An intelligent detection and evaluation system for aero-engine blade repair performance

Abstract

The application relates to the technical field of visual detection, in particular to an intelligent detection and evaluation system for the repair performance of an aero-engine blade, which comprises an aero-engine blade image acquisition module, which is used for acquiring an aero-engine blade repair area image to be detected, performing illumination normalization on the image, and generating a repair area subgraph sequence with coordinates. In the application, the aero-engine blade repair area image is divided into overlapping subareas with spatial coordinate indexes, so that full coverage of detection analysis and the accuracy of flaw positioning are ensured, second-order statistics such as contrast and energy are obtained by calculating a gray level co-occurrence matrix, radial and angular energy distributions extracted by two-dimensional fast Fourier transform are combined, a multi-dimensional texture feature set fusing local pixel relationships in a spatial domain and periodic structures in a frequency domain is constructed, the cladding texture state of the repair area can be more comprehensively described, and the identification ability for texture uniformity and directional abnormality is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual detection, and in particular to an intelligent detection and evaluation system for the repair performance of an aero-engine blade. BACKGROUND

[0002] The technical field of visual detection is a technology that uses machine vision to replace the human eye for observation, measurement and judgment. Through an optical system (such as an industrial camera, a lens, and a light source), a two-dimensional or three-dimensional image of the target is obtained, and then computer image processing and pattern recognition algorithms are used to analyze and understand the information in the image to extract key feature parameters.

[0003] The existing visual detection technology has difficulty in effectively capturing and quantifying the internal complexity and non-periodic changes of the surface micro-topography when evaluating surfaces with complex random textures such as the repair area of an aero-engine blade, because it relies on extracting image features such as edges, contours, or first-order statistics. For example, for the surface formed after laser cladding repair, the quality is not only reflected in the presence or absence of macro-defects such as cracks or pores, but also in the micro-texture and roughness formed by the smoothness of the cladding joint and the uniformity of the energy distribution. When facing two repair areas with different surface complexities but without macro-defects, the traditional method may give similar or even the same detection results due to the single dimension of feature extraction. Therefore, improvement is needed. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art and to provide an intelligent detection and evaluation system for the repair performance of an aero-engine blade.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: an intelligent detection and evaluation system for the repair performance of an aero-engine blade comprises:

[0006] An aero blade image acquisition module is used to acquire an image of the repair area of the aero-engine blade to be detected, normalize the image, and generate a sequence of repair area sub-images with coordinates.

[0007] A cladding texture feature extraction module is used to calculate the gray level co-occurrence matrix of each sub-region image in the sequence according to the sequence of repair area sub-images with coordinates, extract the second-order statistics of contrast, energy, correlation and entropy from the gray level co-occurrence matrix, establish a set of sub-region texture statistical vectors, perform two-dimensional fast Fourier transform on each sub-region image in the sequence of repair area sub-images with coordinates, fuse the set of sub-region texture statistical vectors, and obtain a set of multi-dimensional texture features of the repair area.

[0008] a surface complexity atlas generation module configured to perform Gaussian smoothing on each sub-region image according to the sequence of sub-region images with coordinates, count the number of different size cubic grid covering gray surface, obtain a sequence of local surface fractal dimension values, map each local surface fractal dimension value back to the corresponding position on the aero-engine blade surface according to the spatial position coordinates provided by the sequence of sub-region images with coordinates, and generate a surface fractal dimension atlas of the aero-engine blade.

[0009] a flaw comprehensive evaluation module configured to combine the texture features and the fractal dimension values of each sub-region, construct a regional joint feature description vector, calculate the distance between the joint feature description vector of each region and a preset premium benchmark center, and obtain a flaw positioning map.

[0010] Preferably, the sequence of sub-region images with coordinates is obtained by:

[0011] analyzing an original image input, locating the pixel boundary of the aero-engine blade repair region, cropping the aero-engine blade repair region image, counting the pixel mean and standard deviation of each color channel, linearly scaling according to the difference between the pixel and channel mean, and obtaining a normalized repair region image;

[0012] according to the normalized repair region image, setting the sub-region size and step, cropping row by row and column by column along the horizontal and vertical axes at a fixed step, recording the starting row index and starting column index of each sub-region, and generating a set of overlapping sub-region images;

[0013] according to the set of overlapping sub-region images, calculating the top-left row coordinate and top-left column coordinate in the normalized repair region image for each sub-region image and calculating the center row coordinate and center column coordinate, constructing an index mapping of each sub-region image to the corresponding coordinates, sorting by starting row index from small to large and then by starting column index from small to large, and generating a sequence of sub-region images with coordinates.

[0014] Preferably, the sequence of sub-region images with coordinates is obtained by:

[0015] according to the sequence of sub-region images with coordinates, reading each sub-region image one by one, quantizing the pixel gray value, counting the number of adjacent times between different gray levels, establishing a gray co-occurrence matrix in the horizontal direction, vertical direction, main diagonal direction and auxiliary diagonal direction, respectively, normalizing the matrix, extracting the contrast, energy, correlation and entropy values of each sub-region image, and generating a set of sub-region texture statistical vectors.

[0016] Preferably, the sequence of sub-region images with coordinates is obtained by:

[0017] According to the sub-region texture statistics vector set, the sub-region images of the same index are called, a two-dimensional fast Fourier transform is performed on each sub-region image to convert the spatial domain gray information into a frequency spectrum amplitude map, energy values in different radius ranges of the frequency spectrum amplitude map are extracted, a radial energy distribution is calculated, and a frequency domain energy feature set is generated by partitioning and integrating energy values in different angle ranges to calculate an angular energy distribution.

[0018] According to the frequency domain energy feature set, the sub-region texture statistics vector sets of the same index are matched, and the contrast, energy, correlation, entropy, radial energy distribution, and angular energy distribution are combined according to corresponding positions to form a multi-dimensional texture description vector of each sub-region. The multi-dimensional texture description vectors of all sub-regions are summarized to generate a multi-dimensional texture feature set of the repair area.

