A method for detecting the quality of laser cladding of a water turbine blade

Abstract

The application discloses a kind of water turbine blade laser cladding quality detection method, belong to image processing technical field.The application first executes top hat operation and black hat operation to blade cladding layer gray scale chart respectively, obtains bright part saliency image and dark part saliency image;Then two kinds of images are grid division, by extracting the lateral and longitudinal gray scale difference of adjacent grid, obtain lateral, longitudinal dispersion coefficient, and then construct lateral, longitudinal dispersion matrix, add two to generate dispersion enhancement matrix;At the same time, based on the difference between the gray scale mean of each grid in each row grid and the same row, same column mean, calculate the lateral, longitudinal deviation coefficient and construct the corresponding matrix, add after obtaining deviation enhancement matrix;Finally, the dispersion enhancement matrix and deviation enhancement matrix of two images are processed using multi-fusion neural network, output laser cladding quality score, effectively improve the detection precision.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a method for detecting the quality of laser cladding on turbine blades. Background Technology

[0002] As a core component of hydroelectric power generation equipment, turbine blades are subjected to high-speed water flow erosion, cavitation wear, and alternating loads over long periods. The integrity of their surface directly affects the unit's operating efficiency and safety. Laser cladding technology, with its high coating bonding strength and excellent wear resistance, has become a key method for blade surface repair and strengthening. However, the laser cladding process is affected by many factors, including powder characteristics, energy parameters, and scanning path, and is prone to defects such as overheating, incomplete fusion, porosity, and cracks. If these defects are not detected in time, the cladding layer will be of poor quality, easily corroded, and could lead to serious accidents such as blade breakage.

[0003] Currently, image recognition is used to detect the quality of laser cladding. However, the surface of the laser cladding layer has natural textures, which cause slight differences in grayscale between defects (such as dark spots and bright spots) and normal areas. It is difficult to effectively distinguish them directly from the original image, resulting in low accuracy in laser cladding quality detection. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides a method for detecting the quality of laser cladding on turbine blades, which solves the problem of low accuracy in the detection of laser cladding quality in the prior art.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for detecting the quality of laser cladding on turbine blades, comprising:

[0006] Perform top-hat and black-hat operations on the grayscale image of the blade cladding layer to obtain salient images of bright and dark areas.

[0007] The bright and dark salient images are divided into grids respectively, and the horizontal gray level difference between adjacent grids is extracted to obtain the horizontal discrete coefficient and the vertical gray level difference is extracted to obtain the vertical discrete coefficient, thus obtaining the horizontal discrete matrix and the vertical discrete matrix.

[0008] The discrete enhancement matrix is ​​obtained by adding the horizontal and vertical discrete matrices of the same image.

[0009] In each row of the grid, based on the difference between the average gray value of each grid and the average gray value of the same row and column, the horizontal and vertical deviation coefficients of each grid are obtained, resulting in the horizontal deviation matrix and the vertical deviation matrix.

[0010] The deviation enhancement matrix is ​​obtained by adding the horizontal and vertical deviation matrices of the same image.

[0011] A multi-fusion neural network is used to process the discrete enhancement matrix and the deviation enhancement matrix of the two images to obtain the laser cladding quality score.

[0012] Furthermore, the process of extracting the horizontal and vertical grayscale differences between adjacent grids includes:

[0013] Divide both the highlight salient image and the shadow salient image into N×M grids, where N is the number of rows and M is the number of columns;

[0014] Extract the mean grayscale value of each grid in the bright and dark salient images respectively;

[0015] The horizontal grayscale difference is obtained by the difference in the average grayscale values ​​of adjacent grids in the horizontal direction;

[0016] The vertical grayscale difference is obtained by the difference in the average grayscale values ​​of adjacent grids in the vertical direction.

[0017] Furthermore, the process of obtaining the horizontal grayscale difference includes: in the horizontal direction, subtracting the grayscale mean of the grid at (x+1,y) from the grayscale mean of the grid at (x,y), and taking the absolute value of the subtraction result to obtain the horizontal grayscale difference, where x is the horizontal coordinate and y is the vertical coordinate.

