Metal fracture identification method and device for three-dimensional diagram learning, medium and equipment

Abstract

The invention belongs to the technical field of metal fracture recognition, and provides a metal fracture recognition method and device based on three-dimensional diagram learning, a medium and equipment. The method comprises the following steps: smoothing a three-dimensional image of a metal fracture profile to obtain a smooth curved surface; reserving the part above the minimum value of the wave trough of the three-dimensional image of the metal fracture profile through cutting operation, and then performing projection and flattening operation to obtain a plurality of sections; taking each cross section as a node, extracting the cross section feature of the corresponding connected region and the fractal dimension feature of each connected region from each cross section, and further constructing a node feature matrix and an adjacent matrix corresponding to the reserved three-dimensional image part of the metal fracture profile; and identifying the type of the metal fracture based on the corresponding node feature matrix, the adjacent matrix and a pre-trained graph convolutional network. The microstructure change in the metal fracture process and the relation between the microstructure change and the fracture mechanism can be revealed, and a scientific basis is provided for material design and engineering application.

Description

Technical Field

[0001] The present invention belongs to the technical field of metal fracture recognition, and in particular relates to a metal fracture recognition method, device, medium and equipment based on three-dimensional graph learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Failure fracture analysis of metal components is an important means to study the properties and performance of materials and their performance during service. The fracture morphology not only reflects the internal structure and composition of the material, but also records the stress state and failure fracture mode during the fracture process. It is the key evidence for qualitative analysis and quantitative characterization of metal fracture. Fracture morphology analysis is of great significance for understanding the fracture mechanism of materials, improving material performance and designing new materials.

[0004] Traditional fracture analysis mainly relies on scanning electron microscopy (SEM) to obtain two-dimensional images. However, this method has limitations in obtaining depth information inside the fracture, and the fracture morphology area captured is small, making it difficult to fully reveal the complexity of the fracture morphology. Although two-dimensional image processing can provide valuable information in some aspects, its shortcomings in depth information and three-dimensional structure limit the comprehensive understanding of the fracture mechanism. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a metal fracture recognition method, device, medium and equipment based on three-dimensional graph learning, which can reveal the microstructural changes in the metal fracture process and the relationship between it and the fracture mechanism, and provide a scientific basis for material design and engineering applications.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A first aspect of the present invention provides a metal fracture recognition method based on three-dimensional graph learning.

[0008] In one or more embodiments, a metal fracture recognition method based on three-dimensional graph learning is provided, comprising:

[0009] Obtain three-dimensional images of metal fracture sections;

[0010] Smoothing the three-dimensional image of the metal fracture section to obtain a smooth curved surface;

[0011] The smooth curved surface is moved to the position of the lowest value of the trough of the three-dimensional image of the metal fracture section to perform a clipping operation, and a portion above the lowest value of the trough of the three-dimensional image of the metal fracture section is retained;

[0012] Project and flatten the retained three-dimensional image portion of the metal fracture section relative to the same plane to obtain a plurality of cross sections;

[0013] Taking each section as a node, extracting the corresponding cross-sectional features of the connected area and the fractal dimension features of each connected area from each section, and then constructing the node feature matrix and the neighboring matrix corresponding to the retained three-dimensional image part of the metal fracture section;

[0014] The metal fracture type is identified based on the corresponding node feature matrix, proximity matrix and the pre-trained graph convolutional network.

[0015] As an implementation mode, the retained three-dimensional image portion of the metal fracture section is projected and flattened at equal intervals relative to the same plane.

[0016] As an implementation method, the minimum energy unfolding method is used to project and flatten the retained three-dimensional image portion of the metal fracture section relative to the same plane, so that the angle and area characteristics of the original surface are maintained when the surface is flattened.

[0017] As an implementation method, the minimum energy unfolding method is: constructing an energy function to minimize the deformation generated during the unfolding process, and obtaining the coordinates of each point on the surface on the two-dimensional plane by solving the minimum value of the energy function.