[0019] Preferably, the step of obtaining the sequence of local surface fractal dimension values is:

[0020] According to the sequence of sub-region images with coordinates, each sub-region image is convolved and smoothed by a discrete Gaussian kernel, the kernel size and standard deviation are set, the boundary is filled by mirroring, the blocks are divided according to the size of the cubic grid, the number of blocks is accumulated, and a coverage number sequence is obtained.

[0021] According to the coverage number sequence, the local surface fractal dimension value is calculated to obtain a sequence of local surface fractal dimension values.

[0022] Preferably, the step of obtaining the blade surface fractal dimension map is:

[0023] According to the sequence of local surface fractal dimension values, the top-left row coordinate and the top-left column coordinate of each sub-region image corresponding to the local surface fractal dimension value in the sequence of sub-region images with coordinates are called, each local surface fractal dimension value is written into the corresponding position, and the starting row coordinate and the starting column coordinate are sequentially organized and output to generate a blade surface fractal dimension map.

[0024] Preferably, the step of obtaining the region joint feature description vector is:

[0025] According to the multi-dimensional texture feature set of the repair area and the blade surface fractal dimension map, the second-order statistics of the contrast, energy, correlation, and entropy and the local surface fractal dimension value are spliced into a column vector in a fixed order according to the sub-region coordinates to generate a region joint feature description vector.

[0026] Preferably, the step of obtaining the flaw positioning map is:

[0027] According to the region joint feature description vector, a feature key degree modulation Mahalanobis distance value is calculated.

[0028] According to the feature key degree modulation Markov distance value, the statistical quantile and mean value are counted, a comparison rule is constructed, each sub-region is judged according to a preset threshold, an overall quantitative score is output, and the sub-region coordinates exceeding the preset threshold are extracted and written into a mark layer to generate a defect positioning map.

[0029] Compared with the prior art, the advantages and positive effects of the present application are that:

[0030] In the present application, by dividing the aero-engine blade repair area image into overlapping sub-regions with spatial coordinate indexes, full coverage of detection analysis and accuracy of defect positioning are ensured, by calculating the gray level co-occurrence matrix to obtain the second-order statistics such as contrast and energy, and combining the radial and angular energy distribution extracted by two-dimensional fast Fourier transform, a multi-dimensional texture feature set fusing the local pixel relationship in the spatial domain and the periodic structure in the frequency domain is constructed, which can more comprehensively describe the cladding texture state of the repair area, and improve the recognition ability of the texture uniformity and directionality abnormality, and by calculating the local surface fractal dimension of each sub-region by the difference box counting method, the surface geometric complexity and roughness which are difficult to quantify by traditional methods are measured, and the multi-dimensional texture features and fractal dimension are fused into a unified regional joint feature description vector, not only an objective, multi-dimensional comprehensive evaluation standard is established, but also an intuitive defect positioning map can be generated according to the region coordinates whose distance exceeds the threshold, and the accuracy of detection is enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 The system flowchart of the present application. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0033] Please refer to Figure 1 The present application provides a technical scheme: an aero-engine blade repair performance intelligent detection and evaluation system comprises:

[0034] An aero blade image acquisition module is used to acquire an aero-engine blade repair area image to be detected, perform illumination normalization on the image, and generate a repair area subgraph sequence with coordinates.

[0035] The cladding texture feature extraction module is configured to calculate a gray level co-occurrence matrix of each sub-region image in the sequence according to the sequence of sub-region images with coordinates, extract second-order statistics of contrast, energy, correlation and entropy from the gray level co-occurrence matrix, establish a set of sub-region texture statistical vectors, perform two-dimensional fast Fourier transform on each sub-region image in the sequence of sub-region images with coordinates, fuse the set of sub-region texture statistical vectors, and obtain a multi-dimensional texture feature set of the repair region.

[0036] The surface complexity atlas generation module is configured to perform Gaussian smoothing on each sub-region image according to the sequence of sub-region images with coordinates, count the number of different size cubic grid coverage gray level surfaces, obtain a sequence of local surface fractal dimension values, map each local surface fractal dimension value back to the corresponding surface position of the aero-engine blade according to the spatial position coordinates provided by the sequence of sub-region images with coordinates, and generate a surface fractal dimension atlas of the blade.

[0037] The repair defect comprehensive evaluation module is configured to combine the texture features and fractal dimension values corresponding to each sub-region, construct a regional joint feature description vector, calculate the distance between the joint feature description vector of each region and a preset premium benchmark center, and obtain a defect positioning map.

[0038] The sequence of sub-region images with coordinates is obtained by:

[0039] The original image is parsed, the pixel boundary of the repair region of the aero-engine blade is located, the repair region image of the aero-engine blade is cropped, the pixel mean and standard deviation of each color channel are counted, linear scaling is performed according to the difference between the pixel and channel mean values, and a normalized repair region image is obtained.

[0040] According to the normalized repair region image, the sub-region size and step size are set, the sub-region images are cropped row by row and column by column along the horizontal and vertical axes at a fixed step size, the starting row index and starting column index of each sub-region are recorded, and a set of overlapping sub-region images is generated.

[0041] According to the set of overlapping sub-region images, the top-left row coordinate and top-left column coordinate of each sub-region image in the normalized repair region image are calculated, the center row coordinate and center column coordinate are calculated, the index mapping of each sub-region image to the corresponding coordinates is constructed, the starting row index is sorted from small to large and the starting column index is sorted from small to large, and a sequence of sub-region images with coordinates is generated.

[0042] Specifically, the original image input is analyzed. First, the input aero-engine blade image in RGB format is converted into a grayscale image. Then, a 5x5 Gaussian kernel is used to smooth the grayscale image to suppress the interference of image noise. Next, the Canny edge detection algorithm is used to extract the contour of the repair area. First, a global threshold is calculated by the Otsu method. Then, the high threshold is set to the global threshold, and the low threshold is set to 0.5 times the global threshold. For example, if the calculated global threshold is 110, the high threshold is set to 110, and the low threshold is set to 55. Morphological closing operation is performed on the binary image after edge detection. A 3x3 rectangular structuring element is used for inflation operation and then corrosion operation to connect the broken boundaries that may occur in edge detection, forming a complete closed contour. Then, all contours in the binary image are searched, and the area of each contour is calculated. The contour with the largest area is identified as the pixel boundary of the aero-engine blade repair area. According to the coordinates of the minimum bounding rectangle of the boundary, the aero-engine blade repair area image is cropped from the original color image. Then, the cropped image is subjected to illumination normalization. The mean and standard deviation of the pixel values of the R, G, and B color channels are calculated, respectively. For example, for the R channel, if the pixel value set is {135, 140, 142, 138, 145}, the mean is 140, and the standard deviation is about 3.54. Then, each pixel value is linearly scaled. The target is to make the pixel distribution of each channel have a preset target mean (e.g., 128) and target standard deviation (e.g., 40). The specific operation is to subtract the original mean of the channel from each pixel value, divide by the original standard deviation, multiply by the target standard deviation, and finally add the target mean. By performing this operation on all pixels, the normalized repair area image is obtained.

