A defect detection-based graphite component machining pose auxiliary decision method, device, medium and product

Abstract

The application discloses a graphite component processing pose auxiliary decision method and device based on defect detection, equipment, medium and product, relates to the graphite component manufacturing technical field of high temperature gas cooled reactor, and the method comprises the following steps: acquiring the CT scanning result of a graphite blank, and determining defect point cloud distribution information; based on the geometric model of the graphite component, a plurality of initial candidate processing poses are generated in the graphite blank; for each initial candidate processing pose, the defect point cloud distribution information in the corresponding component region is determined, and a defect evaluation index is constructed; a multi-objective optimization model is solved by using an NSGA-II algorithm to obtain a Pareto optimal candidate processing pose set; based on the defect point cloud distribution information corresponding to each candidate processing pose, an anomaly detection model is used to obtain the anomaly score of each candidate processing pose; and the candidate processing pose with the lowest anomaly score is taken as the optimal processing pose of the tentative processing graphite component. The application improves the processing quality of the graphite component.

Description

Technical Field

[0001] This application relates to the field of high-temperature gas-cooled reactor graphite component manufacturing technology, and in particular to a method, equipment, medium and product for assisting decision-making on the processing posture of graphite components based on defect detection. Background Technology

[0002] High-temperature gas-cooled reactors (HTGRs), as advanced reactor types possessing characteristics of fourth-generation nuclear energy systems, have a core cavity constructed from graphite components that support the fuel spherical bed. During reactor operation, these graphite components are subjected to a complex multi-physics coupling environment. However, due to the naturally porous nature of graphite, inherent defects such as pores or cracks inevitably exist within the components. If these inherent defects induce component failure under complex operating conditions, they will seriously threaten the reactor's operational safety. Therefore, it is urgent to adopt effective measures to control the inherent defects in graphite components to ensure the structural integrity and safety of the reactor from the source.

[0003] Related technologies have proposed non-destructive testing methods for finished graphite components. However, such methods belong to a passive control approach of "discarding after processing". If the testing object can be moved to the graphite blank stage, and the processing posture of the target component inside the blank (i.e., the three-dimensional material picking position and processing tilt angle) can be reasonably planned according to the spatial distribution of defects inside the blank, qualified finished components can be directly processed. Summary of the Invention

[0004] The purpose of this application is to provide a method, device, medium, and product for assisting decision-making on the processing posture of graphite components based on defect detection, so as to improve the processing quality of graphite components.

[0005] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for assisting in the decision-making of graphite component processing pose based on defect detection, including: Obtain CT scan results of graphite billet and determine the defect point cloud distribution information inside the graphite billet; Based on the geometric model of the graphite component to be processed, multiple initial candidate processing poses are generated in the graphite blank to obtain a first set of initial candidate processing poses and a second set of initial candidate processing poses. For each initial candidate processing pose, the defect point cloud distribution information in the region of the graphite component to be processed corresponding to the initial candidate processing pose is determined, and a defect evaluation index is constructed; the defect evaluation index includes the number of defect points on the component surface and the number of defect points inside the component. Based on the first initial candidate processing pose set and the second initial candidate processing pose set, a multi-objective optimization model is constructed with the defect evaluation index as the optimization objective. The NSGA-II algorithm is used for optimization and solution to obtain a Pareto optimal candidate processing pose set. Based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto optimal candidate processing pose set, an anomaly score for each candidate processing pose is obtained using an anomaly detection model based on a graph neural network. Candidate machining poses are screened based on the anomaly score, and candidate machining poses with defective structures are eliminated. The candidate machining pose with the lowest anomaly score among the remaining candidate machining poses is taken as the optimal machining pose for the graphite component to be machined.

[0006] In one embodiment, determining the defect point cloud distribution information inside the graphite blank specifically includes: The CT scan results of the graphite billet were binarized using the Otsu algorithm, with defect points assigned a value of 0 and graphite material assigned a value of 1, to obtain the binarized result. The binarization result is compressed, and only the location index of the defect points is stored to obtain the defect point cloud distribution information inside the graphite billet.

[0007] In one embodiment, based on the geometric model of the graphite component to be processed, multiple initial candidate processing poses are generated in the graphite blank to obtain a first set of initial candidate processing poses and a second set of initial candidate processing poses, specifically including: Based on the CAD design drawings of the graphite components to be processed, design the corresponding geometric model, which is an STL model; Based on the geometric model, within the set parameter search space, a number of initial candidate machining poses are generated using the Latin hypercube sampling method; wherein, the initial candidate machining poses are six-dimensional vectors, including translation parameters and rotation parameters; The multiple initial candidate processing poses are divided into a first set of initial candidate processing poses and a second set of initial candidate processing poses.