[0018] The process of obtaining the vertical grayscale difference includes: in the vertical direction, subtracting the grayscale mean of the grid at (x,y-1) from the grayscale mean of the grid at (x,y), and taking the absolute value of the subtraction result to obtain the vertical grayscale difference.

[0019] Furthermore, the process of obtaining the horizontal and vertical discrete matrices includes:

[0020] Arrange the horizontal grayscale differences according to their corresponding adjacent grid positions to obtain the horizontal grayscale difference matrix;

[0021] Arrange each vertical gray level difference according to the corresponding adjacent grid positions to obtain the vertical gray level difference matrix;

[0022] For each horizontal gray level difference in the horizontal gray level difference matrix, obtain the horizontal discrete coefficients, and arrange the horizontal discrete coefficients according to the corresponding horizontal gray level difference positions to obtain the horizontal discrete matrix.

[0023] For each vertical gray level difference in the vertical gray level difference matrix, obtain the vertical discrete coefficient. Arrange the vertical discrete coefficients according to the corresponding vertical gray level difference to obtain the vertical discrete matrix.

[0024] Furthermore, the process of obtaining the horizontal discrete coefficients includes: taking each horizontal gray-level difference in the same row of the horizontal gray-level difference matrix as the center, extracting multiple horizontal gray-level differences in the left and right neighboring areas to form a horizontal local data segment, extracting the variance of the horizontal local data segment, and normalizing it to obtain the horizontal discrete coefficient of each horizontal gray-level difference.

[0025] The process of obtaining the vertical dispersion coefficients includes: taking each vertical gray level difference in the same column of the vertical gray level difference matrix as the center, extracting multiple vertical gray level differences in the upper and lower neighboring areas to form a vertical local data segment, extracting the variance of the vertical local data segment, and normalizing it to obtain the vertical dispersion coefficient of each vertical gray level difference.

[0026] Furthermore, the process of obtaining the lateral deviation matrix and the longitudinal deviation matrix includes:

[0027] In each row of the grid, the lateral deviation coefficient of each grid is obtained based on the difference between the average gray value of each grid and the average gray value of the same row. The lateral deviation coefficients are arranged according to the grid position to obtain the lateral deviation matrix.

[0028] In each grid column, the vertical deviation coefficient of each grid is obtained based on the difference between the average gray value of each grid and the average gray value of the same column. The vertical deviation coefficients are then arranged according to the grid position to obtain the vertical deviation matrix.

[0029] Furthermore, the multi-fusion neural network includes: a bright area discrete deviation feature fusion unit, a dark area discrete deviation feature fusion unit, a first pooling layer, a second pooling layer, an adder A1, and a fully connected layer;

[0030] The discrete deviation feature fusion unit for bright areas extracts multi-scale features from the discrete enhancement matrix and deviation enhancement matrix of the image with salient bright areas and then fuses them; the discrete deviation feature fusion unit for dark areas extracts multi-scale features from the discrete enhancement matrix and deviation enhancement matrix of the image with salient dark areas and then fuses them.

[0031] The first pooling layer performs pooling processing on the output of the bright area discrete deviation feature fusion unit; the second pooling layer performs pooling processing on the output of the dark area discrete deviation feature fusion unit.

[0032] Adder A1 is used to add the outputs of the first pooling layer element by element;

[0033] The fully connected layer is used to output a laser cladding quality score based on the output characteristics of adder A1.

[0034] Furthermore, both the bright area discrete deviation feature fusion unit and the dark area discrete deviation feature fusion unit include: a first shallow multi-scale convolution module, a second shallow multi-scale convolution module, and a multiplier M1;

[0035] The first shallow multi-scale convolution module extracts shallow multi-scale features from the discrete enhancement matrix;

[0036] The second shallow multi-scale convolution module extracts shallow multi-scale features from the deviation enhancement matrix;

[0037] Multiplier M1 is used to element-wise multiply the outputs of the first shallow multiscale convolution module and the second shallow multiscale convolution module.

