Forging temperature adjustment method and system based on process simulation and neural network model

Abstract

The application provides a forging temperature adjustment method and system based on process simulation and a neural network model, and is applied to the technical field of forging heat treatment processing. The method comprises the following steps: inputting the initial temperature curve of a to-be-processed forging and material thermal physical parameters into a finite element model containing preset boundary conditions to obtain the numerical distribution of a temperature field, a stress field and a strain field; based on the finite element grid data contained in the numerical distribution, a sparse graph is constructed and input into a space-time graph neural network, and the local equivalent plastic strain and the maximum principal stress of the super node in the sparse graph in the future time period are output; when the local equivalent plastic strain is greater than the allowable strain limit of the material and / or the maximum principal stress is greater than the material cracking threshold, the initial temperature curve of the to-be-processed forging is compensated and adjusted based on a preset stress-temperature sensitivity matrix. The method drives a lightweight space-time graph neural network through a multi-physical field finite element model, and realizes dynamic and accurate adjustment of the heating temperature curve of a complex large forging.

Description

Technical Field

[0001] This application relates to the field of forging heat treatment technology, and in particular to a method and system for adjusting the temperature of forgings based on process simulation and neural network model. Background Technology

[0002] In the field of advanced manufacturing, the heat treatment and forming processes of large forgings have extremely high requirements for temperature control. The temperature distribution directly determines the internal structure, mechanical properties and residual stress level of the forgings.

[0003] Traditional temperature control methods primarily rely on human experience or closed-loop control based on proportional-integral-derivative (PID) algorithms, which are typical examples of lag-based control and struggle to predict complex thermo-mechanical coupling effects during the forming process. While numerical simulation techniques such as finite element analysis can provide high-precision stress-strain predictions, their massive computational demands and long processing times make them unsuitable for the real-time dynamic compensation requirements at the second or even millisecond level in production settings. Furthermore, conventional deep learning models struggle to effectively characterize the complex geometric topology of forgings and lack in-depth modeling of the temporal evolution path dependencies during material processing.

[0004] Therefore, how to achieve dynamic and precise adjustment of the heating temperature curve of complex large forgings based on real-time mechanical response prediction is a key problem that urgently needs to be solved. Summary of the Invention

[0005] This application provides a method and system for adjusting the temperature of forgings based on process simulation and neural network models. By driving a lightweight spatiotemporal graph neural network with a multiphysics finite element model, it achieves a leap from offline hourly simulation to online millisecond-level prediction of the mechanical response of complex forgings. It accurately captures the spatiotemporal topological characteristics and path dependence of the forging material evolution. Then, combined with the mapping logic of the stress-temperature sensitivity matrix, it dynamically and accurately compensates and adjusts the temperature curve of the forging, realizing advanced dynamic closed-loop optimization of the forging heating process, which significantly reduces the cracking risk and scrap rate of large forgings during the forming process.

[0006] This application provides a method for adjusting the temperature of forgings based on process simulation and neural network models, including: Using the initial temperature curve and material thermophysical parameters of the forging to be processed as input data, the forming process of the forging to be processed is simulated in a finite element model containing preset boundary conditions, and the numerical distribution of temperature field, stress field and strain field containing time evolution information is obtained. The numerical distribution contains finite element mesh data. Based on the finite element mesh data, a sparse graph with a unified topology is constructed. The super nodes in the sparse graph are used to characterize the geometric aggregation features of the corresponding local regions in the finite element mesh, and carry the average temperature, volume, heat capacity and initial mechanical state features of the local regions. The sparse graph is input into the spatiotemporal graph neural network, and the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in the future time period are output. The spatiotemporal graph neural network is trained based on time-series temperature field feature data and corresponding finite element simulation mechanical response results. When the local equivalent plastic strain exceeds the material's allowable strain limit and / or the maximum principal stress exceeds the material's cracking threshold, the initial temperature curve of the forging to be processed is compensated and adjusted based on a preset stress-temperature sensitivity matrix.

[0007] According to an embodiment of this application, a method for adjusting the temperature of forgings based on process simulation and a neural network model is provided. The spatiotemporal graph neural network includes a spatial graph convolution operator and a time-gated loop unit. The step of inputting the sparse graph into the spatiotemporal graph neural network and outputting the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in a future time period includes: using the spatial graph convolution operator to aggregate spatial features of the supernodes in the sparse graph and extracting the heat-force transfer features in the geometric topology of the forging; using the time-gated loop unit, combined with the heat-force transfer features, to perform temporal evolution modeling of the state of the supernodes in the sparse graph, capturing the path dependence of the forging material in the thermal cycling process, and predicting the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in the future time period.

[0008] According to an embodiment of this application, a method for adjusting the temperature of forgings based on process simulation and neural network model is provided. The method involves constructing a sparse graph with a unified topology based on the finite element mesh data. This includes: converting the finite element mesh data into an initial topology graph, where mesh nodes in the initial topology graph are defined as graph nodes, element connection relationships are defined as graph edges, and the temperature, stress, and strain values ​​of each mesh node are mapped to the initial feature vector of the corresponding graph node; dividing all mesh nodes in the initial topology graph into multiple non-intersecting node clusters using a clustering algorithm based on the geometric distance between the mesh nodes and the material interface properties, with each node cluster defined as a supernode; calculating the weighted average temperature, total volume, total heat capacity, and average initial mechanical state characteristics of all mesh nodes within each node cluster, and using these as the geometric aggregation features of the corresponding supernode; establishing adjacency relationships between supernodes; if there is an original mesh edge connection between two different node clusters, establishing a sparse edge between the corresponding two supernodes, and calculating the edge weight of the corresponding sparse edge based on the contact area and equivalent heat conduction distance between the two different node clusters, thereby generating a sparse graph that retains the original topological features of the forging.

[0009] According to an embodiment of this application, a method for adjusting the temperature of forgings based on process simulation and a neural network model is provided. The method involves using the spatial graph convolution operator to aggregate spatial features of supernodes in the sparse graph and extracting heat-force transfer features from the forging's geometric topology. This includes: targeting supernode v... i According to the supernode v i The connection relationship of the sparse edges determines the supernode v. i The set of adjacent supernodes within the K-hop neighborhood, where the set of adjacent supernodes is used to simulate the physical influence radius of heat conduction and stress transfer, and K is an integer greater than or equal to 1; obtain the supernode v i The neighboring supernode v in the set of neighboring supernodes j And extract the connection to the supernode v i With the adjacent supernode v j edge feature e ij According to the edge feature e ij Including the thermal conduction distance, contact area, and material heterogeneity parameters between supernodes, the adjacent supernode v is calculated. j For the supernode v i Spatial influence weight; based on the influence of each of the adjacent supernodes on the supernode v i The spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i Based on the physical characteristics at the current moment, the heat-force transfer characteristics in the geometric topology of the forging are generated.

[0010] According to an embodiment of this application, a method for adjusting the temperature of a forging based on process simulation and a neural network model is provided, wherein the method adjusts the temperature of the forging based on the temperature of each of the adjacent supernodes. i The spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i Based on the physical characteristics at the current moment, generating the heat-force transfer characteristics in the geometric topology of the forging includes: based on the respective relationships of all adjacent supernodes to the supernode v i Spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i Based on the physical characteristics at the current moment, a node descriptor carrying spatial topological information is generated. According to the weight matrix used to extract the nonlinear mapping law of thermal-mechanical coupling, and the physical characteristics of each of the adjacent supernodes at the current moment, a linear transformation and feature extraction are performed on the node descriptor to obtain the supernode v. i The spatial coupling characteristics at the current moment; the spatial coupling characteristics are defined as the heat-force transfer characteristics in the geometric topology of the forging.

[0011] According to an embodiment of this application, a forging temperature adjustment method based on process simulation and neural network model is provided. The method involves using a time-gated loop unit, combined with the heat-force transfer characteristics, to perform time-series evolution modeling of the state of supernodes in the sparse graph, capturing the path dependence of the forging material during thermal cycling, and predicting the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in the future time period. This includes: constructing an independent time-gated loop unit for each supernode in the sparse graph; and targeting supernode v... i Initialize the supernode v i The hidden state, which is used to encode the supernode v i The accumulated thermodynamic effects, work hardening state, and material damage at historical moments are used to input the spatial coupling characteristics into the supernode v. i In the corresponding time-gated loop unit, the supernode v is updated via an update gate. i The accumulated thermo-mechanical information at the historical moment and the supernode v i The instantaneous thermo-mechanical characteristics at the current moment are proportionally fused, and the supernode v is adjusted through the reset gate. i The influence weight of the historical mechanical state on the current evolution trend is calculated to obtain the result; based on the calculation result, the supernode v is generated. i The local equivalent plastic strain and maximum principal stress during the future time period.