[0008] In one embodiment, for each initial candidate processing pose, the defect point cloud distribution information within the region of the graphite component to be processed corresponding to the initial candidate processing pose is determined, and a defect evaluation index is constructed, specifically including: For each initial candidate machining pose, the geometric model of the graphite component to be machined is copied and transformed according to its translation and rotation parameters to obtain the transformed geometric model. Determine the defect point cloud distribution information within the transformed geometric model region corresponding to the initial candidate processing pose; Based on the defect point cloud distribution information, the Möller-Trumbore ray intersection algorithm is used to emit rays along the positive Z-axis to determine whether each defect point is located inside the graphite component to be processed; if the ray passing through the surface of the graphite component to be processed intersects the triangular mesh an odd number of times, it is determined that the defect point is inside the graphite component to be processed, and the number of defect points inside the component is obtained. The centroid coordinate method is used to determine whether the defect point is located on the surface of the graphite component to be processed; if the defect point is located on any triangular facet, it is determined to be a defect point on the surface of the component, and the number of defect points on the surface of the component is obtained.

[0009] In one embodiment, based on the first initial candidate processing pose set and the second initial candidate processing pose set, a multi-objective optimization model is constructed with the defect evaluation index as the optimization objective. The NSGA-II algorithm is used for optimization and solution to obtain a Pareto optimal set of candidate processing poses, specifically including: The current first initial candidate processing pose set and the current second initial candidate processing pose set are sorted in a non-dominated hierarchical manner to obtain the first sorted initial candidate processing pose set and the second sorted initial candidate processing pose set. Calculate the first crowding distance between each initial candidate processing pose within the same non-dominated front in the first sorted initial candidate processing pose set and the second crowding distance between each initial candidate processing pose within the same non-dominated front in the second sorted initial candidate processing pose set. Tournament selection, simulated binary crossover, and polynomial mutation operators are used to generate the next-generation first initial candidate processing pose set and the next-generation second initial candidate processing pose set; Based on the first crowding distance and the second crowding distance, a multi-objective evolutionary iteration is performed based on a dual-population cooperative mechanism to output an initial Pareto optimal candidate processing pose set; The next-generation first initial candidate processing pose set is used as the current first initial candidate processing pose set, and the next-generation second initial candidate processing pose set is used as the current second initial candidate processing pose set. The program returns "performs non-dominated hierarchical sorting on the current first initial candidate processing pose set and the current second initial candidate processing pose set to obtain the first sorted initial candidate processing pose set and the second sorted initial candidate processing pose set", and repeats this process multiple times. After merging the initial Pareto optimal candidate processing pose sets obtained from multiple runs, non-dominated hierarchical sorting is performed again, and all individuals in the first non-dominated frontier of the merged population are extracted as the Pareto optimal candidate processing pose set.

[0010] In one embodiment, based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto-optimal candidate processing pose set, an anomaly score for each candidate processing pose is obtained using an anomaly detection model based on a graph neural network, specifically including: Based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto optimal candidate processing pose set, construct the attribute defect map corresponding to each candidate processing pose. Based on the attribute defect map corresponding to the candidate processing pose, an anomaly score for the candidate processing pose is determined using a graph neural network-based anomaly detection model. The graph neural network-based anomaly detection model is obtained by training a transductive single-class graph representation learning model using a three-stage mixed data training process based on an unsupervised transductive learning paradigm and a dynamic truncation mechanism. The training process specifically includes: Full initialization training: Initialize the network parameters of the transductive single-class graph representation learning model, input the unlabeled full mixed attribute defect map set into the transductive single-class graph representation learning model for forward propagation, and calculate the mean of the graph embedding features to initialize the hypersphere center; the full mixed attribute defect map set includes attribute defect maps corresponding to all candidate processing poses; Dynamic truncation mechanism training: Before updating parameters in each training round, forward propagation is performed on the full mixed attribute defect map to calculate the single-class classification loss of each sample; the single-class classification losses of all samples are sorted globally in ascending order, and the global truncation threshold for the current round is determined by combining the preset anomalous rate; samples with single-class classification losses greater than the global truncation threshold are identified as anomalous samples, and the gradient backpropagation of anomalous samples is blocked; when calculating the batch average loss, the average is calculated only for normal samples, and gradient descent is performed to update the model parameters; Solidify fine-tuning training: Lock the global truncation threshold to the threshold value of the last round of the second training phase to fix the classification boundary of the hypersphere; use only the normal samples to perform gradient descent fine-tuning on the transductive single-class graph representation learning model at a preset learning rate until convergence; The loss function values ​​corresponding to the defect maps of each attribute at convergence are extracted as anomaly scores.

[0011] In one embodiment, candidate machining poses are screened based on the anomaly score, eliminating candidate machining poses with defective structures, and selecting the candidate machining pose with the lowest anomaly score from the remaining candidate machining poses as the optimal machining pose for the graphite component to be machined. Specifically, this includes: Multiple different preset anomaly rates are set to form an anomaly rate parameter set. For each preset anomaly rate, three-stage mixed data training is repeatedly performed to obtain multiple sets of anomaly score ranking sequences of the full mixed attribute defect map at different preset anomaly rates. An overall consistency test is performed on multiple sets of abnormal score ranking sequences. Abnormal sequences that deviate significantly from the overall ranking distribution are removed, and valid abnormal score ranking sequences that meet the consistency standard are retained. For any attribute defect map, calculate the average value of the anomaly scores corresponding to the attribute defect map in each valid anomaly score ranking sequence, and use the average value as the comprehensive anomaly score of the attribute defect map. All attribute defect maps are sorted in ascending order according to the comprehensive anomaly score. The attribute defect map with the lowest comprehensive anomaly score is selected, and the candidate machining pose corresponding to the attribute defect map with the lowest comprehensive anomaly score is taken as the optimal machining pose for the graphite component to be machined.