[0038] Furthermore, both the first shallow multi-scale convolution module and the second shallow multi-scale convolution module include: a first convolution block, a second convolution block, a third convolution block, and a Concat layer;

[0039] The first convolutional block, the second convolutional block, and the third convolutional block are connected in sequence;

[0040] The Concat layer is used to concatenate the outputs of the first convolutional block, the second convolutional block, and the third convolutional block.

[0041] The beneficial effects of this invention are as follows:

[0042] 1. This invention uses top-hat and black-hat operations to specifically extract salient bright and dark images, respectively, effectively amplifying the grayscale difference between bright defects (such as overheating and metal spatter) and dark defects (such as incomplete fusion and pores) and normal areas in the laser cladding layer, thus overcoming the defect blurring problem caused by natural textures.

[0043] 2. This invention extracts the grayscale difference in the horizontal and vertical directions by dividing the grid, calculates the discrete coefficient and the deviation coefficient, and constructs the discrete enhancement matrix and the deviation enhancement matrix. It captures defect features from two dimensions: local fluctuation (discrete coefficient) and overall deviation (deviation coefficient). It not only pays attention to the subtle changes of adjacent grids, but also considers the difference between the grid and the whole row and column, avoiding the limitations of a single feature dimension. It can comprehensively cover different types and distributions of defects (such as horizontal cracks, vertical pores, etc.).

[0044] 3. This invention adds the horizontal and vertical matrices of the same image to obtain an enhancement matrix, realizing the complementarity and enhancement of two-dimensional features, making the features of the defect area more prominent; the multi-fusion neural network performs deep fusion processing on the four enhancement matrices of the two images, improving the accuracy of defect recognition. Attached Figure Description

[0045] Figure 1 A flowchart of a method for inspecting the quality of laser cladding on turbine blades;

[0046] Figure 2 A schematic diagram of a high-quality blade cladding layer in grayscale;

[0047] Figure 3This is a schematic diagram of the structure of a multi-fusion neural network;

[0048] Figure 4 This is a schematic diagram of the structure of the first shallow multi-scale convolutional module and the second shallow multi-scale convolutional module. Detailed Implementation

[0049] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0050] like Figure 1 As shown, a method for inspecting the quality of laser cladding on turbine blades includes:

[0051] Perform top-hat and black-hat operations on the grayscale image of the blade cladding layer to obtain salient images of bright and dark areas.

[0052] The bright and dark salient images are divided into grids respectively, and the horizontal gray level difference between adjacent grids is extracted to obtain the horizontal discrete coefficient and the vertical gray level difference is extracted to obtain the vertical discrete coefficient, thus obtaining the horizontal discrete matrix and the vertical discrete matrix.

[0053] The discrete enhancement matrix is ​​obtained by adding the horizontal and vertical discrete matrices of the same image.

[0054] In each row of the grid, based on the difference between the average gray value of each grid and the average gray value of the same row and column, the horizontal and vertical deviation coefficients of each grid are obtained, resulting in the horizontal deviation matrix and the vertical deviation matrix.

[0055] The deviation enhancement matrix is ​​obtained by adding the horizontal and vertical deviation matrices of the same image.

[0056] A multi-fusion neural network is used to process the discrete enhancement matrix and the deviation enhancement matrix of the two images to obtain the laser cladding quality score.

[0057] In this invention, discrete enhancement matrices and deviation enhancement matrices are obtained for images with significant highlights, and discrete enhancement matrices and deviation enhancement matrices are also obtained for images with significant shadows using the same technical steps.

[0058] like Figure 2 As shown, the grayscale image of a normal blade cladding layer shows cladding texture, but it is uniform overall and there are no cracks or spots.

[0059] Top-hat operation = grayscale image - grayscale image opening operation result; Black-hat operation = grayscale image closing operation result - grayscale image. Top-hat and black-hat operations are existing technologies. Top-hat operation amplifies the grayscale difference of defect areas that are brighter than the overall image by subtracting the opening operation result from the original image. Black-hat operation amplifies the grayscale difference of defect areas that are darker than the overall image by subtracting the closing operation result from the original image.