[0043] According to the normalized repair area image, first need to set the size of the sub-region and the step length of sliding, the setting of the two parameters is based on the statistical analysis of the size of typical defects (such as cracks, pores, unfused) in the historical repair blade image, the principle of setting the size of the sub-region is that its size should be sufficient to contain a complete minimum texture unit or a typical tiny defect, for example, through statistical analysis, it is found that most of the micro cracks are between 20 to 50 pixels in length, therefore, the size of the sub-region is set to a fixed 64x64 pixel square region, while the setting of the step length determines the overlap rate between adjacent sub-regions, higher overlap rate can provide more detailed local feature change information, but will increase the calculation amount, here a strategy of 75% overlap rate is set, that is, for a size of 64 pixels, the step length is set to 16 pixels, which means that the starting position of the next sub-region will move 16 pixels in the horizontal or vertical direction compared with the starting position of the previous sub-region, after setting the parameters, the sub-region image is cut along the horizontal and vertical axes with a fixed step length, the process is realized through a double loop, the outer loop controls the movement in the vertical direction, the starting row index starts from 0 and increases by 16 each time, until the starting row index plus the sub-region height 64 is greater than the total height of the normalized repair area image, the inner loop controls the movement in the horizontal direction, the starting column index starts from 0 and increases by 16 each time, until the starting column index plus the sub-region width 64 is greater than the total width of the image, in each loop, a sub-region image is extracted from the normalized repair area image according to the current starting row index and starting column index and the size of 64x64, and the sub-region image and its corresponding starting row index and starting column index are stored together, after traversing the entire normalized repair area image, the overlapping sub-region image set is generated.

[0044] According to the set of overlapping sub-region images, the accurate spatial coordinates of each sub-region image in the original normalized repair region image are calculated. First, the starting row index and starting column index of each sub-region recorded during the generation process are directly used as the top-left row coordinate and top-left column coordinate, for example, a sub-region recorded starting index (32, 48) during generation, the top-left row coordinate is 32, and the top-left column coordinate is 48. Then, based on the top-left coordinates and the known sub-region size (for example, 64x64 pixels), the center coordinates are calculated. The center row coordinate is calculated by adding half of the sub-region height to the top-left row coordinate, that is, 32 + (64 / 2) = 64. The center column coordinate is calculated by adding half of the sub-region width to the top-left column coordinate, that is, 48 + (64 / 2) = 80. Thus, the center coordinates of the sub-region are (64, 80). Subsequently, a structured data list is constructed, in which each element contains the image data of a sub-region, its corresponding top-left row coordinate, top-left column coordinate, center row coordinate and center column coordinate. This process essentially establishes a clear index mapping relationship between each sub-region image and its corresponding coordinate information. After completing the coordinate calculation and mapping construction of all sub-regions, the structured data list is sorted. The sorting rule is to first arrange the starting row index (i.e. the top-left row coordinate) in ascending order, and then arrange the starting column index (i.e. the top-left column coordinate) in ascending order in the case of the same starting row index. This sorting method ensures that the final generated sequence is organized in a raster scan order from top to bottom and from left to right, generating a repair region sub-graph sequence with coordinates.

[0045] The sub-region texture statistical vector set acquisition step is:

[0046] According to the repair region sub-graph sequence with coordinates, each sub-region image is read one by one, the pixel gray value is quantized and graded, the number of adjacent times between different gray levels is counted, and the gray level co-occurrence matrix is established in the horizontal direction, vertical direction, main diagonal direction and auxiliary diagonal direction. The matrix is symmetrized and normalized, and the contrast, energy, correlation and entropy values of each sub-region image are extracted to generate a sub-region texture statistical vector set.

[0047] Specifically, according to the sequence of sub-region images with coordinates, the pixel gray value of each sub-region image is quantized and classified first, and the original gray scale range of 0 to 255 is compressed to 16 levels. The specific method is to divide the pixel value by 16 and take the integer part. For example, the pixel with a gray value in the range of 0 to 15 is quantized to level 0, the pixel with a gray value in the range of 16 to 31 is quantized to level 1, and so on. After quantization, the number of times of different gray levels appearing simultaneously in a certain spatial relationship, i.e. the number of adjacent times, is counted. The spatial relationship here is defined as four directions, which are horizontal direction (0 degree angle, pixel pair offset (0, 1)), vertical direction (90 degree angle, pixel pair offset (-1, 0)), main diagonal direction (45 degree angle, pixel pair offset (-1, 1)) and auxiliary diagonal direction (135 degree angle, pixel pair offset (-1, -1)). The distance of the pixel pair is fixed at 1. A 16x16 gray co-occurrence matrix is constructed for each direction. The value of the (i, j) element in the matrix represents the frequency of the pixel with gray level i and the pixel with gray level j appearing in the specified direction and distance. Then, the four matrices are symmetrized respectively, i.e. each matrix is added to its transpose matrix, and then normalized, i.e. each element value in the matrix is divided by the sum of all element values in the matrix, so that it is converted into a probability matrix. Finally, based on the four normalized gray co-occurrence matrices, four second-order statistics, contrast, energy, correlation and entropy, are calculated respectively, and the four values calculated for each direction are averaged to obtain the final contrast, energy, correlation and entropy values of the sub-region image. The four values constitute a four-dimensional texture statistical vector. Repeat this process for all sub-region images to generate a set of sub-region texture statistical vectors.