[0012] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for assisting decision-making in the processing posture of graphite components based on defect detection.

[0013] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described defect detection-based graphite component machining pose-assisted decision-making method.

[0014] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described defect detection-based graphite component machining pose-assisted decision-making method.

[0015] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, device, medium, and product for auxiliary decision-making of graphite component processing pose based on defect detection. The method involves acquiring CT scan results of graphite blanks and determining the defect point cloud distribution information within the graphite blanks; generating multiple initial candidate processing poses in the graphite blanks based on the geometric model of the graphite component to be processed; determining the defect point cloud distribution information within the region of the graphite component to be processed corresponding to each initial candidate processing pose and constructing a defect evaluation index; constructing a multi-objective optimization model with the defect evaluation index as the optimization objective, and using the NSGA-II algorithm for optimization to obtain a set of Pareto optimal candidate processing poses; based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto optimal candidate processing pose set, using an anomaly detection model based on graph neural networks to obtain an anomaly score for each candidate processing pose; filtering candidate processing poses based on the anomaly score, eliminating candidate processing poses with abnormal defect structures, and selecting the candidate processing pose with the lowest anomaly score among the remaining candidate processing poses as the optimal processing pose for the graphite component to be processed. Based on the obtained spatial distribution of defects in the billet, this application constructs a multi-objective optimization model with the goal of minimizing the number of defect points in the component, and introduces graph anomaly detection to perform a secondary evaluation of the defect structure. Finally, it outputs a processing pose scheme that takes into account both the number of defects and the rationality of the structure, so as to improve the processing quality of graphite components. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a graphite component processing pose-assisted decision-making method based on defect detection, provided in an embodiment of this application; Figure 2 A point cloud distribution diagram of internal defects in a graphite blank provided in an embodiment of this application; Figure 3 A distribution diagram of internal and surface defects of a component provided in an embodiment of this application; Figure 4 A flowchart illustrating the solution process of the NSGA-II algorithm provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] This application provides a defect detection-based auxiliary decision-making method for the machining pose of graphite components, including a preliminary construction process of the machining pose set based on the NSGA-II optimization algorithm and a final decision-making process for the machining pose based on a graph neural network anomaly recognition framework. Based on the defect point cloud information within the graphite billet obtained by CT technology, this application initially solves for the machining pose set with the objective of minimizing the number of defect points on the surface and inside the machined component. Then, using the corresponding defect structure of the component as the evaluation object, the machining pose with the most normal structure is selected as the auxiliary decision for actual machining. Compared with traditional component defect detection methods and centering machining methods, this application ensures the reliability of the finished component from the source and also improves the economy of the machining process.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Based on the obtained spatial distribution of defects in the billet, this application constructs a multi-objective optimization model with the goal of minimizing the number of defect points on the surface and inside of the component, and introduces graph anomaly detection to perform a secondary evaluation of the defect structure, and finally outputs a processing pose scheme that takes into account both the number of defects and the rationality of the structure.

[0022] In one exemplary embodiment, such as Figure 1 As shown, a method for assisting decision-making on the processing pose of graphite components based on defect detection is provided, including the following steps: S1: Obtain the CT scan results of the graphite billet and determine the defect point cloud distribution information inside the graphite billet.

[0023] In this embodiment, the spatial distribution information of defects inside the billet is obtained. The spatial distribution information of defects can be obtained through non-destructive testing methods such as computed tomography (CT), and the spatial location and geometric features of the defects are extracted. The point cloud distribution of defects inside the graphite billet is shown below. Figure 2 As shown.

[0024] In one embodiment, determining the defect point cloud distribution information inside the graphite blank specifically includes: S11: The CT scan results of the graphite billet are binarized using the Otsu algorithm. The defect points are assigned a value of 0, and the graphite material is assigned a value of 1 to obtain the binarized result.

[0025] In this embodiment, the three-dimensional digital matrix composed of voxels obtained from CT scans is binarized based on the Otsu algorithm. After binarization, defect points, i.e., points with an actual line attenuation coefficient of 0, are assigned a value of 0; graphite material, i.e., points with an actual line attenuation coefficient equal to the line attenuation coefficient corresponding to graphite, are assigned a value of 1. The number of points with a value of 0 is far less than the number of points with a value of 1.

[0026] S12: Compress the binarization result and store only the location index of the defect points to obtain the defect point cloud distribution information inside the graphite billet.

[0027] In this embodiment, the binarization result is compressed, and only the location index corresponding to the defect point is stored (for example, the coordinates of a point with a value of 0 in the three-dimensional matrix are (20, 15, 30), and this coordinate is stored), and the three-dimensional dimensions of the graphite billet are stored in h5 format.

[0028] S2: Based on the geometric model of the graphite component to be processed, multiple initial candidate processing poses are generated in the graphite blank to obtain a first set of initial candidate processing poses and a second set of initial candidate processing poses.

[0029] In this embodiment, the processing posture refers to the spatial position and orientation of the component in the blank, which is characterized by the material pick-up position parameters and the orientation parameters.