[0060] In this embodiment, the process of extracting the horizontal and vertical grayscale differences between adjacent grids includes:

[0061] Divide both the highlight salient image and the shadow salient image into N×M grids, where N is the number of rows and M is the number of columns;

[0062] Extract the mean gray value of each grid in the bright and dark salient images respectively. Here, the mean gray value is the average gray value of each pixel in the grid.

[0063] The horizontal grayscale difference is obtained by the difference in the average grayscale values ​​of adjacent grids in the horizontal direction;

[0064] The vertical grayscale difference is obtained by the difference in the average grayscale values ​​of adjacent grids in the vertical direction.

[0065] In this embodiment, the resolution of the grayscale image of the blade cladding layer is 1024×768 pixels (width×height, i.e., 1024 pixels horizontally and 768 pixels vertically). Typical defects in laser cladding layers (such as pores and microcracks) are usually 50-200 micrometers in size. When the physical resolution of the image is 10 micrometers / pixel, M (number of columns, number of horizontal grids) = 32 and N (number of rows, number of vertical grids) = 24 are set. A 32×32 grid can cover an area of ​​320×320 micrometers (converted to 10 micrometers / pixel) and can capture defects of 50-200 micrometers.

[0066] This invention quantifies the gradation features of texture in the horizontal and vertical directions by dividing an image into N×M grids and calculating the mean grayscale difference (horizontal / vertical grayscale difference) between adjacent grids. For a uniform cladding layer, the grayscale difference between adjacent grids is small and the change is gradual, avoiding the misjudgment of defects caused by local grayscale fluctuations due to texture details when directly using the original pixel values.

[0067] In this embodiment, the process of obtaining the horizontal grayscale difference includes: in the horizontal direction, subtracting the average grayscale value of the grid at (x+1,y) from the average grayscale value of the grid at (x,y), and taking the absolute value of the subtraction result to obtain the horizontal grayscale difference, where x is the horizontal coordinate and y is the vertical coordinate.

[0068] The process of obtaining the vertical grayscale difference includes: in the vertical direction, subtracting the grayscale mean of the grid at (x,y-1) from the grayscale mean of the grid at (x,y), and taking the absolute value of the subtraction result to obtain the vertical grayscale difference.

[0069] In this embodiment, the process of obtaining the horizontal discrete matrix and the vertical discrete matrix includes:

[0070] Arrange the horizontal grayscale differences according to their corresponding adjacent grid positions to obtain the horizontal grayscale difference matrix;

[0071] Arrange each vertical gray level difference according to the corresponding adjacent grid positions to obtain the vertical gray level difference matrix;

[0072] For each horizontal gray level difference in the horizontal gray level difference matrix, obtain the horizontal discrete coefficients, and arrange the horizontal discrete coefficients according to the corresponding horizontal gray level difference positions to obtain the horizontal discrete matrix.

[0073] For each vertical gray level difference in the vertical gray level difference matrix, obtain the vertical discrete coefficient. Arrange the vertical discrete coefficients according to the corresponding vertical gray level difference to obtain the vertical discrete matrix.

[0074] In this invention, the horizontal grayscale difference is calculated based on "horizontally adjacent grids"—for an M-column grid (horizontal), each row can form (M-1) pairs of horizontally adjacent grids (e.g., columns 1 and 2, columns 2 and 3, ..., columns M-1 and M), for a total of N rows. The horizontal grayscale difference matrix should have a size of N×(M-1) (N rows, M-1 columns). In this embodiment, to maintain consistent matrix size, 0 pairs are used to fill the M-th column of the horizontal grayscale difference matrix, making the size of the horizontal grayscale difference matrix N×M.

[0075] The vertical grayscale difference is calculated based on "vertical adjacent grid pairs"—for an N-row grid (vertical), each column can form (N-1) vertical adjacent grid pairs (e.g., rows 1 and 2, rows 2 and 3, ..., rows N-1 and N), for a total of M columns. The size of the vertical grayscale difference matrix should be (N-1) × M (N-1 rows, M columns). In this embodiment, to maintain a consistent matrix size, zero pairs are used to fill the Nth row of the vertical grayscale difference matrix, making the size of the vertical grayscale difference matrix N × M.

