A method and system for precise control of the nominal force generation point of a large forging press

Abstract

The present application relates to a kind of large-scale forging press nominal force generation point's accurate control method and system, to solve the problem that nominal force generation point space position deviation is difficult to real-time correction due to structural parameter uncertainty, complex coupling and abnormal state interference in the process of forging and pressing.The three-dimensional coordinates of nominal force generation point and its covariance ellipsoid are obtained by finite element discrete modeling, parameter space error sensitivity analysis, dynamic comparison of multi-dimensional sensor array measured data and digital twin model, automatic screening of sensitive parameters and rapid reconstruction of local submodel, and efficient response of physical behavior based on pre-verified alternative model.Combined with the output regulation of multi-axis linkage hydraulic system, the dynamic adjustment of the parallelism of the slide and the table is achieved, the accurate closed-loop control of the nominal force generation point is achieved, and the precision, response speed and robustness of the forging press are improved.

Description

Technical Field

[0001] This invention relates to the field of digital twin modeling and adaptive control technology for large forging forming equipment, and in particular to a precise control method and system for the nominal force generation point of a large forging press. Background Technology

[0002] Large forging presses, as critical forming equipment, are widely used in high-end manufacturing fields such as aviation, shipbuilding, and energy. Their forming accuracy and nominal force control quality directly determine the geometric accuracy and mechanical properties of downstream products. Currently, mainstream digital twin technologies for large forging presses in the industry include approaches based on high-precision finite element global modeling, multi-physics field strong coupling simulation, and integration of traditional reduced-order proxy models. Their goal is to comprehensively reflect complex physical processes such as elastic deformation of the press body structure, die interface compression, worktable deflection, and hydraulic servo coupling, achieving accurate spatial prediction and optimized output control of the nominal force generation point. With the continuous increase in process complexity and precision requirements, the next generation of digital twin models for forging presses is gradually developing towards high-speed closed-loop, scene-adaptive, lightweight modeling, and high robustness.

[0003] Existing technologies still suffer from the following prominent problems: First, the uncertainty of physical parameters (such as the Young's modulus of the fuselage material, the equivalent compressive stiffness of the mold, and the support stiffness of the worktable area) is difficult to fully identify and correct in real time. Compensation can only be achieved through global model re-simulation or periodic manual calibration, resulting in the inability of digital twin models to make timely and efficient accuracy updates to minor perturbations in operating conditions. Second, the non-uniformity of parameter response in the model space is ignored, and existing computing resources are also invested in low-sensitivity areas, resulting in low overall computational efficiency. Third, when encountering sudden abnormal operating conditions such as mold damage or hydraulic system failure, mainstream models can only perform overall verification with a delayed response, failing to achieve "key protection and rapid reset" for important parameters and highly coupled areas, affecting the self-stabilization capability of the entire nominal force closed-loop control system. In addition, mainstream order reduction methods and offline proxy model paths have problems such as weak physical representation capabilities of parameters, insufficient generalization and interpretability, which limit the actual performance of digital twin models under new operating conditions and extreme events. Summary of the Invention

[0004] This application provides a method and system for precise control of the nominal force generation point of a large forging press, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0005] This application provides a method for precisely controlling the nominal force generation point of a large forging press, specifically including: S1: Based on the structural dimensions and typical working condition sequence of a large forging press, the physical parameters of the machine body Young's modulus, the equivalent compressive stiffness of the die, and the support stiffness of the worktable are disturbed in sequence. The magnitude and direction of the influence of each parameter on the three-dimensional coordinate offset of the center point of the lower surface of the slider are recorded, and the parameter space error sensitivity gradient matrix is ​​generated.

[0006] S2: The measured slider posture data of the current forging cycle is collected using a displacement sensor array, and the measured slider posture data is compared with the predicted value of the digital twin model to calculate the spatial residual vector of the nominal force occurrence point.

[0007] S3: Map the spatial residual vector of the nominal force generation point to the parameter space error sensitivity gradient matrix, filter out the physical parameter channels with sensitivity weights higher than a preset threshold, and generate a high-sensitivity parameter calibration set.

[0008] S4: Based on the physical parameter type in the high-sensitivity parameter calibration set, locate the local geometric domain that is strongly coupled with it inside the digital twin model, and logically separate the local geometric domain from the global reference finite element mesh model to generate the local sub-model region to be reconstructed.

[0009] S5: Based on the boundary conditions of the local sub-model region to be reconstructed, a simplified sub-model template verified by real forging load spectrum excitation is matched from the fast response template library to generate a local physical behavior replacement model with millisecond-level response.

[0010] S6: Determine whether the real-time acquired sensor data exhibits a step-like jump abnormal mode. If an abnormal mode is determined to exist, freeze all current parameter updates and input the spatial residual vector of the nominal force occurrence point into the abnormal mode recognizer to generate a local physical behavior instantaneous reset command containing the constitutive characteristics of interface damage.

[0011] S7: Execute the output result projection operation of the local physical behavior replacement model or the local physical behavior instantaneous reset command, directly project the output results of all local sub-models onto the nominal force generation point definition plane, and generate the corrected three-dimensional coordinates and covariance ellipsoid of the nominal force generation point.

[0012] S8: Based on the corrected three-dimensional coordinates and covariance ellipsoid of the nominal force generation point, dynamically adjust the output ratio of multiple sets of hydraulic cylinders or connecting rods to correct the parallelism between the slider and the worktable, and complete the precise closed-loop control of the nominal force generation point.

[0013] This application also provides a precise control system for the nominal force generation point of a large forging press, which uses the above-mentioned precise control method for the nominal force generation point of a large forging press to correct the position of the nominal force generation point.

[0014] This application provides a method and system for precise control of the nominal force generation point of a large forging press, which has the following beneficial effects: (1) To address the problems of heavy computational burden, high response delay, and amplification of parameter uncertainty caused by the traditional digital twin model in the nominal force generation point control of large forging presses due to the "unified modeling-overall solution" paradigm, this invention significantly improves the real-time performance and robustness of the model operation by constructing a parameter sensitivity-driven hierarchical fidelity mechanism. Traditional methods require recalculation or iterative correction of the entire model after each change in operating conditions, which is difficult to meet the timeliness requirements under high-frequency control cycles, and is prone to numerical disturbances due to redundant updates of non-sensitive parameters. In the offline stage, this scheme constructs a parameter-space error gradient matrix through a systematic sensitivity probe experiment to accurately identify the key physical parameters affecting the nominal force positioning accuracy; in the online stage, it only activates the high-sensitivity parameter channel based on the measured residual to achieve directional calibration of 3-5 dominant parameters, avoiding ineffective adjustments to low-sensitivity parameters, effectively suppressing the propagation and diffusion of uncertainty in the multidimensional parameter space, thereby significantly reducing computational overhead while ensuring output accuracy, enabling the digital twin model to complete state updates in hundreds of milliseconds, fully meeting the real-time response requirements of closed-loop control of large forging presses.

[0015] (2) To overcome the shortcomings of existing technologies, such as weak model adaptability and delayed fault response under abnormal working conditions, this invention introduces a local physical behavior instantaneous reset mechanism based on abnormal pattern recognition, which significantly enhances the system's fault tolerance and dynamic adaptability. When sensor data undergoes a step jump, traditional models often rely on the accumulation of data from the next complete cycle to trigger the correction process, leading to continuous accumulation of control deviations or even loss of control. This solution quickly matches typical fault categories such as "nonlinear degradation of mold interface" through an abnormal pattern recognizer, immediately freezes global parameter updates, and then calls a dedicated simplified sub-model template (such as a constitutive model containing interface damage) verified by real loads to forcibly reset the initial boundary values ​​of the relevant regions to the physical response characteristics under the degraded state, achieving millisecond-level reconstruction of key local behaviors. This mechanism does not require waiting for a complete working condition sequence, nor does it rely on complex surrogate model training or neural network inference processes, and can complete the smooth transition from normal to degraded state, ensuring that the prediction of the nominal force occurrence point still has high reliability under sudden faults, and effectively improving the stable operation capability of the digital twin system in complex industrial sites.

[0016] (3) Furthermore, this invention abandons the traditional technical paths that rely on large amounts of computing resources and prior parameter tuning, such as multi-precision fusion, domain decomposition parallel computing, and Kriging fitting. Instead, it innovatively adopts a "physical cognition focusing" strategy, unifying model lightweighting and accuracy within the logical stripping and projection output mechanism of the local geometric domain. By pre-building 127 sets of verified fast-response sub-model templates for strongly coupled areas such as the fuselage column connection area and the mold mounting flange surface, the system performs local solutions only on key areas during operation and directly projects its output onto the nominal force definition plane to generate corrected coordinates and covariance ellipsoids, skipping the full-field stitching and global interpolation stages, greatly reducing intermediate calculation steps and potential error sources. This design not only significantly reduces the dependence of model deployment on hardware computing power and supports long-term stable operation on embedded edge controllers, but also makes the entire digital twin model system have good interpretability and engineering maintainability—every correction can be traced back to the specific physical mechanism and structural area, providing a clear diagnostic basis for on-site operation and maintenance.

[0017] In summary, this solution constructs a digital twin operating system that combines high precision, strong robustness, and low latency through a parameter-sensitive hierarchical fidelity architecture, an anomaly-driven local physical reset mechanism, and a projection-based output strategy oriented towards control objectives. This achieves a paradigm shift from "computation-intensive simulation" to "knowledge-guided decision-making," reliably ensuring the precise controllability of the nominal force generation point of large forging presses without increasing hardware costs. It provides a scalable technical paradigm for the intelligent upgrading of high-end manufacturing equipment. Attached Figure Description

[0018] Figure 1 This is the main flowchart of a method for precisely controlling the nominal force generation point of a large forging press.

[0019] Figure 2 This is a sub-flowchart of a method for precisely controlling the nominal force generation point of a large forging press.

[0020] Figure 3 This is another sub-flowchart of a method for precisely controlling the nominal force generation point of a large forging press. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] like Figure 1 As shown, this application provides a method for precisely controlling the nominal force generation point of a large forging press, specifically including: S1: Based on the structural dimensions and typical working condition sequence of a large forging press, the physical parameters of the machine body Young's modulus, the equivalent compressive stiffness of the die, and the support stiffness of the worktable are disturbed in sequence. The magnitude and direction of the influence of each parameter on the three-dimensional coordinate offset of the center point of the lower surface of the slider are recorded, and the parameter space error sensitivity gradient matrix is ​​generated.

[0024] S2: The measured slider posture data of the current forging cycle is collected using a displacement sensor array, and the measured slider posture data is compared with the predicted value of the digital twin model to calculate the spatial residual vector of the nominal force occurrence point.

[0025] S3: Map the spatial residual vector of the nominal force generation point to the parameter space error sensitivity gradient matrix, filter out the physical parameter channels with sensitivity weights higher than a preset threshold, and generate a high-sensitivity parameter calibration set.

[0026] S4: Based on the physical parameter type in the high-sensitivity parameter calibration set, locate the local geometric domain that is strongly coupled with it inside the digital twin model, and logically separate the local geometric domain from the global reference finite element mesh model to generate the local sub-model region to be reconstructed.

[0027] S5: Based on the boundary conditions of the local sub-model region to be reconstructed, a simplified sub-model template verified by real forging load spectrum excitation is matched from the fast response template library to generate a local physical behavior replacement model with millisecond-level response.

[0028] S6: Determine whether the real-time acquired sensor data exhibits a step-like jump abnormal mode. If an abnormal mode is determined to exist, freeze all current parameter updates and input the spatial residual vector of the nominal force occurrence point into the abnormal mode recognizer to generate a local physical behavior instantaneous reset command containing the constitutive characteristics of interface damage.

[0029] S7: Execute the output result projection operation of the local physical behavior replacement model or the local physical behavior instantaneous reset command, directly project the output results of all local sub-models onto the nominal force generation point definition plane, and generate the corrected three-dimensional coordinates and covariance ellipsoid of the nominal force generation point.

[0030] S8: Based on the corrected three-dimensional coordinates and covariance ellipsoid of the nominal force generation point, dynamically adjust the output ratio of multiple sets of hydraulic cylinders or connecting rods to correct the parallelism between the slider and the worktable, and complete the precise closed-loop control of the nominal force generation point.

[0031] This application also provides a precise control system for the nominal force generation point of a large forging press, which uses the above-mentioned precise control method for the nominal force generation point of a large forging press to correct the position of the nominal force generation point.

[0032] Step S1: Based on the structural dimensions and typical operating condition sequence of the large forging press, the physical parameters of the machine body Young's modulus, the equivalent compressive stiffness of the die, and the support stiffness of the worktable are disturbed sequentially. The amplitude and direction of the influence of each parameter on the three-dimensional coordinate offset of the center point of the lower surface of the slider are recorded, and a parameter space error sensitivity gradient matrix is ​​generated. Specifically, this includes: S1.1: Obtain the structural dimension parameters and typical working condition sequence data of the large forging press. Based on the finite element mesh generation technology, discretize the machine body column, mold mounting surface and workbench support area to generate a global reference finite element mesh model containing node coordinates and element topology relationships, so as to establish the geometric calculation domain of multi-physics coupling simulation.