[0030] In one embodiment, S2 specifically includes: S21: Based on the CAD design drawing of the graphite component to be processed, design the corresponding geometric model, which is an STL model.

[0031] S22: Based on the geometric model, within the set parameter search space, a number of initial candidate machining poses are generated using the Latin hypercube sampling method; wherein, the initial candidate machining poses are six-dimensional vectors, including translation parameters and rotation parameters.

[0032] S23: Divide the multiple initial candidate processing poses into a first set of initial candidate processing poses and a second set of initial candidate processing poses.

[0033] In this embodiment, an initial set of candidate machining poses is generated within a given parameter search space, as follows: Let the dimension of the parameters to be optimized be n (including translation parameters x0, y0, z0 and rotation parameters) The search interval for each dimension is [lower_i, upper_i], where the lower bound of each of the six dimensions is 0; the upper bound of the first three dimensions corresponds to the size of the graphite billet in the X, Y, and Z directions, respectively, and the upper bound of the last three dimensions is 180 degrees. The population size is size.

[0034] The initial population was generated using the Latin hypercube sampling method: For each dimension i, divide its search interval into size non-overlapping sub-intervals; A sampling point is randomly and uniformly selected within each sub-interval; The sampling point sequences of each dimension are randomly shuffled and combined to obtain size n-dimensional vectors as initial candidate processing poses. To achieve dual-race evolution, two sets of initial candidate processing poses are generated: a first set of initial candidate processing poses and a second set of initial candidate processing poses.

[0035] S3: For each initial candidate processing pose, determine the defect point cloud distribution information in the region of the graphite component to be processed corresponding to the initial candidate processing pose, and construct a defect evaluation index; the defect evaluation index includes the number of defect points on the component surface and the number of defect points inside the component.

[0036] In one embodiment, S3 specifically includes: S31: For each of the initial candidate machining poses, the geometric model of the graphite component to be machined is copied and transformed according to its translation and rotation parameters to obtain the transformed geometric model.

[0037] In this embodiment, for each initial candidate machining pose, its parameters include translation parameters (x0, y0, z0) and rotation parameters (x0, y0, z0). First, copy the component's STL model, and then rotate it around the X-axis, Y-axis, and Z-axis by angles respectively. θ, ψ, and then translate (x0, y0, z0).

[0038] S32: Determine the defect point cloud distribution information within the transformed geometric model region corresponding to the initial candidate processing pose.

[0039] S33: Based on the defect point cloud distribution information, the Möller-Trumbore ray intersection algorithm is used to emit rays along the positive Z-axis to determine whether each defect point is located inside the graphite component to be processed; if the ray passing through the surface of the graphite component to be processed intersects the triangular mesh an odd number of times, it is determined that the defect point is inside the graphite component to be processed, and the number of defect points inside the component is obtained.

[0040] In this embodiment, the Möller–Trumbore ray intersection algorithm is used to emit rays along the positive Z-axis and determine whether each defect point is located inside the component. The algorithm solves for the intersection parameters (t, m, n) for each triangular facet and each ray. A ray is considered to intersect a triangle if t > 0, m > 0, n > 0, and (m + n) < 1. If a ray passing through the component surface intersects the triangular mesh an odd number of times, the defect point is determined to be inside the component.

[0041] S34: Use the centroid coordinate method to determine whether the defect point is located on the surface of the graphite component to be processed; if the defect point is located on any triangular facet, it is determined to be a defect point on the surface of the component, and the number of defect points on the surface of the component is obtained.

[0042] In this embodiment, the centroid coordinate method is used to determine whether a defect point is located on the surface of the component. The centroid coordinates (s, t) of the defect point relative to each triangular facet are calculated. When s≥0, t≥0 and (s+t)≤1, the defect point is determined to be located on the triangular facet, which is a defect point on the surface of the component.

[0043] Defects identified as internal to the component but not on the surface are counted as internal defects, while defects identified as surface defects are counted as surface defects. The surface defect count represents the number of defects located on the outer surface of the component, while the internal defect count represents the remaining number of defects completely enclosed within the component. The distribution of internal and surface defects is as follows: Figure 3 As shown.

[0044] S4: Based on the first initial candidate machining pose set and the second initial candidate machining pose set, and with the defect evaluation index as the optimization objective, a multi-objective optimization model is constructed. The NSGA-II algorithm is used for optimization and solution to obtain a Pareto optimal set of candidate machining poses. The solution process is as follows: Figure 4 As shown.

[0045] In one embodiment, S4 specifically includes: S41: Perform non-dominated hierarchical sorting on the current first initial candidate processing pose set and the current second initial candidate processing pose set to obtain the first sorted initial candidate processing pose set and the second sorted initial candidate processing pose set.

[0046] In this embodiment, before performing non-dominated hierarchical sorting, the crossover and mutation indices of the multi-objective optimization algorithm (NSGA-II algorithm) are dynamically adjusted to balance the algorithm's global search capability and local exploitation capability.