[0048] The acquisition step of the multi-dimensional texture feature set of the repair area is:

[0049] According to the set of sub-region texture statistical vectors, the same index sub-region image is called to perform two-dimensional fast Fourier transform on each sub-region image, to convert the spatial domain gray scale information into a frequency amplitude map, to extract the energy values in different radius ranges of the frequency amplitude map, to calculate the radial energy distribution, and to perform partition integration on the energy values in different angle ranges to calculate the angular energy distribution, to generate a frequency energy feature set;

[0050] According to the frequency energy feature set, the set of sub-region texture statistical vectors with the same index is matched, and the contrast, energy, correlation, entropy, radial energy distribution and angular energy distribution are combined according to the corresponding positions to form a multi-dimensional texture description vector of each sub-region. The multi-dimensional texture description vectors of all sub-regions are summarized to generate a multi-dimensional texture feature set of the repair area.

[0051] Specifically, according to the sub-region texture statistical vector set, the sub-region image with the same index in the coordinate-based repair area subgraph sequence is called by using the index information thereof, a two-dimensional fast Fourier transform is applied to each sub-region image (for example, 64x64 pixels) to convert the image from a spatial domain to a frequency domain, a complex matrix is obtained after the transformation, then the modulus of each element in the matrix is calculated, that is, the amplitude, the amplitude is added by 1 and then a natural logarithm is taken to obtain a frequency spectrum amplitude graph, for the convenience of analysis, the zero frequency component of the frequency spectrum amplitude graph is moved to the center of the image, then the frequency domain energy features are extracted on the frequency spectrum amplitude graph, first, the radial energy distribution is calculated, the frequency spectrum graph is divided into 5 concentric ring regions with no overlap, the radius range of these ring regions is pre-set according to the size of the sub-region image, for example, the radius range can be divided into 1-8 pixels, 9-16 pixels, 17-24 pixels, 25-32 pixels and a region greater than 32 pixels, the sum of the square of the frequency spectrum amplitude of all pixel points in each ring region is calculated to obtain 5 radial energy values, and this group of values constitutes a radial energy distribution feature vector, then the angular energy distribution is calculated, the frequency spectrum graph is divided into 8 fan-shaped regions with an angle range of 45 degrees, for example, 0-45 degrees, 45-90 degrees and the like, the sum of the square of the frequency spectrum amplitude of all pixel points in each fan-shaped region is calculated to obtain 8 angular energy values, and this group of values constitutes an angular energy distribution feature vector, and the process is repeated for all sub-region images to generate a frequency domain energy feature set.

[0052] According to the frequency domain energy feature set, the texture statistical vector in the sub-region texture statistical vector set corresponding to the same index is matched through the index information of the sub-region, the spatial domain texture feature and the frequency domain energy feature are combined for each sub-region, and the specific operation is to splice the vectors in a fixed order, first, 4 second-order statistics in the sub-region texture statistical vector set, that is, contrast, energy, correlation and entropy, then 5 radial energy values in the radial energy distribution feature in the frequency domain energy feature set, and finally 8 angular energy values in the angular energy distribution feature, through this splicing mode, a 4+5+8=17-dimensional multi-dimensional texture description vector is formed for each sub-region, which comprehensively describes the texture fineness, complexity and directionality of the sub-region in the spatial domain and the energy distribution characteristics in the frequency domain, all multi-dimensional texture description vectors corresponding to the sub-regions in the coordinate-based repair area subgraph sequence are summarized to form a large feature matrix, wherein each row represents a sub-region and each column represents a feature dimension, and finally a repair area multi-dimensional texture feature set is generated.

[0053] The acquisition steps of the local surface fractal dimension value sequence are as follows:

[0054] Based on the sequence of sub-images of the repaired area with coordinates, each sub-region image is smoothed by convolution with a discrete Gaussian kernel. The kernel size and standard deviation are set, and the boundary is mirrored. The area is divided into blocks according to the cube grid size. The number of each block is accumulated to obtain the coverage number sequence.

[0055] Based on the coverage quantity sequence, the local surface fractal dimension value is calculated, resulting in a local surface fractal dimension value sequence. The calculation formula is as follows:

[0056] ;

[0057] in, For the first Local surface fractal dimension values ​​of each sub-region image. The number of scales used. For the first Each cube grid size For the first The values ​​of the independent variable for each dimension, For the first The sub-region image in the first Number of units covered by each size For the first The sub-region image in the first The values ​​of the dependent variable for each dimension.

[0058] Specifically, according to the sub-subgraph sequence of the coordinate repair area, each sub-region image in the sequence is processed by discrete Gaussian kernel convolution smoothing. Specifically, a discrete Gaussian kernel with a size of 5x5 is used, and the standard deviation is set to 1.5. The selection of the standard deviation value is based on the statistical analysis of the texture characteristics of a large number of qualified repair area images, aiming to effectively filter out high-frequency noise while retaining the inherent medium-low frequency texture structure information of the cladding layer. Before performing the convolution operation, the boundary of the sub-region image is mirror filled, that is, the inside of the image boundary is reflected to the outside of the boundary, creating a virtual extended boundary. The filling width matches the radius of the 5x5 kernel, so that the convolution kernel can be completely located within the image data when processing the boundary pixels, avoiding calculation errors introduced by boundary effects. After smoothing, the gray value of each sub-region image is regarded as the height (z-axis) in three-dimensional space, and the pixel coordinates (x, y-axis) form the bottom surface. The difference box counting method is used to calculate the surface coverage. For this purpose, a set of gradually increasing cubic grid sizes is set, for example, for a 64x64 pixel sub-region, four different scales are selected, that is, the edge length of the cubic grid is 2, 4, 8, and 16 pixels. For each set grid size, for example, the edge length of 8 pixels, the 64x64 pixel bottom surface of the sub-region image is divided into 8x8 non-overlapping 8x8 pixel grids. In each 8x8 grid, find the maximum and minimum gray values of all 64 pixels, subtract the minimum value from the maximum value to get the range of gray values in the grid, then divide the range by the current grid edge length 8 and take the integer part to get the number of cubes needed to cover the gray surface in the grid area. Add up the number of cubes calculated by all 8x8 grids to get the total number of cubes needed to cover the entire sub-region image gray surface under the current grid size. Repeat this calculation process for all four grid sizes to finally generate a sequence of four coverage numbers for each sub-region image, and obtain the coverage sequence.