[0033] Obtain the structural dimension parameter dataset from the large forging press equipment design data management system and the typical working condition sequence confirmed by the process department. Classify the structural dimension parameter dataset according to the spatial area attributes of the machine body column, mold mounting surface and workbench support area, and remove redundant non-mechanical related redundant fields to form the original input matrix that can be used for geometric modeling.

[0034] The CAD geometry parsing interface is called to convert the key dimensional parameters in the original input matrix into three-dimensional geometric feature entities, and a continuous geometric model is established, including the column cross-sectional shape, mounting surface positioning surface and workbench support features. It is also ensured that key process loading path identifiers consistent with typical working condition sequences are introduced into the model.

[0035] For the continuous geometric model, a finite element meshing method based on a mixture of eight-node hexahedral elements and four-node shell elements is implemented. A fine meshing strategy is preferentially used for the fuselage column area to improve the analytical accuracy of local stress concentration areas. A transition mesh is used for the mold mounting surface and the workbench support area to achieve shape quality control of cross-regional elements.

[0036] The mesh quality assessment module is used to evaluate and calculate the shape factor, warpage, and twist rate of the initially divided finite element elements. Elements that fail to meet the quality threshold are locally re-meshed, and node smoothing is performed on the boundary transition region to avoid stress distortion during the numerical simulation process.

[0037] The quality-optimized finite element set is mapped to the node coordinate data through topological relationships to generate a global reference finite element mesh data structure containing unique node numbers, three-dimensional coordinate vectors, and element connection arrays. The data structure is then labeled with material property placeholders and boundary condition interface identifiers required for multiphysics coupling analysis.

[0038] By normalizing node coordinates and standardizing element types, the size and working condition data from the previous step were successfully transformed into a global reference finite element mesh model that can be used for multiphysics coupled simulation. This enabled the accurate discretization of the geometric domain and the establishment of the spatial computational domain for subsequent parameter perturbation analysis.

[0039] For example, in the application scenario of a GK-5000 large forging press, the structural dimension parameter dataset includes a frame column height of 4500 mm and a cross-sectional dimension of 600 × 700 mm; a circular positioning hole array on the die mounting surface with a diameter of 200 mm and a spacing of 620 mm; and a three-point support structure for the worktable support area with a support point spacing of 3000 mm. The typical working condition sequence includes three nominal load levels, with the main working condition load being 50,000 kN and a loading duration of 0.8 seconds. In the geometric modeling stage, the frame column is modeled as a hexahedral element with a refinement of 5 mm, the die mounting surface area has an 8 mm mesh element size, and the worktable support area gradually transitions to a 12 mm mesh size. During mesh quality evaluation, elements with a warpage exceeding 0.4 are removed, and the node coordinates of adjacent elements are iteratively smoothed three times, improving the average element quality index to 0.92. The final generated global baseline finite element mesh model contains a total of 1.25 × 10⁻⁶ nodes. 6 The total number of units is 1.1 × 10 6 All nodes are equipped with material property indexes and load boundary condition identifiers. This model can be directly input into subsequent steps for multiphysics simulation calculations to apply disturbances to the fuselage Young's modulus, mold equivalent compressive stiffness, and workbench support stiffness, significantly improving the calculation accuracy and efficiency of parameter sensitivity analysis.

[0040] S1.2: Based on the global benchmark finite element mesh model, differential perturbations with preset amplitudes are applied to key physical parameters such as Young's modulus of fuselage, equivalent compressive stiffness of mold and support stiffness of workbench in sequence. Multiple sets of independent static simulation calculations are performed to generate a dataset of nodal displacement fields of the lower surface of the slider under different parameter perturbation states.

[0041] Based on the obtained global benchmark finite element mesh model and its node coordinates and element topology relationship data, the physical parameters of fuselage Young's modulus, mold equivalent compressive stiffness and workbench support stiffness are input as independent variables into the static simulation control module and set as an adjustable parameter set.

[0042] Apply a preset amplitude offset that meets the differential perturbation condition to a single parameter in the adjustable parameter set, while keeping the other parameters at their reference values ​​unchanged, to construct a boundary condition input set for single-parameter perturbation.

[0043] The multiphysics statics solver is invoked, and for each set of single-parameter disturbance inputs, the corresponding typical working condition load distribution model and constraint conditions are loaded. A global finite element single independent simulation operation is performed to obtain the overall displacement field solution of the large forging press under the disturbance state.

[0044] The displacement components of the node set on the lower surface of the slider are extracted from the overall displacement field solution and arranged in order of node index to generate a two-dimensional structured node displacement vector.

[0045] Perform index consistency verification on the two-dimensional nodal displacement vectors generated under all different parameter perturbation states, and store them uniformly as a nodal displacement field dataset on the lower surface of the slider according to the perturbation parameter type and amplitude.

[0046] By performing parametric perturbation static simulation and extracting nodal displacement fields based on a global benchmark finite element mesh model, the geometric discretization results of the previous step are transformed into multiple sets of nodal displacement field data of the lower surface of the slider that can be used for sensitivity analysis, thereby realizing offline evaluation of the spatial response characteristics of various physical parameter perturbations to the slider attitude.

[0047] For example, in the global reference finite element mesh model of the GK-5000 large forging press, the Young's modulus of the machine body is set to 210 GPa, the differential perturbation amplitude is set to 5 GPa, the equivalent compressive stiffness of the die is set to 380 GPa, the perturbation amplitude is set to 10 GPa, and the table support stiffness is set to 300 GPa, the perturbation amplitude is set to 8 GPa. Three sets of input conditions are constructed sequentially, one parameter is modified while the other parameters are kept at the reference values, a typical working load of 15000 kN is applied, the base is fixed and the contact boundary conditions between the slider and the guide rail are limited. In the first simulation, the Young's modulus of the machine body is modified to 215 GPa, and a static solution is performed to obtain the displacement vector of all nodes on the lower surface of the slider in the Z direction, with node indices from 1 to 2048; in the second simulation, the equivalent compressive stiffness of the die is changed to 390 GPa, and the above solution is repeated to obtain the corresponding node displacement field; in the third simulation, the table support stiffness is changed to 308 GPa, and the same process is performed. For each simulation result, the nodal displacement vectors are extracted and a node order consistency check is performed. The output two-dimensional matrix of nodal displacements is stored in the nodal displacement field dataset on the lower surface of the slider. For example, the Z-direction displacement of node 100 is 0.0245 mm in the first simulation, 0.0187 mm in the second, and 0.0201 mm in the third. These values ​​will be used for subsequent centroid calculation and gradient calculation to achieve quantitative analysis of parameter sensitivity and significantly improve the stability and physical interpretability of the sensitivity gradient matrix.

[0048] S1.3: Perform centroid calculation on the nodal displacement field dataset of the lower surface of the slider, extract the three-dimensional coordinate offset vector of the center point of the lower surface of the slider under each parameter disturbance condition, and associate the three-dimensional coordinate offset vector with the corresponding physical parameter disturbance amplitude to generate the original mapping record table of parameter disturbance to spatial response.

[0049] Based on the dataset of nodal displacement fields of the lower surface of the slider under different physical parameter disturbance states output in step S1.2, a data pointer index for the set of nodal coordinates is constructed so as to maintain the one-to-one correspondence between the spatial distribution of nodals and the displacement vectors during the centroid calculation.

[0050] The three-dimensional centroid calculation method is invoked to map all nodal displacement components on the lower surface of the slider under each disturbance condition to a unified Cartesian coordinate system. The centroid position of the nodal group is calculated by weighted average. The nodal weights are determined based on the area distribution of the finite element units to ensure that the centroid position can truly represent the overall translation and tilting state of the slider after being subjected to force.

[0051] Perform component-wise difference operations between the three-dimensional coordinates of the centroid under each disturbance condition and the centroid coordinates under the reference state (undisturbed) to obtain the three-dimensional coordinate offset vector of the corresponding condition. Each component represents the positional change amplitude of the slider along the X, Y, and Z directions.

[0052] Using data association mapping logic, corresponding mapping record entries are established for the disturbance amplitude of each physical parameter and the resulting three-dimensional coordinate offset vector according to the disturbance parameter type, forming an original mapping record table with "parameter identifier - disturbance amplitude - three-dimensional coordinate offset vector" as the fields.

[0053] The original mapping record table is checked for data integrity and consistency. Abnormal records caused by local mesh distortion or abnormal calculation convergence are removed to ensure the accuracy and availability of input data for subsequent sensitive gradient calculations.

[0054] By using centroid calculation and differential mapping, a quantitative correlation is established between perturbations of different physical parameters and the spatial response of the slider centroid, and a quantifiable parameter-spatial response original mapping record table is output, thus providing complete and high-precision basic data support for S1.4 gradient calculation.

[0055] S1.4: The numerical difference method is used to perform gradient calculation on the parameter perturbation amplitude and the three-dimensional coordinate offset vector in the original mapping record table. The deflection rate of the center point coordinate of the lower surface of the slider caused by the unit change of each physical parameter is calculated, and the parameter space error sensitivity gradient component characterizing the propagation characteristics of the uncertainty of the physical parameters is generated.

[0056] S1.5: Assemble the parameter space error sensitivity gradient components corresponding to all physical parameters into a matrix according to the parameter type index, construct a complete parameter space error sensitivity gradient matrix, so as to quantitatively store the control weight of each physical parameter on the spatial position of the nominal force occurrence point, and complete the data solidification of the offline sensitivity probe experiment.

[0057] Step S2: The measured slider posture data of the current forging cycle is collected using a displacement sensor array, and the measured slider posture data is compared with the predicted values ​​of the digital twin model to calculate the spatial residual vector of the nominal force occurrence point. Specifically, this includes: The digital twin model is the core virtual simulation platform for achieving precise control of the nominal force generation point of the forging press in this invention. Its core functions include: ① offline stage: completing high-fidelity finite element modeling and parameter calibration; ② online operation stage: receiving real-time sensor data and rapidly deriving the theoretical attitude vector of the lower surface of the slider and the spatial position of the nominal force generation point through a multi-physics coupling solution engine. This theoretical attitude vector is the "digital twin model prediction value," used to compare with the measured slider attitude vector to calculate the spatial residual vector of the nominal force generation point; ③ in step S4, based on the high-sensitivity parameter calibration set, rapidly locating strongly coupled local geometric domains through an internally pre-stored parameter-geometric domain mapping table, and providing geometric domain topological feature data for local sub-model reconstruction. This model adopts a "global-local" hierarchical architecture, ensuring the accuracy of the global mechanical response while supporting rapid reconstruction and millisecond-level response of local highly sensitive areas.

[0058] The digital twin model uses a global baseline finite element mesh model as the geometric and physical discretization basis, and integrates the following four core modules on this basis: Material Constitutive Parameter Library: This library stores the material mechanical parameters of various components of the forging press (machine body, die, worktable), including elastic modulus, Poisson's ratio, density, yield strength, hardening index, etc. For nonlinear behaviors (such as elastoplasticity, creep, and damage) under different temperatures and loading rates, the library also stores the corresponding constitutive model equations and coefficients. The library supports offline calibration and online updates.

[0059] Boundary condition mapping table: Records the constraint types (fixed supports, sliding guides, contact surfaces), force load distribution patterns (concentrated forces, uniformly distributed pressures, equivalent nodal forces), and thermal boundary conditions (temperature field, heat flux density) of various parts of the forging press under typical working conditions. This mapping table quickly converts real-time acquired process parameters (such as total pressure and loading position) into boundary condition vectors required for finite element solutions.

[0060] Multiphysics Coupled Solving Engine: Employing a finite element method-multibody dynamics joint solver, this engine can simultaneously consider the coupling effects of multiple physics fields, such as elastic deformation, thermal expansion, and contact friction in mechanical structures. The engine incorporates an explicit time integration algorithm, enabling simulation of a complete forging cycle within millisecond-level time steps. The engine also integrates the parameter space error sensitivity gradient matrix generated in step S1.4, used to quickly evaluate the influence weights of various physical parameter perturbations on the slider's attitude.

[0061] Parameter-Geometric Domain Mapping Table: This table is a knowledge component within the digital twin model specifically designed to support the reconstruction of local sub-models. It associates highly sensitive physical parameter types (such as fuselage Young's modulus, mold compressive stiffness, and table support stiffness) with their strongly coupled local geometric domains within the global finite element mesh. Each mapping record contains: a physical parameter type identifier (e.g., "fuselage Young's modulus"). 左前立柱 The corresponding local geometric domain name (e.g., "LF-Col"); geometric domain topological feature data (node ​​number array, cell identifier code set, boundary characteristic descriptor); spatial extent of the geometric domain (bounding box coordinates, center point coordinates).