[0076] In this embodiment, the process of obtaining the horizontal discrete coefficient includes: taking each horizontal gray difference in the same row of the horizontal gray difference matrix as the center, extracting multiple horizontal gray differences in the left and right neighboring areas to form a horizontal local data segment, extracting the variance of the horizontal local data segment, and normalizing it to obtain the horizontal discrete coefficient of each horizontal gray difference.

[0077] The process of obtaining the vertical dispersion coefficients includes: taking each vertical gray level difference in the same column of the vertical gray level difference matrix as the center, extracting multiple vertical gray level differences in the upper and lower neighboring areas to form a vertical local data segment, extracting the variance of the vertical local data segment, and normalizing it to obtain the vertical dispersion coefficient of each vertical gray level difference.

[0078] In this embodiment, the lengths of the horizontal local data segment and the vertical local data segment can be set to 3 or 4.

[0079] In this embodiment, the horizontal discrete matrix and the horizontal gray-level difference matrix have the same size, and the vertical discrete matrix and the vertical gray-level difference matrix have the same size.

[0080] In this embodiment, the formula for calculating the lateral dispersion coefficient is: θ σ,x,i = σ x,i / (1 + σ x,i ), where θ σ,x,i Let σ be the horizontal dispersion coefficient of the i-th horizontal grayscale difference. x,i Let be the variance of the horizontal local data segment of the i-th horizontal grayscale difference, where i is a positive integer.

[0081] In this embodiment, the formula for calculating the longitudinal dispersion coefficient is: θ σ,y,i = σ y,i / (1 + σ y,i ), where θ σ,y,i Let σ be the vertical dispersion coefficient of the i-th vertical grayscale difference. y,i Let be the variance of the vertical local data segment of the i-th vertical grayscale difference.

[0082] The calculation of the horizontal / vertical dispersion coefficients is based on the variance of local neighborhood data segments, which can reflect the fluctuation trend of gray-scale differences within a local range. For uniform cladding layers, the gray-scale differences within the same neighborhood are small (small variance). After processing with the normalization formula, the dispersion coefficients approach 0, and the dispersion coefficients of different regions are stable, forming the characteristic of "low dispersion".

[0083] In this embodiment, the process of obtaining the lateral deviation matrix and the longitudinal deviation matrix includes:

[0084] In each row of the grid, the lateral deviation coefficient of each grid is obtained based on the difference between the average gray value of each grid and the average gray value of the same row. The lateral deviation coefficients are arranged according to the grid position to obtain the lateral deviation matrix.

[0085] In each grid column, the vertical deviation coefficient of each grid is obtained based on the difference between the average gray value of each grid and the average gray value of the same column. The vertical deviation coefficients are then arranged according to the grid position to obtain the vertical deviation matrix.

[0086] In this embodiment, the formula for calculating the lateral deviation coefficient is: , where ε j,k G is the lateral deviation coefficient of the k-th grid in the j-th row. j,k Let G be the average gray value of the k-th grid in the j-th row. j,avg Let | be the average gray value of all grid cells in the j-th row, and | be the absolute value.

[0087] The formula for calculating the longitudinal deviation coefficient is: , where ε r,k G is the vertical deviation coefficient of the k-th grid in the r-th column. r,k G is the gray mean of the k-th grid in the r-th column. r,avg is the average gray value of all grid cells in column r.

[0088] The dimensions of the lateral deviation matrix and the longitudinal deviation matrix are N×M.

[0089] When cladding textures exist in the blade cladding layer, the difference between the average gray value of the mesh and the average gray value of the row / column can be calculated row by row / column to accurately quantify the "degree of deviation" of the local mesh relative to the whole of the row or column. For example, the deviation coefficient of a mesh with uniform cladding texture will be lower because the difference from the average gray value of the row / column is small.

[0090] like Figure 3 As shown, the multi-fusion neural network includes: a bright area discrete deviation feature fusion unit, a dark area discrete deviation feature fusion unit, a first pooling layer, a second pooling layer, an adder A1, and a fully connected layer;

[0091] The discrete deviation feature fusion unit for bright areas extracts multi-scale features from the discrete enhancement matrix and deviation enhancement matrix of the image with salient bright areas and then fuses them; the discrete deviation feature fusion unit for dark areas extracts multi-scale features from the discrete enhancement matrix and deviation enhancement matrix of the image with salient dark areas and then fuses them.