[0047] Specifically, let the current generation number be g, the maximum generation number be G, the minimum value of the parameter be min, and the maximum value be max. The parameter value adaptively adjusts with an exponential decay law as the number of generations increases, as follows: First, calculate the current evolutionary progress percentage. ; Then, the decay factor is constructed based on the exponential decay function. , where k is the preset attenuation coefficient (usually taken as 3); Finally, calculate the parameter values ​​corresponding to the current generation. The result is then rounded to the nearest integer.

[0048] The current first initial candidate processing pose set and the current second initial candidate processing pose set are subjected to non-dominated hierarchical sorting, specifically including: First, each initial candidate processing pose is considered as an individual, and each individual contains two attributes: the objective function value vector (i.e., the number of surface defect points and the number of internal defect points) and the constraint violation value.

[0049] Secondly, using the constraint rule that feasible solutions take precedence over infeasible solutions, pairwise comparisons are performed on all individuals in the population: If both individuals are feasible solutions, the Pareto dominance relation is used to determine whether individual a dominates v. That is, when all objective function values ​​of individual a are not inferior to those of individual b and there is at least one objective function value that is superior to that of individual b, then a is determined to dominate v. If individual a is a feasible solution and individual v is an infeasible solution, then a is determined to dominate v; If both entities are infeasible solutions, the entity with the smaller constraint violation value dominates the entity with the larger constraint violation value.

[0050] Then, based on the above domination relationship, count the number of times each individual is dominated by other individuals (referred to as the domination count).

[0051] Finally, all individuals with a dominance count of zero are assigned to the first non-dominated frontier; then, individuals in this frontier are removed, the dominance counts of the remaining individuals are updated, and newly generated individuals with a dominance count of zero are assigned to the second non-dominated frontier; and so on, until all individuals have completed stratification.

[0052] S42: Calculate the first crowding distance between each initial candidate processing pose within the same non-dominated front in the first sorted initial candidate processing pose set and the second crowding distance between each initial candidate processing pose within the same non-dominated front in the second sorted initial candidate processing pose set.

[0053] In this embodiment, the crowding distance of each initial candidate processing pose within the same non-dominated front is calculated to characterize the density distribution of individuals in the target space. Individuals with higher crowding are more sparsely distributed around them and are preferentially retained in subsequent selection operations to maintain population diversity.

[0054] S43: Using tournament selection, simulated binary crossover, and polynomial mutation operators, generate the next-generation first initial candidate processing pose set and the next-generation second initial candidate processing pose set.

[0055] In this embodiment, tournament selection, simulated binary crossover, and polynomial mutation operators are used to perform evolutionary operations on the two current initial processing pose populations to generate the next generation of initial candidate processing pose populations, namely the next generation first initial candidate processing pose set and the next generation second initial candidate processing pose set.

[0056] S44: Based on the first crowding distance and the second crowding distance, perform multi-objective evolutionary iteration based on the dual-population cooperative mechanism to output an initial Pareto optimal candidate processing pose set.

[0057] In this embodiment, firstly, in each generation of evolution, steps S41 to S43 are performed independently for population A and population B, namely, fitness assessment, parameter adaptive adjustment, non-dominated sorting, crowding calculation and genetic operator operation are performed respectively, parent and offspring generations are merged and individuals of a specified size are truncated according to the relationship between dominance and crowding, thereby completing the independent update of the two populations in a single generation.

[0058] Secondly, a fixed population information exchange cycle should be established. With exchange quantity When the current generation... for When the value is an integer multiple of the population size, a dual-population collaborative interaction mechanism is triggered: First, based on the principles of prioritizing the non-dominant frontier level in step S41 and prioritizing the population with the larger crowding distance in step S42, the populations A and B are extracted respectively. The optimal individuals form an elite subset A. and elite subset B Secondly, the elite subset A Immigrating into population B will result in the elite subset B being... Before the migration to population A, duplicate solutions identical to the migrating individuals in the target population are removed to maintain population diversity. Then, a corresponding number of non-elite individuals are randomly selected from the target population and replaced with the migrating elite individuals. After the replacement is completed, step S3 is called again to update the objective function value.

[0059] make Repeat the above process of independent evolution and cooperative interaction until... The iteration terminates at the specified time. For the final population A, perform the non-dominated stratification of step S41, extract all individuals at the first non-dominated front, and these non-dominated individuals constitute the final Pareto optimal solution set.

[0060] S45: Take the next generation first initial candidate processing pose set as the current first initial candidate processing pose set, take the next generation second initial candidate processing pose set as the current second initial candidate processing pose set, and return "perform non-dominated hierarchical sorting on the current first initial candidate processing pose set and the current second initial candidate processing pose set to obtain the first sorted initial candidate processing pose set and the second sorted initial candidate processing pose set", and repeat the execution multiple times.

[0061] S46: After merging the initial Pareto optimal candidate processing pose sets obtained from multiple runs, perform non-dominated hierarchical sorting again, and extract all individuals in the first non-dominated frontier of the merged population as the Pareto optimal candidate processing pose set.

[0062] In this embodiment, the above steps are repeated three times, with each independent run outputting a set of Pareto optimal solutions. After merging all Pareto solutions obtained from the three runs, the non-dominated hierarchical sorting in step S41 is performed again to extract all individuals in the first non-dominated front of the merged population as the final output set of Pareto optimal candidate processing poses.

