CN-122020878-A - Knowledge-graph-based tee joint pouring cooling design method

CN122020878ACN 122020878 ACN122020878 ACN 122020878ACN-122020878-A

Abstract

The invention discloses a three-way pipe joint pouring cold design method based on a knowledge graph, which aims to solve the problems of the three-way pouring system that the arrangement and the size depend on experience, the trial and error are more, the efficiency is low and the stability is poor, the hot-junction driving and process constraint mask pouring cold-drawing generating network is adopted, and the similar subgraph size statistics priori is combined to perform local Bayesian optimization, so that automatic intelligent recommendation of arrangement positions and sizes of the pouring gate, the dead head and the chill is realized, the trial production period is shortened, the stability and consistency are improved, the shrinkage cavity and shrinkage cavity risks are reduced, and the technical effect of casting quality is improved.

Inventors

ZHU XIAOGANG
ZHU HAIMIN

Assignees

苏州创瑞机电科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (8)

1.A three-way pipe joint pouring cooling design method based on a knowledge graph is characterized by comprising the following steps: S1, data acquisition and map construction, namely acquiring three-dimensional models, pouring chilling process schemes and quality inspection results of a plurality of historical tee joint castings, and constructing a pouring chilling knowledge map of a tee joint comprising a geometric view, a pouring chilling view and a defect view, wherein the pouring chilling knowledge map comprises geometric nodes, pouring gate nodes, riser nodes, chill nodes, defect nodes and connection relations of the defect nodes; S2, training a sub-graph representation and generation network, namely identifying hot node in a geometric view according to a preset hot node judgment rule, constructing a hot sub-graph around the hot node, taking a corresponding sub-graph of each hot sub-graph under three views as a multi-view training sample, carrying out multi-view comparison learning to obtain a sub-graph representation model for mapping the hot sub-graph into a hot sub-graph representation vector, taking the hot sub-graph in the geometric view as input, and taking a corresponding sub-graph in a casting cold view as a supervision training casting cold graph generation network, so that a gate node, a riser node, a cold iron node and a connection relation thereof can be generated; S3, analyzing the piece to be designed, namely identifying hot node and constructing a hot node sub-graph according to a hot node judging rule based on an initial structural diagram of the three-way pipe joint casting to be designed, and inputting a sub-graph representation model to obtain characterization vectors of all the hot node sub-graphs; s4, similar retrieval and size statistics, namely calculating a hot section sub-graph representation vector of a historical hot section sub-graph in the knowledge graph by utilizing a sub-graph representation model, selecting a plurality of similar historical hot section sub-graphs for each hot section sub-graph to be designed according to the similarity, and counting size statistical information of a gate node, a riser node and a chill node in the pouring cold view; S5, initial arrangement generation, namely performing feature aggregation on an initial structure diagram, similar historical hot-node subgraphs and characterization vectors of the hot-node subgraphs in a pouring cold-drawing generation network by taking each hot-node as a drive, constructing a process constraint mask according to preset casting process constraint, and generating a pouring cold-drawing initial scheme diagram comprising a gate node, a riser node, a chill node and connection relations thereof under the limitation of the process constraint mask; S6, local Bayesian optimization, namely based on a pouring cold arrangement initial scheme and the size statistical information, taking size parameters of a gate node, a riser node and a cold iron node as optimization variables, establishing a local Bayesian optimization model, taking shrinkage defect risk indexes and shrinkage defect risk indexes as optimization targets, optimizing the size parameters under the condition that modulus constraint, metal liquid fluidity constraint and filling time constraint are met, obtaining updated size parameters of the gate node, the riser node and the cold iron node, and forming a pouring cold arrangement scheme comprising updated size parameters and connection relations thereof; and S7, outputting a design scheme, namely determining the arrangement positions and the sizes of the pouring gates, the dead heads and the chill on the tee joint casting to be designed based on the pouring chill arrangement scheme diagram, and generating a pouring chill design scheme.
2. The knowledge-graph-based tee joint pouring cold design method according to claim 1, wherein the step S1 comprises: collecting and preprocessing a three-dimensional model of a plurality of historical tee joint castings, a corresponding pouring chilling process scheme and a corresponding quality inspection result; Extracting geometric features of a three-dimensional model of each historical three-way pipe joint casting, respectively constructing each pipe section representing a main pipe and a branch pipe and a junction part representing connection between the pipe sections as geometric nodes, recording wall thickness, section size, spatial position and orientation geometric attributes in each geometric node, and establishing connection edges between the corresponding geometric nodes according to actual connection relations between each pipe section and the junction part to form a geometric view; Analyzing a pouring chilling process scheme corresponding to each historical