[0018] As an implementation method, the expression of the energy function is:

[0019]

[0020] Where E is a surface; u and v are the coordinates of all vertices on the two-dimensional plane, (i, j) represents the edge on the surface E; x and y represent the horizontal and vertical directions respectively.

[0021] As an implementation manner, the smoothing operation includes:

[0022] Noise filtering and surface smoothing operations are performed sequentially.

[0023] As an implementation method, the fractal dimension features include: voxel density, geometric morphology features, fractal dimension, surface roughness and local geometric features;

[0024] The cross-sectional features of the connected region include perimeter, area and shape features.

[0025] The second aspect of the present invention provides a metal fracture recognition device based on three-dimensional graph learning.

[0026] In one or more embodiments, a metal fracture recognition device based on three-dimensional graph learning includes:

[0027] A section three-dimensional image acquisition module, which is used to acquire a three-dimensional image of a metal fracture section;

[0028] A three-dimensional image smoothing processing module, which is used to smooth the three-dimensional image of the metal fracture section to obtain a smooth curved surface;

[0029] A section three-dimensional image cropping module, which is used to move the smooth curved surface to the position of the lowest value of the trough of the metal fracture section three-dimensional image to perform a cropping operation, and retain the portion above the lowest value of the trough of the metal fracture section three-dimensional image;

[0030] A three-dimensional image projection flattening module is used to project and flatten the retained three-dimensional image portion of the metal fracture section relative to the same plane to obtain a plurality of cross sections;

[0031] A node feature matrix construction module is used to take each section as a node, extract the corresponding connected area section features and the fractal dimension features of each connected area from each section, and then construct a node feature matrix and a neighbor matrix corresponding to the retained metal fracture section three-dimensional image portion;

[0032] The metal fracture type recognition module is used to identify the metal fracture type based on the corresponding node feature matrix and proximity matrix and a pre-trained graph convolutional network.

[0033] A third aspect of the present invention provides a computer-readable storage medium.

[0034] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the metal fracture recognition method based on three-dimensional graph learning as described above.

[0035] A fourth aspect of the present invention provides an electronic device.

[0036] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the metal fracture recognition method based on three-dimensional graph learning as described above are implemented.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] (1) The present invention intercepts the portion above the lowest value of the trough of the three-dimensional image of the metal fracture profile, and then performs a projection flattening operation relative to the same plane. The cross-sectional features of the connected areas of the projected flattened section and the fractal dimension features of each connected area are extracted, and then aggregated to construct a node feature matrix. Finally, the graph convolutional network is combined to realize accurate identification and classification of metal fractures, revealing the microstructural changes in the metal fracture process and their relationship with the fracture mechanism, providing a scientific basis for material design and engineering applications.

[0039] (2) The present invention smoothes the fracture surface, moves the fracture surface to the plane reference line by calculating the lowest value of the trough, and then performs projection flattening operations, thereby better presenting the details of the fracture, providing an accurate data basis for constructing a node feature matrix, and ultimately improving the accuracy of metal fracture recognition and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0041] Figure 1 It is a schematic flow chart of a metal fracture recognition method based on three-dimensional graph learning according to an embodiment of the present invention;

[0042] Figure 2 It is a metal fracture profile image;

[0043] Figure 3 It is a schematic diagram of the reserved metal fracture profile image section being cut at equal intervals;

[0044] Figure 4 It is a node schematic diagram;

[0045] Figure 5 It is a schematic diagram of the structure of a metal fracture recognition device using three-dimensional graph learning according to an embodiment of the present invention;

[0046] Figure 6 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0049] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0050] Figure 1 is a flow chart of a metal fracture recognition method based on three-dimensional graph learning in an embodiment of the present invention, such as Figure 1 The metal fracture recognition method using three-dimensional graph learning in the embodiment shown may include:

[0051] S101, obtaining a three-dimensional image of a metal fracture section;