[0059] In the local surface fractal dimension value calculation formula, the local surface fractal dimension is calculated by least squares linear regression, which is the slope of the straight line describing the linear relationship between the scale and the coverage number in the logarithmic coordinate system. Fractal theory points out that for a fractal surface, the number of small boxes needed to cover it and the size of the small box have a power law relationship, that is where is the fractal dimension, taking the logarithm of both sides of the equation gives which shows that and have a linear relationship, and the slope is the fractal dimension The formula is used to calculate the fractal dimension of the local surface at multiple scales ) and the corresponding number of coverage ( ) to form a plurality of data points, and the best slope is fitted by linear regression This method is more robust than using only two points to calculate the slope.

[0060] The acquisition step is, representing the number of scales used to calculate the fractal dimension, is determined according to the dimension of the coverage number sequence. In the previous step, a set of cubic grid sizes is set for each sub-region image, and the coverage number under each size is calculated. The number of this set of sizes is the value of In this embodiment, the side length of the cubic grid is selected as 2, 4, 8, and 16 pixels, a total of 4 different scales, so The value of is 4. The selection of this value needs to balance the calculation accuracy and efficiency. If the number of selected scales is too small, for example, only 2, it will lead to unstable fitting of linear regression and be easily affected by noise. If the number of selected scales is too large, it will increase a large amount of calculation burden. Moreover, when the grid size is too large (close to the size of the sub-region image) or too small (close to the pixel size), the power-law relationship may no longer hold, leading to distortion of the fitting. Through experimental analysis of a large number of leaf repair samples, it is found that selecting 4 to 6 scales spanning an order of magnitude for calculation can ensure the accuracy of the estimation of the fractal dimension while maintaining high calculation efficiency. Therefore, the value of is set to 4.

[0061] The acquisition step is, is the independent variable in the linear regression model, and its calculation formula is wherein is the normalized dimensionless value of the th cubic grid size , and the specific method is to divide it by a unit reference length , for example pixels, so as to obtain the dimensionless ratio In this embodiment, the grid sizes are set to pixels, pixels, pixels, and pixels, respectively. After normalization, the values of , , , are obtained, and then a group of corresponding values are calculated:

[0062] ;

[0063] ;

[0064] ;

[0065] .

[0066] The acquisition step is, is the dependent variable in the linear regression model, and its calculation formula is , wherein is obtained from the coverage number sequence obtained from the previous step, representing the coverage number of the i-th sub-area image under the j-th grid size, for example, for the i-th sub-area, the four coverage numbers obtained from the coverage number sequence are , , , , , , , The calculation process of the corresponding value is:

[0067] ;

[0068] ;

[0069] ;

[0070] .

[0071] This set of values is specific to the first sub-area, and each sub-area will have its own independent set of values. According to the parameters, the calculation is:

[0072] Take the calculation of the local surface fractal dimension value of the i-th sub-area as an example, and input the obtained parameter value:

[0073] ;

[0074] ;

[0075] The sequence is ;

[0076] The sequence is ;

[0077] First, calculate the cumulative sum of each term required in the formula:

[0078] ; ​​​

[0079] ;

[0080] ;

[0081] ;

[0082] The above cumulative sum is brought into the formula:

[0083] ;

[0084] ;

[0085] The results show that the local surface fractal dimension value of the first sub-region is 2.651, which is between 2 and 3, reflecting that the surface of this region has high complexity and irregularity, and may correspond to a region with a small flaw or a region with a sharp change in texture in the repair area. Compared with the normal region (for example, the fractal dimension value is 2.2), the surface of this region is more rough. By performing the same calculation on all sub-regions, a complete sequence of local surface fractal dimension values can be obtained.

[0086] The steps for obtaining the leaf surface fractal dimension map are as follows:

[0087] According to the sequence of local surface fractal dimension values, the top-left row coordinate and the top-left column coordinate of the sub-region image corresponding to each local surface fractal dimension value in the coordinate repair area sub-image sequence are retrieved, each local surface fractal dimension value is written to the corresponding position, and the starting row coordinate and the starting column coordinate are sequentially organized and output to generate the leaf surface fractal dimension map.

[0088] Specifically, according to the sequence of local surface fractal dimension values, for each calculated local surface fractal dimension value, the index of the value in the sequence is used to call back the sub-region information with the same index in the coordinate repair area sub-image sequence, and the top-left row coordinate and the top-left column coordinate of the sub-region in the normalized repair area image are extracted. Next, a two-dimensional floating-point matrix with the same size as the normalized repair area image and all initial values of zero is created, which will serve as the carrier of the leaf surface fractal dimension map. Then, each value in the sequence of local surface fractal dimension values is traversed, for example, for the first fractal dimension value in the sequence and its corresponding top-left coordinates , since the sub-regions are overlapped and cropped, a pixel point may be covered by multiple sub-regions, so The writing of a single coordinate point will lose information and result in a non-smooth result, therefore a weighted average strategy is adopted to generate the atlas, specifically, two auxiliary matrices of the same size as the atlas are created, one for accumulating the fractal dimension values (accumulation matrix), and the other for recording the number of times each pixel is covered (count matrix), for each fractal dimension value , it is added to the accumulation matrix to all the pixels in the region with the upper left corner at 64x64, while the values of all the pixels in the corresponding region in the count matrix are added by 1, after all the local surface fractal dimension values are traversed, the atlas matrix is finally processed, the value of each pixel position is set to the value of the corresponding position in the accumulation matrix divided by the value of the corresponding position in the count matrix, through this pixel-by-pixel average method, a smooth and accurate fractal dimension value is obtained for each pixel point, finally, this two-dimensional matrix filled with fractal dimension values is organized according to its row and column order to generate the leaf surface fractal dimension atlas.