tee joint casting, constructing each pouring gate as a pouring gate node, each riser as a riser node and each chiller as a chiller node, recording the type, the section size, the volume, the spatial position and the technological parameters related to the metal liquid flow in each pouring gate node, each riser node and each chiller node, and establishing connecting edges between the corresponding geometric nodes and the pouring gate nodes, the riser nodes and the chiller nodes according to the actual arrangement positions of the pouring gates, the risers and the chills in the three-dimensional model so as to form a pouring chilling view; Analyzing quality inspection results corresponding to each historical tee joint casting, constructing detected shrinkage cavity defects and shrinkage cavity defects as defect nodes, recording defect types, defect sizes and space positions of defects in a three-dimensional model in each defect node, establishing connection edges between corresponding geometric nodes and the defect nodes according to corresponding pipe sections or intersection parts of the defects in the three-dimensional model, and establishing connection edges between corresponding geometric nodes, gate nodes, riser nodes or chill nodes and the defect nodes when the defects are in process association with the gate nodes, the riser nodes or the chill nodes so as to form defect views; And integrating the geometric view, the pouring cold view and the defect view in the same knowledge graph structure to obtain a pouring cold knowledge graph of the tee joint comprising geometric nodes, gate nodes, riser nodes, chill nodes, defect nodes and relation edges thereof.
3. The knowledge-graph-based tee joint pouring cold design method of claim 1, wherein step S2 comprises: In a geometric view of a pouring cold knowledge graph of a three-way pipe joint, calculating a heat accumulation index of each geometric node based on the local wall thickness and the local modulus of the geometric node and the connection relation between the geometric node and a plurality of pipe sections and three-way junction parts, marking the geometric node with the heat accumulation index larger than a preset threshold value and the geometric node positioned at the three-way junction part as candidate hot node nodes, screening out the geometric nodes meeting the conditions of a preset wall thickness range, a modulus range and a neighborhood range from the candidate hot node nodes according to a preset hot node judging rule, and determining the geometric node as the hot node; Selecting a hot node and a neighborhood geometrical node thereof and connecting edges among the hot node and the neighborhood geometrical node thereof in a geometrical view by taking each hot node as a center to form a hot node sub-graph part in the geometrical view, selecting a gate node, a riser node and a chill node which are directly connected with the hot node and the neighborhood geometrical node thereof in a pouring cold view and connecting edges among the chill node and the geometrical node thereof to form a hot node sub-graph part in the pouring cold view, and selecting a defect node and connecting edges thereof which are related with the hot node and the neighborhood geometrical node thereof or the gate node, the riser node and the chill node in a defect view to form a hot node sub-graph part in the defect view, thereby obtaining a plurality of hot node sub-graphs which are jointly defined by the geometrical view, the pouring cold view and the defect view; Inputting a hot sub-graph part of each hot sub-graph in the geometric view, a hot sub-graph part in the pouring cold view and a hot sub-graph part in the defect view as multi-view training samples into a multi-view contrast learning model, and training to obtain a sub-graph representation model for mapping any hot sub-graph into a hot sub-graph representation vector by maximizing the similarity of the same hot sub-graph represented under the geometric view, the pouring cold view and the defect view and minimizing the similarity of the different hot sub-graphs; And training the pouring cold graph generating network by taking a hot node sub-part in the geometric view as input and a corresponding hot node sub-part in the pouring cold view as supervision output, so that the pouring cold graph generating network can generate corresponding gate nodes, riser nodes, chill nodes and connection relations thereof under the condition of given hot node nodes, neighborhood geometric nodes and connection edges thereof, and a sub-graph representation model and the pouring cold graph generating network are obtained.
4. The knowledge-graph-based tee joint pouring cold design method according to claim 1, wherein the step S3 comprises: Dividing grids and performing geometric analysis on the three-dimensional model by utilizing a sub-graph representation model and a three-dimensional model of the three-way pipe joint casting to be designed, constructing each pipe section representing a main pipe and a branch pipe and a junction part representing connection between the pipe sections as geometric nodes, recording wall thickness, section size, space position and orientation geometric attributes in each geometric node, and establishing connection edges between the corresponding geometric nodes according to the actual connection relation between each pipe section and the junction part in the three-dimensional model so as to form an initial structural diagram representing the overall geometric topological structure and wall thickness distribution of the three-way pipe joint casting to be designed; In the initial structure diagram, according to the same hot node judging rule as in the step S2, calculating a heat accumulation index based on the local wall thickness, the local modulus and the connection relation between each geometric