[0062] This mapping table is constructed offline based on the spatial topology and physical influence range analysis of the finite element model. This ensures that in step S4.1, the corresponding local geometric domain feature index can be quickly retrieved based on the calibration set of highly sensitive parameters, thereby generating the local geometric domain feature index set to be located.

[0063] The process of building a digital twin model: 1. Load the global reference finite element mesh model: Read the generated global reference finite element mesh model from the construction results of S1.1, including node coordinates, element topology relationships, material property placeholders and boundary condition interface identifiers, as the geometric and physical discretization basis of the digital twin model.

[0064] 2. Constitutive Parameter Calibration: Using material test data (uniaxial tensile, compression, and fatigue tests) and field-measured force-displacement curves, the constitutive parameters of each component under different temperatures and strain rates are calibrated using a reverse identification method. For contact areas that are difficult to measure directly (such as the mold-workpiece interface), the equivalent contact stiffness and friction coefficient are derived by comparing and iterating offline finite element simulation with measured data.

[0065] 3. Boundary Condition Mapping Rule Establishment: Based on the loading position, pressure distribution, and constraint methods of typical forging conditions, mapping rules are established from real-time process parameters (such as total pressure, slide stroke, and die temperature) to finite element boundary conditions. These mapping rules are stored using interpolation tables or polynomial regression to ensure millisecond-level conversion during online operation.

[0066] 4. Construction of the Parameter-Geometric Domain Mapping Relationship Table: Based on the topological relationships of the global baseline finite element mesh model, a spatial influence range analysis is performed on each type of physical parameter (e.g., changes in the Young's modulus of the fuselage mainly affect the stiffness matrix of the columns and their connected regions). Through finite element sensitivity analysis or analytical methods, the main influence region corresponding to each physical parameter is determined, and the node set, element set, and boundary features of this region are extracted to form a mapping record. All records are summarized into a parameter-geometric domain mapping relationship table and stored in the memory of the digital twin model for rapid retrieval in step S4.

[0067] 5. Multiphysics Coupling Solution Engine Integration: The aforementioned mesh model, material parameter library, boundary condition mapping table, and parameter-geometric domain mapping table are integrated into the solution engine, and full-condition verification is performed. Verification method: Measured slider attitude data from multiple real forging cycles are selected and compared with the prediction results of the solution engine. Parameters such as numerical damping and time step within the engine are adjusted to control the prediction error (root mean square deviation) within a preset threshold (e.g., 0.05 mm).

[0068] 6. Real-time Data Synchronization Interface Deployment: A high-speed data channel is established between the digital twin model and the field sensor network (displacement sensor array, pressure sensor, temperature sensor) to achieve microsecond-level data synchronization. The interface module is responsible for converting the real-time acquired slider attitude data, load data, temperature data, etc., into input vectors for the solution engine, and feeding back the simulation results to the theoretical prediction slider attitude vector generation module in S2.3.

[0069] "Digital twin model prediction value" refers to the theoretically predicted slider attitude vector output by the digital twin model after performing forward simulation deduction through its internal multiphysics coupling solution engine based on the working condition parameters of the current forging cycle (including the elastic deformation parameters of the machine body, the compression parameters of the die, and the deflection parameters of the worktable). This vector has the same time base and spatial coordinate system as the measured slider attitude vector and is used to compare and calculate the spatial residual vector of the nominal force occurrence point in step S2.4.

[0070] Specifically, in step S2.3, the digital twin model, based on the current operating parameters, calls a multiphysics coupling solution engine to perform a joint forward solution of finite element and multibody dynamics, obtaining the displacement, velocity, and acceleration time series of the slider nodes within the simulation period. The six-degree-of-freedom motion components of the measurement points corresponding to the measured slider attitude vector are extracted from these components, and a theoretically predicted slider attitude vector that is strictly aligned with the measured time series is constructed; this is the "digital twin model predicted value".

[0071] For example, taking the GK-5000 large forging press as an example, its digital twin model's global reference finite element mesh contains 1.25 million nodes and 1.1 million elements. The parameter-geometry domain mapping table pre-stores the "fuselage Young's modulus". 左前立柱 The mapping between the "LF-Col" geometric domain and the "LF-Col" geometric domain, which contains an array of node numbers [1024, 1025, 1048, ...] and the corresponding tetrahedral element identifiers. When S3 outputs the highly sensitive parameter "Fuse Young's Modulus",... 左前立柱 Within 0.1ms, the digital twin model returns the topological feature data of the geometric domain, which is then used by S4.2 to generate a set of local geometric domain spatial coordinates. Simultaneously, during one forging cycle, the digital twin model, based on the measured load of 24500kN, calls the solution engine to predict the slider's Z-axis displacement as 2.10mm (i.e., the digital twin model's predicted value). This prediction is compared with the measured value of 2.35mm, yielding a residual of 0.25mm. This residual is used for subsequent correction of the nominal force occurrence point.

[0072] Through the digital twin model and internal parameter-geometric domain mapping table constructed above, this invention realizes real-time, high-precision virtual mapping of the dynamic behavior of the forging press, and provides a reliable data foundation for residual calculation in S2 and rapid local geometric domain positioning in S4.

[0073] S2.1: Based on the multi-channel synchronous sampling displacement sensor array deployed at key measuring points of the machine body column and workbench of a large forging press, the original voltage analog signal within the current forging cycle is acquired at high frequency and noise filtered to generate a multi-dimensional original displacement time series dataset with a unified timestamp reference.

[0074] S2.2: Using the multidimensional original displacement time series dataset as input conditions, execute the spatial coordinate reconstruction method based on rigid body kinematic constraints to map the linear displacement of each measuring point into the measured slider attitude vector representing the six degrees of freedom attitude state of the slider, so as to eliminate the sensor installation angle error and establish the actual pose reference of the slider in three-dimensional space.

[0075] Using the multidimensional raw displacement time-series dataset generated by S2.1 as input, a spatial geometric calibration coefficient library containing the reference frames of the fuselage column and the worktable is invoked to uniformly map the displacements of different measuring points to a common computational coordinate system, thereby eliminating the installation reference differences in the raw outputs of the sensors from each channel. For the time-series data of each measuring point, a zero-bias vector and proportional gain coefficient in the static state of the measuring point are constructed, and linear normalization is performed to restore the physical quantities to millimeter-level true displacement values. Utilizing rigid body kinematic constraints, the reference plane composed of the three displacement measuring points is defined as the local reference coordinate system of the slider. The normal vector direction of this plane is fitted using the least squares method to obtain a preliminary estimate of the attitude angle. Based on the preliminary estimated attitude angle and the three-dimensional positional relationship of each measuring point, a rigid body pose matrix T=[R|t] is constructed, where R is the rotation matrix and t is the translation vector. The precise rotation and translation parameters are obtained by solving the overdetermined equations. The rotation matrix is ​​decomposed using Euler angles to generate an attitude triplet containing roll, pitch, and yaw angles. This triplet is then combined with the three components of the translation vector to form a six-degree-of-freedom attitude vector. This attitude vector is used to eliminate sensor mounting angle errors and establish the actual pose reference of the slider in three-dimensional space.

[0076] By using the spatial coordinate reconstruction processing method based on rigid body kinematic constraints, the multidimensional original displacement time series dataset is transformed into a measured slider attitude vector representing the six-degree-of-freedom attitude state of the slider, thereby achieving high-precision recovery of the three-dimensional pose reference and a unified coordinate reference for subsequent simulation comparison.

[0077] S2.3: Based on the elastic deformation parameters of the fuselage, the compression amount of the mold, and the deflection parameters of the worktable under the current working conditions, the multiphysics coupling solution engine in the digital twin model is called to perform forward simulation and generate a theoretically predicted slider attitude vector that strictly corresponds to the time axis of the measured slider attitude vector, so as to construct a slider motion trajectory reference system under ideal conditions.

[0078] Based on the measured slider attitude vector representing the six-degree-of-freedom attitude state of the slider output by S2.2 and the working condition identification information of the current forging cycle, the elastic deformation parameters of the machine body, such as the elastic modulus of the machine body, damping coefficient, and thermal expansion coefficient, corresponding to the working condition are called from the process database. At the same time, the die material stiffness, contact interface friction coefficient, and other die compression parameters, as well as the flexural deformation parameters such as the moment of inertia of the worktable structure section and the stiffness of the support spring, are loaded.

[0079] The above-mentioned multiple physical field parameters are normalized without dimensions and converted to physical dimensions according to the requirements of the digital twin model input interface. The converted parameter sequence is then bound to the current operating condition timestamp to form a multi-physics joint input vector.

[0080] The multiphysics field joint input vector is imported into the multiphysics field coupled solution engine inside the digital twin model. The thermo-mechanical-structural field is integrally derived in the whole time domain through the finite element-multibody dynamics joint forward solution method to obtain the displacement, velocity and acceleration time sequence of the slider node in the simulation period.

[0081] Based on the obtained full-field displacement information of the slider, the six-degree-of-freedom motion components of the measurement points corresponding to the measured slider attitude vector are extracted, and a theoretically predicted slider attitude vector that is strictly aligned with the measured time series is constructed.

[0082] Time-step interpolation and phase correction are used to align the sampling points of the theoretically predicted slider attitude vector to ensure the matching of its time axis with the measured slider attitude vector in terms of sampling frequency and phase delay.

[0083] By calling the coupled solution engine through the above multiphysics input parameters, the theoretical attitude data corrected by the time axis is output, and the time axis corresponding to the measured slider attitude vector in the previous step is mapped to the slider motion trajectory reference system under ideal conditions, so as to achieve accurate comparability between the two in the same computational domain and time domain.

[0084] S2.4: Input the measured slider attitude vector and the theoretically predicted slider attitude vector into the spatial residual calculation module, perform component-wise vector difference operation and time-domain integral correction processing to generate a nominal force generation point spatial residual vector that characterizes the degree of deviation between the actual point of action and the theoretical center point.

[0085] The measured slider attitude vector obtained by the attitude reconstruction method and the theoretically predicted slider attitude vector output by the multiphysics coupling solution engine are synchronously input into the spatial residual calculation module to establish the correspondence between the two sets of vectors under the same time reference. Component-wise vector difference operations are performed on the two sets of attitude vectors to calculate the instantaneous deviation of each degree of freedom component between the measured and predicted values, resulting in a time-series attitude difference vector set. The obtained attitude difference vector set is input into the time-domain integration correction processing unit, where the difference results are weighted and integrated over the entire forging cycle. The weighting coefficients are determined based on the load amplitude distribution function at the sampling instant to highlight the deviation contribution during the peak forming force stage. The following vector integration formula is used to calculate the three-dimensional spatial residual:

[0086] in, Let be the measured three-dimensional attitude vector of the slider at time t. The predicted 3D attitude vector of the slider at time t. ΔP is the load amplitude normalization weighting coefficient, and ΔP is the three-dimensional spatial residual vector of the nominal force occurrence point obtained by integration. This residual vector is output to the data buffer to ensure that it completes data exchange with the subsequent sensitivity screening module within the same control cycle. Through the above differential and time-domain weighted integration processing method, the instantaneous attitude deviation is transformed into a stable spatial displacement error index characterizing the cumulative effect of the entire forging cycle, realizing the accurate quantification of the deviation between the actual point of action and the theoretical center point.

[0087] For example, during the online operation of the GK-5000 large forging press, eight laser displacement sensors are deployed at the top of the machine body column and the four corners of the worktable to acquire the slider displacement data of the current cycle at a sampling frequency of 1kHz. The measured attitude vector obtained by rigid body kinematic attitude calculation has a Z-axis component of 2.35mm at the moment of forming, while the theoretical predicted value is 2.10mm. The component deviations on the X and Y horizontal axes are 0.12mm and -0.08mm, respectively. The load weighting coefficient calculated by the hydraulic load sensor is 0.95 at the moment of forming and decreases to 0.15 during the unloading stage. Substituting the above time series data into the weighted integral formula and performing numerical integration, the three-dimensional components of ΔP are obtained as (0.105mm, -0.072mm, 0.212mm). This result is stored as the spatial residual vector of the nominal force occurrence point of the current cycle. Subsequent sensitivity analysis showed that the displacement residual in the Z-axis direction was dominant in the mechanical deviation of this cycle, triggering the high-frequency calibration process of the local sub-model related to the table deflection, and achieving a significant improvement and alignment of the position of the Z-axis action point in the next cycle.

[0088] S2.5: Based on the generated nominal force generation point spatial residual vector, perform modulus calculation and direction vector normalization to output standardized nominal force generation point spatial residual feature values ​​containing spatial offset amplitude and off-center load direction characteristics, which serve as the direct driving basis for triggering subsequent high-sensitivity parameter calibration processes.