[0012] According to an embodiment of this application, a method for adjusting the temperature of forgings based on process simulation and a neural network model is provided, wherein the supernode v is generated based on the calculation results. i The local equivalent plastic strain and maximum principal stress in the future time period include: generating the supernode v based on the calculation results. i The candidate hidden state; update the supernode v according to the candidate hidden state. i The hidden state of the supernode v is obtained. i The target hidden state at the current moment is used to achieve deep decoupling and fusion characterization of the local geometric topological constraints and the global temporal thermal cycling path of the forging; the target hidden state is input into a preset nonlinear prediction head, and the supernode v is calculated through nonlinear mapping in high-dimensional space. i The corresponding local equivalent plastic strain and maximum principal stress of the local region in the future time period.

[0013] According to an embodiment of this application, a method for adjusting the temperature of a forging based on process simulation and a neural network model is provided. The method involves compensating and adjusting the initial temperature curve of the forging based on a preset stress-temperature sensitivity matrix. This includes: obtaining a first deviation between the local equivalent plastic strain and the allowable strain limit of the material, a second deviation between the maximum principal stress and the cracking threshold of the material, and the preset stress-temperature sensitivity matrix. The stress-temperature sensitivity matrix is ​​obtained through partial differential analysis using a "temperature-stress-deformation" sample library generated offline. Based on the inverse or transpose of the stress-temperature sensitivity matrix, the first deviation and the second deviation are mapped back from the mechanical space to the temperature space, and the temperature correction vector required by the supernodes in the sparse graph for the next control period is calculated. The temperature correction vector is superimposed on the initial temperature curve to generate a dynamic temperature command. Based on the dynamic temperature command, the initial temperature curve of the forging is compensated and adjusted, including peak shaving and valley filling of the heating power or extension of the temperature equalization time.

[0014] This application also provides a forging temperature adjustment system based on process simulation and neural network model, including: The process simulation module is used to simulate the forming process of the forging by taking the initial temperature curve and material thermophysical parameters of the forging to be processed as input data and inputting them into a finite element model containing preset boundary conditions. The simulation results in the numerical distribution of temperature field, stress field and strain field containing time evolution information. The numerical distribution includes finite element mesh data. The model prediction module is used to construct a sparse graph with a unified topology based on the finite element mesh data. The supernodes in the sparse graph are used to characterize the geometric aggregation features of the corresponding local regions in the finite element mesh, and carry the average temperature, volume, heat capacity and initial mechanical state features of the local regions. The sparse graph is input into a spatiotemporal graph neural network, which outputs the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods. The spatiotemporal graph neural network is trained based on time-series temperature field feature data and the corresponding finite element simulation mechanical response results. The temperature adjustment module is used to compensate and adjust the initial temperature curve of the forging to be processed based on a preset stress-temperature sensitivity matrix when the local equivalent plastic strain is greater than the material's allowable strain limit and / or the maximum principal stress is greater than the material's cracking threshold.

[0015] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the forging temperature adjustment method based on process simulation and neural network model as described above.

[0016] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the forging temperature adjustment method based on process simulation and neural network model as described above.

[0017] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the forging temperature adjustment method based on process simulation and neural network model as described above.

[0018] The forging temperature adjustment method and system based on process simulation and neural network model provided in this application simulates the forming process of the forging by inputting the initial temperature curve and material thermophysical parameters of the forging to be processed into a finite element model containing preset boundary conditions. This simulation yields numerical distributions of temperature, stress, and strain fields containing temporal evolution information, where the numerical distributions include finite element mesh data. Based on the finite element mesh data, a sparse graph with a unified topology is constructed. The supernodes in the sparse graph characterize the geometric aggregation features of corresponding local regions in the finite element mesh and carry... The method incorporates the average temperature, volume, heat capacity, and initial mechanical state characteristics of the local region. The sparse graph is input into a spatiotemporal graph neural network, which outputs the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph for future time periods. This spatiotemporal graph neural network is trained based on time-series temperature field characteristic data and corresponding finite element simulation mechanical response results. When the local equivalent plastic strain exceeds the material's allowable strain limit, and / or the maximum principal stress exceeds the material's cracking threshold, the initial temperature curve of the forging to be processed is compensated and adjusted based on a preset stress-temperature sensitivity matrix. This method, driven by a multi-physics finite element model, utilizes a lightweight spatiotemporal graph neural network to achieve a leap from offline hourly simulation to online millisecond-level prediction of the mechanical response of complex forgings. It accurately captures the spatiotemporal topological characteristics and path dependence of the forging material evolution. Furthermore, combined with the mapping logic of the stress-temperature sensitivity matrix, it dynamically and accurately compensates and adjusts the forging temperature curve, achieving advanced dynamic closed-loop optimization of the forging heating process and significantly reducing the cracking risk and scrap rate of large forgings during the forming process. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart of the forging temperature adjustment method based on process simulation and neural network model provided in the embodiments of this application; Figure 2 This is a schematic diagram of the forging temperature adjustment system based on process simulation and neural network model provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions 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, 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.

[0022] To better understand the embodiments of this application, the application scenarios of the forging temperature adjustment method based on process simulation and neural network model provided in the embodiments of this application are first described: The forging temperature adjustment method based on process simulation and neural network model described above can be applied to cutting-edge manufacturing fields that have dual requirements for high density, high reliability, and complex structures. For example, it can be used in the heating process of forgings such as load-bearing components, load-bearing rods, and high-strength plates with complex ribs and deep cavities in the aerospace field; in the heating process of forgings such as high-performance aluminum alloy wheels, steering knuckles, complex shells, and shock-absorbing structural components in the automotive and rail transportation fields; and in the heating process of forgings such as precision structural components and various irregularly shaped high-strength aluminum alloy components that require high fatigue life in the high-end machinery manufacturing field.

[0023] It should be noted that the execution entity involved in the embodiments of this application can be a forging temperature adjustment system based on process simulation and neural network model, or it can be an electronic device. Optionally, the electronic device may include: computer / laptop, mobile terminal, server, electronic assembly equipment and electrical production equipment, etc.

[0024] The following section uses electronic equipment as an example to elaborate on the forging temperature adjustment method based on process simulation and neural network model provided in this application: Figure 1 This is a schematic flowchart of the forging temperature adjustment method based on process simulation and neural network model provided in the embodiments of this application. Figure 1 As shown, the method includes the following steps 101-104.

[0025] Step 101: Using the initial temperature curve and material thermophysical parameters of the forging to be processed as input data, input them into the finite element model containing preset boundary conditions to simulate the forming process of the forging to be processed, and obtain the numerical distribution of temperature field, stress field and strain field containing time evolution information. The numerical distribution includes finite element mesh data.

[0026] The forging to be processed refers to metal billets or semi-finished products undergoing heating, heat treatment, or other forming processes. Optionally, the forging to be processed may include at least: aero-engine turbine disks, shaft forgings, ring forgings, large nuclear power plant main shafts, and various irregularly shaped die forgings.

[0027] The initial temperature curve refers to the preset target heating temperature T that changes over time. set (t) is used to guide heating equipment (such as heating furnace, induction furnace or resistance furnace) to heat, hold or cool the above-mentioned forgings to be processed.

[0028] Material thermophysical parameters refer to the set of parameters reflecting the changes in the physical properties of a forging material with temperature under thermo-mechanical coupling conditions. Optionally, these material thermophysical parameters may include at least: density ρ, specific heat capacity c. p Thermal conductivity λ, coefficient of thermal expansion α, Poisson's ratio ν, and yield strength σ s wait.