node and the adjacent geometric nodes, marking the geometric nodes with the heat accumulation index larger than a preset threshold value and the geometric nodes positioned at the junction part of the tee joint as candidate hot node, screening out the geometric nodes meeting the conditions of the preset wall thickness range, the modulus range and the neighborhood range from the candidate hot node, and determining the geometric nodes as the hot node; And taking each hot node as a center, selecting the geometric nodes of the hot node and the neighborhood thereof and connecting edges among the hot nodes in the initial structural diagram, constructing a corresponding hot node sub-graph, inputting each hot node sub-graph into a sub-graph representation model to obtain a corresponding hot node sub-graph representation vector, and thus obtaining the initial structural diagram, a plurality of hot node sub-graphs and the hot node sub-graph representation vectors thereof.
5. The knowledge-graph-based tee joint pouring cold design method of claim 1, wherein step S4 comprises: Based on the pouring cold knowledge graph of the three-way pipe joint, calculating each historical hot node sub graph determined by a hot node judgment rule in the pouring cold knowledge graph of the three-way pipe joint one by utilizing a sub graph representation model to obtain a hot node sub graph representation vector corresponding to each historical hot node sub graph; Then, for each hot sub-graph to be designed, obtaining a hot sub-graph representation vector corresponding to the hot sub-graph representation vector, carrying out similarity calculation on the hot sub-graph representation vector and the hot sub-graph representation vectors of all the historical hot sub-graphs, sorting the historical hot sub-graphs according to the similarity, and selecting a plurality of historical hot sub-graphs from the sorting result as similar historical hot sub-graphs corresponding to the hot sub-graph to be designed according to a preset similarity threshold value and/or a preset sorting quantity; For each similar historical hot node sub-graph, searching a pouring cold view of a pouring cold knowledge graph of a three-way pipe joint for a pouring node, a riser node, a cold iron node and a connecting edge thereof which are directly connected with a hot node and a neighborhood geometrical node in the similar historical hot node sub-graph, forming the pouring cold sub-graph corresponding to the similar historical hot node sub-graph by the pouring node, the riser node, the cold iron node and the connecting edge thereof, and reading the size parameters of each pouring node, the riser node and the cold iron node from the pouring cold sub-graph; And carrying out statistical calculation on the size parameters of the gate node, the riser node and the chill node corresponding to all similar historical hot node sub-graphs belonging to the same hot node sub-graph to be designed to obtain the size statistical information of the gate node, the riser node and the chill node corresponding to the hot node sub-graph to be designed, wherein the size statistical information at least comprises the average value, the value range and the discrete degree of each size parameter, and the similar historical hot node sub-graph, the corresponding pouring chill sub-graph and the size statistical information of the gate node, the riser node and the chill node obtained for each hot node sub-graph to be designed.
6. The knowledge-graph-based tee joint pouring cold design method according to claim 1, wherein the step S5 comprises: Based on the geometric attributes of each hot node in an initial structure diagram, neighborhood geometric nodes and connecting edges of the hot node, similar historical hot node sub-graphs corresponding to each hot node and pouring cold sub-graphs of the hot node sub-graphs, constructing a characteristic vector containing the hot node and the neighborhood geometric nodes of the hot node sub-graphs, a hot node sub-graph characteristic vector of the similar historical hot node sub-graphs corresponding to the hot node sub-nodes and joint input characteristics of pouring gate nodes, riser nodes and cold iron node characteristic vectors in the corresponding pouring cold sub-graphs in a pouring cold graph generating network, inputting the joint input characteristics into an attention mechanism module of the pouring cold graph generating network, obtaining attention weights by calculating correlations among the geometric node characteristic vectors, the similar historical hot node sub-graph characteristic vectors and the pouring cold sub-graph characteristic vectors, and weighting and aggregating the characteristic vectors based on the attention weights to obtain characteristic aggregation results for each hot node; According to the feature aggregation result, in a preset candidate arrangement area surrounding each hot node in the initial structure diagram, generating a plurality of candidate gate nodes, candidate riser nodes and candidate chill nodes according to a preset node generation rule, and generating candidate connecting edges according to the spatial positions of each candidate gate node, candidate riser node and candidate chill node relative to the corresponding geometric node and a preset runner topology rule, so as to form a candidate pouring chill arrangement diagram comprising the candidate gate nodes, the candidate riser nodes, the candidate chill nodes and the candidate connecting edges thereof; Performing process constraint judgment on the candidate pouring chill layout according to casting process rules, calculating the wall thickness, modulus, arrangement direction relative to the gravity direction and influence on the filling sequence of the parts where each candidate pouring gate node, candidate riser node and candidate chill node are