[0063] S5: Based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto optimal candidate processing pose set, an anomaly score for each candidate processing pose is obtained using an anomaly detection model based on a graph neural network.

[0064] In this embodiment, the point cloud distribution of internal defects in the component corresponding to each candidate processing pose is input into an anomaly detection model based on a graph neural network to obtain an anomaly score for each candidate processing pose; the candidate poses are sorted from low to high according to the anomaly score; and the processing pose with the lowest anomaly score is selected as the reference pose for actual processing.

[0065] In one embodiment, S5 specifically includes: S51: Based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto optimal candidate processing pose set, construct the attribute defect map corresponding to each candidate processing pose.

[0066] In this embodiment, an inverse matrix of point cloud coordinate transformation is constructed for each pose to obtain the distribution of defect point clouds inside the component processed for each pose.

[0067] To characterize the defect structure and its inherent relationships, the defect point cloud distribution of the component is constructed into an attribute defect map. The specific steps are as follows: S511: The DBSCAN algorithm is used to cluster the defect point cloud, and each defect point cluster is used as a graph node.

[0068] S512: Connect the nodes in the graph with edges. The specific process is as follows: First, for any given graph node, obtain the three-dimensional coordinates of all discrete defect points contained within it, calculate the mean of the three-dimensional coordinates, and define the mean of the coordinates as the spatial centroid of the corresponding graph node; then calculate the Euclidean distance between the spatial centroids of any two graph nodes, and sort all node pairs composed of graph nodes in ascending order according to the Euclidean distance. Secondly, select the node pairs with the shortest Euclidean distance according to a pre-defined ratio, and establish local neighboring undirected edges between the selected node pairs. Then, for the remaining unconnected node pairs, globally connected undirected edges are constructed based on the minimum spanning tree strategy. That is, an undirected edge is established between two graph nodes if and only if they belong to different connected regions, until all the graph nodes constitute a fully connected graph. Finally, based on the connection relationships constructed above, an adjacency matrix containing the spatial distance information of the graph nodes is generated.

[0069] S513: Extract the feature data of each graph node to construct a feature vector, assign node attributes to the attribute defect graph, and complete the construction of the attribute defect graph; the specific process is as follows: First, for any given graph node, the total number of discrete defect points contained within the node is counted, and the total number is extracted as the first node feature of the graph node to characterize the defect volume. Secondly, obtain the pre-marked defect location information, count the number of surface defect points inside the graph node that belong to the outer surface of the target component, and calculate the ratio of the number of surface defect points to the total number of discrete defect points inside the graph node. Extract the ratio as the second node feature of the graph node to characterize the proportion of surface defect points. Then, based on the STL model of the target component, the Euclidean distance from all discrete defect points within the graph node to the surface of the target component is calculated, and the minimum value among all the calculated Euclidean distances is extracted. The minimum value is extracted as the third node feature of the graph node to characterize the defect depth. Finally, the first node feature, the second node feature, and the third node feature are combined to construct a three-dimensional feature vector for the corresponding graph node, and the three-dimensional feature vector is assigned to the corresponding graph node; the above operation is repeated to traverse all the graph nodes and obtain the attribute defect graph containing the node feature vector and the graph connected edges corresponding to each candidate processing pose.

[0070] S52: Based on the attribute defect map corresponding to the candidate processing pose, an anomaly score for the candidate processing pose is determined using an anomaly detection model based on a graph neural network; wherein, the anomaly detection model based on a graph neural network is obtained by training a transductive single-class graph representation learning model with three stages of mixed data based on an unsupervised transductive learning paradigm and a dynamic truncation mechanism, and the training process specifically includes: Full initialization training: Initialize the network parameters of the transductive single-class graph representation learning model, input the unlabeled full mixed attribute defect map set into the transductive single-class graph representation learning model for forward propagation, and calculate the mean of the graph embedding features to initialize the hypersphere center; the full mixed attribute defect map set includes attribute defect maps corresponding to all candidate processing poses; Dynamic truncation mechanism training: Before updating parameters in each training round, forward propagation is performed on the full mixed attribute defect map to calculate the single-class classification loss of each sample; the single-class classification losses of all samples are sorted globally in ascending order, and the global truncation threshold for the current round is determined by combining the preset anomalous rate; samples with single-class classification losses greater than the global truncation threshold are identified as anomalous samples, and the gradient backpropagation of anomalous samples is blocked; when calculating the batch average loss, the average is calculated only for normal samples, and gradient descent is performed to update the model parameters; Solidify fine-tuning training: Lock the global truncation threshold to the threshold value of the last round of the second training stage to fix the classification boundary of the hypersphere; use only the normal samples to perform gradient descent fine-tuning on the transductive single-class graph representation learning model at a preset minimum learning rate (preset learning rate) until convergence; extract the loss function value corresponding to each attribute defect map at convergence as anomaly score, and sort and output all attribute defect maps in the graph set according to the size of the anomaly score.

[0071] S6: Based on the anomaly score, the candidate machining poses are screened, and the candidate machining poses with abnormal defect structures are eliminated. The candidate machining pose with the lowest anomaly score among the remaining candidate machining poses is taken as the optimal machining pose for the graphite component to be machined.