[0029] Preset boundary conditions refer to the mathematical constraints used in finite element simulation to simulate the actual production environment of forgings. Optionally, these preset boundary conditions may include at least: ambient temperature T. amb Factors such as convective heat transfer coefficient h, surface emissivity ε, mold contact thermal resistance, displacement constraints, and the distribution of externally applied loads are considered.

[0030] The finite element model is a multiphysics simulation model that discretizes a continuous forging geometry into a finite number of elements and solves the heat conduction equation, mechanical equilibrium equation, and material deformation constitutive equation using numerical calculation methods. Different forgings require different finite element models.

[0031] Temporal evolution information refers to the dynamic evolution of each physical quantity at discrete time steps.

[0032] The numerical distribution of the temperature field refers to the temperature value distribution at various spatial coordinate points inside and on the surface of a forging at a specific moment.

[0033] The numerical distribution of stress field refers to the distribution of tension, pressure and shear force generated inside a forging under thermal stress or external load, and is usually characterized by equivalent stress or maximum principal stress.

[0034] The numerical distribution of the strain field refers to the distribution of geometric deformation caused by the expansion of a forging under stress or heat, including elastic strain, plastic strain, and thermal strain.

[0035] Finite element mesh data refers to a digital collection that includes mesh node coordinates, element topology connections, and physical quantity values ​​(such as temperature, stress, strain, etc.) mapped onto the mesh nodes.

[0036] In step 101, the electronic device first suggests a corresponding three-dimensional solid model based on the actual dimensions of the forging to be processed, and performs non-uniform finite element mesh generation. The mesh is then refined at sharp corners, thin walls, or areas prone to stress concentration in the forging to ensure the convergence and accuracy of the numerical calculation. Then, the electronic device acquires the initial temperature curve and material thermal property parameters of the forging to be processed, as well as a finite element model containing preset boundary conditions. The initial temperature curve and material thermal property parameters are then input into the finite element model. The forming process of the forging to be processed is simulated through the finite element model, simulating the dynamic response of the forging to be processed within a complete thermal cycle. The preset heat conduction equation and material deformation constitutive equation are solved in real time to obtain time-series evolution data at the full time step, i.e., the numerical distribution of temperature field, stress field, and strain field containing time-series evolution information, which serves as the input source for subsequent graph coarsening and neural network training.

[0037] Step 102: Based on the finite element mesh data, construct a sparse graph with a unified topology. The super nodes in the sparse graph are used to characterize the geometric aggregation features of the corresponding local regions in the finite element mesh, and carry the average temperature, volume, heat capacity and initial mechanical state features of the local regions.

[0038] Sparse graphs are topological models derived from the original high-density finite element mesh (i.e., the numerical distribution mentioned above includes finite element mesh data) after order reduction. This sparse graph significantly reduces the total number of nodes (e.g., from 10) by clustering spatially adjacent, physically similar groups of original mesh nodes into a single supernode. 5 ~10 6 The magnitude was reduced to 10 3 (on a scale of magnitude), while preserving the global geometric features and physical connectivity of the forgings to be processed by reconstructing sparse edges.

[0039] In step 102, the electronic device can employ a graph coarsening algorithm. Based on the grid node coordinates and element connection relationships in the finite element mesh data, the grid nodes in the finite element mesh data are divided into multiple non-overlapping local regions. Each local region is logically abstracted as a supernode, thereby achieving spatial scale transformation and dimensionality reduction from micro-units to macro-regions. Combined with the sparse edges between the supernodes, a sparse graph that retains the original geometric features of the forging and has a unified topological structure is constructed. This step, by performing topological reduction and physical equivalent aggregation on massive mesh nodes, significantly compresses the computational load of artificial intelligence (AI) while fully preserving the geometric constraints of the forging, achieving a deep balance between computational accuracy and real-time prediction.

[0040] In some embodiments, the electronic device constructs a sparse graph with a unified topology based on finite element mesh data. This can include: the electronic device converting the finite element mesh data into an initial topology graph, wherein the mesh nodes in the initial topology graph are defined as graph nodes, the element connection relationships are defined as graph edges, and the temperature, stress, and strain values ​​of each mesh node are mapped to the initial feature vector of the corresponding graph node; the electronic device, based on the geometric distance between each mesh node and the material interface properties, uses a clustering algorithm to divide all mesh nodes in the initial topology graph into multiple non-intersecting node clusters, each node cluster being defined as a supernode; for each node cluster, the electronic device calculates the weighted average temperature, total volume, total heat capacity, and average initial mechanical state characteristics of all mesh nodes within the node cluster, and uses these as the geometric aggregation characteristics of the corresponding supernode; the electronic device establishes the adjacency relationships between each supernode, and if there is an original mesh edge connection between two different node clusters, a sparse edge is established between the corresponding two supernodes, and the edge weight of the corresponding sparse edge is calculated based on the contact area and thermal conduction equivalent distance between the two different node clusters, generating a sparse graph that retains the original topology characteristics of the forging.

[0041] In this embodiment of the application, during the construction of a sparse graph, the electronic device can first convert the grid node coordinates in the finite element mesh data into grid nodes in the initial topology graph, and define the grid nodes in the initial topology graph as graph nodes; then, the element topology connection relationship in the finite element mesh data is converted into the element connection relationship definition in the initial topology graph, and the element connection relationship in the initial topology graph is defined as the edge of the graph; at the same time, the physical quantity values ​​mapped on the grid nodes in the finite element mesh data are converted into physical quantity values ​​in the initial topology graph, and the physical quantity values ​​in the initial topology graph are mapped to the initial feature vector of the corresponding graph node.

[0042] Then, the electronic device analyzes all grid nodes in the initial topology graph, obtaining the geometric distance and material interface properties between each grid node. Since the geometric distance determines the physical spatial correlation strength between grid nodes, and the material interface properties determine the boundary continuity of heat-force field transfer, the electronic device performs a clustering operation based on the geometric distance and material interface properties between each grid node. Specifically, the electronic device constructs a weighting function for comprehensive similarity, mapping the geometric distance to the topological attraction between nodes and the material interface properties to physical barrier constraints; when adjacent grid nodes belong to different material regions, for The weighting function assigns a preset interface penalty factor to enforce physical interface isolation during clustering. Based on this weighting function, clustering algorithms (such as Graclus, spectral clustering, or K-means) are used to segment the initial topology graph, aggregating spatially adjacent and physically homogeneous grid nodes into multiple independent node clusters, ensuring that no grid node cluster crosses its corresponding material interface. Finally, the energy centroid of each node cluster is extracted as a supernode, thus completing the spatial reduction from a high-dimensional grid to a sparse topology, i.e., achieving a reduced-order representation of the forging geometry. This step, through dual constraints on geometric distance and material interface, ensures that the sparse graph after topology reduction strictly follows the abrupt boundary of the physical field, effectively avoiding AI prediction bias caused by the erroneous merging of heterogeneous material nodes.

[0043] Next, for each node cluster, the electronic device calculates the weighted average temperature, total volume, total heat capacity, and average initial mechanical state characteristics of all mesh nodes within the cluster, and uses these as the geometric aggregation characteristics of the corresponding supernodes. Based on this, the electronic device can determine the geometric aggregation characteristics of multiple supernodes, ensuring the consistency of energy and mass information during the sparsification process. Finally, the electronic device establishes the adjacency relationships between each supernode. If there is no original mesh edge connection between two different node clusters, it indicates that the local regions corresponding to these two node clusters are not physically adjacent or do not have a thermodynamic interaction interface. In this case, no corresponding connection edge is established in the sparse graph to maintain the sparsity of the structure. If there is an original mesh edge connection between two different node clusters, it indicates that there is a direct heat conduction or stress transfer path between the local regions corresponding to these two node clusters. In this case, a sparse edge is established between the corresponding two supernodes, and the edge weight of the corresponding sparse edge is calculated based on the contact area and equivalent heat conduction distance between the two different node clusters, generating a sparse graph that retains the original topological characteristics of the forging.

[0044] The contact area between two different node clusters is the sum of the areas of all the original unit surfaces of these two node clusters; the equivalent thermal conductivity distance between two different node clusters is the distance between the centroids of these two node clusters. The edge weight of the sparse edge corresponding to these two different node clusters is the ratio of the contact area to the equivalent thermal conductivity distance.