[0052] S102, smoothing the three-dimensional image of the metal fracture section to obtain a smooth curved surface;

[0053] S103, moving the smooth curved surface to the position of the lowest value of the trough of the three-dimensional image of the metal fracture section to perform a clipping operation, and retaining the portion above the lowest value of the trough of the three-dimensional image of the metal fracture section;

[0054] S104, performing a projection and flattening operation on the retained three-dimensional image portion of the metal fracture section relative to the same plane to obtain a plurality of cross sections;

[0055] S105, taking each cross section as a node, extracting the cross-sectional features of the corresponding connected regions and the fractal dimension features of each connected region from each cross section, and then constructing a node feature matrix and a neighbor matrix corresponding to the retained three-dimensional image portion of the metal fracture section;

[0056] S106, identifying the metal fracture type based on the corresponding node feature matrix and the proximity matrix and the pre-trained graph convolutional network.

[0057] In industrial production, fracture analysis is an important part of quality control. By applying three-dimensional morphology technology and fractal theory, it is possible to more accurately monitor and analyze fracture problems that occur during the production process, timely discover and solve material defects, and ensure the stability and reliability of product quality. The microstructure of metal fracture has self-similar characteristics, which is consistent with the concept of statistical self-similarity in fractal theory. By introducing fractal theory, the fracture of metal components can be quantitatively analyzed to reveal the relationship between fractal dimension and material microstructure information. As an important parameter in fractal theory, fractal dimension can reflect the complexity and fracture characteristics of the fracture.

[0058] This embodiment intercepts the part above the lowest value of the trough of the three-dimensional image of the metal fracture profile, and then performs a projection flattening operation relative to the same plane. By extracting the cross-sectional features of the connected areas of the projected flattened section and the fractal dimension features of each connected area, a node feature matrix is ​​constructed by aggregation. Finally, a graph convolutional network is combined to realize accurate recognition and classification of metal fractures, revealing the microstructural changes in the metal fracture process and their relationship with the fracture mechanism, providing a scientific basis for material design and engineering applications.

[0059] In step S101, a three-dimensional image of a metal fracture section can be obtained by a three-dimensional stereo microscope, such as Figure 2 shown.

[0060] In step S102, the smoothing operation includes: performing a noise filtering operation and a surface smoothing operation in sequence.

[0061] Specifically, a variety of filtering methods are used including mean filtering, median filtering, etc. The calculation method is as follows:

[0062] Mean filtering is a smoothing method that reduces noise by updating the position of each vertex to the average of the positions of its neighboring vertices. The specific operation is as follows:

[0063] For each vertex P i , calculate its neighborhood vertex P j The position average of:

[0064]

[0065] Where N represents the number of neighboring vertices, P j Represents vertex P i This operation can effectively smooth out sharp burrs, but it may also cause the surface to become too smooth.

[0066] Median filtering is a nonlinear filtering technique that reduces noise by updating the position of each vertex to the median of its neighboring vertex positions. The specific operation is as follows:

[0067] For each vertex P i , calculate its neighborhood vertex P j The median position of:

[0068] P′ i = median(P j )

[0069] Among them, median(P j ) represents the vertex P i The median filter effectively removes noise and burrs while retaining surface details.

[0070] Surface smoothing further improves surface quality by removing noise and burrs, making the surface smoother and more continuous, thereby improving the overall quality and application effect of the surface.

[0071] In one or more embodiments, multiple smoothing methods are used, including Laplacian smoothing, Taubin smoothing, and bilateral smoothing.

[0072] Laplacian smoothing is a smoothing method based on differential geometry. It smoothes the surface by selectively adjusting the position of the vertex to make it closer to the average position of its neighboring vertices.

[0073] The specific operations are as follows:

[0074] For each vertex P i , calculate the average of its neighborhood vertex positions and adjust the vertex positions:

[0075]

[0076] Among them, λ is a smoothing factor, which usually takes a value between 0 and 1. By iterating this process multiple times, the entire surface can be gradually smoothed.