located, marking the candidate pouring gate node, candidate riser node, candidate chill node and candidate connecting edges thereof which do not meet any one of preset wall thickness requirements, modulus requirements, gravity direction requirements and filling sequence requirements as invalid nodes and invalid connecting edges, constructing a process constraint mask in a pouring chill generating network for shielding the invalid nodes and the invalid connecting edges, and screening and topologically adjusting the residual valid candidate pouring gate nodes, candidate riser nodes, candidate chill nodes and the candidate connecting edges thereof by the pouring chill generating network under the action of the process constraint mask; And outputting a pouring cold arrangement initial scheme diagram only comprising an effective pouring gate node, an effective dead head node, an effective chill node and connecting edges thereof.
7. The knowledge-graph-based tee joint pouring cold design method according to claim 1, wherein the step S6 comprises: Based on the initial scheme diagram of pouring cold arrangement, reading neck diameter, height, section size and volume size parameters of each pouring gate node, each riser node and each chill node one by one from the initial scheme diagram of pouring cold arrangement, combining the size parameters of each node into corresponding size parameter vectors, and taking each size parameter vector as an optimization variable; According to the dimension statistical information of the gate node, the riser node and the chill node, respectively calculating the mean value, the variance and the value range of various dimension parameters, constructing prior distribution for various dimension parameters based on the mean value, the variance and the value range, and taking the prior distribution as prior information of dimension parameter vectors in a local Bayesian optimization model; Establishing an agent model of an objective function with a size parameter vector as input and a shrinkage cavity defect risk index as output by adopting Gaussian process regression, and establishing a corresponding constraint probability model based on modulus constraint, molten metal fluidity constraint and filling time constraint, so that the feasibility probability of the shrinkage cavity defect risk index and the shrinkage cavity defect risk index corresponding to the size parameter vector and meeting the modulus constraint, molten metal fluidity constraint and the filling time constraint can be obtained when the size parameter vector is given; selecting expected improvement or constraint expected improvement from a local Bayesian optimization model as an acquisition function, iteratively selecting candidate size parameter vectors from a size parameter space based on prior distribution and a current agent model in a feasible domain meeting modulus constraint, metal liquid fluidity constraint and filling time constraint, substituting each candidate size parameter vector into a calculation process of a shrinkage defect risk index and a shrinkage defect risk index to obtain a corresponding true evaluation value of the shrinkage defect risk index and the shrinkage defect risk index, and updating a mean function, a covariance function and a constraint probability model of Gaussian process regression according to the actual evaluation value; Repeating the candidate size parameter vector selection, risk index calculation and model updating processes until the improvement of the shrinkage cavity defect risk index and the shrinkage cavity defect risk index is smaller than a preset threshold value or the iteration number reaches a preset upper limit, obtaining a target size parameter vector set which enables the shrinkage cavity defect risk index and the shrinkage cavity defect risk index to be minimum and meets the modulus constraint, the metal liquid fluidity constraint and the filling time constraint, and writing the target size parameter vector set into corresponding pouring gate nodes, riser nodes and chill nodes in the pouring chill setting initial scheme diagram respectively to form a pouring chill setting scheme diagram containing updated pouring gate nodes, riser nodes and chill node size parameters and connection relations thereof.
8. The knowledge-graph-based tee joint pouring cold design method according to claim 1, wherein the step S7 comprises: traversing each gate node, each riser node and each chill node in the casting cold arrangement scheme, reading geometric node identifiers related to each gate node, each riser node and each chill node, spatial positions of the three-dimensional model coordinate system of the three-way pipe joint casting to be designed and dimensional parameters obtained through partial Bayesian optimization, combining the geometric node identifiers and the spatial positions to determine specific arrangement positions of each gate node, each riser node and each chill node on the outer surface or in a cavity of the three-way pipe joint casting to be designed, taking the arrangement position and the size of each gate node and the sectional size of each gate node, the neck diameter and the height dimensional parameters of each riser node as the arrangement position and the size of a riser, taking the arrangement position and the size of each riser node and the modulus, the sectional size and the height dimensional parameters of each riser node as the arrangement position and the size of a chill, and determining the topology relation among each gate, each gate and each riser and each chill in the casting system according to the connection relation among the gate node, the riser node and the chill nodes in the casting cold arrangement scheme, and the arrangement position and the size of the casting cold, and the arrangement position and the size of the riser and the design plan.