[0072] In one embodiment, S6 specifically includes: S61: Set multiple different preset anomaly rates to form an anomaly rate parameter set. For each preset anomaly rate, repeatedly perform three-stage mixed data training to obtain multiple sets of anomaly score ranking sequences of the full mixed attribute defect map under different preset anomaly rates.

[0073] S62: Perform an overall consistency test on multiple sets of abnormal score ranking sequences, remove abnormal sequences that significantly deviate from the overall ranking distribution, and retain valid abnormal score ranking sequences that meet the consistency standard.

[0074] S63: For any attribute defect map, calculate the average value of the abnormal scores corresponding to the attribute defect map in each valid abnormal score ranking sequence, and use the average value as the comprehensive abnormal score of the attribute defect map.

[0075] S64: Sort all attribute defect maps in ascending order according to the comprehensive anomaly score, select the attribute defect map with the lowest comprehensive anomaly score, and take the candidate machining pose corresponding to the attribute defect map with the lowest comprehensive anomaly score as the optimal machining pose of the graphite component to be machined.

[0076] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described defect detection-based graphite component processing pose-assisted decision-making method.

[0077] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described defect detection-based graphite component machining pose-assisted decision-making method.

[0078] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described defect detection-based graphite component machining pose-assisted decision-making method.

[0079] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a defect detection-based pose-assisted decision-making method for graphite component processing.

[0080] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0081] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0082] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0083] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0084] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for assisting in the decision-making of graphite component processing pose based on defect detection, characterized in that, include: Obtain CT scan results of graphite billet and determine the defect point cloud distribution information inside the graphite billet; Based on the geometric model of the graphite component to be processed, multiple initial candidate processing poses are generated in the graphite blank to obtain a first set of initial candidate processing poses and a second set of initial candidate processing poses. For each initial candidate processing pose, the defect point cloud distribution information in the region of the graphite component to be processed corresponding to the initial candidate processing pose is determined, and a defect evaluation index is constructed; the defect evaluation index includes the number of defect points on the component surface and the number of defect points inside the component. Based on the first initial candidate processing pose set and the second initial candidate processing pose set, a multi-objective optimization model is constructed with the defect evaluation index as the optimization objective. The NSGA-II algorithm is used for optimization and solution to obtain a Pareto optimal candidate processing pose set. Based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto optimal candidate processing pose set, an anomaly score for each candidate processing pose is obtained using an anomaly detection model based on a graph neural network. Candidate machining poses are screened based on the anomaly score, and candidate machining poses with defective structures are eliminated. The candidate machining pose with the lowest anomaly score among the remaining candidate machining poses is taken as the optimal machining pose for the graphite component to be machined.

2. The graphite component processing pose-assisted decision-making method based on defect detection according to claim 1, characterized in that, Determining the defect point cloud distribution information inside the graphite blank specifically includes: The CT scan results of the graphite billet were binarized using the Otsu algorithm, with defect points assigned a value of 0 and graphite material assigned a value of 1, to obtain the binarized result. The binarization result is compressed, and only the location index of the defect points is stored to obtain the defect point cloud distribution information inside the graphite billet.

3. The graphite component processing pose-assisted decision-making method based on defect detection according to claim 1, characterized in that, Based on the geometric model of the graphite component to be processed, multiple initial candidate processing poses are generated in the graphite blank, resulting in a first set of initial candidate processing poses and a second set of initial candidate processing poses, specifically including: Based on the CAD design drawings of the graphite components to be processed, design the corresponding geometric model, which is an STL model; Based on the geometric model, within the set parameter search space, a number of initial candidate machining poses are generated using the Latin hypercube sampling method; wherein, the initial candidate machining poses are six-dimensional vectors, including translation parameters and rotation parameters; The multiple initial candidate processing poses are divided into a first set of initial candidate processing poses and a second set of initial candidate processing poses.

4. The graphite component processing pose-assisted decision-making method based on defect detection according to claim 1, characterized in that, For each initial candidate processing pose, the defect point cloud distribution information within the region of the graphite component to be processed corresponding to the initial candidate processing pose is determined, and a defect evaluation index is constructed, specifically including: For each initial candidate machining pose, the geometric model of the graphite component to be machined is copied and transformed according to its translation and rotation parameters to obtain the transformed geometric model. Determine the defect point cloud distribution information within the transformed geometric model region corresponding to the initial candidate processing pose; Based on the defect point cloud distribution information, the Möller-Trumbore ray intersection algorithm is used to emit rays along the positive Z-axis to determine whether each defect point is located inside the graphite component to be processed. If the ray passing through the surface of the graphite component to be processed intersects the triangular mesh an odd number of times, it is determined that the defect point is inside the graphite component to be processed, and the number of defect points inside the component is obtained. The centroid coordinate method is used to determine whether the defect point is located on the surface of the graphite component to be processed; if the defect point is located on any triangular facet, it is determined to be a defect point on the surface of the component, and the number of defect points on the surface of the component is obtained.