[0077] Taubin smoothing is an improvement on Laplace smoothing. It reduces the volume shrinkage effect by performing reverse smoothing after Laplace smoothing. The specific operation is as follows:

[0078] (1) First, perform a standard Laplace smoothing:

[0079]

[0080] (2) Then perform a reverse smoothing:

[0081]

[0082] Among them, μ usually takes a negative value to offset the volume shrinkage effect caused by the first smoothing.

[0083] Bilateral smoothing:

[0084] Bilateral smoothing combines spatial distance and normal vector similarity to smooth the surface while preserving important geometric details and edges. The specific operation is as follows:

[0085] For each vertex P i , calculate its neighborhood vertex P j The weighted average of:

[0086]

[0087] Among them, W c is the spatial distance weight, defined as:

[0088]

[0089] W s is the normal vector similarity weight, defined as:

[0090]

[0091] σ c and σ s is a parameter that controls weight decay.

[0092] This embodiment uses mean filtering and median filtering to denoise and smooth the fracture morphology data to improve data quality. Subsequently, the three-dimensional calibration technology is applied to smooth the fracture surface based on multiple methods such as Laplace smoothing, Taubin smoothing and bilateral smoothing. The calculation of the fracture surface using the subsequent trough minimum value can better present the details of the fracture, provide an accurate data basis for constructing the node feature matrix, and ultimately improve the recognition and classification accuracy of metal fractures.

[0093] In step S103, the smoothed surface is moved to the position of the lowest value of the trough of the three-dimensional image data of the metal fracture, and a clipping operation is performed, that is, the three-dimensional image of the non-smooth metal fracture is subtracted from the three-dimensional pixels of the smooth surface and the position below.

[0094] First, assuming that the height function of the surface is z = f(x, y), we need to find the lowest point of the surface, that is, the trough. The lowest height value of the trough can be found using the following formula: min =min(f(x,y));

[0095] Then, the entire surface is moved downward so that the height of the trough becomes zero. The new surface height function can be expressed as: result =f(x,y)-z min .

[0096] In step S104, the retained three-dimensional image portion of the metal fracture section is flattened with equal interval projection relative to the same plane.

[0097] Specifically, flattening the smooth curved surface at the lowest value of the trough of the three-dimensional image of the metal fracture section;

[0098] The voxels corresponding to the three-dimensional image of the metal fracture section above the smooth surface are projected into the new three-dimensional space according to equal displacement, such as Figure 3 shown.

[0099] Assume that the smooth surface at the bottom of the voxel is S, the point on the surface is P(x, y, z), and the point after projection onto the two-dimensional plane is P ′ (u,v). The voxel point above is P i (x i ,y i ,z i ), the projected point is P i ′ (u i ,v i ,h i ). First, project the surface S onto a two-dimensional plane: (u,v)=f(x,y,z); then, project the voxel points above into the new three-dimensional space with equal displacement:

[0100] In this embodiment, the minimum energy unfolding method is used to project and flatten the retained three-dimensional image portion of the metal fracture section relative to the same plane, so that the angle and area characteristics of the original curved surface are maintained when the curved surface is flattened.

[0101] The minimum energy unfolding method is as follows: constructing an energy function to minimize the deformation generated during the unfolding process, and obtaining the coordinates of each point on the surface on the two-dimensional plane by solving the minimum value of the energy function.

[0102] As an implementation method, the expression of the energy function is:

[0103]

[0104] Where E is a surface; u and v are the coordinates of all vertices on the two-dimensional plane, (i, j) represents the edge on the surface E; x and y represent the horizontal and vertical directions respectively.

[0105] At the beginning of the optimization, the coordinates u0 and v0 of all vertices on the two-dimensional plane are initialized. The initial coordinates can be obtained by simple projection or other methods.