[0089] The acquisition steps of the region joint feature description vector are:

[0090] According to the multi-dimensional texture feature set of the repair area and the leaf surface fractal dimension atlas, the second-order statistics of contrast, energy, correlation and entropy and the local surface fractal dimension value are spliced into a column vector in a fixed order according to the alignment of the sub-region coordinates one by one, to generate the region joint feature description vector.

[0091] Specifically, according to the multi-dimensional texture feature set of the repair region and the fractal dimension atlas of the blade surface, a data alignment and feature fusion process is started. The process traverses each sub-region index in the sequence of sub-regions with coordinates of the repair region. First, the 17-dimensional multi-dimensional texture description vector corresponding to the sub-region is extracted from the multi-dimensional texture feature set of the repair region using the index information of the sub-region. The first four elements of the vector are the second-order statistics of the required contrast, energy, correlation and entropy. Then, according to the same index, the center row coordinate and the center column coordinate of the sub-region in the normalized repair region image are obtained. The coordinates are used as indexes to query the fractal dimension atlas of the blade surface. Since the fractal dimension atlas of the blade surface is a two-dimensional matrix as large as the original image, each pixel point stores a smoothed fractal dimension value. Therefore, by directly reading the value at the center coordinate position, the local surface fractal dimension value representing the surface complexity of the sub-region can be obtained. For example, for the sub-region with index 100, the center coordinates are (80, 112). The texture vector is extracted from the multi-dimensional texture feature set of the repair region to obtain a contrast of 0.75, an energy of 0.52, a correlation of 0.91, and an entropy of 2.33. At the same time, the value read at the 80th row and 112th column of the fractal dimension atlas of the blade surface is 2.68. Then, the five values are concatenated in a fixed order to form a five-dimensional column vector. The order is: the first dimension is the contrast, the second dimension is the energy, the third dimension is the correlation, the fourth dimension is the entropy, and the fifth dimension is the local surface fractal dimension value. In the aforementioned example, the generated column vector is the transpose of [0.75, 0.52, 0.91, 2.33, 2.68]. The alignment, extraction and concatenation process is repeated for all sub-regions to generate a five-dimensional feature vector for each sub-region. Finally, all the feature vectors of the sub-regions are combined to generate a regional joint feature description vector.

[0092] The flaw positioning map acquisition step is:

[0093] According to the regional joint feature description vector, the feature key degree modulation Mahalanobis distance value is calculated, and the calculation formula is:

[0094] ;

[0095] wherein, is the feature key degree modulation Mahalanobis distance value of the i-th sub-region, is the regional joint feature description vector of the i-th sub-region, is the preset superior product reference center vector, is the covariance matrix of the preset superior product reference center, is the inverse matrix of is the i-th sub-region, is the i-th sub-region, is the i-th sub-region, is the i-th sub-region. The components of the joint feature description vector of each sub-region that correspond to the local surface fractal dimension value. This refers to the components of the local surface fractal dimension value in the preset benchmark center vector of the superior product. The standard deviation of the fractal dimension values ​​of the local surface. Modulation intensity coefficient;

[0096] Modulate Mahalanobis distance values ​​based on feature criticality, statistically analyze quantiles and mean values, construct comparison rules, determine each sub-region according to a preset threshold, output the overall quantization score, extract the coordinates of sub-regions exceeding the preset threshold and write them into the labeling layer, and generate a defect location map.

[0097] Specifically, in the formula for calculating the modulated Mahalanobis distance value of feature criticality, the standard Mahalanobis distance (the square root part of the formula) can effectively measure the distance from a point to a distribution center and consider the correlation between features. However, it treats deviations from all feature dimensions equally. In the repair and inspection of aero-engine blades, the surface geometric complexity represented by the local surface fractal dimension is a decisive indicator for judging key defects such as cracks and porosity. Its deviation from the normal range often indicates serious quality problems. Therefore, this formula introduces a modulation term (the multiplicative part of the formula), which specifically amplifies the degree of deviation of the local surface fractal dimension value through the hyperbolic tangent function. The deviation value is smoothly mapped to a range of 0 to 1, and then the modulation intensity coefficient is used. Controlling its amplification weights, the larger the difference between the fractal dimension and the reference value, the larger the value of the modulation term, thus increasing the total distance value. It is more sensitive to changes in surface complexity.

[0098] The steps to obtain it are as follows: Representing the The joint feature description vector for each sub-region is directly generated from the previous step. It is a five-dimensional column vector containing the texture and surface complexity information of the sub-region. Its five components are contrast, energy, correlation, entropy, and local surface fractal dimension, all of which are dimensionless pure numbers. For example, for a sub-region to be detected... The calculated joint feature description vector of the region is This vector will be used as input for calculating the distance value.

[0099] The steps to obtain it are as follows: The pre-defined high-quality benchmark center vector is obtained through statistical analysis of a large number of qualified repaired blade samples. The specific acquisition steps are as follows: First, at least 100 images of the repaired areas of aero-engine blades, confirmed as flawless by experts or non-destructive testing techniques, are collected as a high-quality sample library. Then, all steps from image acquisition to the generation of the region joint feature description vector, as described in this application, are performed on these 100 sample images, resulting in a high-quality feature set containing tens of thousands (hundreds per sample image) of five-dimensional region joint feature description vectors. Finally, the average value of all vectors in this high-quality feature set is calculated component by component, and the resulting five-dimensional vector is the high-quality feature set. For example, after statistical calculation, the benchmark center vector of the superior product is obtained as follows: .

[0100] The steps to obtain it are as follows: The covariance matrix representing the preset benchmark center is related to the benchmark center vector. Together, they were calculated from the set of superior product features. Then, the joint feature description vector of each region in the premium feature set is used. and mean vector Calculate the covariance matrix The element at position (m, n) The covariance matrix is ​​calculated by multiplying the difference between the m-th feature component and the mean of all samples by the average of the differences between the n-th feature component and the mean. This 5x5 matrix describes the variance of the five feature dimensions in the premium sample and the pairwise correlations between them. For example, the calculated covariance matrix is:

[0101] .