[0045] The above-mentioned process of constructing sparse graphs achieves topology depth reduction driven by physical information, which not only significantly reduces the computational load of the model but also rigorously restores the thermal evolution characteristics of complex forgings, providing key data closed-loop support for achieving millisecond-level high-precision process compensation.

[0046] Step 103: Input the sparse graph into the spatiotemporal graph neural network and output the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods. The spatiotemporal graph neural network is trained based on the time-series temperature field feature data and the corresponding finite element simulation mechanical response results.

[0047] Among them, the spatiotemporal graph neural network is an AI model that combines a spatial graph convolution operator (used to extract topological features of non-Euclidean space) with a deep learning architecture of a recurrent neural network unit (i.e., a time-gated recurrent unit used to extract time series features). It is specifically designed to learn the nonlinear mapping law of the spatial thermal field distribution and time evolution path of forgings to the mechanical response during complex thermodynamic cycles.

[0048] Local equivalent plastic strain is a scalar physical quantity used to characterize the degree of permanent deformation in a local area of ​​a forging to be processed. It is usually calculated based on the deviatoric strain tensor and is used to evaluate the forming quality and microstructure evolution of that local area.

[0049] The maximum principal stress refers to the maximum normal stress acting on a point in a local area of ​​the forging to be processed. It is the core mechanical indicator for determining whether microcracks or macrocracks occur in the forging material.

[0050] In step 103, the electronic device inputs the sparse graph as input data into the spatiotemporal graph neural network. The spatial graph convolution operator captures the thermal conduction and stress correlation features between each supernode, and the time-gated cyclic unit memorizes the work hardening and thermal softening trends at historical moments. This allows for the prediction of the mechanical field distribution vectors of the supernodes in the sparse graph in future time periods, namely the local equivalent plastic strain and maximum principal stress. This step, through feature decoupling and parallel computation in the spatiotemporal dimension, transforms the finite element iterative calculation, which originally took several hours, into a tensor mapping at the millisecond level, achieving instantaneous perception of the internal evolution state of the forging to be processed.

[0051] Optionally, prior to step 103, the spatiotemporal graph neural network is trained based on the following steps: The electronic device first acquires a historical simulation database, which uses time-series temperature field feature data as training input and the corresponding finite element simulation mechanical response results (i.e., historical local equivalent plastic strain and historical maximum principal stress) as true value labels; The electronic device inputs the time-series temperature field feature data into the original spatiotemporal graph neural network to obtain the predicted finite element simulation mechanical response results output by the original spatiotemporal graph neural network; The electronic device updates the model weight parameters in the original spatiotemporal graph neural network according to the loss value (i.e., deviation) between the predicted finite element simulation mechanical response results and the true value labels, until a preset number of iterations is reached or the current prediction accuracy reaches a preset accuracy threshold, thus obtaining the trained spatiotemporal graph neural network.

[0052] The entire training process constructs a proxy model that combines physical rigor and computational efficiency by deeply transferring physical knowledge to the weights of the neural network. This ensures that the AI ​​model can strictly follow the constitutive relationship of thermo-mechanical coupling during inference, thereby achieving extremely high prediction reliability and process robustness under complex working conditions in the production site.

[0053] In some embodiments, the spatiotemporal graph neural network includes a spatial graph convolution operator and a time-gated recurrent unit. The electronic device inputs a sparse graph into the spatiotemporal graph neural network and outputs the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods. This may include: the electronic device performing spatial feature aggregation on the supernodes in the sparse graph through the spatial graph convolution operator to extract the thermal-mechanical transfer features in the forging geometry; the electronic device performing temporal evolution modeling on the state of the supernodes in the sparse graph through the time-gated recurrent unit, combined with the thermal-mechanical transfer features, to capture the path dependence of the forging material during thermal cycling and predict the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods.

[0054] In this embodiment, the electronic device first uses a spatial graph convolution operator to aggregate spatial features of the supernodes in the sparse graph, extracting features that characterize the spatial coupling correlation of the physical field or the spatial interaction between local regions, namely, the thermal-mechanical transfer features in the geometric topology of the forging. These thermal-mechanical transfer features characterize the instantaneous mechanical response potential of the corresponding local region under the current temperature field. Then, the electronic device uses a time-gated loop unit, combined with the aforementioned thermal-mechanical transfer features, to perform temporal evolution modeling of the state of the supernodes in the sparse graph, capturing the path dependence of the forging material during thermal cycling, and predicting the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods. The entire process achieves a deep fit of the complex "thermal-mechanical-time" multidimensional evolution path of the forging material through the dual decoupling of spatial convolution to extract topological correlation and time loop to capture evolution law. Compared with traditional black-box models, this process can accurately capture the stress concentration phenomenon caused by geometric abrupt changes or non-uniform temperature fields inside the forging to be processed, providing a physically realistic prediction benchmark for preventing cracking of the forging to be processed.

[0055] In some embodiments, the electronic device performs spatial feature aggregation on supernodes in a sparse graph using a spatial graph convolution operator to extract heat-force transfer features from the geometric topology of the forging. This may include: targeting supernode v i Electronic devices according to supernode v i Determine the connection relationships of the sparse edges to identify the supernode v. i The set of adjacent supernodes within the K-hop neighborhood, used to simulate the physical influence radius of heat conduction and stress transfer, where K is an integer greater than or equal to 1; the electronic device acquires supernode v i The neighboring supernode v in the set of neighboring supernodes j And extract the connection supernode v i With adjacent supernode v j edge feature e ij The electronic device is based on the edge feature e ij Including the thermal conduction distance, contact area, and material heterogeneity parameters between supernodes, the v of adjacent supernodes is calculated. j For supernode v i Spatial influence weights; the electronic device is based on the respective influence of all neighboring supernodes on supernode v. i The spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i The physical characteristics at the current moment are used to generate the heat-force transfer characteristics in the geometric topology of the forging.

[0056] The number of supernodes in the sparse graph is V, where V is an integer greater than 1.

[0057] In this embodiment of the application, during the process of spatial feature aggregation of supernodes in a sparse graph using a spatial graph convolution operator, the electronic device can perform the following operations on each supernode. Specifically, for the i-th supernode v among V supernodes... i First, based on the supernode v i Determine the connection relationship of the sparse edges to identify the supernode v. i Within a K-hop neighborhood, multiple adjacent supernodes are identified, and a set of adjacent supernodes is constructed to define the physical influence radius of heat conduction and stress transfer. Then, the electronic device traverses each of the multiple adjacent supernodes, targeting each adjacent supernode v. j Extract the corresponding edge features e ij And combined with the edge feature e ij Including contact area, equivalent distance, and material heterogeneity parameters, quantitative calculation of adjacent supernode v j For supernode v i The spatial coupling contribution, i.e., the spatial influence weight, allows the electronic device to obtain the contribution of each of the multiple adjacent supernodes to supernode v. i The spatial influence weights are then combined with the physical characteristics of each of these multiple adjacent supernodes at the current moment, and the supernode v i Based on the physical characteristics at the current moment, heat-force transfer features in the geometric topology of the forging are generated. The entire process transforms complex geometric topological relationships into high-fidelity heat-force transfer features by constructing a spatial weight aggregation mechanism with physical realism. This effectively improves the recognition accuracy of the spatiotemporal graph neural network model for local thermal stress concentration in irregularly shaped parts (i.e., the forgings to be processed mentioned above), ensuring the physical realism of process control decisions.

[0058] In some embodiments, the electronic device, based on the respective supernode v of all neighboring supernodes, i The spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i The physical characteristics at the current moment, generating the heat-force transfer characteristics in the forging geometry topology, can include: electronic devices based on the respective interactions of all adjacent supernodes with supernode v. i Spatial influence weights, the physical characteristics of all adjacent supernodes at the current time, and supernode v i Based on the physical characteristics at the current moment, a node descriptor carrying spatial topological information is generated. The electronic device performs linear transformation and feature extraction on the node descriptor according to the weight matrix used to extract the nonlinear mapping law of thermal-mechanical coupling, and the physical characteristics of all adjacent supernodes at the current moment, to obtain the supernode v. i The spatial coupling characteristics at the current moment; the electronic device will interpret the spatial coupling characteristics as the heat-force transfer characteristics in the geometry of the forging.