[0106] Perform iterative updates. In each iteration, the coordinates of all vertices are updated simultaneously. The specific steps are as follows:

[0107] (1) Calculate the gradient g of the energy function:

[0108] The gradient of the energy function is calculated for all vertex coordinates u and v. The gradient vector contains the partial derivatives of each vertex in the u and u directions.

[0109] (2) Calculate the search direction d:

[0110] Initially, the search direction is the negative direction of the gradient d0 = -g0.

[0111] In subsequent iterations, the search direction is updated according to the conjugate gradient method.

[0112] (3) Calculate the step length a:

[0113] The step length a is determined by minimizing the energy function along the search direction, as follows:

[0114]

[0115] Where H is the Hessian matrix of the energy function.

[0116] Update the vertex coordinates u and v, and update the coordinates of all vertices:

[0117] u k+1 =u k +α k d k

[0118] v k+1 =v k +α k d k

[0119] Calculate the new gradient g k+1 , recalculate the gradients for the coordinates of all vertices.

[0120] (4) Update search direction:

[0121] Update the search direction according to the conjugate gradient method:

[0122]

[0123] d k+1 =-g k+1 +β k d k

[0124] Check the convergence condition. If the norm of the gradient is less than the predetermined threshold, it is considered converged and the iteration is stopped.

[0125] When the algorithm converges, the coordinates (u, v) of all vertices are the optimal solution of the energy function, which means that the entire surface has been flattened and the coordinates of each vertex maintain the geometric characteristics of the original three-dimensional surface as much as possible.

[0126] When the algorithm converges, the coordinates (u, v) of all vertices are the optimal solution of the energy function. This means that the entire surface has been flattened, and the coordinates of each vertex maintain the geometric characteristics of the original three-dimensional surface as much as possible. These coordinates (u, v) are obtained by minimizing the energy function E, so that the unfolded surface maintains the geometric characteristics (angle and area) of the original three-dimensional surface as much as possible. The energy function E takes into account the geometric structure of the entire surface. The optimization process calculates different gradients and update directions based on the position of each vertex on the surface and the positions of adjacent vertices. In each iteration, the update amount α for each vertex k d k It is calculated based on the current gradient and search direction. The gradient and search direction of different vertices are different. In order to maintain the angle and area relationship of the original surface as much as possible, the optimization algorithm will adjust each vertex to different degrees according to the local characteristics of the surface, rather than simply making the same displacement.

[0127] The above steps cut the flattened three-dimensional voxels at equal intervals to obtain multiple cross-sectional images, and identify connected regions for each cross-sectional image. First, the cross-sectional images are binarized according to a threshold value to identify independent islands (connected regions).

[0128] Through the above steps, the three-dimensional voxels are projected and flattened. After the three-dimensional voxels are flattened to a two-dimensional plane, the island features can more easily identify and separate connected areas. The noise in the three-dimensional data may be suppressed after flattening to a two-dimensional plane, making the island features more stable and accurate when performing connected area analysis on the two-dimensional plane. The flattening operation helps to smooth the data and reduce the impact of noise on connectivity analysis. By flattening the surface by methods such as minimum energy expansion, the geometric characteristics of the original three-dimensional data (such as the relationship between angles and areas) can be maintained as much as possible. This means that the flattened two-dimensional data can still accurately reflect the structure of the original three-dimensional data, making the connected area analysis more reliable. The flattened data as above can more conveniently extract features on a two-dimensional plane, such as calculating the area, perimeter, shape characteristics, etc. of the connected area. These features are of great significance for subsequent classification, recognition and analysis.

[0129] Among them, the fractal dimension features include: voxel density (density value of each voxel area), geometric morphological features (such as length, width, height, voxel volume, etc.), fractal dimension (fractal dimension of each voxel area), surface roughness (surface roughness of each voxel area) and local geometric features (such as curvature, inclination angle, etc.); the cross-sectional features of the connected area include: perimeter (perimeter of each "island"), area (area of ​​each "island") and shape features (such as compactness and roundness of the shape, etc.).