Description

Knowledge-graph-based tee joint pouring cooling design method Technical Field The invention relates to the field of intelligent design of metal casting technology, in particular to a three-way pipe joint pouring cold design method based on a knowledge graph. Background In the prior art, the three-way pipe joint casting has the defects of easily generating hot joints and inducing shrinkage holes and shrinkage porosity because the main pipe and the branch pipe form a three-way thick large section at the junction. The arrangement and the size design of a pouring gate, a dead head and a chill of a pouring system are long-term dependent on an analog-to-digital method, an empirical formula and the experience judgment of a field master, and are matched with filling and solidification simulation to carry out multiple trial-and-error iteration. Under the push of digital trend, the industry gradually adopts simulation software, a rule base and partial parameterized templates to evaluate and adjust schemes and tries to utilize a data driving method to predict parameters, but the whole is still mainly based on manual leading, and lacks an intelligent design flow for complex three-way topology. The shortcomings of the prior art are mainly manifested in: 1. the scheme generation relies on experience and simulation repeated trial change, and lacks an end-to-end method capable of automatically calculating and recommending arrangement and size of a pouring gate, a dead head and a chill, and has low efficiency and poor stability; 2. geometry, process and quality data are fractured, knowledge carriers which are used for uniformly modeling geometry, pouring cooling parameters and defect results and can be reused are lacked, and similar case retrieval and migration are difficult to carry out based on hotspots and neighborhood semantics thereof; 3. Most of the existing generation and optimization are regular post-inspection or global coarse-grain adjustment, casting process constraint is difficult to embed in the generation stage, the parameter optimization lacks statistical prior based on similar cases, and comprehensive coordination of shrinkage cavity and shrinkage cavity risks and constraint such as modulus, fluidity and filling time is insufficient. Therefore, a tee joint cold design method capable of solving the defects of the prior art is a problem which needs to be solved by the person skilled in the art. Disclosure of Invention The invention aims to provide a three-way pipe joint pouring cold design method based on a knowledge graph, and provides a technical scheme of multi-view modeling, similar subgraph retrieval, graph generation and local optimization by taking a hot node as a core, aiming at the problems of depending on experience, trial and error and lack of intelligence in the prior art. The method comprises the steps of constructing a knowledge graph of a geometric view, a pouring cold view and a defect view, extracting a hot node sub-graph according to a hot node judging rule, carrying out multi-view comparison learning to obtain a sub-graph representation model, training a pouring cold graph to generate a network, searching similar historical hot node sub-graphs on a casting to be designed, counting a priori in size, combining attention feature aggregation and a process constraint mask to generate a pouring cold arrangement initial scheme, and optimizing the sizes of a pouring gate, a dead head and a cold iron by taking shrinkage cavity risks and shrinkage cavity risks as targets under modulus constraint, metal liquid fluidity constraint and filling time constraint through local Bayesian optimization, and outputting arrangement positions and sizes. The invention has the technical effects of automatic recommendation, shortened trial production period, improved stability, reduced defect risk and improved casting quality. In order to achieve the above purpose, the invention provides a three-way pipe joint pouring cold design method based on a knowledge graph, which comprises the following steps: S1, data acquisition and map construction, namely acquiring three-dimensional models, pouring chilling process schemes and quality inspection results of a plurality of historical tee joint castings, and constructing a pouring chilling knowledge map of a tee joint comprising a geometric view, a pouring chilling view and a defect view, wherein the pouring chilling knowledge map comprises geometric nodes, pouring gate nodes, riser nodes, chill nodes, defect nodes and connection relations of the defect nodes; S2, training a sub-graph representation and generation network, namely identifying hot node in a geometric view according to a preset hot node judgment rule, constructing a hot sub-graph around the hot node, taking a corresponding sub-graph of each hot sub-graph under three views as a multi-view training sample, carrying out multi-view comparison learning to obtain a sub-graph representation model for mapping the hot sub