[0092] The first pooling layer performs pooling processing on the output of the bright area discrete deviation feature fusion unit; the second pooling layer performs pooling processing on the output of the dark area discrete deviation feature fusion unit.

[0093] Adder A1 is used to add the outputs of the first pooling layer element by element;

[0094] The fully connected layer is used to output a laser cladding quality score based on the output characteristics of adder A1.

[0095] This invention extracts features from the discrete enhancement matrix and the deviation enhancement matrix at multiple scales through a bright area discrete deviation feature fusion unit and a dark area discrete deviation feature fusion unit, and then fuses them. The fused features are then pooled using a pooling layer, and the features of the two images are added together by an adder A1. This allows the fully connected layer to comprehensively evaluate the laser cladding quality by combining the features of the two images, thereby improving the accuracy of laser cladding quality evaluation.

[0096] Both the highlight discrete deviation feature fusion unit and the dark discrete deviation feature fusion unit include: a first shallow multi-scale convolution module, a second shallow multi-scale convolution module, and a multiplier M1;

[0097] The first shallow multi-scale convolution module extracts shallow multi-scale features from the discrete enhancement matrix;

[0098] The second shallow multi-scale convolution module extracts shallow multi-scale features from the deviation enhancement matrix;

[0099] Multiplier M1 is used to element-wise multiply the outputs of the first shallow multiscale convolution module and the second shallow multiscale convolution module.

[0100] This invention uses a discrete enhancement matrix to focus on local grayscale fluctuations (such as abrupt changes in discrete coefficients that may be caused by tiny pores), and a deviation enhancement matrix to focus on the differences between the grid and the row / column mean (such as local deviations caused by pores). The dual-matrix input allows the network to simultaneously mine both "local abrupt changes" and "global shifts" features, making defect features more prominent.

[0101] like Figure 4 As shown, both the first shallow multi-scale convolution module and the second shallow multi-scale convolution module include: a first convolution block, a second convolution block, a third convolution block, and a Concat layer;

[0102] The first convolutional block, the second convolutional block, and the third convolutional block are connected in sequence;

[0103] The Concat layer is used to concatenate the outputs of the first convolutional block, the second convolutional block, and the third convolutional block.

[0104] In this embodiment, the kernel of the first convolution block is set to 1×1, the kernel of the second convolution block is set to 3×3, and the kernel of the third convolution block is set to 5×5.

[0105] The first, second, and third convolutional blocks each consist of: a convolutional layer, an activation function ReLU, and a batch normalization layer.

[0106] This invention improves the richness of features by extracting features at different scales using convolutional blocks of different sizes.

[0107] In this embodiment, a laser cladding quality score of 90-100 points is considered excellent cladding; 80-89 points is considered good cladding with minor defects; 70-79 points is considered acceptable cladding with local defects; and <70 points is considered unacceptable cladding.

[0108] This invention uses top-hat and black-hat operations to specifically extract salient images of bright and dark areas, effectively amplifying the grayscale differences between bright defects (such as overheating and metal spatter) and dark defects (such as incomplete fusion and pores) and normal areas in the laser cladding layer, thus overcoming the problem of defect blurring caused by natural textures.

[0109] This invention extracts the grayscale difference in the horizontal and vertical directions by dividing the grid, calculates the discrete coefficient and the deviation coefficient, and constructs the discrete enhancement matrix and the deviation enhancement matrix. It captures defect features from two dimensions: local fluctuation (discrete coefficient) and overall deviation (deviation coefficient). It not only pays attention to the subtle changes of adjacent grids, but also considers the difference between the grid and the whole row and column, avoiding the limitations of a single feature dimension. It can comprehensively cover different types and distributions of defects (such as horizontal cracks, vertical pores, etc.).

[0110] This invention adds the horizontal and vertical matrices of the same image to obtain an enhancement matrix, achieving complementarity and enhancement of two-dimensional features, making the features of the defect area more prominent; the multi-fusion neural network performs deep fusion processing on the four enhancement matrices of the two images, improving the accuracy of defect identification.