[0130] Connected regions are identified for each cross-sectional image. First, the cross-sectional image is binarized according to a threshold value to identify independent islands (connected regions).

[0131] Use a connectivity algorithm (such as the flood fill algorithm) to mark all connected regions C i ;

[0132] Calculate the morphological characteristics of each connected region, such as area, perimeter, and shape eigenvalues;

[0133] Area A i :

[0134] Perimeter P i :

[0135] Shape characteristics can include shape factor (circularity, slenderness ratio, etc.)

[0136] Compute the geometric features of each connected region:

[0137] Compute geometric features of each connected region, such as centroid and principal axis orientation.

[0138] Center of gravity (u i ,v i ):

[0139]

[0140] Spindle direction:

[0141] Compute the covariance matrix for each connected region:

[0142]

[0143] The covariance matrix is ​​decomposed into eigenvalues, and the principal axis direction corresponds to the eigenvector with the largest eigenvalue.

[0144] Among them, u and v are the horizontal and vertical coordinate values ​​of each pixel in the connected domain.

[0145] In step S105, the flattened three-dimensional voxels are cut in parallel at equal intervals. The present invention cuts the flattened three-dimensional voxel data at multiple different intervals. The node features and adjacency matrix of the graph convolutional neural network are constructed for different cutting intervals, thereby constructing multiple different graph neural networks.

[0146] The features extracted from the fractal dimension and the island features are combined, and the feature vector of each node is composed of two parts of features. The features are standardized or normalized so that different features have the same scale, which is convenient for subsequent modeling and training.

[0147] For Figure 4 For each node ("island") shown, its feature vector can be expressed as:

[0148]

[0149] in:

[0150] Density i : Density value of each voxel area;

[0151] Geometric Properties i : Such as the geometric properties of voxels, such as length, width, height, and volume of voxels;

[0152] Fractal Dimension i : The fractal dimension of each voxel area;

[0153] Surface Roughness i : Surface roughness of each voxel area;

[0154] Local Geometric Features i : Local geometric features such as curvature and tilt angle;

[0155] Perimeter i :The circumference of each "island";

[0156] Area i :The area of ​​each "island";

[0157] 2D Fractal Dimension i : The two-dimensional fractal dimension of each "island";

[0158] Shape Features i :Such as shape characteristics such as compactness and roundness.

[0159] This vector x i It represents all the features of the i-th node, combining the features of the 3D voxel method and the island feature.

[0160] The adjacency matrix A is used to represent the connection relationship between nodes, and can be constructed based on the spatial proximity of the nodes. The present invention uses the method of constructing the adjacency matrix based on spatial proximity, but it does not mean that this is the only method to construct the adjacency matrix.

[0161] First, define the distance metric: Euclidean distance: a commonly used spatial distance metric. For the position coordinates p of node i and node j i and p j , the Euclidean distance is defined as: dij =∥p i -p i ∥.

[0162] To set the proximity threshold:

[0163] Threshold d threshold : If the distance between nodes is less than the threshold d threshold , they are considered neighbors.

[0164] Construct the adjacency matrix: The adjacency matrix A is defined as

[0165] In step S106, the graph convolutional network includes an input layer, a graph convolutional layer, a pooling layer (optional), a fully connected layer and an output layer.

[0166] The input layer inputs the node feature matrix X and the adjacency matrix A;

[0167] Graph convolution layer, which performs graph convolution operation:

[0168]

[0169] in:

[0170] H (l) is the node feature matrix of the lth layer (for the input layer, H (0) =X).

[0171] is the adjacency matrix with self-connection added.

[0172] yes The degree matrix of (diagonal matrix, each diagonal element is the degree of the corresponding node)

[0173] W (l) is the weight matrix of the lth layer.

[0174] σ is the activation function (such as ReLU).

[0175] Global pooling: All node features are pooled, such as average pooling or maximum pooling, to obtain a graph-level feature representation. In which, the node features are pooled, such as global average pooling or maximum pooling, to obtain a graph-level feature representation. Global average pooling averages the features of all nodes to obtain a graph-level feature representation. Through global average pooling or global maximum pooling, all node features are aggregated into a fixed-size vector.

