixel value of the same position point in the neighborhood block of B;

[0087] The coordinates of point P in the integral graph are P(x+half_ksize, y+half_ksize). The integral graph neighborhood block centered on point P is obtained in the integral graph. The coordinates of the four vertices of the integral graph neighborhood block are P1(x, y), P2(x+half_ksize, y-half_ksize), P3(x-half_ksize, y+half_ksize), and P4(x+half_ksize, y+half_ksize). According to the pixel values ​​IP1, IP2, IP3, and IP4 of points P1, P2, P3, and P4 on the integral graph, the calculation formula of the integral graph is:

[0088] IP=IP4-IP2-IP3+IP1

[0089] The mean square error similarity between the neighborhood block and the integral graph neighborhood block is as follows:

[0090]

[0091] Where: MSE is the mean square error similarity.

[0092] S208. Calculate weights by mean square error similarity;

[0093] The formula for calculating the mean square error similarity weight is:

[0094]

[0095] Where: sum is the sum of Gaussian coefficients, MES(P, P') is the neighborhood block similarity between point P and any point P' in the search window, h is the filter coefficient that controls the smoothness of the filter. The larger the filter coefficient that controls the smoothness of the filter, the better the denoising effect, but the more blurred the image. On the contrary, the smaller h is, the worse the denoising effect is, but the distortion after denoising is smaller.

[0096] S209. Calculate the filter value and output the denoised fan blade image;

[0097] The filter value of point P is obtained by weighted average of the pixel values ​​of all points in the search window. The formula is:

[0098] NLmeans(P)=∑W(P,P')*I(P')

[0099] Where: W(P,P') is the Gaussian weight of point P and any point P' in the search window, which is calculated by the MSE similarity of the neighborhood blocks of the two points, and I(P') is the integral image of any point P' in the search window.

[0100] S3. Extracting feature vectors from the denoised fan blade image by using a width learning method;

[0101] Extracting the feature vector of the fan blade image by the width learning method includes: connecting the initial fan blade image data through a linear mapping function and inputting a weight matrix transformation to obtain a mapping node, and the mapping feature is recorded as the i-th group of mapping features, i=1, 2, ..., n, obtained through linear mapping and activation function, and the formula is:

[0102]

[0103] Where: X is the input data for model training, X∈R a×b , R is a set of real numbers, a is the number, b is the dimension, is the activation function of the feature node, is the weight matrix of the i-th feature node, For the input layer to map feature layer and The corresponding bias matrix, weight The dimension of determines the mapping feature set Z i The number of nodes in ,;

[0104] The mapping node group obtained by n transformations is:

[0105] Z n =[Z1,Z2,Z3,......,Z n ]

[0106] The enhancement node is obtained by transforming the mapping node through nonlinear mapping and activation function. The formula is:

[0107]

[0108] Where: j is the activation function of the enhanced node, is the random connection weight matrix that maps the feature nodes to the enhanced node layer of the jth group, is the j-th group bias matrix;

[0109] The enhanced node group obtained by m transformations is:

[0110] H m =[H1,H2,H3,......,H m ]

[0111] The mapping node and the enhancement node are used together as the input of the width learning neural network to obtain the feature vector of the fan blade image. The formula is:

[0112] Y=[Z n |H m ]W m

[0113] Where: Y is the feature vector of the input fan blade image, [Z n |H m ] matrix is ​​the actual input of the width learning network, W m is the weight matrix connecting the system input to the output layer;

[0114] The characteristic vector of the fan blade image is solved by solving the ridge regression pseudo-inverse method. The formula is:

[0115] W m =(λI+A T A) -1 A T Y

[0116] Where A=[Z n |H m ], λ is the regularization coefficient, I is the unit matrix, and T is the symbol for matrix transposition in mathematical operations;

[0117] S4. Inputting the feature vector of the fan blade image into the classifier to generate a fan blade damage recognition model;

[0118] S5. After the fan blade image to be identified is subjected to denoising processing using a non-local mean filtering algorithm, the image is input into a fan blade damage identification model, and a damage identification result of the fan blade image is output;

[0119] The damage identification results corresponding to the output data to be identified may include four types: normal blades, surface cracks, surface damage and corrosion.