[0089] like Figure 2 As shown, step S3 involves mapping the nominal force generation point spatial residual vector to the parameter space error sensitivity gradient matrix, filtering out physical parameter channels with sensitivity weights higher than a preset threshold, and generating a high-sensitivity parameter calibration set. Specifically, this includes: S3.1: Obtain the spatial residual vector of the nominal force occurrence point and the parameter space error sensitivity gradient matrix. Use the tensor shrinking method to perform weighted projection operations on each component of the spatial residual vector of the nominal force occurrence point and the corresponding row vector of the parameter space error sensitivity gradient matrix to generate a parameter sensitivity response scalar sequence that characterizes the theoretical contribution of each physical parameter to the current spatial deviation.

[0090] Based on the standardized nominal force generation point spatial residual eigenvalues ​​output by S2.5 and the parameter space error sensitivity gradient matrix constructed in S1.5, residual vectors and sensitivity matrices with the same spatial component dimensions are selected as input datasets to establish their correspondence in the tensor algebraic sense.

[0091] The spatial residual vector of the nominal force generation point is paired with the parameter row vector of the sensitivity matrix in turn according to the coordinate components, so as to ensure that the direction of action of each spatial component is consistent with the parameter corresponding to the physical semantics in mathematical operation.

[0092] The tensor shrinking method is invoked to perform a weighted projection operation on each pair of paired data, and the spatial residual components are used as weight coefficients to apply to each component of the corresponding parameter row vector, thereby realizing a quantitative estimate of the theoretical contribution of each physical parameter under the current spatial deviation conditions.

[0093] The weighted projection process is expressed in the form of a vector inner product, and the calculation formula is as follows:

[0094] in, Let j be the sensitivity response scalar for the j-th physical parameter. Let i be the i-th component of the spatial residual vector. This represents the gradient component corresponding to the i-th row and j-th column of the sensitivity gradient matrix. Let be the dimension of the spatial residual vector.

[0095] The scalars of sensitivity responses to each physical parameter obtained by weighted projection are stored sequentially into the parameter sensitivity response scalar sequence to form a quantitative dataset that corresponds one-to-one with the physical parameter index.

[0096] Through the above processing method, the standardized spatial residual features are mapped into sensitive quantitative indicators that can directly characterize the current importance of physical parameters, thereby providing highly relevant and high-resolution input data for the subsequent S3.2 standardization and screening process.

[0097] S3.2: Based on the parameter sensitivity response scalar sequence, perform absolute value normalization to eliminate the influence of the difference in the dimensions of different physical parameters, and map the normalized values ​​to the dimensionless interval from zero to one to generate a standardized parameter sensitivity weight distribution vector.

[0098] S3.3: Receives the standardized parameter sensitivity weight distribution vector, calls the threshold dynamic judgment logic, and calculates the dynamic truncation threshold by combining the load level of the current forging cycle and the historical noise baseline to generate a sensitivity screening critical value for distinguishing between parameters with significant and insignificant influence.

[0099] The system receives the standardized parameter sensitivity weight distribution vector after absolute value normalization as input, and load level data of the current forging cycle and historical noise baseline parameters are loaded into the threshold calculation unit. The weighted combination module in the dynamic threshold determination logic is called to modulate the historical noise baseline amplitude using the load level as a dynamic gain factor, constructing a noise reference curve that adjusts with changing operating conditions. Based on this noise reference curve, an element-wise signal-to-noise ratio evaluation is performed on the standardized weight distribution vector to obtain a dynamic weight correction sequence characterizing the effective contribution of each physical parameter under the current load and noise background. The correction sequence is input to the statistical analysis module to calculate the mean μ and standard deviation σ as representations of the central tendency and dispersion of the weight distribution. A dynamic truncation threshold is generated using the mean-standard deviation method, where the dynamic truncation threshold T is defined by the following formula:

[0100] Where μ is the arithmetic mean of the dynamically weighted correction sequence, σ is the standard deviation of the sequence, and k is the overshoot coefficient optimized based on historical false positive and false negative rates. By adaptively adjusting the value of k, the cutoff threshold can be automatically increased under high load impact or high noise interference conditions to avoid selecting noise-driven false high-sensitivity parameters; under low load and low noise conditions, the cutoff threshold is appropriately reduced to retain more channels of true high-sensitivity parameters. Using the above dynamic cutoff threshold, the sensitivity scores formed by the standardized weight distribution vector are compared, and parameters greater than T are identified as significantly influential parameters, providing a reliable critical value basis for subsequent high-sensitivity parameter screening. Through this adaptive threshold calculation processing method based on load level and noise baseline, the standardized weight data of the previous step is transformed into dynamically adjustable sensitivity screening critical values, realizing adaptive discrimination of high-sensitivity physical parameters under different operating conditions.

[0101] S3.4: Perform an element-by-element comparison operation on the standardized parameter sensitivity weight distribution vector using the sensitivity screening threshold, extract the index identifiers of all physical parameters whose weight values ​​are greater than the sensitivity screening threshold, and generate a preliminary candidate list of highly sensitive physical parameters.

[0102] Based on the normalized parameter sensitivity weight distribution vector and the dynamically calculated sensitivity screening threshold, the element-by-element calculation module performs amplitude judgment operations on each value in the weight distribution vector, comparing the current element value with the sensitivity screening threshold to form a Boolean judgment result sequence. This sequence is matched one-to-one with the physical parameter index mapping table, filtering out all physical parameter index identifiers whose judgment results are true and whose corresponding weight values ​​are greater than the threshold, thus obtaining a high-sensitivity parameter index set. The index aggregation function is called on the high-sensitivity parameter index set to arrange the indices in order according to their position in the original parameter list, removing duplicate indices to ensure the uniqueness of candidate parameters. The deduplicated high-sensitivity parameter indices are associated and bound to their weight values ​​in the standardized parameter sensitivity weight distribution vector, generating a preliminary high-sensitivity physical parameter candidate list containing parameter identifiers and corresponding weight values. Through the above element-level comparison and index mapping processing method, the continuous weight vector information output in the previous step is transformed into discrete, structured candidate parameter data that can be directly used for subsequent priority sorting and reconstruction domain positioning, achieving the preliminary extraction of key physical parameters that significantly affect the deviation of the nominal force occurrence point.

[0103] S3.5: Based on the preliminary candidate list of highly sensitive physical parameters, perform parameter type deduplication and priority sorting processing, arrange the selected physical parameters in descending order of sensitivity weight and encapsulate them into structured data objects, and generate the final set of highly sensitive parameter calibrations as the input basis for local sub-model reconstruction.

[0104] Based on the initial candidate list of highly sensitive physical parameters, the parameter type identification module is invoked to parse the physical attribute codes and category labels of each candidate parameter, establishing a parameter type index table to support subsequent type deduplication operations. A type conflict detection method is executed to merge physical parameters in the index table that have the same category label but originate from different measurement points or different calculation paths. During the merging process, the highest-valued weight item is retained according to the peak retention principle of sensitivity weights, while redundant records are deleted to reduce the set size. A sorting subroutine is invoked to sort the deduplicated physical parameter set in descending order of sensitivity weight values, ensuring that the physical parameter with the highest weight is given priority in the calculation queue during subsequent local sub-model construction. The physical parameter index generated during the sorting process is bound and encapsulated with the corresponding sensitivity weight value to form a multi-field structured data record containing parameter name, unit, weight value, and type label, meeting the requirement that the data can be directly parsed and called in subsequent steps. Through the above chained processing method, the initial candidate list is refined into a limited number of final highly sensitive parameter calibration sets with ordered weights and unique types, providing a clear input basis for local sub-model reconstruction and achieving the technical effect of focusing computational resources on the parameter channels that have the most impact on the control target.

[0105] For example, in a calibration task for the nominal force application point deviation of a GK-5000 device, the initial candidate list of highly sensitive physical parameters includes five items: fuselage Young's modulus (column X1), fuselage Young's modulus (column X2), workbench support stiffness (left side), mold equivalent compressive stiffness (upper mold), and mold equivalent compressive stiffness (lower mold). Among them, fuselage Young's modulus (column X1) and fuselage Young's modulus (column X2) both belong to the fuselage Young's modulus type. After parsing the physical attribute encoding, type deduplication is performed, retaining only the fuselage Young's modulus (column X1) with a higher sensitivity weight, which has a weight value of 0.92. The four parameters after deduplication are sorted in descending order of sensitivity weight, resulting in an ordered set of fuselage Young's modulus (column X1, 0.92), mold equivalent compressive stiffness (upper mold, 0.88), workbench support stiffness (left side, 0.74), and mold equivalent compressive stiffness (lower mold, 0.71). During encapsulation, each parameter's name, physical unit (MPa, N / μm, etc.), weight value, and type label are combined into a record entry, forming a structured JSON data object for direct use by the S4 step. In this example, the set of highly sensitive parameters arranged in descending order effectively ensures that the local geometric domains of the fuselage column and upper mold contact area are prioritized for sub-model reconstruction within the digital twin model. Through local calculations driven by this set, the nominal force application point offset converges from 2.3mm to 0.4mm, significantly improving position control accuracy.

[0106] like Figure 3 As shown, step S4 involves locating a strongly coupled local geometric domain within the digital twin model based on the physical parameter type in the high-sensitivity parameter calibration set, and logically separating the local geometric domain from the global reference finite element mesh model to generate a local sub-model region to be reconstructed. Specifically, this includes: S4.1: Based on the physical parameter type identifier in the high-sensitivity parameter calibration set, retrieve the pre-stored parameter-geometric domain mapping relationship table, obtain the local geometric domain topological feature data that is strongly coupled with each high-sensitivity physical parameter, and generate the local geometric domain feature index set to be located.

[0107] Based on the physical parameter type identifiers in the high-sensitivity parameter calibration set, the pre-established parameter-geometric domain mapping relationship table index mechanism within the digital twin model is invoked. Each physical parameter index in the input set is used as a retrieval key to perform an exact matching operation on the mapping relationship table to lock the corresponding geometric domain candidate record. The geometric topological feature fields contained in each geometric domain candidate record returned by the mapping relationship table are parsed, and a multi-dimensional topological feature vector composed of domain name, node feature code, unit shape identifier, boundary characteristic descriptor, etc., is extracted to form a mapping pair from high-sensitivity physical parameters to geometric topological features. Redundancy removal processing is performed on the mapping pair set, and duplicate topological feature vectors are eliminated using hash comparison to ensure the uniqueness and efficiency of subsequent spatial retrieval. The redundancy-removed topological feature vector set is then transformed by feature encoding, and a fixed-length topological feature index format is used to unify the encoding method of different types of geometric domain features, generating a structured local geometric domain feature index set to be located. Through the construction of the above index set, the high-sensitivity parameter calibration results are accurately projected in the topological space into explicit index data that can be used for subsequent grid node spatial matching calculations, realizing the accurate association between physical parameters and grid geometric entities.

[0108] For example, during the nominal force application point deviation correction process for a GK-5000 large forging press, the high-sensitivity parameter calibration set includes three physical parameter type identifiers: Young's modulus of the left front column of the machine body, die compressive stiffness, and table front edge support stiffness. The parameter-geometric domain mapping table pre-stores the "LF-Col" geometric domain corresponding to the Young's modulus of the left front column of the machine body. Its topological feature fields include an array of column section node numbers [1024, 1025, 1048, ...] and a tetrahedral element identifier code set. The die compressive stiffness corresponds to the "Mould-Contact" geometric domain, whose topological features include contact surface node identifiers and local normal force boundary distribution descriptors. The table front edge support stiffness corresponds to the "FW-Table" geometric domain, whose topological features include a support rib node group index and node stiffness coefficient labels. After hash deduplication, the three records returned by the matching retrieval retain independent topological feature vectors for the three geometric domains. Subsequently, these vectors are encoded into fixed-length 128-bit feature index codes, such as "0x3A7F…" for the LF-Col field, "0x9B42…" for the Mould-Contact field, and "0x58CC…" for the FW-Table field. The final output feature index set of the local geometric domain to be located contains three structured records, which are used in step S4.2 to perform spatial matching queries in the node coordinate data of the global reference finite element mesh model. This significantly improves the accuracy and retrieval efficiency of local geometric domain location, ensuring that the model reconstruction process focuses on the physical region most significantly affected by off-center loading. This achieves the technical effect of reducing computational latency while ensuring the accuracy of multiphysics coupling.

[0109] S4.2: Use the feature index set of the local geometric domain to be located to perform spatial matching query on the node coordinate data of the global reference finite element mesh model, calculate the Euclidean distance weight between each node and the feature topology center, and generate a set of local geometric domain spatial coordinates containing the influence range of highly sensitive parameters.

[0110] Based on the generated local geometric domain feature index set to be located, the three-dimensional spatial coordinate data matrix indexed by node number in the global reference finite element mesh model node coordinate database is retrieved as input conditions.