[0059] In this embodiment of the application, the electronic device associates the physical characteristics of multiple adjacent supernodes at the current time with supernode v. i The physical features at the current moment are concatenated to obtain multiple target physical features; then, these multiple target physical features are combined with the respective supernodes of multiple adjacent supernodes v. i The spatial influence weights are weighted and summed to obtain the supernode v. i The corresponding target physical characteristics, the supernode v i The corresponding target physical feature is the node descriptor carrying spatial topological information. Then, the electronic device retrieves the weight matrix obtained through training for extracting the nonlinear mapping law of thermo-mechanical coupling, performs matrix multiplication between the node descriptor and the weight matrix, and performs feature extraction in combination with a nonlinear activation function. This allows the device to uncover the deep thermo-mechanical field coupling correlation between the extracted feature and multiple adjacent supernodes, thus obtaining the supernode v. i The spatial coupling characteristics at the current moment are identified and defined as the heat-force transfer characteristics in the forging geometry. The entire process, through a step-by-step approach of first establishing local physical correlations and then extracting deep nonlinear logic, enables the AI ​​model to accurately identify stress concentration risks caused by geometric abrupt changes from complex mesh data, much like a seasoned process expert, significantly improving the reliability of predictions.

[0060] The formula for calculating the heat-force transfer characteristics is as follows: .

[0061] Indicates the characteristics of heat and force transfer. Indicates the current moment. This indicates the current layer of the spatial graph convolution operator; This represents the weight matrix obtained during training for extracting the nonlinear mapping law of thermo-mechanical coupling; This represents the physical constraints of the process of weighted summation based on spatial influence weights, ensuring that the feature aggregation path strictly follows the physical laws of heat conduction and stress diffusion within the forging; Indicates adjacent supernode v j Physical characteristics at the current moment; Indicates supernode v i Physical characteristics at the current moment; Indicates a splicing operation; Represents the set of adjacent supernodes; This represents a non-linear activation function.

[0062] In some embodiments, the electronic device, through a time-gated cyclic unit and in conjunction with thermo-mechanical transfer characteristics, models the temporal evolution of the state of supernodes in a sparse graph, captures the path dependence of the forging material during thermal cycling, and predicts the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods. This may include: the electronic device constructing an independent time-gated cyclic unit for each supernode in the sparse graph; and for supernode v... i The electronic device initializes the supernode v i The hidden state is used to encode the supernode v. i The accumulated thermodynamic effects, work hardening state, and material damage at historical moments; this electronic device inputs spatial coupling characteristics to the supernode v. i In the corresponding time-gated loop unit, the supernode v is updated via the update gate. i The accumulated thermal information and supernode v at historical moments i The instantaneous thermo-mechanical properties at the current moment are proportionally fused, and the supernode v is adjusted by resetting the gate. i The influence weight of the historical mechanical state on the current evolution trend is calculated to obtain the result; based on the calculation result, the electronic device generates a supernode v. i Local equivalent plastic strain and maximum principal stress in the future time period.

[0063] In this embodiment of the application, the electronic device generates a supernode v i During the process of local equivalent plastic strain and maximum principal stress in the future time period, an independent time-gated cyclic unit can be constructed for each supernode in the sparse graph, and feature capture can be performed on the corresponding supernode using each time-gated cyclic unit. Specifically, for supernode v i The electronic device first initializes the supernode v. i The hidden state; then, the electronic device will... i Spatial coupling characteristics Input to the supernode v i In the corresponding time-gated loop unit, the supernode v is updated via the update gate. i The accumulated thermal information and supernode v at historical moments i The instantaneous thermo-mechanical characteristics at the current moment are proportionally fused to simulate the nonlinear process of rapid stress accumulation and plastic work-to-heat conversion in the forging workpiece during the heating and deformation stages, and this is achieved through a reset gate. Adjust the supernode v i The historical mechanical state is weighted in relation to the current evolution trend, capturing the stress relaxation and microstructure recovery characteristics of forging materials during high-temperature creep or holding stages, and obtaining the calculation results for the update gate and the reset gate; based on these two calculation results, the electronic device generates a supernode v.i The local equivalent plastic strain and maximum principal stress in the future time period. The entire process accurately captures the path-dependent characteristics of forging metal materials in complex thermal cycling by introducing a time-gated operator with a memory mechanism (i.e., time-gated loop unit). This not only solves the problem that traditional AI models are difficult to simulate the dynamic evolution of material hardening / softening, but also provides second-level real-time prediction capability to prevent the generation of microcracks inside large forgings.

[0064] Optionally, the door update process is as follows: .

[0065] Optionally, the door reset procedure is as follows: .

[0066] in, This indicates the calculation result of the update gate; This represents the first learnable weight matrix; This represents the initialized supernode v. i The hidden state; This indicates the calculation result for resetting the door; This represents the learnable second weight matrix.

[0067] In some embodiments, the electronic device generates a supernode v based on the calculation results. i The local equivalent plastic strain and maximum principal stress in the future time period may include: Electronic devices generating supernode v based on the calculation results. i The candidate hidden state; the electronic device updates the supernode v based on the candidate hidden state. i The hidden state is obtained by getting the supernode v. i The target hidden state at the current moment is used to achieve deep decoupling and fusion characterization of the local geometric topological constraints and the global temporal thermal cycling path of the forging. The electronic device inputs the target hidden state into a preset nonlinear prediction head, and calculates the supernode v through nonlinear mapping in high-dimensional space. i This corresponds to the local equivalent plastic strain and maximum principal stress of the local area in the future time period.

[0068] In this embodiment of the application, the electronic device calculates the result of the reset gate. and the supernode v i Spatial coupling characteristics Generate the supernode v i Candidate hidden state Then, the electronic device determines the candidate hiding state. Combine the calculation results of the update gate Update supernode v i Hidden state , obtain supernode vi The target's hidden state at the current moment. This will hide the target. As input data, it is fed into a pre-defined nonlinear prediction head (such as a fully connected neural network layer) to determine the hidden state of the target. Perform a nonlinear mapping in high-dimensional space to calculate the supernode v. i This corresponds to the local equivalent plastic strain and maximum principal stress of the local region in the future time period. The entire process constructs an intelligent evolutionary closed loop with physical memory capabilities through deep temporal fusion and high-dimensional nonlinear mapping of spatial coupling features. This not only achieves second-level ultra-fast reasoning of the stress evolution process of the forging to be processed, but also accurately captures the nonlinear hardening and failure characteristics of forging materials under complex thermal cycling through a gating mechanism, significantly improving the reliability of crack prediction under dynamic conditions. Furthermore, the extraction of spatial coupling features effectively solves the problem of the difficulty in coordinating the spatial interaction and temporal cumulative effects of the thermal and stress fields during the heating process of forging materials.

[0069] Optionally, candidate hidden state The calculation formula is: .

[0070] Optionally, the target is hidden. The calculation formula is: .

[0071] in, This represents the product of corresponding elements.

[0072] It should be noted that by using the cascaded architecture of spatial graph convolution operator and time-gated cyclic unit, the nonlinear effects of temperature evolution over time on material properties during the heating process can be effectively learned, significantly improving the accuracy of dynamic strain prediction.

[0073] Step 104: When the local equivalent plastic strain is greater than the material's allowable strain limit and / or the maximum principal stress is greater than the material's cracking threshold, the initial temperature curve of the forging to be processed is compensated and adjusted based on the preset stress-temperature sensitivity matrix.

[0074] Among them, the allowable strain limit of a material is used to characterize the critical value (i.e., the plastic limit) at a specific forging temperature and strain rate at which the forging material can withstand the maximum amount of plastic deformation without macroscopic failure, reflecting the intrinsic formability of the forging material.

[0075] The material cracking threshold is used to characterize the critical energy density or stress state value required for macroscopic cracks to initiate inside the forging material based on a specific mechanical damage criterion (such as the Cockcroft-Latham criterion).

[0076] The preset stress-temperature sensitivity matrix is ​​a mathematical matrix that describes the response gradient of the internal stress field of a forging to changes in the temperature field. It is used to quantify the mapping relationship of how local temperature fluctuations ΔT cause corresponding changes in mechanical characteristic values ​​(such as principal stress σ) at each location. It is also used to characterize the contribution rate of temperature disturbances in each local supernode region of the forging to the target mechanical response characteristics.