5. The graphite component processing pose-assisted decision-making method based on defect detection according to claim 1, characterized in that, Based on the first initial candidate processing pose set and the second initial candidate processing pose set, and with the defect evaluation index as the optimization objective, a multi-objective optimization model is constructed. The NSGA-II algorithm is used for optimization and solution to obtain a Pareto optimal set of candidate processing poses, specifically including: The current first initial candidate processing pose set and the current second initial candidate processing pose set are sorted in a non-dominated hierarchical manner to obtain the first sorted initial candidate processing pose set and the second sorted initial candidate processing pose set. Calculate the first crowding distance between each initial candidate processing pose within the same non-dominated front in the first sorted initial candidate processing pose set and the second crowding distance between each initial candidate processing pose within the same non-dominated front in the second sorted initial candidate processing pose set. Tournament selection, simulated binary crossover, and polynomial mutation operators are used to generate the next-generation first initial candidate processing pose set and the next-generation second initial candidate processing pose set; Based on the first crowding distance and the second crowding distance, a multi-objective evolutionary iteration is performed based on a dual-population cooperative mechanism to output an initial Pareto optimal candidate processing pose set; The next-generation first initial candidate processing pose set is used as the current first initial candidate processing pose set, and the next-generation second initial candidate processing pose set is used as the current second initial candidate processing pose set. The program returns "the current first initial candidate processing pose set and the current second initial candidate processing pose set are sorted in a non-dominated hierarchical manner to obtain the first sorted initial candidate processing pose set and the second sorted initial candidate processing pose set", and repeats this process multiple times. After merging the initial Pareto optimal candidate processing pose sets obtained from multiple runs, non-dominated hierarchical sorting is performed again, and all individuals in the first non-dominated frontier of the merged population are extracted as the Pareto optimal candidate processing pose set.

6. The graphite component processing pose-assisted decision-making method based on defect detection according to claim 1, characterized in that, Based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto-optimal candidate processing pose set, an anomaly score for each candidate processing pose is obtained using an anomaly detection model based on a graph neural network, specifically including: Based on the defect point cloud distribution information corresponding to each candidate processing pose in the Pareto optimal candidate processing pose set, construct the attribute defect map corresponding to each candidate processing pose. Based on the attribute defect map corresponding to the candidate processing pose, an anomaly score for the candidate processing pose is determined using a graph neural network-based anomaly detection model. The graph neural network-based anomaly detection model is obtained by training a transductive single-class graph representation learning model using a three-stage mixed data training process based on an unsupervised transductive learning paradigm and a dynamic truncation mechanism. The training process specifically includes: Full initialization training: Initialize the network parameters of the transductive single-class graph representation learning model, input the unlabeled full mixed attribute defect map set into the transductive single-class graph representation learning model for forward propagation, and calculate the mean of the graph embedding features to initialize the hypersphere center; the full mixed attribute defect map set includes attribute defect maps corresponding to all candidate processing poses; Dynamic truncation mechanism training: Before updating parameters in each training round, forward propagation is performed on the full mixed attribute defect map to calculate the single-class classification loss of each sample; the single-class classification losses of all samples are sorted globally in ascending order, and the global truncation threshold for the current round is determined by combining the preset anomalous rate; samples with single-class classification losses greater than the global truncation threshold are identified as anomalous samples, and the gradient backpropagation of anomalous samples is blocked; when calculating the batch average loss, the average is calculated only for normal samples, and gradient descent is performed to update the model parameters; Solidify fine-tuning training: Lock the global truncation threshold to the threshold value of the last round of the second training phase to fix the classification boundary of the hypersphere; use only the normal samples to perform gradient descent fine-tuning on the transductive single-class graph representation learning model at a preset learning rate until convergence; The loss function values ​​corresponding to the defect maps of each attribute at convergence are extracted as anomaly scores.

7. The graphite component processing pose-assisted decision-making method based on defect detection according to claim 6, characterized in that, Candidate machining poses are screened based on the anomaly score. Candidate machining poses with abnormal defect structures are eliminated, and the candidate machining pose with the lowest anomaly score among the remaining candidate machining poses is selected as the optimal machining pose for the graphite component to be machined. Specifically, this includes: Multiple different preset anomaly rates are set to form an anomaly rate parameter set. For each preset anomaly rate, three-stage mixed data training is repeatedly performed to obtain multiple sets of anomaly score ranking sequences of the full mixed attribute defect map at different preset anomaly rates. An overall consistency test is performed on multiple sets of abnormal score ranking sequences. Abnormal sequences that deviate significantly from the overall ranking distribution are removed, and valid abnormal score ranking sequences that meet the consistency standard are retained. For any attribute defect graph, calculate the average value of the anomaly scores corresponding to the attribute defect graph in each valid anomaly score ranking sequence, and use the average value as the comprehensive anomaly score of the attribute defect graph. All attribute defect maps are sorted in ascending order according to the comprehensive anomaly score. The attribute defect map with the lowest comprehensive anomaly score is selected, and the candidate machining pose corresponding to the attribute defect map with the lowest comprehensive anomaly score is taken as the optimal machining pose for the graphite component to be machined.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the graphite component processing pose-assisted decision-making method based on defect detection as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the graphite component machining pose-assisted decision-making method based on defect detection as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the graphite component machining pose-assisted decision-making method based on defect detection as described in any one of claims 1-7.

Images