[0111] For each geometric feature center point in the feature index set, the spatial geometry calculation module is invoked to perform a subtraction operation between the position vectors of all nodes in the global grid and the position vector of the center point, forming a set of relative displacement vectors from the node to the feature center.

[0112] Based on the Euclidean distance formula, the summation of square terms and the square root operation are performed on each relative displacement vector to obtain the absolute distance scalar value from the node to the feature center.

[0113] The Euclidean distance scalar is input into the weighting function generation module, and combined with the preset radial decay weight model, the distance value is mapped to the influence weight coefficient of the node to form a node-weight correspondence table.

[0114] By using a weight threshold filtering mechanism, for nodes whose weight coefficient is greater than or equal to a preset lower limit, their corresponding original spatial coordinates are extracted to construct a primary spatial coordinate set of the local geometric domain.

[0115] Perform connectivity analysis on the spatial coordinate set, eliminate discrete node clusters that are discontinuous and do not meet the topological requirements of the geometric domain, and retain nodes that meet the continuity requirements as the effective local geometric domain node set under the influence of highly sensitive parameters.

[0116] By using the spatial matching and weight calculation methods described above, the geometric domain index information obtained in the previous step is transformed into a set of three-dimensional spatial coordinates that include the range of action of highly sensitive parameters, thereby achieving accurate positioning and high-fidelity extraction of the local geometric domain in the global reference finite element mesh model.

[0117] S4.3: Extract the corresponding element connection relationships and material property data from the global reference finite element mesh model based on the set of spatial coordinates of the local geometric domain, and construct the initial sub-mesh model of the local geometric domain containing complete physical information.

[0118] Based on the set of local geometric domain spatial coordinates output in step S4.2 and the corresponding unique node identifiers, the element-node connection relationship index table within the global reference finite element mesh model database is invoked. The element connection topology data, containing all nodes within the set, is retrieved and summarized according to the node identifier to establish a complete list of elements belonging to the local geometric domain. Using this element connection topology data, the material property parameter table of the global reference finite element mesh model is linked, and the material parameter values ​​such as elastic modulus, Poisson's ratio, density, and yield strength corresponding to each selected element are extracted to form a local material property mapping set. The local material property mapping set is indexed and bound to the element connection topology data to generate an element entity set data structure that carries geometric and physical features. A node coordinate reconstruction method is performed on this element entity set to map the node positions in the global coordinate system to relative position vectors in the local geometric domain reference coordinate system, ensuring coordinate consistency requirements in subsequent sub-model solution processes. The node coordinate data, element connection relationships, and material property sets already described in a unified coordinate system are comprehensively encapsulated to construct an initial sub-mesh model of the local geometric domain with complete geometric information, element topological relationships, and physical parameter characteristics. This model serves as the direct input benchmark for subsequent boundary node identification and logical cutting. Through the above data extraction and mapping encapsulation processing, the local geometric domain coordinate set from the previous step is transformed into an initial sub-mesh model with all the physical information required for solution. This effectively transforms the location data of local sensitive areas into a computable model, laying the foundation for the rapid construction of subsequent local alternative models.

[0119] For example, in the digital twin model of the GK-5000 large forging press, the high-sensitivity parameter calibration set includes local Young's modulus correction targets for the machine body columns and worktable support areas, and the spatial coordinate set output in step S4.2 contains 258 node identifiers. The system calls the global finite element database to map these nodes to the connection relationships of the corresponding 74 four-node solid elements and 36 six-node solid elements, and extracts their default elastic modulus from the material library. Poisson's ratio 0.29, density Yield strength The material properties of the nodes were analyzed. During the node coordinate reconstruction process, the average node coordinate range in the original global coordinate system was 0~600 mm for the X-axis, 0~700 mm for the Y-axis, and 0~1200 mm for the Z-axis. After being transformed to the local geometric domain reference system, these ranges were translated and normalized to -300~300 mm for the X-axis, -350~350 mm for the Y-axis, and 0~1200 mm for the Z-axis, ensuring the numerical stability of the stiffness matrix assembly during local model calculations. After this processing, the initial local sub-mesh model contained a total of 110 elements and 258 nodes, possessing element topology encoding and material property indexes consistent with the global model. These were imported into the boundary node identification module for detection, confirming that they met the input conditions for subsequent logical cutting and fast template matching. Ultimately, this significantly improved the solution efficiency and ensured the consistency of local physical behavior with the real device during the replacement model generation stage with millisecond-level response.

[0120] S4.4: Perform boundary node identification on the initial sub-mesh model of the local geometry, separate internal free nodes from external constraint nodes, and generate a list of boundary condition nodes that define the interaction interface between the local geometry and the global environment.

[0121] The initial sub-mesh model of the local geometric domain, which contains complete physical information, is built based on S4.3. Its node coordinate set and element connection information are loaded as input data, and the boundary node identification module is called to perform node attribute classification calculation.

[0122] Read the topological correlation parameters of all nodes in the local submesh, and mark the nodes that share elements with the global reference finite element mesh model and belong to the edge of the extraction region as the candidate boundary node set.

[0123] For each candidate boundary node set, calculate its degree-of-freedom constraint state vector in the local domain, and use the displacement constraint criterion to identify nodes subject to constraints imposed by the external global environment and mark them as external constraint nodes.

[0124] Perform connectivity analysis on the remaining unlabeled candidate nodes and all non-edge nodes within the local domain to confirm that they are only associated with the cells within the local domain and that all their constraint states are free, and label them as internal free nodes.

[0125] The external constraint nodes are generated into a list of boundary condition nodes that define the interface between the local geometric domain and the global environment according to the topological sequence, thus establishing accurate boundary conditions for subsequent logical cutting and independent solving.

[0126] By using the above node classification and interface definition processing methods, the mesh geometry information from the previous step is transformed into a data structure with clear physical boundary constraints, achieving the expected technical effect of preparing standardized interface conditions for independent solution of local sub-models.

[0127] S4.5: Based on the boundary condition node list, perform logical cutting operations on the initial sub-mesh model of the local geometric domain, remove redundant connections with the global mesh, encapsulate independent data structures, and generate a local sub-model region to be reconstructed with independent solution capabilities.

[0128] Based on the boundary condition node list output from step S4.4, which includes the classification results of internal free nodes and external constraint nodes, constraint node identifier matching and retrieval are performed on the node-element relationship matrix within the initial sub-mesh model of the local geometric domain to identify the set of external constraint nodes that maintain direct connections with the global baseline finite element mesh model. Using this set of external constraint nodes, boundary truncation is performed on the corresponding element connection table, removing all elements crossing the local domain boundary from the local sub-mesh element index set and simultaneously deleting their global references in the adjacency matrix to eliminate cross-domain coupled computational paths. The truncated local sub-mesh node set is renumbered, and the node coordinate array and element topology array are reconstructed according to the local node index sequence, achieving data independence between the local domain finite element metadata structure and the global mesh. Based on the renumbering results, a sparse matrix storage format containing only local internal elements and their boundary condition node relationships is constructed. Material parameters, geometric properties, and boundary loading information are bound by node and element indices and encapsulated as cohesive data objects, ensuring that matrix assembly and boundary application can be completed without accessing global mesh information when calling the numerical solver. The encapsulated independent data objects undergo initialization verification using numerical examples to validate the consistency of their node coordinate systems, the legality of their element geometry, and the integrity of their boundary loading, ensuring that the local sub-model to be reconstructed has independent solution capabilities. Through the above logical segmentation and data encapsulation processing, the initial sub-mesh model of the local geometric domain from the previous step is transformed into a local sub-model region to be reconstructed with completely independent numerical solution capabilities and boundary physical self-consistency, thus providing a highly consistent input foundation for subsequent rapid response template matching.

[0129] For example, in the mold mounting flange area of ​​a large hydraulic forging press, the boundary condition node list output by S4.4 contains 48 external constraint nodes and 312 internal free nodes. When performing boundary cutoff, the element indices associated with 142 cross-domain elements of the global mesh in the external constraint nodes are locked, and these cross-domain element records are deleted from the element connection table. The remaining local sub-mesh node set is renumbered so that the node numbers are consecutively arranged from 1 to 312, and the node coordinate array and element topology array are reconstructed accordingly. For example, the original node number 58 is adjusted to node 12 in the new numbering system and is synchronously replaced in the node index of the local element connection. When constructing the local sparse matrix, the elastic modulus of the mold flange material is set to 210 GPa, the Poisson's ratio is set to 0.3, and the boundary pressure vector in the actual forging load is bound to the boundary condition node index in node number order. After initialization and verification, it was found that the geometric distortion rate of the elements of the local sub-model is less than 0.02, the node coordinate system maintains a Cartesian orthogonal structure, and the length of the boundary load vector is completely matched with the number of external constraint nodes. It can be directly used as an independent dynamic system for millisecond-level solution when calling the fast response template library in the future, achieving the technical effect of significantly improving the efficiency of local response calculation and maintaining accuracy.

[0130] Step S5: Based on the boundary conditions of the local sub-model region to be reconstructed, a simplified sub-model template verified by real forging load spectrum excitation is matched from the pre-stored fast response template library to generate a local physical behavior replacement model with millisecond-level response. Specifically, this includes: The rapid response template library is a core knowledge component in this invention used to accelerate the reconstruction of local sub-models. Its core function is to quickly match pre-stored simplified sub-model templates based on the geometric topology, boundary constraints, and load characteristics of the local sub-model region to be reconstructed during online runtime. It then generates an instantiated local physical behavior replacement model adapted to the current working conditions through projection transformation, thereby avoiding the need for re-meshing the finite element mesh and assembling the global stiffness matrix each time, achieving millisecond-level response. In the offline phase, this template library is built based on a large amount of finite element simulation data, modal analysis results, and measured load spectra under typical working conditions, and supports online incremental updates.

[0131] The rapid response template library consists of multiple simplified sub-model templates. Each template corresponds to a typical local geometric domain (such as the fuselage column connection area, mold mounting flange surface, workbench support rib area, etc.) and includes the following data: Reduced basis function matrix Where N is the number of degrees of freedom in the high-dimensional finite element method, and r is the number of principal modes after order reduction (usually...). This matrix projects the high-dimensional finite element displacement field to a low-dimensional generalized coordinate space and is the core operator for model order reduction. The reduction basis functions are obtained offline by performing eigenvalue decomposition or orthogonal eigenvalue decomposition (POD) on the local sub-model.

[0132] Constitutive parameter set: Includes fundamental parameters of the local geometric domain material such as elastic modulus, Poisson's ratio, density, and yield strength, as well as constitutive model parameters (such as hardening exponent, damage initiation threshold, and evolution coefficient) under elastoplastic, contact, and damage states. The parameter set is stored in categories according to material type and operating condition range.

[0133] Boundary condition mapping rules: Define the displacement constraint types (fixed, symmetric, periodic) and force load application modes (concentrated force, distributed pressure, equivalent nodal force) of the boundary nodes of the simplified sub-model. Each mapping rule also includes a transformation matrix from the global coordinate system to the local coordinate system.

[0134] Historical load spectrum fingerprint: Stores the feature vectors (such as peak value, rising slope, frequency domain dominant frequency, etc.) of the typical forging load spectrum (load-time curve) used by the template in the offline verification stage. It is used to compare the similarity with the load characteristics of the current working condition during online matching to ensure the applicability of the selected template.

[0135] Geometric topology descriptor: includes typical size range of local geometric domain, node distribution density, cell type, etc., used to quickly exclude templates that are too different from the shape of the current region to be reconstructed.

[0136] The template library building process: 1. Classification of typical local geometric domains: Based on the global baseline finite element mesh model of the digital twin model of a large forging press, common local sub-model reconstruction regions (driven by the S4 high-sensitivity parameter calibration set) are analyzed and summarized into several typical categories, such as: Connection area between fuselage column and crossbeam (type code: COL-01); Mold mounting flange face contact area (type code: DIE-02); T-slot support rib area of ​​workbench (type code: TBL-03); Guide rail slider mating surface area (type code: RAIL-04).

[0137] For each type of geometric domain, select representative samples of size, shape, and boundary conditions.

[0138] 2. Offline finite element simulation and data acquisition: For each typical local geometric domain, forging load spectra with different amplitudes and loading positions (covering various working conditions that may be encountered in actual production lines) are applied, and high-precision nonlinear finite element analysis is performed to record the displacement, stress, and strain responses of each node. At the same time, the load spectrum is extracted into feature vectors (peak value, slope, and frequency components) as load spectrum fingerprints for templates.

[0139] 3. Extraction of reduced-order basis functions: For each typical geometric domain, collect displacement field snapshots under all working conditions to form a matrix. (M represents the total number of snapshots). The singular value decomposition of S is performed using the Orthogonal Eigenvalue Decomposition (POD) method. The first r left singular vectors (corresponding to cumulative energy contributions exceeding 99%) are used to form the reduced-order basis function matrix Φ. The value of r is typically between 10 and 50, much smaller than the original degrees of freedom.