[0077] In step 104, the electronic device, having determined the local equivalent plastic strain and maximum principal stress of each supernode in the sparse graph in step 103, analyzes all local equivalent plastic strains and all maximum principal stresses. Specifically, if all local equivalent plastic strains are less than or equal to the material's allowable strain limit and all maximum principal stresses are less than or equal to the material's cracking threshold, it indicates that the overall forming state of the forging under the current process path is within a safe range and there is no risk of failure. In this case, there is no need to compensate or adjust the initial temperature curve of the forging. If there is a local equivalent plastic strain greater than the material's allowable strain limit and / or the maximum principal stress greater than the material's cracking threshold, it indicates that the forging has potential plastic instability or cracking in a local area and requires process optimization through thermal response intervention. In this case, the initial temperature curve of the forging is compensated and adjusted based on the preset stress-temperature sensitivity matrix. This step transforms offline prediction into online compensation, establishing a real-time closed-loop optimization loop based on an AI model. This eliminates quality defects such as tearing and cracking in complex forgings caused by uneven temperature fields at the source. In addition, by extracting the sensitivity matrix and introducing an online compensation mechanism, the electronic equipment can correct the heating power or holding time in real time, controlling residual stress and warpage deformation within a very small range, significantly reducing the scrap rate and extending the service life of the mold.

[0078] In some embodiments, the electronic device compensates and adjusts the initial temperature curve of the forging to be processed based on a preset stress-temperature sensitivity matrix. This may include: the electronic device acquiring a first deviation between the local equivalent plastic strain and the material's allowable strain limit, a second deviation between the maximum principal stress and the material's cracking threshold, and a preset stress-temperature sensitivity matrix, which is obtained through partial differential analysis of a "temperature-stress-deformation" sample library generated in the offline stage; the electronic device mapping the first and second deviations from the mechanical space back to the temperature space according to the inverse or transpose of the stress-temperature sensitivity matrix, and calculating the temperature correction vector required by the supernodes in the sparse graph for the next control period; the electronic device superimposing the temperature correction vector onto the initial temperature curve to generate a dynamic temperature command; and the electronic device compensating and adjusting the initial temperature curve of the forging to be processed according to the dynamic temperature command, including peak shaving and valley filling of heating power or extension of the temperature equalization time.

[0079] In this embodiment, during the compensation and adjustment of the initial temperature curve of the forging to be processed, the electronic device can first obtain the first deviation between the local equivalent plastic strain and the material's allowable strain limit, the second deviation between the maximum principal stress and the material's cracking threshold, and a preset stress-temperature sensitivity matrix. The first deviation and the second deviation are used to characterize the degree of failure risk under the current process parameters and quantify the distance between the current preset process and the safety boundary. Then, the electronic device maps the first deviation and the second deviation from the mechanical space back to the temperature space according to the inverse matrix or transpose matrix of the stress-temperature sensitivity matrix, and calculates the temperature correction vector (i.e., the aforementioned local temperature fluctuation ΔT) required by the supernode corresponding to the first deviation and the second deviation in the next control period. Next, the electronic device superimposes the temperature correction vector onto the initial temperature curve to generate a dynamic temperature command to achieve the compensation and adjustment of the initial temperature curve of the forging to be processed. This allows for thermal stress verification and equipment power constraint limitation on the corrected temperature curve, ensuring that the adjusted temperature change rate is within the tolerable range of the forging material, thereby suppressing stress concentration by changing the temperature field distribution. The entire process constructs a linearized mapping relationship between mechanical failure risk and thermal energy input through a sensitivity matrix, transforming the invisible material failure hazard into a quantifiable temperature compensation command. This allows for real-time correction of deficiencies in the original process (i.e., the initial temperature curve) before processing, ensuring that the forging to be processed always undergoes thermoplastic deformation within a physically safe range.

[0080] Optionally, the method may further include: an electronic device outputting a dynamic temperature command to the actuator of the heating control system, thereby dynamically adjusting the induction heating frequency or furnace temperature setpoint to achieve online real-time compensation for forging deformation and stress. The entire process utilizes the sensitivity of the induction frequency to the heating depth to achieve directional intervention in the three-dimensional thermal field distribution inside the forging to be processed, fundamentally eliminating the stress conditions that cause cracks. Simultaneously, it effectively offsets the impact of environmental fluctuations or raw material composition deviations on forming quality, ensuring the consistency of each forging.

[0081] It should be noted that, in actual operation, the technical solutions described in steps 101-104 only need to acquire the sparse temperature measurement point data collected by the current sensor (i.e., the temperature of several points on the surface of the forging), and map it to the nodes of the coarsening diagram through interpolation. The deformation prediction of the entire workpiece can be completed within <50ms. If the predicted value is greater than the safety threshold, the control module is triggered to correct the temperature curve, thereby correcting the heating power or holding time of the next step in reverse.

[0082] In this embodiment, the technical solution described in steps 101-104 above uses a lightweight spatiotemporal graph neural network driven by a multiphysics finite element model to achieve a leap from offline hourly simulation to online millisecond-level prediction of the mechanical response of complex forgings. It accurately captures the spatiotemporal topological characteristics and path dependence of the forging material evolution, and then combines the mapping logic of the stress-temperature sensitivity matrix to dynamically and accurately compensate and adjust the temperature curve of the forging, realizing advanced dynamic closed-loop optimization of the forging heating process, which significantly reduces the cracking risk and scrap rate of large forgings during the forming process.

[0083] The forging temperature adjustment system based on process simulation and neural network model provided in the embodiments of this application is described below. The forging temperature adjustment system based on process simulation and neural network model described below can be referred to in correspondence with the forging temperature adjustment method based on process simulation and neural network model described above.

[0084] Figure 2 This is a schematic diagram of the forging temperature adjustment system based on process simulation and neural network model provided in an embodiment of this application. Figure 2 As shown, the system includes: a process simulation module 201, a model prediction module 202, and a temperature adjustment module 203.

[0085] The process simulation module 201 is used to simulate the forming process of the forging with the initial temperature curve and material thermophysical parameters of the forging to be processed as input data into a finite element model containing preset boundary conditions, and obtain the numerical distribution of temperature field, stress field and strain field containing time evolution information. The numerical distribution contains finite element mesh data. The model prediction module 202 is used to construct a sparse graph with a unified topology based on the finite element mesh data. The super nodes in the sparse graph are used to characterize the geometric aggregation features of the corresponding local regions in the finite element mesh, and carry the average temperature, volume, heat capacity and initial mechanical state features of the local regions. The sparse graph is input into the spatiotemporal graph neural network, which outputs the local equivalent plastic strain and maximum principal stress of the super nodes in the sparse graph in future time periods. The spatiotemporal graph neural network is trained based on the time-series temperature field feature data and the corresponding finite element simulation mechanical response results. The temperature adjustment module 203 is used to compensate and adjust the initial temperature curve of the forging to be processed based on a preset stress-temperature sensitivity matrix when the local equivalent plastic strain is greater than the material's allowable strain limit and / or the maximum principal stress is greater than the material's cracking threshold.

[0086] Optionally, the spatiotemporal graph neural network includes a spatial graph convolution operator and a time-gated recurrent unit. The model prediction module 202 is specifically used to perform spatial feature aggregation on the supernodes in the sparse graph through the spatial graph convolution operator to extract the thermal-mechanical transfer features in the geometric topology of the forging; and to perform temporal evolution modeling on the state of the supernodes in the sparse graph by combining the thermal-mechanical transfer features through the time-gated recurrent unit, to capture the path dependence of the forging material in the thermal cycling process, and to predict the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in the future time period.

[0087] Optionally, the model prediction module 202 is specifically used to convert the finite element mesh data into an initial topology graph, wherein the mesh nodes in the initial topology graph are defined as graph nodes, the element connection relationships are defined as graph edges, and the temperature, stress, and strain values ​​of each mesh node are mapped to the initial feature vectors of the corresponding graph nodes; based on the geometric distance between each mesh node and the material interface properties, a clustering algorithm is used to divide all mesh nodes in the initial topology graph into multiple non-intersecting node clusters, and each node cluster is defined as a supernode; for each node cluster, the weighted average temperature, total volume, total heat capacity, and average initial mechanical state characteristics of all mesh nodes in the node cluster are calculated and used as the geometric aggregation characteristics of the corresponding supernode; the adjacency relationship between each supernode is established, and if there is an original mesh edge connection between two different node clusters, a sparse edge is established between the corresponding two supernodes, and the edge weight of the corresponding sparse edge is calculated based on the contact area and thermal conduction equivalent distance between the two different node clusters, generating a sparse graph that retains the original topology characteristics of the forging.