[0102] The steps to obtain it are as follows: The standard deviation of the local surface fractal dimension values ​​in the premium feature set is also a statistic calculated when constructing the premium benchmark. Specifically, the fifth component (i.e., the local surface fractal dimension values) of the joint feature description vector of all regions in the premium feature set is extracted to form a one-dimensional array. Then, the standard deviation of this array is calculated. This value reflects the normal fluctuation range of the fractal dimension of qualified repaired surfaces. For example, after calculation, the standard deviation of the fractal dimension of the premium sample is: .

[0103] The steps to obtain it are as follows: To modulate the intensity coefficient, the value of which determines the amplification degree of the final distance by the deviation of the fractal dimension, the coefficient is determined by optimizing the experiment on the verification set containing known small defects. The specific method is: prepare a verification set containing 50 sub-regional samples with small (just above the acceptance standard) defects and 50 excellent sub-regional samples, set the candidate value range of , for example , for each candidate value, calculate the value of all samples in the verification set, then calculate the ratio of the average distance of the defect samples to the average distance of the excellent samples, select the that makes the ratio the largest as the final value, for example, when , the average distance of the defect samples is 4.5, the average distance of the excellent samples is 1.2, the ratio is 3.75, which is the highest among all candidate values, so set .

[0104] According to the parameters, calculate:

[0105] Take the first sub-region as an example, and input the obtained parameter values:

[0106] ;

[0107] ;

[0108] , ;

[0109] , ;

[0110] Calculate the inverse matrix of the covariance matrix ;

[0111] Calculate the square term of Mahalanobis distance ;

[0112] Calculate the modulation term ;

[0113] Calculate the final feature key modulation Mahalanobis distance value :

[0114] ;

[0115] The result shows that the feature key modulation Mahalanobis distance value of the 50th sub-region is 13.66, which indicates that the feature vector of the sub-region has a great difference with the excellent benchmark center after considering the correlation between features, strongly indicating that the sub-region has defects.

[0116] ​According to the characteristic key degree modulation Mahalanobis distance value, firstly, the distance value set of all sub-regions of the current to-be-detected blade repair area is statistically analyzed, the mean value and the standard deviation of the set are calculated, then, a dynamic comparison rule is constructed based on the statistical quantities to set the threshold value of the defect judgment, the setting method of the threshold value is to take the mean value of the distance value plus 3 times the standard deviation, for example, for a repair area, the distance mean value of all sub-regions is calculated as 1.8, and the standard deviation is 0.9, then the defect judgment threshold value of the repair area is set as After setting the threshold value, each sub-region is traversed, and the corresponding characteristic key degree modulation Mahalanobis distance value is compared with the threshold value, if the distance value is greater than 4.5, the sub-region is judged as a defect region, otherwise it is a qualified region, after the judgment of all sub-regions is completed, a whole quantitative score is calculated, the calculation method of the score is that the number of qualified sub-regions accounts for the percentage of the total number of sub-regions, for example, there are 2500 sub-regions in total, and 50 of them are judged as defect regions, then the whole quantitative score is Finally, the left upper corner coordinates and size information of all sub-regions judged as defect regions in the subgraph sequence of the repair area with coordinates are extracted, a transparent marking layer equal to the original repair region image is created, for each extracted defect region coordinates, a semi-transparent red rectangle is drawn on the corresponding position of the marking layer, the color depth of the rectangle can be proportional to the size of the distance value, that is, the larger the distance value, the deeper the red color, finally, the marking layer is superimposed on the original repair region image to generate a defect positioning map.

[0117] The aviation engine blade repair performance intelligent detection and evaluation system further comprises an internal defect CT detection module configured to identify internal defects of the blade according to data obtained by CT scanning of the to-be-detected aviation engine blade repair area.

[0118] Specifically, according to the data obtained by CT scanning of the repair area of the aero-engine blade to be detected, the data is reconstructed into a three-dimensional voxel data set, wherein the gray value of each voxel represents the material density of the corresponding spatial position, then a three-dimensional median filter with a 3x3x3 kernel is applied to the three-dimensional voxel data set for processing to eliminate the salt and pepper noise generated in the CT scanning process, and then a three-dimensional U-Net convolutional neural network model is used to perform voxel-by-voxel semantic segmentation on the processed voxel data, the network structure of the three-dimensional U-Net model includes an encoder path and a decoder path, the encoder path is composed of four consecutive down-sampling modules, each module includes two 3x3x3 convolution layers, each convolution layer is followed by a ReLU activation function, and finally a 2x2x2 max-pooling layer, the decoder path also includes four up-sampling modules, each module is up-sampled by 2x2x2 deconvolution, and is connected and spliced with the feature map of the corresponding level in the encoder path, and then passes through two 3x3x3 convolution layers and ReLU activation functions, finally the network outputs the probability of each voxel belonging to each of the three categories of background, qualified base material or internal defect through a 1x1x1 convolution layer and a Softmax activation function, the training process of the model is based on a database of more than 500 CT scanning samples of blade repair areas verified by metallographic analysis, tens of thousands of 32x32x32 size voxel blocks and their corresponding expert manually labeled class labels are extracted from these samples as a training set, classification cross-entropy is used as the loss function, and the Adam optimizer is used to iteratively optimize the network weights with an initial learning rate of 0.0001 until convergence, in the identification stage, the three-dimensional voxel data set to be detected is divided into overlapping 32x32x32 voxel blocks by the sliding window method and input into the trained model for inference, and the output probability atlas is spliced and restored to obtain the voxel-level classification result of the entire repair area, then the segmentation result is post-processed, all adjacent voxels marked as defects are aggregated into independent defect objects by applying a three-dimensional connected component labeling algorithm, and the geometric and physical characteristics of each defect object are calculated, including its volume, centroid three-dimensional coordinates and sphericity, then each defect object is evaluated according to a set of preset defect judgment rules, the rules are based on aviation manufacturing standards, for example, set the volume threshold to 0.01 cubic millimeters, any defect object with a volume less than this threshold is classified as an acceptable small pore and ignored, and any defect object with a volume greater than this threshold is further judged according to its sphericity, set a sphericity threshold of 0.6, if the sphericity of the defect is greater than 0.6, it is judged as a spherical pore, and if it is less than 0.6, it is a more dangerous crack or un-melted defect that needs to be further marked, and finally the internal defects of the blade are identified.