[0140] 4. Constitutive Parameter Calibration and Verification: Using offline simulation results, the constitutive parameters of the simplified sub-model are calibrated in reverse, ensuring that the relative error between the reduced-order model's response under typical working conditions and the high-precision finite element solution is controlled within a preset threshold (e.g., 5%). For each template, its applicable parameter range (e.g., peak load range, temperature range) is recorded.

[0141] 5. Template Encapsulation and Indexing: The reduced-order basis function matrix, constitutive parameter set, boundary condition mapping rules, load spectral fingerprint, and geometric topology descriptor are uniformly encapsulated into structured template objects and stored in the template library. Simultaneously, a multi-dimensional index (based on geometric type, size range, load characteristics, etc.) is established to support fast online retrieval.

[0142] For example, in the digital twin system of the GK-5000 forging press, the fast response template library contains 127 simplified sub-model templates. When S4.5 outputs a local sub-model of the "workbench T-slot support rib area" to be reconstructed, its boundary condition descriptor is a 324-dimensional feature vector. The system selects 12 candidate templates based on the geometric domain type "TBL-03". The Euclidean distance between the current feature vector and the load spectrum fingerprint of each template is calculated, and the template with the highest matching score (0.93) is "TBL-03". V2 The reduced-order basis function matrix Φ has a size of 5600×25, and the constitutive parameter set includes an elastic modulus of 210 GPa and a hardening exponent of 0.12. The low-dimensional dynamic equations generated after the projection transformation have only 25 degrees of freedom, and the single-step solution time is only 0.5 ms, which is about 500 times faster than the original finite element model (5600 degrees of freedom), while maintaining more than 95% accuracy consistency with high-fidelity simulation.

[0143] By using the pre-built fast response template library, this invention achieves millisecond-level generation of local physical behavior substitution models, significantly improving the computational efficiency and response speed of digital twin models in real-time closed-loop control.

[0144] S5.1: Obtain the geometric topology and current stress-strain state of the local sub-model region to be reconstructed, perform mesh node mapping processing on the geometric topology, extract the displacement constraint vector and force load distribution characteristics of key boundary nodes, and generate local region boundary condition descriptors.

[0145] After receiving the independently solvable local sub-model region data output by S4.5, the geometric topology information of this region and the stress-strain state at the current moment are imported into the local boundary condition analysis module as input conditions. For the local geometric topology, a mesh node mapping process is performed through a dual association between the node index table and the spatial coordinate set, ensuring a one-to-one correspondence between the original finite element node numbers and their geometric spatial positions, guaranteeing the spatial accuracy and topological consistency of subsequent boundary condition extraction. Based on the established mapping relationship, the boundary condition node list is retrieved, and displacement constraint vector extraction is performed on key nodes at the region boundary. This operation is based on the concatenation of the three-dimensional displacement results of each node to form a fixed-length constraint vector sequence, rigorously describing the kinematic constraints of the boundary nodes. After completing the displacement constraint feature extraction, the load identification subroutine is called to analyze the distribution pattern of the stress field and nodal reaction force data of the boundary nodes and their adjacent elements. The reaction force components of each node are decomposed and integrated into a force load distribution feature vector according to the global coordinate system, capturing the force characteristic pattern under the current physical state. Displacement constraint vectors and force load distribution features are sequentially concatenated and normalized under a parameterized interface to form a unified high-dimensional feature description structure. Numerical anomalies are removed and missing data is filled during the feature vector preprocessing stage. After the feature structure is constructed, it is encapsulated according to a preset descriptor specification to generate a local region boundary condition descriptor containing geometric boundary morphology, displacement constraint conditions, and force load patterns. This descriptor transforms the geometric and physical state information of the local sub-model in stage S4 into a standardized data object that can be directly called by feature similarity matching in stage S5.2. This achieves the conversion from geometric and physical information to characteristic boundary conditions, significantly improving the accuracy and rapid response capability of subsequent template matching.

[0146] For example, during a nominal force generation point calibration process on a GK-5000 forging press, the connection area of ​​the machine frame column was identified as a local geometric domain strongly coupled with the highly sensitive parameter "Young's modulus of the machine frame." The reconstructed region after stripping contained 480 finite element nodes and 912 elements. After acquiring the geometric topology and real-time stress-strain state of this region, node mapping was performed to ensure precise correspondence between node indices and spatial coordinates. The triaxial displacements of 54 key boundary nodes were extracted from the boundary node identification results, forming a displacement constraint vector of length 162. The associated reaction forces were calculated for the same batch of boundary nodes, obtaining the load components in three directions for each node, and integrated into a force load distribution feature vector of length 162. The two vectors were concatenated and normalized to eliminate unit differences and remove an outlier caused by a sensor malfunction. The final generated local region boundary condition descriptor is a 324-dimensional normalized feature vector, encapsulating the geometric contour, displacement constraints, and force patterns of the current local domain. In the subsequent S5.2 step, the descriptor is input into the multidimensional similarity matching and distance calculation is performed with the load spectrum fingerprint in the template library. The matching time is less than 2 milliseconds, which verifies the effectiveness of the boundary condition descriptor in ensuring millisecond-level template retrieval performance.

[0147] S5.2: Based on the local region boundary condition descriptor, multi-dimensional feature similarity matching is used to search and traverse the pre-stored fast response template library, calculate the Euclidean distance between the current boundary feature and the historical load spectrum fingerprint of each simplified sub-model template in the library, and generate a template matching score sequence.

[0148] Based on the local region boundary condition descriptor generated in step S5.1, the boundary node displacement constraint vectors and force load distribution features contained in the descriptor are organized into a multi-dimensional feature vector input structure, ensuring that each feature corresponds to a unique physical quantity. For this feature vector, a multi-dimensional feature space index covering geometric parameters, boundary constraint types, load spectrum morphology, and historical deformation response modes is constructed for comparison with the feature fingerprint vectors of each simplified sub-model template stored in the fast response template library. In the matching operation of each pair of feature fingerprint vectors, Euclidean distance is used to measure their difference, calculating the Euclidean distance value between the current boundary condition feature vector and the template fingerprint feature vector in the feature space. Each calculated Euclidean distance is used as an inverse proportional index of the template matching degree. After reciprocal operation and normalization mapping, the template matching degree score is obtained and stored in the matching degree score sequence. The above measurement and transformation process is repeated for all simplified sub-model templates in the fast response template library to ensure that comparable matching quantification indicators are established between templates with different size specifications, boundary morphologies, and load spectrum conditions. Through the above multidimensional feature similarity calculation and score sequence generation, the boundary condition description data of the previous step is transformed into template matching metric results, so as to achieve the technical effect of quickly selecting the best simplified sub-model based on the score value in S5.3.

[0149] For example, during the operation of a large forging press with a nominal force of 5000kN, the acquired local geometric domain boundary condition descriptor includes four types of features: the displacement constraint amplitude of key boundary nodes (3.2mm), the load direction vector (0.85, 0.50, 0.12), the total load amplitude (420kN), and the frequency domain peak (0.8Hz). A four-dimensional feature vector is constructed from these features, corresponding to the historical load spectrum fingerprint features of a template in the template library: 3.0mm, (0.83, 0.54, 0.10), 400kN, and 0.75Hz. The squared differences of each dimension are calculated, summed, and squared. Applying the above formula, the Euclidean distance is obtained as 0.258. Taking the reciprocal of this distance and linearly normalizing it to the 0-1 interval yields a matching score of 0.87. The same calculation was performed on all 127 templates in the library, generating a set of matching scores. The template ranked first in the sorted list achieved a matching score of 0.93. This result was output as a scoring sequence, providing a quantitative basis for the optimal template selection in S5.3. This process was completed in milliseconds, significantly improving the real-time performance and matching accuracy of local model reconstruction.

[0150] S5.3: Based on the template matching score sequence, perform threshold filtering and optimal index positioning operations, select the simplified sub-model template with the highest score and verified by real forging load spectrum excitation as the target candidate, and generate the target simplified sub-model template identifier.

[0151] After receiving the template matching score sequence output from the previous sub-step S5.2, the score sequence is loaded into the threshold determination module as input data. Simultaneously, a physical validity flag table bound to the historical load spectrum fingerprint verification status of each template is loaded. The score value of each template is matched one-to-one with its corresponding physical validity flag to construct an initial screening set. Threshold filtering is performed on the initial screening set. By calling the dynamic score threshold calculation unit, and combining the load peak value, cycle frequency characteristics, and historical matching deviation distribution of the current working condition, a screening threshold suitable for different working conditions is calculated. An element-by-element comparison operation is performed on the score sequence, removing template entries with scores below the threshold or with invalid physical validity flags, resulting in a reduced subset of candidate templates. Optimal index positioning is performed within the candidate template subset. A maximum value search method is used to extract the template entry with the highest score value while satisfying physical validity, and its index is output as the candidate optimal template index value to the index binding module. The index binding module invokes the template metadata manager to retrieve template entity records within the template library index space that match the candidate optimal template index value. It then reads and extracts the globally unique identifier, version number, model accuracy level, and other additional attributes, encapsulating them into a structured template identification information object. Through the aforementioned chain-like filtering, positioning, and identification binding process, the quantitative evaluation results of the matching score sequence are transformed into target simplified sub-model template identifiers that can be directly invoked in subsequent reduced-order basis function matrix projection and dynamic equation reconstruction, achieving millisecond-level template invocation accuracy at millimeter-level geometric resolution.

[0152] For example, under the forming conditions of a GK-5000 large forging press, the length of the input template matching score sequence is 127, the value of a single score ranges from 0 to 1, and there are 119 records in the physical validity flag table that are valid flags. The threshold determination module uses the peak load monitored under the current working conditions as 24500kN, the cycle frequency as 0.5Hz, and the root mean square value of the historical matching deviation as 0.014mm, and calculates a screening threshold of 0.85 through the dynamic scoring threshold calculation unit. During the element-by-element screening process, templates with scores below 0.85 or invalid validity flags are eliminated, reducing the size of the candidate template subset to 9. The optimal index positioning process locks the template with a score of 0.947 and extracts its index value of 58 in the template library. The index binding module retrieves the entity of index 58 in the template metadata manager and reads its globally unique identifier as "GK5000_TMPL_58_VER3.2", and the model accuracy level is high accuracy level. After the identifier is structured and stored, it is output to the S5.4 stage to call the corresponding constitutive parameter set and reduced basis function matrix, so as to realize the construction of the local physical behavior substitution model for subsequent millisecond-level response.

[0153] S5.4: Using the target simplified sub-model template identifier, call the corresponding constitutive relation parameter set and reduced basis function matrix, substitute the local region boundary condition descriptor into the reduced basis function matrix for projection transformation, reconstruct the local dynamic equation set adapted to the current working condition, and generate an instantiated local physical behavior alternative model.

[0154] Based on the obtained target simplified sub-model template identifier and its associated constitutive relation parameter set and reduced-order basis function matrix, the boundary condition descriptor of the local region is used as the initial input data.

[0155] The constitutive relation parameter set pointed to by the template index is invoked to determine the values ​​of the stress-strain response equation parameters of the local region element material in elastic, plastic and contact states, and the initial state of the material stiffness matrix and damping matrix is ​​locked accordingly.

[0156] Load the reduced-order basis function matrix of the target simplified sub-model template, and use this matrix as a projection operator to map the high-dimensional finite element degrees of freedom to a low-dimensional coordinate system, ensuring that only the main modal components that have a significant dynamic contribution to the current working condition are retained.

[0157] Perform boundary condition mapping operation, substitute the displacement constraint vector and force load distribution characteristics in the local region boundary condition descriptor into the reduced-order basis function matrix, and calculate the low-dimensional generalized coordinate initial value vector through matrix multiplication.

[0158] Based on the constitutive equations in the reduced coordinate system, a reduced-order dynamic equation system for the local domain is generated.

[0159] in For the reduced quality matrix, For the reduced order damping matrix, For the reduced stiffness matrix, For generalized acceleration vector, For generalized velocity vector, For generalized displacement vector, This is the generalized load vector.

[0160] Numerical stability analysis is performed on the generated reduced-order dynamic equations to verify that the eigenvalue distribution under the current constraints and loads satisfies the stability condition of millisecond-level integration step size. The reduced-order state equations are associated with the physical boundary conditions and encapsulated into an instantiated local physical behavior replacement model.

[0161] Through the above projection transformation and structural reconstruction processing methods, the template matching result of the previous step is transformed into an instantiated physical sub-model that can be solved in real time and coupled to the main loop of the digital twin model, which is adapted to the current working condition. This achieves the expected technical effect of meeting the real-time response while maintaining high fidelity.

[0162] S5.5: Perform millisecond-level response performance verification on the instantiated local physical behavior substitution model, execute single-step time integration derivation to output the displacement field increment and reaction torque in the local domain, and form the output result of the local physical behavior substitution model that can be directly projected onto the plane defining the nominal force occurrence point.