[0176] The features of the global average pooling layer of the graph convolutional network of all cut sections (graph-level features) will be weighted fused once, and the fusion formula is as follows: Let H iis the global average pooling layer output feature (graph-level feature) of the graph convolutional network of the i-th cut section, and its weight is α i Then, the weighted fusion of graph-level features of all cut sections can be expressed as:

[0177] H fused Fusion is the feature vector after weighted fusion, and N is the number of cut sections. Through such feature fusion, a richer feature representation can be formed.

[0178] Then the feature vector H fused Input to the fully connected network for final classification or application. Input the fused graph-level feature vector to the fully connected network for final classification, judgment of fracture properties, or providing features for subsequent applications.

[0179] Fully connected layer: The graph-level features are input into the fully connected layer for further processing and classification.

[0180] Output layer: Output classification results and use the Softmax activation function to convert the output into a probability distribution.

[0181] The loss function of the present invention is as follows: the output is converted into classification probability using the Softmax activation function. Then the loss is calculated. For classification tasks, the cross entropy loss function is used:

[0182]

[0183] Where: N is the number of training samples; C is the number of categories; Y ic is the true label of the i-th sample; is the predicted probability of the ith sample.

[0184] Figure 5 Schematic diagram of the structure of a metal fracture recognition device based on three-dimensional graph learning in an embodiment of the present invention. Figure 1 The metal fracture recognition method based on three-dimensional graph learning corresponds to Figure 5 As shown, the metal fracture recognition device using three-dimensional graph learning in this embodiment may include:

[0185] A section three-dimensional image acquisition module 501 is used to acquire a metal fracture section three-dimensional image;

[0186] A three-dimensional image smoothing processing module 502 is used to smooth the three-dimensional image of the metal fracture section to obtain a smooth curved surface;

[0187] A section three-dimensional image clipping module 503 is used to move the smooth curved surface to the position of the lowest value of the trough of the metal fracture section three-dimensional image to perform a clipping operation, and retain the portion above the lowest value of the trough of the metal fracture section three-dimensional image;

[0188] A three-dimensional image projection flattening module 504 is used to project and flatten the retained three-dimensional image portion of the metal fracture section relative to the same plane to obtain a plurality of cross sections;

[0189] A node feature matrix construction module 505 is used to take each cross section as a node, extract the corresponding connected region cross-sectional features and the fractal dimension features of each connected region from each cross section, and then construct a node feature matrix and a neighbor matrix corresponding to the retained metal fracture section three-dimensional image portion;

[0190] The metal fracture type identification module 506 is used to identify the metal fracture type based on the corresponding node feature matrix and the proximity matrix and the pre-trained graph convolutional network.

[0191] It should be noted here that Figure 5 The various modules in the metal fracture recognition device based on three-dimensional graph learning are Figure 1 The various steps in the metal fracture recognition method based on three-dimensional graph learning in correspond to each other one by one, and the specific implementation process is the same, which will not be repeated here.

[0192] Reference Figure 6 , a schematic diagram of an electronic device is given. It should be noted that, Figure 6 The electronic device 600 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0193] like Figure 6 As shown, electronic device 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage part 608 into a random access memory (RAM) 603. In RAM 603, various programs and data required for system operation are also stored. Central processing unit 601, ROM 602 and RAM 603 are connected to each other via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0194] The following components are connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a local area network (LAN) card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 610 as needed, so that a computer program read therefrom is installed into the storage section 608 as needed.

[0195] When the central processing unit 601 in the electronic device of this embodiment executes the program, the following is achieved: Figure 1 The steps in the metal fracture recognition method based on three-dimensional graph learning are shown.

[0196] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer readable medium, the computer program including a computer program for executing Figure 1 In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 609, and / or installed from the removable medium 611. When the computer program is executed by the central processing unit 601, various functions defined in the apparatus of the present application are executed.