Claims

1. An intelligent detection and evaluation system for the repair performance of aero-engine blades, characterized in that, The system includes: The aircraft blade image acquisition module is used to acquire images of the repair area of ​​the aircraft engine blade to be inspected, perform illumination normalization on the images, and generate a sequence of repair area sub-images with coordinates. The cladding texture feature extraction module is used to calculate the gray-level co-occurrence matrix of each sub-region image in the coordinate-based repair area sub-image sequence, extract second-order statistics of contrast, energy, correlation and entropy from the gray-level co-image, establish a sub-region texture statistical vector set, perform a two-dimensional fast Fourier transform on each sub-region image in the coordinate-based repair area sub-image sequence, and fuse it with the sub-region texture statistical vector set to obtain a multi-dimensional texture feature set of the repair area. The surface complexity map generation module is used to perform Gaussian smoothing on each sub-region image based on the coordinated repair area sub-image sequence, count the number of cubic meshes of different sizes required to cover the gray-scale surface, obtain the local surface fractal dimension value sequence, and map each local surface fractal dimension value back to the corresponding aero-engine blade surface position based on the spatial position coordinates provided by the coordinated repair area sub-image sequence, thereby generating a blade surface fractal dimension map. The defect repair comprehensive evaluation module is used to combine the texture features and fractal dimension values ​​of each sub-region according to the multi-dimensional texture feature set of the repair area and the fractal dimension map of the blade surface, construct the region joint feature description vector, calculate the distance between the joint feature description vector of each region and the preset excellent benchmark center, and obtain the defect location map. The steps for obtaining the sub-region texture statistical vector set are as follows: Based on the coordinate-based repair area sub-image sequence, each sub-region image is read one by one, the pixel grayscale values ​​are quantized and graded, the number of adjacencies between different grayscale levels is counted, grayscale co-occurrence matrices are established in the horizontal, vertical, main diagonal, and secondary diagonal directions, the matrices are symmetricized and normalized, the contrast, energy, correlation, and entropy values ​​of each sub-region image are extracted, and a set of sub-region texture statistical vectors is generated. The steps for obtaining the joint feature description vector of the region are as follows: Based on the multidimensional texture feature set of the repair area and the fractal dimension map of the blade surface, the coordinates of each sub-region are aligned and recorded one by one. The second-order statistics of contrast, energy, correlation and entropy and the local surface fractal dimension values ​​are concatenated into a column vector in a fixed order to generate a joint feature description vector of the region.

2. The intelligent detection and evaluation system for the repair performance of aero-engine blades according to claim 1, characterized in that, The steps for obtaining the coordinate-based repair area sub-map sequence are as follows: The original image input is parsed, the pixel boundaries of the aero-engine blade repair area are located, the aero-engine blade repair area image is cropped, the pixel mean and standard deviation of each color channel are calculated, and the normalized repair area image is obtained by linear scaling according to the difference between the pixel and channel mean. Based on the normalized repaired region image, the size and step size of the sub-region are set, and the sub-regions are cropped row by row and column by column with a fixed step size along the horizontal and vertical axes. The starting row index and starting column index of each sub-region are recorded to generate a set of overlapping sub-region images. Based on the set of overlapping sub-region images, calculate the top-left row coordinates and top-left column coordinates of each sub-region image in the normalized repaired region image, and calculate the center row coordinates and center column coordinates. Construct an index mapping from each sub-region image to the corresponding coordinates, sort them by starting row index in ascending order and then by starting column index in ascending order, and generate a sequence of repaired region sub-images with coordinates.

3. The intelligent detection and evaluation system for the repair performance of aero-engine blades according to claim 1, characterized in that, The steps for obtaining the multidimensional texture feature set of the repair area are as follows: Based on the set of texture statistics vectors of the sub-regions, sub-region images with the same index are called, and a two-dimensional fast Fourier transform is performed on each sub-region image to convert the spatial domain grayscale information into a spectrum amplitude map. Energy values ​​of different radius ranges in the spectrum amplitude map are extracted, radial energy distribution is calculated, and then the energy values ​​of different angle ranges are partitioned and integrated to calculate the angular energy distribution and generate a set of frequency domain energy features. Based on the frequency domain energy feature set, the texture statistical vector set of the sub-region with the same index is matched, and the contrast, energy, correlation, entropy, radial energy distribution and angular energy distribution are merged according to the corresponding position to form a multi-dimensional texture description vector for each sub-region. The multi-dimensional texture description vectors of all sub-regions are summarized to generate a multi-dimensional texture feature set for the repair area.

4. The intelligent detection and evaluation system for the repair performance of aero-engine blades according to claim 1, characterized in that, The steps for obtaining the local surface fractal dimension numerical sequence are as follows: Based on the coordinate-based repair area sub-image sequence, each sub-region image is smoothed by discrete Gaussian kernel convolution, the kernel size and standard deviation are set, the boundary is mirror-filled, the image is divided into blocks according to the cube grid size, and the number of each block is accumulated to obtain the coverage number sequence. Based on the coverage quantity sequence, the local surface fractal dimension value is calculated to obtain the local surface fractal dimension value sequence.

5. The intelligent detection and evaluation system for the repair performance of aero-engine blades according to claim 1, characterized in that, The steps for obtaining the fractal dimension map of the blade surface are as follows: Based on the local surface fractal dimension numerical sequence, the top-left row coordinates and top-left column coordinates of the sub-region image corresponding to each local surface fractal dimension value are retrieved from the coordinate-bound repair area sub-image sequence. Each local surface fractal dimension value is written to the corresponding position, and the output is organized in the order of the starting row coordinates and the starting column coordinates to generate the blade surface fractal dimension map.

6. The intelligent detection and evaluation system for the repair performance of aero-engine blades according to claim 1, characterized in that, The steps for obtaining the defect location map are as follows: Based on the joint feature description vector of the region, calculate the feature key modulated Mahalanobis distance value; based on the feature key modulated Mahalanobis distance value, count the quantiles and mean, construct comparison rules, judge each sub-region according to the preset threshold, output the overall quantization score, extract the coordinates of sub-regions exceeding the preset threshold and write them into the labeling layer to generate a defect location map.

Images