[0088] Optionally, the model prediction module 202 is specifically used for supernode v i According to the supernode v i Determine the connection relationship of the sparse edges to identify the supernode v. i The set of neighboring supernodes within the K-hop neighborhood, used to simulate the physical influence radius of heat conduction and stress transfer, where K is an integer greater than or equal to 1; obtain the supernode v i The neighboring supernode v in the set of neighboring supernodes j And extract the connection to the supernode v i With the adjacent supernode v j edge feature e ij According to the edge feature e ij Including the thermal conduction distance, contact area, and material heterogeneity parameters between supernodes, calculate the v of the adjacent supernode. j For the supernode v i Spatial influence weight; based on the respective influence of all adjacent supernodes on the supernode v iThe spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i The physical characteristics at this current moment generate the heat-force transfer characteristics in the geometry of the forging.

[0089] Optionally, the model prediction module 202 is specifically used to predict the value of each of the neighboring supernodes based on the value of the supernode v. i Spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i Based on the physical characteristics at the current moment, a node descriptor carrying spatial topological information is generated. According to the weight matrix used to extract the nonlinear mapping law of thermal-mechanical coupling, and the physical characteristics of all adjacent supernodes at the current moment, a linear transformation and feature extraction are performed on the node descriptor to obtain the supernode v. i The spatial coupling characteristics at this current moment; the spatial coupling characteristics are defined as the heat-force transfer characteristics in the geometric topology of the forging.

[0090] Optionally, the model prediction module 202 is specifically used to construct an independent time-gated cyclic unit for each supernode in the sparse graph; for supernode v i Initialize the supernode v i The hidden state, which is used to encode the supernode v i The accumulated thermodynamic effects, work hardening state, and material damage at historical moments are used to input this spatial coupling characteristic into the supernode v. i In the corresponding time-gated loop unit, the supernode v is updated via the update gate. i The accumulated thermo-mechanical information at this historical moment and the supernode v i The instantaneous thermo-mechanical properties at the current moment are proportionally fused, and the supernode v is adjusted by resetting the gate. i The influence weight of the historical mechanical state on the current evolution trend is calculated, and the result is obtained; based on the calculation result, the supernode v is generated. i The local equivalent plastic strain and maximum principal stress during this future period.

[0091] Optionally, the model prediction module 202 is specifically used to generate the supernode v based on the calculation result. i The candidate hidden state; update the supernode v based on the candidate hidden state. i The hidden state of the supernode v is obtained. i The target hidden state at the current moment is used to achieve deep decoupling and fusion characterization of the local geometric topological constraints and the global temporal thermal cycling path of the forging. This target hidden state is input into a preset nonlinear prediction head, and the supernode v is calculated through nonlinear mapping in high-dimensional space.i The corresponding local equivalent plastic strain and maximum principal stress in the local area during this future time period.

[0092] Optionally, the temperature adjustment module 203 is specifically used to obtain the first deviation between the local equivalent plastic strain and the allowable strain limit of the material, the second deviation between the maximum principal stress and the cracking threshold of the material, and the preset stress-temperature sensitivity matrix. The stress-temperature sensitivity matrix is ​​obtained by partial differential analysis through a "temperature-stress-deformation" sample library generated in the offline stage. Based on the inverse or transpose of the stress-temperature sensitivity matrix, the first deviation and the second deviation are mapped from the mechanical space back to the temperature space, and the temperature correction vector required by the supernodes in the sparse graph for the next control period is calculated. The temperature correction vector is superimposed on the initial temperature curve to generate a dynamic temperature command. Based on the dynamic temperature command, the initial temperature curve of the forging to be processed is compensated and adjusted, including peak shaving and valley filling of heating power or extension of the temperature equalization time.

[0093] Figure 3 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device may include a processor 310, a communication interface 320, a memory 330, and a communication bus 340. The processor 310, communication interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a forging temperature adjustment method based on process simulation and a neural network model. This method includes: using the initial temperature curve and material thermophysical parameters of the forging to be processed as input data, inputting them into a finite element model containing preset boundary conditions to simulate the forming process of the forging to be processed, obtaining numerical distributions of temperature field, stress field, and strain field containing time-series evolution information, wherein the numerical distributions include finite element mesh data; and constructing a sparse graph with a unified topology based on the finite element mesh data, wherein the supernodes in the sparse graph are used to characterize corresponding local regions in the finite element mesh. The geometric aggregation features are carried, along with the average temperature, volume, heat capacity, and initial mechanical state characteristics of the local region. The sparse graph is input into a spatiotemporal graph neural network, which outputs the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods. The spatiotemporal graph neural network is trained based on time-series temperature field feature data and corresponding finite element simulation mechanical response results. When the local equivalent plastic strain is greater than the material's allowable strain limit, and / or the maximum principal stress is greater than the material's cracking threshold, the initial temperature curve of the forging to be processed is compensated and adjusted based on a preset stress-temperature sensitivity matrix.

[0094] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the forging temperature adjustment method based on process simulation and neural network model provided by the above methods. The method includes: using the initial temperature curve and material thermophysical parameters of the forging to be processed as input data, inputting them into a finite element model containing preset boundary conditions to simulate the forming process of the forging to be processed, and obtaining the numerical distribution of temperature field, stress field and strain field containing time-series evolution information, wherein the numerical distribution includes finite element mesh data; and constructing a unified topology based on the finite element mesh data. A sparse graph of the structure is generated, in which supernodes characterize the geometric aggregation features of corresponding local regions in the finite element mesh, and carry the average temperature, volume, heat capacity, and initial mechanical state characteristics of the local regions. The sparse graph is input into a spatiotemporal graph neural network, which outputs the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods. The spatiotemporal graph neural network is trained based on time-series temperature field feature data and corresponding finite element simulation mechanical response results. When the local equivalent plastic strain is greater than the material's allowable strain limit, and / or the maximum principal stress is greater than the material's cracking threshold, the initial temperature curve of the forging to be processed is compensated and adjusted based on a preset stress-temperature sensitivity matrix.

[0096] Furthermore, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements a forging temperature adjustment method based on process simulation and a neural network model, as provided by the methods described above. This method includes: inputting the initial temperature curve and material thermophysical parameters of the forging to be processed as input data into a finite element model containing preset boundary conditions to simulate the forming process of the forging to be processed, obtaining numerical distributions of temperature field, stress field, and strain field containing time-series evolution information, wherein the numerical distributions include finite element mesh data; and constructing a sparse graph with a unified topology based on the finite element mesh data, wherein the sparse graph contains... Supernodes are used to characterize the geometric aggregation features of corresponding local regions in the finite element mesh, and carry the average temperature, volume, heat capacity, and initial mechanical state features of the local region. The sparse graph is input into a spatiotemporal graph neural network, which outputs the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods. The spatiotemporal graph neural network is trained based on time-series temperature field feature data and corresponding finite element simulation mechanical response results. When the local equivalent plastic strain is greater than the material's allowable strain limit, and / or the maximum principal stress is greater than the material's cracking threshold, the initial temperature curve of the forging to be processed is compensated and adjusted based on a preset stress-temperature sensitivity matrix.

[0097] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for adjusting the temperature of forgings based on process simulation and neural network models, characterized in that, include: Using the initial temperature curve and material thermophysical parameters of the forging to be processed as input data, the forming process of the forging to be processed is simulated in a finite element model containing preset boundary conditions, and the numerical distribution of temperature field, stress field and strain field containing time evolution information is obtained. The numerical distribution contains finite element mesh data. Based on the finite element mesh data, a sparse graph with a unified topology is constructed. The super nodes in the sparse graph are used to characterize the geometric aggregation features of the corresponding local regions in the finite element mesh, and carry the average temperature, volume, heat capacity and initial mechanical state features of the local regions. The sparse graph is input into the spatiotemporal graph neural network, and the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in the future time period are output. The spatiotemporal graph neural network is trained based on time-series temperature field feature data and corresponding finite element simulation mechanical response results. When the local equivalent plastic strain exceeds the material's allowable strain limit and / or the maximum principal stress exceeds the material's cracking threshold, the initial temperature curve of the forging to be processed is compensated and adjusted based on a preset stress-temperature sensitivity matrix.