[0111] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for inspecting the quality of laser cladding on turbine blades, characterized in that, include: Perform top-hat and black-hat operations on the grayscale image of the blade cladding layer to obtain salient images of bright and dark areas. The bright and dark salient images are divided into grids respectively, and the horizontal gray level difference between adjacent grids is extracted to obtain the horizontal discrete coefficient and the vertical gray level difference is extracted to obtain the vertical discrete coefficient, thus obtaining the horizontal discrete matrix and the vertical discrete matrix. The formula for calculating the lateral dispersion coefficient is: θ σ,x,i = σ x,i / (1 + σ x,i ), where θ σ,x,i Let σ be the horizontal dispersion coefficient of the i-th horizontal grayscale difference. x,i Let be the variance of the horizontal local data segment of the i-th horizontal gray level difference, where i is a positive integer. Taking each horizontal gray level difference in the same row of the horizontal gray level difference matrix as the center, multiple horizontal gray level differences in the left and right neighboring areas are extracted to form a horizontal local data segment. The formula for calculating the longitudinal dispersion coefficient is: θ σ,y,i = σ y,i / (1 + σ y,i ), where θ σ,y,i Let σ be the vertical dispersion coefficient of the i-th vertical grayscale difference. y,i Let V be the variance of the vertical local data segment of the i-th vertical gray level difference. Taking each vertical gray level difference in the same column of the vertical gray level difference matrix as the center, multiple vertical gray level differences in the upper and lower neighborhoods are extracted to form a vertical local data segment. The discrete enhancement matrix is ​​obtained by adding the horizontal and vertical discrete matrices of the same image. In each row of the grid, based on the difference between the average gray value of each grid and the average gray value of the same row and column, the horizontal and vertical deviation coefficients of each grid are obtained, resulting in the horizontal deviation matrix and the vertical deviation matrix. The formula for calculating the lateral deviation coefficient is: , where ε j,k G is the lateral deviation coefficient of the k-th grid in the j-th row. j,k Let G be the average gray value of the k-th grid in the j-th row. j,avg Let | be the average gray value of all grid cells in the j-th row, and | be the absolute value. The formula for calculating the longitudinal deviation coefficient is: , where ε r,k G is the vertical deviation coefficient of the k-th grid in the r-th column. r,k G is the gray mean of the k-th grid in the r-th column. r,avg Let be the average gray value of all grid cells in column r; The deviation enhancement matrix is ​​obtained by adding the horizontal and vertical deviation matrices of the same image. A multi-fusion neural network is used to process the discrete enhancement matrix and the deviation enhancement matrix of the two images to obtain the laser cladding quality score; The multi-fusion neural network includes: a bright area discrete deviation feature fusion unit, a dark area discrete deviation feature fusion unit, a first pooling layer, a second pooling layer, an adder A1, and a fully connected layer. The bright area discrete deviation feature fusion unit and the dark area discrete deviation feature fusion unit extract multi-scale features from the discrete enhancement matrix and the deviation enhancement matrix, and then fuse them. The fused features are pooled using a pooling layer, and the features of the two images are added together by the adder A1, so that the fully connected layer can evaluate the laser cladding quality by combining the features of the two images.

2. The method for detecting the quality of laser cladding on turbine blades according to claim 1, characterized in that, The process of extracting the horizontal and vertical grayscale differences between adjacent grids includes: Divide both the highlight salient image and the shadow salient image into N×M grids, where N is the number of rows and M is the number of columns; Extract the mean grayscale value of each grid in the bright and dark salient images respectively; The horizontal grayscale difference is obtained by the difference in the average grayscale values ​​of adjacent grids in the horizontal direction; The vertical grayscale difference is obtained by the difference in the average grayscale values ​​of adjacent grids in the vertical direction.