[0163] Based on the instantiated local physical behavior replacement model generated by S5.4 and its constitutive relation parameter set and reduced-order basis function matrix output results, during the millisecond-level response performance verification process, the initial state vector and boundary condition matrix of the replacement model are first loaded, and a high-precision single-step time integration operator is called to ensure that numerical stability and physical consistency are maintained simultaneously within a solution step. The current time step Δt is combined with the local mass matrix, damping matrix, and stiffness matrix of the replacement model to construct an explicit central difference integral scheme, and the predicted displacement vector of the local degrees of freedom is calculated from it. The velocity increment vector is obtained through the inverse calculation process of the mass matrix, and it is substituted into the balance equation of the damping term and the external load input to update the displacement and velocity states. The updated displacement vector is multiplied with the stiffness matrix to obtain the local internal force vector, and the difference with the externally applied load is accumulated in reverse to obtain the reaction torque vector, realizing the instantaneous solution of the constraint reaction force of the boundary node. For the numerical results of the displacement field increment, a geometric projection transformation from spatial nodes to the defined plane is performed. This generates a two-dimensional component mapping of the three-dimensional displacement vector field within the plane defined by the nominal force occurrence point, and a weighted superposition operation based on node weights is added to form a dataset that can be directly input into the subsequent S7 step. Through appropriate tolerance verification logic, step error evaluation calculations are performed on the integral operator. This error value is compared with a preset response accuracy threshold. If it does not exceed the threshold, the physical consistency and numerical stability of the millisecond-level single-step integration are confirmed, and the local displacement field increment and reaction torque are output as the final output results of the local physical behavior substitution model. Through this chain integral and reaction force inversion verification method, the instantiated substitution model from the previous step is transformed into displacement and torque output data with high accuracy and high real-time performance, achieving rapid quantitative characterization of local deformation behavior.

[0164] For example, within the T-slot region of the worktable of a GK-5000 large forging press, for the local sub-model caused by the change in worktable support stiffness determined by a high-sensitivity parameter calibration set, 1024 finite element nodes are selected as the response monitoring set. The single-step time integration step size Δt is set to 0.5 milliseconds, and the mass matrix and stiffness matrix are provided by pre-stored templates. The vertical concentrated load vector of the actually loaded mold mounting surface is input into the local dynamic equation. The norm of the velocity increment vector obtained after inverting the mass matrix is ​​approximately 2.13 × 10⁻⁶. -3 The sum of the boundary node reaction moments after substituting the damping term into the equation m / s is 4.52 × 10 m / s. 3N·m. The calculated value of e using the above formula is 0.0047, which is lower than the preset threshold of 0.01, confirming that the single-step response meets the accuracy requirements. After projecting the displacement field onto the plane of the nominal force generation point, the average value of the two-dimensional displacement component field obtained by weighted superposition of node weights is 0.032 mm. This value is used in subsequent step S7 to generate the corrected coordinates of the nominal force generation point, achieving a high-precision closed-loop control response within a millisecond timescale.

[0165] Step S6: Determine whether the real-time acquired sensor data exhibits a step-like aberration anomaly pattern. If an anomaly pattern is detected, freeze all current parameter updates and input the spatial residual vector of the nominal force occurrence point into the anomaly pattern recognizer to generate a local physical behavior instantaneous reset command containing the constitutive characteristics of interface damage. Specifically, this includes: The anomaly pattern recognizer is an intelligent component in this invention used to quickly identify sudden abnormal operating conditions of a forging press and guide local model reconstruction. Its core function is to determine the fault type of the current anomaly (e.g., "embrittlement and peeling of the die interface," "sudden drop in stiffness of the workbench support," "local buckling of the machine frame column") after the digital twin detects an abnormal step jump in the rate of attitude change. It then outputs the corresponding interface damage constitutive feature label (including damage initiation strain threshold, stiffness reduction evolution curve parameters, and nonlinear hysteresis coefficient) for subsequent instantaneous local physical behavior reset commands. This recognizer is trained and built based on historical fault data in the offline phase and supports online millisecond-level inference.

[0166] Inputs and outputs of the recognizer: Input: Spatial residual vector ΔP (3D vector, unit: mm) at the nominal force occurrence point, which represents the spatial deviation between the measured slider posture and the predicted value of the digital twin model in the current forging cycle.

[0167] Output: Interface damage constitutive feature labels, including structured information such as fault type identifier, damage initiation strain threshold, stiffness reduction evolution curve parameters (such as exponential decay coefficient), and nonlinear hysteresis coefficient.

[0168] The anomaly pattern detector uses a multi-class support vector machine (SVM) or random forest as its core classification model, and its internal components include the following: Feature normalization layer: The input three-dimensional residual vector is standardized with zero mean and unit variance to eliminate the scale effect caused by the difference in load amplitude in different forging cycles. The standardized parameters (mean, standard deviation) are statistically obtained from historical normal working condition data.

[0169] Feature expansion layer: To improve classification accuracy, the three-dimensional residual vector is expanded into a higher-dimensional feature space. This recognizer uses a fixed expansion method: calculating the residual magnitude. , direction angle cosine , This is then concatenated with the original three-dimensional residual components to form a five-dimensional or higher-dimensional feature vector. If necessary, operating parameters such as peak load and loading rate can also be introduced.

[0170] Predefined Fault Mode Feature Template Set: This is the core knowledge base stored internally by the recognizer. It is used to perform similarity matching between the feature vector to be matched and the feature templates of various typical fault modes after a step jump anomaly is triggered. Each template record corresponds to a known interface damage or structural anomaly mode and includes the following elements: Feature Template Vector: Fixed-dimensional normalized feature components (such as residual modulus, orientation angle, deviation from normal operating conditions, load-related features, etc.); Fault Type Identifier (such as "FRACTURE_DIE_001"); Interface Damage Constitutive Feature Label (damage initiation strain threshold, stiffness reduction evolution coefficient, nonlinear hysteresis coefficient); Template Confidence (based on historical matching success rate statistics).

[0171] The core of the classifier is a Support Vector Machine (SVM) with radial basis function (RBF) kernels or a random forest composed of multiple decision trees. After offline training, the model internally stores support vectors or decision tree splitting rules for various faults. During the matching process, the classifier outputs the similarity (or distance) between the feature to be matched and each template, and selects the template with the smallest distance as the matching result.

[0172] Fault Mode Library Mapping Layer: Maps the fault type identifiers corresponding to the successfully matched templates to predefined fault mode library records, and extracts complete interface damage constitutive feature labels.

[0173] Construction of a predefined set of fault mode feature templates: 1. Failure Mode Definition and Sample Collection: For typical failure modes of forging presses (such as die interface embrittlement and peeling, sudden drop in worktable support stiffness, local buckling of machine body columns, uneven wear of guide rails, etc.), characteristic samples are obtained through the following methods: Historical failure cases: Extract the spatial residual vector sequence of nominal force occurrence points corresponding to confirmed failure events from the attribution log library; Offline simulation: Specific damage constitutive parameters (such as reducing local stiffness or introducing interface softening) are artificially injected into the digital twin model to simulate various fault conditions and generate corresponding residual vectors; Experimental verification: Typical faults were reproduced and measured data were collected using a scaled-down model or test bench.

[0174] At least 200 valid samples were collected for each type of failure mode, covering different load levels and loading rates.

[0175] 2. Feature Extraction and Normalization: For each sample, extract the following feature components: residual magnitude, residual orientation angle, Mahalanobis distance (or Euclidean distance) between the residual and the mean under normal operating conditions, peak load, and loading rate. Normalize all feature components using Z-score (subtract the mean and divide by the standard deviation) to eliminate the influence of different operating conditions and different dimensions, resulting in a standardized feature vector.

[0176] 3. Clustering and Template Generation: For all sample feature vectors within each fault mode, the K-means clustering algorithm (with 3-5 clusters) is used to identify typical subclasses within that mode. For each cluster, the mean of the sample feature vectors within the cluster is calculated and used as the feature template vector for that subclass. Simultaneously, the standard deviation within the cluster is calculated as the tolerance radius for matching (used for distance threshold determination). Multiple subclass templates are grouped together to belong to the same fault type.

[0177] 4. Template Validation: Leave-one-out cross-validation is used to test the matching success rate of each feature template for similar fault samples and the mismatch rate for dissimilar fault samples. Templates with a matching success rate below 85% or a mismatch rate above 10% are discarded. Validated templates and their associated constitutive feature labels are finally stored in the read-only memory of the anomaly pattern recognizer.

[0178] For example, during a single run of the GK-5000 forging press, a frozen parameter after a step jump was detected, and the residual vector was... ΔP = (0.48, -0.35, 0.27) mm. The calculated module length r = 0.68 mm, and the direction angle cosine are (0.706, -0.515, 0.397), are compared with the peak load F. max =24500kN concatenation, after normalization, yields the feature vector to be matched. The "mold interface embrittlement and peeling" feature template vector stored internally by the anomaly pattern recognizer is (after normalization): residual modulus 0.52, orientation angle θ x Cosine 0.38, θ z The cosine is 0.71, the residual abnormality deviation from the normal operating condition is 0.63, and the load correlation characteristic is 0.45. The weighted Euclidean distance calculation yields d=0.18, which is less than the preset threshold of 0.3, indicating a successful match. The system outputs the interface damage constitutive feature labels bound to this template: damage initiation strain threshold 0.002, stiffness reduction evolution coefficient 0.85, and nonlinear hysteresis coefficient 0.3. These labels are used in S6.5 to generate instantaneous reset commands for local physical behavior, guiding the local sub-model to switch to the template containing the interface damage constitutive structure.

[0179] By using the pre-built anomaly pattern recognizer and the internal predefined fault mode feature template set, this invention achieves rapid identification and accurate classification of sudden abnormal working conditions, provides a reliable constitutive feature basis for instantaneous reset of local physical behavior, and significantly improves the adaptive control capability of forging presses under abnormal conditions.

[0180] S6.1: Acquire real-time slider attitude time series data collected by displacement sensor array, and use the sliding window difference method to calculate the first derivative of the real-time slider attitude time series data to obtain the attitude change rate vector characterizing the change rate of slider motion state.

[0181] S6.2: Based on the attitude change rate vector, execute adaptive threshold comparison logic to compare the magnitude of the attitude change rate vector with a preset dynamic noise baseline to generate a step jump judgment flag that identifies the characteristics of data mutation.

[0182] Based on the attitude change rate vector output by sub-step S6.1, its amplitude within the entire sampling window is selected as the basic input for anomaly detection. The amplitude is defined as the square root of the sum of squares of each component to simultaneously reflect the comprehensive intensity of multi-degree-of-freedom attitude changes. The dynamic noise baseline generation module is invoked, and based on the historical statistics of attitude changes during the slide's unloaded phase in the current forging cycle, the allowable fluctuation range of the multi-dimensional change rate is adaptively calculated, generating the corresponding dynamic noise baseline vector. A comparison operation between the amplitude and the corresponding components of the noise baseline is used to form an instantaneous over-limit residual vector to capture the characteristic of the attitude change rate significantly deviating from the normal fluctuation level. During the comparison process, single-channel over-limit flags for each degree-of-freedom channel are established and synthesized into a global step jump candidate flag through logical OR operation to avoid misjudgment caused by a single sensor channel failure. With the global candidate flag activated, the ratio of the amplitude to the noise baseline is input into a smoothing filter to suppress instantaneous error triggering caused by high-frequency measurement jitter, outputting the final step jump judgment flag. By using this adaptive threshold comparison logic, the attitude change rate data is mapped into discrete trigger signals that characterize instantaneous load disturbances, enabling high-confidence identification of sudden anomalies.

[0183] S6.3: In response to the trigger signal of the step transition determination flag, perform a global freeze operation on the parameter calibration process in the current digital twin model, block the iterative update path of all physical parameters, and output a set of frozen parameters in a locked state.

[0184] Upon receiving the trigger signal output from the step transition judgment logic and the context state description of the current digital twin model parameter calibration process, the trigger signal is input to the locking control channel of the global calibration controller to initiate the execution entry point of the parameter freeze process. Based on this trigger event, the iterative scheduling table of the global parameter manager is invoked to retrieve all physical parameter iterative update threads in the running state, generating a real-time running mapping table containing thread identifiers and memory data pointers. Using this running mapping table, a thread suspension operation is performed, removing the iterative calculation routine of the corresponding parameter from the scheduling queue and writing the snapshot value of the current iteration step into the parameter storage unit to form a frozen snapshot vector. For the frozen snapshot vector, a read-only protection mechanism is invoked to modify its memory page access permissions, marking the physical parameter data area as read-only to prevent any direct or indirect calculation process from modifying its data content. Combined with the state check module of the freeze process, path blocking is performed on all parameter update paths, including the real-time sensor calibration path, the model prediction residual feedback path, and the external intervention input path, ensuring that all channels of iterative updates are stopped. The above-mentioned freeze control strategy transforms the received trigger signal into a global parameter update blocking and freeze data output, forming a set of frozen parameters in a completely locked state, thereby achieving stable solidification and traceability assurance of the model's internal state under sudden abnormal situations.