2. The forging temperature adjustment method based on process simulation and neural network model according to claim 1, characterized in that, The spatiotemporal graph neural network includes a spatial graph convolution operator and a time-gated recurrent unit. The step of inputting the sparse graph into the spatiotemporal graph neural network and outputting the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in future time periods includes: The spatial graph convolution operator is used to perform spatial feature aggregation on the supernodes in the sparse graph to extract the heat-force transfer features in the geometric topology of the forging. By using a time-gated loop unit and combining the heat-force transfer characteristics, the state of the supernodes in the sparse graph is modeled in a time-series manner to capture the path dependence of the forging material during thermal cycling and predict the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in the future time period.

3. The forging temperature adjustment method based on process simulation and neural network model according to claim 1, characterized in that, The construction of a sparse graph with a unified topology based on the finite element mesh data includes: The finite element mesh data is transformed into an initial topology graph, wherein the mesh nodes in the initial topology graph are defined as graph nodes, the element connection relationships are defined as graph edges, and the temperature, stress, and strain values ​​of each mesh node are mapped to the initial feature vector of the corresponding graph node. Based on the geometric distance between each grid node and the material interface properties, a clustering algorithm is used to divide all grid nodes in the initial topology into multiple non-overlapping node clusters, and each node cluster is defined as a supernode. For each node cluster, the weighted average temperature, total volume, total heat capacity, and average initial mechanical state characteristics of all grid nodes within the node cluster are calculated and used as the geometric aggregation characteristics of the corresponding supernodes. Establish the adjacency relationship between each supernode. If there is an original mesh edge connection between two different node clusters, then establish a sparse edge between the corresponding two supernodes. Calculate the edge weight of the corresponding sparse edge based on the contact area and thermal conduction equivalent distance between the two different node clusters to generate a sparse graph that retains the original topological features of the forging.

4. The forging temperature adjustment method based on process simulation and neural network model according to claim 2, characterized in that, The step of using the spatial graph convolution operator to aggregate spatial features of the supernodes in the sparse graph and extracting heat-force transfer features from the forging geometry includes: For a supernode v i , according to the connection relationship of the sparse edge where the supernode v i is located, a set of adjacent supernodes within a K-hop neighborhood of the supernode v i is determined, the set of adjacent supernodes is used to simulate a physical influence radius of heat conduction and stress transmission, and K is an integer greater than or equal to 1. obtaining the super node v i a neighboring super node v j in the set of neighboring super nodes i extracting the edge feature e j connecting the super node v ij and the neighboring super node v According to the edge feature e ij Including the thermal conduction distance, contact area, and material heterogeneity parameters between supernodes, the adjacent supernode v is calculated. j For the supernode v i Spatial influence weight; Based on the respective relationships of all adjacent supernodes to the supernode v i The spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i Based on the physical characteristics at the current moment, the heat-force transfer characteristics in the geometric topology of the forging are generated.

5. The forging temperature adjustment method based on process simulation and neural network model according to claim 4, characterized in that, The process is based on each of the adjacent supernodes for the supernode v. i The spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i Based on the physical characteristics at the current moment, the heat-force transfer characteristics in the geometry of the forging are generated, including: Based on the respective relationships of all adjacent supernodes to the supernode v i Spatial influence weights, the physical characteristics of each of the adjacent supernodes at the current time, and the supernode v i Based on the physical characteristics at the current moment, a node descriptor carrying spatial topology information is generated; Based on the weight matrix used to extract the nonlinear mapping law of thermo-mechanical coupling, and the physical characteristics of all adjacent supernodes at the current time, the node descriptor is linearly transformed and features are extracted to obtain the supernode v. i Spatial coupling characteristics at the current moment; The spatial coupling feature is defined as the heat-force transfer feature in the geometry of the forging.

6. The forging temperature adjustment method based on process simulation and neural network model according to claim 2, characterized in that, The time-gated loop unit, combined with the heat-force transfer characteristics, models the temporal evolution of the supernodes in the sparse graph, captures the path dependence of the forging material during thermal cycling, and predicts the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in the future time period, including: Construct an independent time-gated loop unit for each supernode in the sparse graph; For supernode v i Initialize the supernode v i The hidden state, which is used to encode the supernode v i The accumulated thermodynamic effects, work hardening state, and degree of material damage at historical moments; The spatial coupling feature is input to the supernode v. i In the corresponding time-gated loop unit, the supernode v is updated via an update gate. i The accumulated thermo-mechanical information at the historical moment and the supernode v i The instantaneous thermo-mechanical characteristics at the current moment are proportionally fused, and the supernode v is adjusted through the reset gate. i The weight of the influence of historical mechanical states on the current evolutionary trend is used to obtain the calculation results; Based on the calculation results, the supernode v is generated. i The local equivalent plastic strain and maximum principal stress during the future time period.

7. The forging temperature adjustment method based on process simulation and neural network model according to claim 6, characterized in that, The supernode v is generated based on the calculation results. i The local equivalent plastic strain and maximum principal stress in the future time period include: Based on the calculation results, the supernode v is generated. i The candidate hidden state; Update the supernode v according to the candidate hidden state. i The hidden state of the supernode v is obtained. i The target hiding state at the current moment is used to achieve deep decoupling and fusion characterization of the local geometric topology constraints and the global temporal thermal cycle path of the forging; The target hidden state is input into a preset nonlinear prediction head, and the supernode v is calculated through nonlinear mapping in high-dimensional space. i The corresponding local equivalent plastic strain and maximum principal stress of the local region in the future time period.

8. The forging temperature adjustment method based on process simulation and neural network model according to any one of claims 1-7, characterized in that, The compensation adjustment of the initial temperature curve of the forging to be processed based on the preset stress-temperature sensitivity matrix includes: The first deviation between the local equivalent plastic strain and the allowable strain limit of the material, the second deviation between the maximum principal stress and the cracking threshold of the material, and the preset stress-temperature sensitivity matrix are obtained. The stress-temperature sensitivity matrix is ​​obtained by partial differential analysis through the "temperature-stress-deformation" sample library generated in the offline stage. Based on the inverse or transpose of the stress-temperature sensitivity matrix, the first deviation and the second deviation are mapped back from the mechanical space to the temperature space, and the temperature correction vector required by the supernode in the sparse graph for the next control period is calculated. The temperature correction vector is superimposed on the initial temperature curve to generate a dynamic temperature command; According to the dynamic temperature command, the initial temperature curve of the forging to be processed is compensated and adjusted, including peak shaving and valley filling of heating power or extension of the temperature equalization time.

9. A forging temperature adjustment system based on process simulation and neural network model, characterized in that, include: The process simulation module is used to simulate the forming process of the forging by taking the initial temperature curve and material thermophysical parameters of the forging to be processed as input data and inputting them into a finite element model containing preset boundary conditions. The simulation results in the numerical distribution of temperature field, stress field and strain field containing time evolution information. The numerical distribution includes finite element mesh data. The model prediction module is used to construct a sparse graph with a unified topology based on the finite element mesh data. The super nodes in the sparse graph are used to characterize the geometric aggregation features of the corresponding local regions in the finite element mesh, and carry the average temperature, volume, heat capacity and initial mechanical state features of the local regions. The sparse graph is input into the spatiotemporal graph neural network, and the local equivalent plastic strain and maximum principal stress of the supernodes in the sparse graph in the future time period are output. The spatiotemporal graph neural network is trained based on time-series temperature field feature data and corresponding finite element simulation mechanical response results. The temperature adjustment module is used to compensate and adjust the initial temperature curve of the forging to be processed based on a preset stress-temperature sensitivity matrix when the local equivalent plastic strain is greater than the material's allowable strain limit and / or the maximum principal stress is greater than the material's cracking threshold.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the forging temperature adjustment method based on process simulation and neural network model as described in any one of claims 1-8.

Images