3. The method for detecting the quality of laser cladding on turbine blades according to claim 2, characterized in that, The process of obtaining the horizontal grayscale difference includes: in the horizontal direction, subtracting the grayscale mean of the grid at (x+1,y) from the grayscale mean of the grid at (x,y), and taking the absolute value of the subtraction result to obtain the horizontal grayscale difference, where x is the horizontal coordinate and y is the vertical coordinate. The process of obtaining the vertical grayscale difference includes: in the vertical direction, subtracting the grayscale mean of the grid at (x,y-1) from the grayscale mean of the grid at (x,y), and taking the absolute value of the subtraction result to obtain the vertical grayscale difference.

4. The method for detecting the quality of laser cladding on turbine blades according to claim 1, characterized in that, The process of obtaining the horizontal and vertical discrete matrices includes: Arrange the horizontal grayscale differences according to their corresponding adjacent grid positions to obtain the horizontal grayscale difference matrix; Arrange each vertical gray level difference according to the corresponding adjacent grid positions to obtain the vertical gray level difference matrix; For each horizontal gray level difference in the horizontal gray level difference matrix, obtain the horizontal discrete coefficients, and arrange the horizontal discrete coefficients according to the corresponding horizontal gray level difference positions to obtain the horizontal discrete matrix. For each vertical gray level difference in the vertical gray level difference matrix, obtain the vertical discrete coefficient. Arrange the vertical discrete coefficients according to the corresponding vertical gray level difference to obtain the vertical discrete matrix.

5. The method for detecting the quality of laser cladding on turbine blades according to claim 4, characterized in that, The process of obtaining the horizontal dispersion coefficients includes: extracting the variance of the horizontal local data segments and normalizing it to obtain the horizontal dispersion coefficients of each horizontal gray level difference. The process of obtaining the longitudinal dispersion coefficients includes: extracting the variance of the longitudinal local data segments and normalizing it to obtain the longitudinal dispersion coefficients of each longitudinal gray level difference.

6. The method for detecting the quality of laser cladding on turbine blades according to claim 1, characterized in that, The process of obtaining the lateral deviation matrix and the longitudinal deviation matrix includes: In each row of the grid, the lateral deviation coefficient of each grid is obtained based on the difference between the average gray value of each grid and the average gray value of the same row. The lateral deviation coefficients are arranged according to the grid position to obtain the lateral deviation matrix. In each grid column, the vertical deviation coefficient of each grid is obtained based on the difference between the average gray value of each grid and the average gray value of the same column. The vertical deviation coefficients are then arranged according to the grid position to obtain the vertical deviation matrix.

7. The method for detecting the quality of laser cladding on turbine blades according to claim 1, characterized in that, The discrete deviation feature fusion unit for bright areas extracts multi-scale features from the discrete enhancement matrix and deviation enhancement matrix of the image with salient bright areas and then fuses them; the discrete deviation feature fusion unit for dark areas extracts multi-scale features from the discrete enhancement matrix and deviation enhancement matrix of the image with salient dark areas and then fuses them. The first pooling layer performs pooling processing on the output of the bright part discrete deviation feature fusion unit; The second pooling layer performs pooling processing on the output of the dark area discrete deviation feature fusion unit; Adder A1 is used to add the outputs of the first pooling layer element by element; The fully connected layer is used to output a laser cladding quality score based on the output characteristics of adder A1.

8. The method for detecting the quality of laser cladding on turbine blades according to claim 7, characterized in that, Both the highlight discrete deviation feature fusion unit and the dark discrete deviation feature fusion unit include: a first shallow multi-scale convolution module, a second shallow multi-scale convolution module, and a multiplier M1; The first shallow multi-scale convolution module extracts shallow multi-scale features from the discrete enhancement matrix; The second shallow multi-scale convolution module extracts shallow multi-scale features from the deviation enhancement matrix; Multiplier M1 is used to element-wise multiply the outputs of the first shallow multiscale convolution module and the second shallow multiscale convolution module.

9. The method for detecting the quality of laser cladding on turbine blades according to claim 8, characterized in that, Both the first and second shallow multi-scale convolutional modules include: a first convolutional block, a second convolutional block, a third convolutional block, and a Concat layer; The first convolutional block, the second convolutional block, and the third convolutional block are connected in sequence; The Concat layer is used to concatenate the outputs of the first convolutional block, the second convolutional block, and the third convolutional block.

Images