[0120] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. A method for classifying and identifying surface damage of wind turbine blades based on width learning, characterized in that: The steps include: S1. Collecting images of wind turbine blades; S2. Preprocessing the collected fan blade image, and denoising the fan blade image by a non-local mean filtering algorithm; the step S2 includes: S201. Adjust the fan blade image to m rows * n columns, take point P (x, y) in the image as the point to be filtered, take point P as the center, take ksize rows * ksize columns as the neighborhood block, take ssize rows * ssize columns as the search window, and expand the edge of the fan blade image; S202. Clip half_ssize rows from the upper edge, half_ssize rows from the lower edge, half_ssize columns from the left edge, and half_ssize columns from the right edge of the fan blade image after edge expansion to obtain image A; S203. offset the position of the current calculation point and the point to be filtered in the search window by (x_offset, y_offset) to obtain image B with the same rows and columns as image A; S204. Calculate the pixel difference C of the same coordinate points, pixel difference C = AB; S205. Calculate the square of the pixel difference of each pair of points with the same coordinates, the square graph of C = C*C; S206. Calculate the integral graph of the position offset of the square graph D by a block parallel method; the step S206 comprises: S2061. Divide each row of the wind turbine blade image of m rows and n columns into a data block of 4 data, start m*(n / 4) threads, each thread calculates the prefix sum of 4 data in a small block, and all threads calculate in parallel; S2062. Start m threads, each thread corresponds to a row of data, and calculate the prefix sum of the last data in all small blocks in the row; S2063. Start m*(n / 4) threads, each thread corresponds to a small block, and each row starts from the second small block. Except for the last number, add the number in each small block to the last number of its previous small block to get the prefix sum of all rows; S2064. Divide each column of data into blocks of 4 data, start (m / 4)*n threads, each thread calculates the prefix sum of 4 data in a small block, and all threads calculate in parallel; S2065. Start n threads, each thread corresponds to a column of data, and calculate the prefix sum of the last data in all small blocks in the column; S2066. Start (m / 4)*n threads, each thread corresponds to a small block, and each column starts from the second small block. Except for the last number, add the number in each small block to the last number of its previous small block to obtain the integral graph D of the original image; S207. After the integral image is obtained, the neighborhood block of the integral image is obtained according to the corresponding position of the point to be filtered on the integral image, and the mean square error similarity between the neighborhood block and the neighborhood block of the integral image is calculated; S208. Calculate weights by mean square error similarity; S209. Calculate the filter value and output the denoised fan blade image; S3. Extracting feature vectors from the denoised fan blade image by using a width learning method; S4. Inputting the feature vector of the fan blade image into the classifier to generate a fan blade damage recognition model; S5. After performing denoising processing on the fan blade image to be identified using a non-local mean filtering algorithm, the image is input into the fan blade damage identification model, and a damage identification result of the fan blade image is output.

2. According to claim 1, a method for classifying and identifying wind turbine blade surface damage based on width learning is characterized in that: In the step S1, a video of the wind blades is acquired by aerial photography using a drone, an image of the wind blades in a partial frame of the video of the wind blades is extracted, the extracted image is detected, and the area where the wind blades are located is located.

3. The method for classifying and identifying surface damage of wind turbine blades based on width learning according to claim 1 is characterized in that: In step S201, the upper edge of the wind turbine blade image is expanded by half_ssize+half_ksize rows, the lower edge is expanded by half_ssize+half_ksize rows, the left edge is expanded by half_ssize+half_ksize columns, and the right edge is expanded by half_ssize+half_ksize columns.

4. The method for classifying and identifying wind turbine blade surface damage based on width learning according to claim 1 is characterized in that: In step S207, the coordinates of point P in the integral graph are P(x+half_ksize, y+half_ksize), and the integral graph neighborhood block centered on point P is obtained in the integral graph, and the coordinates of the four vertices of the integral graph neighborhood block are P1(x, y), P2(x+half_ksize, y-half_ksize), P3(x-half_ksize, y+half_ksize), and P4(x+half_ksize, y+half_ksize). According to the pixel values ​​IP1, IP2, IP3, and IP4 of points P1, P2, P3, and P4 on the integral graph, the calculation formula of the integral graph is: IP=IP4-IP2-IP3+IP1 The mean square error similarity between the neighborhood block and the integral graph neighborhood block is as follows: Where: MSE is the mean square error similarity.

5. The method for classifying and identifying wind turbine blade surface damage based on width learning according to claim 1 is characterized in that: In step S208, the formula for calculating the mean square error similarity weight is: Where: sum is the sum of Gaussian coefficients, MES(P, P') is the neighborhood block similarity between point P and any point P' in the search window, and h is the filter coefficient that controls the degree of filter smoothness.

6. The method for classifying and identifying wind turbine blade surface damage based on width learning according to claim 1 is characterized in that: In step S209, the filter value of point P is obtained by weighted average of the pixel values ​​of all points in the search window, and the formula is: NLmeans(P)=∑W(P,P')*I(P') Where: W(P,P') is the Gaussian weight of point P and any point P' in the search window, and I(P') is the integral graph of any point P' in the search window.

7. The method for classifying and identifying wind turbine blade surface damage based on width learning according to claim 1 is characterized in that: In step S3, extracting the feature vector of the fan blade image by the width learning method includes: connecting the initial fan blade image data through a linear mapping function and inputting a weight matrix transformation to obtain a mapping node, and the formula is: Where: X is the input data for model training, X∈R a×b , R is a set of real numbers, a is the number, b is the dimension, is the activation function of the feature node, is the weight matrix of the i-th feature node, For the input layer to map feature layer and The corresponding bias matrix; The mapping node group obtained by n transformations is: WITH n =[Z1,Z2,Z3,......,Z n ] The enhancement node is obtained by transforming the mapping node through nonlinear mapping and activation function. The formula is: Where: j is the activation function of the enhanced node, is the random connection weight matrix that maps the feature nodes to the enhanced node layer of the jth group, is the j-th group bias matrix; The enhanced node group obtained by m transformations is: H m =[H1,H2,H3,......,H m ] The mapping node and the enhancement node are used together as the input of the width learning neural network to obtain the feature vector of the fan blade image. The formula is: Y=[Z n |H m ]W m Where: Y is the feature vector of the input fan blade image, [Z n |H m ] matrix is ​​the actual input of the width learning network, W m is the weight matrix connecting the system input to the output layer; The characteristic vector of the fan blade image is solved by solving the ridge regression pseudo-inverse method. The formula is: W m =(λI+A T A) -1 A T Y Where A=[Z n |H m ], λ is the regularization coefficient, I is the unit matrix, and T is the symbol for matrix transposition in mathematical operations.

8. The method for classifying and identifying wind turbine blade surface damage based on width learning according to claim 1 is characterized in that: In the step S5, the damage identification results corresponding to the data to be identified are outputted, including four types: normal blades, surface cracks, surface damage and corrosion.

Images