[0197] in, Figure 1 The computer program instructions corresponding to the method shown may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0198] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0199] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A metal fracture recognition method based on three-dimensional graph learning, characterized in that: include: Obtain three-dimensional images of metal fracture sections; Smoothing the three-dimensional image of the metal fracture section to obtain a smooth curved surface; The smooth curved surface is moved to the position of the lowest value of the trough of the three-dimensional image of the metal fracture section to perform a clipping operation, and a portion above the lowest value of the trough of the three-dimensional image of the metal fracture section is retained; Project and flatten the retained three-dimensional image portion of the metal fracture section relative to the same plane to obtain a plurality of cross sections; Taking each section as a node, extracting the corresponding cross-sectional features of the connected area and the fractal dimension features of each connected area from each section, and then constructing the node feature matrix and the neighboring matrix corresponding to the retained three-dimensional image part of the metal fracture section; The metal fracture type is identified based on the corresponding node feature matrix, proximity matrix and the pre-trained graph convolutional network.

2. The metal fracture recognition method based on three-dimensional graph learning as claimed in claim 1, characterized in that: The retained three-dimensional image portion of the metal fracture section is flattened by equally spaced projection relative to the same plane.

3. The metal fracture recognition method based on three-dimensional graph learning as claimed in claim 1, characterized in that: The minimum energy unfolding method is used to project and flatten the retained three-dimensional image portion of the metal fracture section relative to the same plane, so that the angle and area characteristics of the original surface are maintained when the surface is flattened.

4. The metal fracture recognition method based on three-dimensional graph learning as claimed in claim 3, characterized in that: The minimum energy unfolding method is: constructing an energy function to minimize the deformation generated during the unfolding process, and obtaining the coordinates of each point on the surface on the two-dimensional plane by solving the minimum value of the energy function.

5. The metal fracture recognition method based on three-dimensional graph learning as claimed in claim 4, characterized in that: The expression of the energy function is: Where E is a surface; u and v are the coordinates of all vertices on the two-dimensional plane, (i, j) represents the edge on the surface E; x and y represent the horizontal and vertical directions respectively.

6. The metal fracture recognition method based on three-dimensional graph learning as claimed in claim 1, characterized in that: The smoothing operation includes: Noise filtering and surface smoothing operations are performed sequentially.

7. The metal fracture recognition method based on three-dimensional graph learning as claimed in claim 1, characterized in that: The fractal dimension features include: voxel density, geometric morphology features, fractal dimension, surface roughness and local geometric features; The cross-sectional features of the connected region include perimeter, area and shape features.

8. A metal fracture recognition device based on three-dimensional graph learning, characterized in that: include: A section three-dimensional image acquisition module, which is used to acquire a three-dimensional image of a metal fracture section; A three-dimensional image smoothing processing module, which is used to smooth the three-dimensional image of the metal fracture section to obtain a smooth curved surface; A section three-dimensional image cropping module, which is used to move the smooth curved surface to the position of the lowest value of the trough of the metal fracture section three-dimensional image to perform a cropping operation, and retain the portion above the lowest value of the trough of the metal fracture section three-dimensional image; A three-dimensional image projection flattening module is used to project and flatten the retained three-dimensional image portion of the metal fracture section relative to the same plane to obtain a plurality of cross sections; A node feature matrix construction module is used to take each section as a node, extract the corresponding connected area section features and the fractal dimension features of each connected area from each section, and then construct a node feature matrix and a neighbor matrix corresponding to the retained metal fracture section three-dimensional image portion; The metal fracture type recognition module is used to identify the metal fracture type based on the corresponding node feature matrix and proximity matrix and a pre-trained graph convolutional network.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the metal fracture recognition method based on three-dimensional graph learning as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the metal fracture recognition method based on three-dimensional graph learning as described in any one of claims 1-7 are implemented.

Images