[0185] S6.4: Using the nominal force occurrence point spatial residual vector in the frozen parameter set as input features, drive the pre-trained abnormal pattern recognizer to perform feature matching operation, and map the nominal force occurrence point spatial residual vector to a predefined fault mode library to generate interface damage constitutive feature labels that characterize specific failure mechanisms.

[0186] The system receives the nominal force occurrence point spatial residual vector contained in the frozen parameter set. This residual vector is then input into the input layer of the offline-calibrated and trained anomaly pattern recognizer in a preset dimensional order, ensuring a one-to-one correspondence between the input features and the training feature space. The feature normalization module is invoked to perform zero-mean unit variance standardization on each component of the residual vector, eliminating weight biases of different dimensional components in subsequent similarity calculations. The standardized residual feature vector is then compared one-to-one with the predefined fault mode feature template set stored internally by the anomaly pattern recognizer, using a weighted Euclidean distance method to quantify the matching degree between the current residual feature and each template vector. Based on the matching score sequence, the maximum value selection logic is invoked to extract the fault mode index with the highest matching degree. A matching degree threshold is then used to determine whether the mode is valid; if the matching degree is below the threshold, an unknown mode labeling mechanism is triggered. The successfully matched fault mode index is mapped to the corresponding interface damage constitutive feature metadata record in the fault mode library. A set of labels, including the damage initiation strain threshold, stiffness reduction evolution curve parameters, and nonlinear hysteresis coefficient, is extracted as the output interface damage constitutive feature labels. By performing residual feature mapping and pattern recognition operations based on the frozen parameter set, the original spatial deviation information is transformed into interface damage constitutive feature labels that characterize specific failure mechanisms, thereby enabling accurate acquisition of the boundary conditions required for the instantaneous reset strategy of local physical behavior.

[0187] S6.5: Based on the interface damage constitutive feature label, call the corresponding sub-model reconstruction strategy to generate a local physical behavior instantaneous reset instruction containing specific nonlinear degradation boundary conditions, so as to guide the subsequent local sub-model library to load the interface damage constitutive template to complete the instantaneous correction of the model state.

[0188] After obtaining the constitutive feature label of the interface damage, the label is used as the index key to retrieve the predefined sub-model reconstruction strategy library and extract the nonlinear boundary condition parameter set and constitutive evolution equation system structure definition information that strictly correspond to the failure mechanism.

[0189] Based on the extracted set of nonlinear boundary condition parameters, a boundary condition matrix is ​​constructed to constrain the degrees of freedom of the boundary nodes of the local geometric domain. Combined with the material degradation path defined by the constitutive evolution equations, it is converted into an adjustment coefficient vector of the time-varying stiffness matrix and damping matrix.

[0190] These adjustment coefficient vectors are used to perform parameter embedding on the local sub-model in the frozen state. The original linear stiffness term is replaced with a local stiffness matrix containing nonlinear degradation terms by matrix superposition.

[0191] A transient integral pre-calculation strategy is adopted to perform forward calculations on the corrected local stiffness matrix and damping matrix within a single time step, generating a transient state vector containing displacement, velocity and stress states, which serves as the instantaneous response prediction value of local physical behavior.

[0192] The transient state vector is re-bound to the interface damage label and encapsulated into an instruction data package. This instructs the local sub-model library to directly load the corresponding "constitutive model containing interface damage" template and immediately replace the current running instance, so as to achieve physical consistency reset of the local model state without interrupting the global simulation process.

[0193] The local sub-model hot-switching processing method triggered by the instruction data packet transforms the parameters locked in the previous step and the current abnormal pattern recognition results into reset control information with targeted nonlinear degradation boundary conditions, thereby realizing rapid state correction and anomaly isolation of relevant areas within the digital twin model.

[0194] For example, in a single forming cycle of a large forging press, when the anomaly pattern recognizer outputs the "die interface embrittlement and peeling" tag, the system retrieves boundary condition parameters from the strategy library that include settings such as a contact surface friction coefficient decreasing to 0.3 and a local contact stiffness attenuation coefficient of 0.65. Simultaneously, the constitutive evolution equations define a nonlinear softening curve when the stress threshold exceeds 350 MPa. These parameters are embedded in the frozen local sub-model. After superimposing correction terms onto the original stiffness matrix K (dimension 180×180), a transient state vector with a displacement response increment of 0.12 mm and a stress peak decreasing to 320 MPa is obtained within a time step of 0.5 ms using Euler single-step integration. This state vector, along with the "die interface embrittlement and peeling" tag, is encapsulated into a reset command package. Upon issuance, the local sub-model library immediately switches to the pre-stored constitutive template containing interface damage. In subsequent simulations, the force and displacement predictions for this region are calculated based on the new degradation characteristics, significantly improving the rapid response capability to interface failure and ensuring the stability and reliability of the nominal force occurrence point calculation.

[0195] Step S7: Execute the output result projection operation of the local physical behavior replacement model or the local physical behavior instantaneous reset command, directly projecting the output results of all local sub-models onto the nominal force generation point definition plane, generating the corrected three-dimensional coordinates and covariance ellipsoid of the nominal force generation point. Specifically, this includes: S7.1: Obtain the displacement field increment output data or constitutive feature label of the instantaneous reset command of the local physical behavior of the instantiated local physical behavior replacement model, and perform geometric mapping processing on the displacement field increment output data or constitutive feature label using the spatial coordinate projection transformation method to generate an initial three-dimensional coordinate offset vector located in the plane defined by the nominal force occurrence point.

[0196] Obtain the displacement field increment data output by the instantiated local physical behavior replacement model or the interface damage constitutive feature label identified by the instantaneous reset command. Call the preset data parsing module to convert the displacement field node information or damage location identifier into a unified three-dimensional spatial coordinate expression form, and establish the index relationship between node number and global coordinate system.

[0197] Based on the established index relationship, the displacement field increment data is solved by nodal components, and vector superposition operation is performed in combination with the global geometric reference coordinates to obtain the set of absolute spatial displacement vectors of each node under the current working condition; when the input is the interface damage constitutive feature label, the equivalent set of displacement vectors is obtained by parsing from the built-in displacement mode of the damage template.

[0198] Obtain the normal vector of the plane defining the nominal force generation point and the absolute spatial coordinates of a reference point on the plane. Construct an orthogonal transformation matrix from the global coordinate system to the local coordinate system of the plane. Perform a linear coordinate transformation on the set of absolute spatial displacement vectors of the nodes based on the matrix to map the global displacement data to the local coordinate system of the plane.

[0199] For the mapped local coordinate system node displacement data, the geometric projection operator is called to project the three-dimensional displacement components of all nodes onto the defined plane, eliminating the influence of the plane normal component and retaining the tangential component, thus realizing the pure geometric distribution expression of the displacement components in the plane.

[0200] In the projected displacement distribution, a set of reference nodes corresponding to the physical meaning of the nominal force occurrence point is selected, and the mean value is calculated to form an initial three-dimensional coordinate offset vector representing the comprehensive displacement of the entire action point area. During the calculation process, the displacement contribution of different nodes is weighted using a weight coefficient matrix, and the weight is set according to the mechanical dominance ratio of the node in the load distribution.

[0201] Project the nodal displacements onto the defined plane.

[0202] Through the above geometric mapping and weighted projection processing, the displacement field increment result or damage feature command of the previous step is transformed into an initial three-dimensional coordinate offset vector that defines the trend of change of the action point in the plane, thereby realizing the docking of the local analytical result with the global control plane.

[0203] For example, in the digital twin control system of the GK-5000 large forging press, the instantiated local sub-model outputs displacement field incremental data containing 5120 nodes, with node coordinates recorded based on the machine body's global coordinate system (mm-level accuracy). The nominal force generation point is defined by the central section of the mold cavity, with a normal vector of (0,0,1) and a reference point of (0,0,0). The system analyzes and obtains a set of 20 reference nodes in this region, with their weighting coefficients set between 0.05 and 0.08 based on the finite element load sensitivity. When performing projection calculations, for example, a reference node u=(0.12, -0.08, 0.05)mm, multiplied by the normal vector n=(0,0,1), results in 0.05, yielding a projected displacement u'=(0.12, -0.08, 0)mm, after removing the 0.05mm normal component. The weighted average of the projected displacements of the 20 nodes yields the initial three-dimensional coordinate offset vector as (0.115, -0.082, 0) mm. This result is directly used in the downstream S7.2 step for multi-source error propagation evaluation, ensuring the spatial consistency and physical meaning of the displacement response during the correction of the forging press's action point.

[0204] S7.2: Based on the initial three-dimensional coordinate offset vector, the multi-source error propagation evaluation mechanism is invoked to perform variance synthesis calculation on the residual uncertainties of the fuselage elastic deformation parameters, mold compression parameters, and workbench deflection parameters, so as to generate the original covariance matrix characterizing the reliability of spatial position prediction.

[0205] Based on the initial three-dimensional coordinate offset vector located in the plane defined by the nominal force occurrence point, a multi-source error propagation evaluation mechanism is introduced to collaboratively quantify the residual uncertainties of the fuselage elastic deformation parameters, mold compression parameters, and worktable deflection parameters. A parameter uncertainty database is invoked to retrieve the residual variance values ​​and their cross-correlation matrices of physical parameters matching the current working condition, forming a multi-source variance input set for error calculation. A biased guide vector array for each physical parameter is constructed using the three-dimensional coordinate prediction function to ensure the completeness and orthogonality of the gradient matrix in the three physical domains of fuselage elasticity, mold compression, and worktable deformation. Matrix multiplication is performed between the gradient matrix and the multi-source variance input set, and cross-correlation terms are superimposed to achieve the mapping and transfer of uncertainties in each physical domain to the three-dimensional spatial position prediction error. Variance synthesis is employed, and the synthesized covariance matrix is ​​numerically stabilized to suppress ill-conditioned condition numbers caused by cross-correlation heights approaching ±1, thus achieving the numerical convergence requirement for subsequent eigenvalue decomposition. Through the matrix operation process described above, the initial three-dimensional coordinate offset from the previous step is expanded into an original covariance matrix with uncertainty quantification features, thereby achieving a quantitative characterization of the reliability of spatial location prediction.

[0206] For example, in one downward cycle of a GK-5000 large forging press, the initial three-dimensional coordinate offset vector is [0.18mm, -0.12mm, 0.05mm], and the residual variance of the elastic modulus of the press body is taken as... The residual variance of the equivalent compressive stiffness of the mold is taken as The residual variance of the workbench flexural stiffness is taken as The cross-correlation matrix elements among the three are 0.12, -0.08, and 0.05, respectively. The gradient matrix J obtained from finite element sensitivity analysis has components of 0.85, 0.63, and 0.47 in the three physical domain directions, respectively. Substituting this gradient matrix and the variance-covariance matrix into the synthesis formula, and performing matrix multiplication and transpose multiplication, the diagonal elements of the original covariance matrix are... , , This matrix is ​​used as input for eigenvalue decomposition in the subsequent S7.3 sub-step to support the geometric modeling of the uncertainty of the nominal force occurrence point. Verification results show that the condition number of the covariance matrix is ​​controlled within 200 under this working condition, ensuring the stability and high accuracy of the prediction confidence calculation.

[0207] S7.3: Receive the original covariance matrix, and perform eigenvalue decomposition and principal axis alignment optimization to orthogonalize the original covariance matrix to generate a standard covariance ellipsoid parameter set describing the distribution of uncertainty in the location of the nominal force occurrence point.

[0208] The system receives the original covariance matrix generated by substep S7.2 as input data and treats it as a mathematical representation matrix describing the correlation of multi-source uncertainties in the initial 3D coordinate prediction. The eigenvalue decomposition processing unit is invoked to perform matrix decomposition on the original covariance matrix, decomposing it into an eigenvector matrix and an eigenvalue diagonal matrix, ensuring that each eigenvector satisfies orthogonality normalization constraints. The eigenvalue diagonal matrix and eigenvector matrix are correlated and calculated, using the square root of the eigenvalues ​​as the length parameters of the three principal axes of the covariance ellipsoid, and using the combination of eigenvector column vectors as the attitude direction set of the covariance ellipsoid in 3D space. A principal axis alignment optimization method is introduced. Based on the normal characteristics of the center plane of the mold cavity, the principal axis direction components in the eigenvector matrix are rotated and mapped to an orthogonal basis consistent with the coordinate system of the defined plane, thereby achieving geometric alignment between the covariance ellipsoid and the plane defining the nominal force generation point. The aligned eigenvalues ​​and eigenvectors are recombined and encapsulated into standard ellipsoidal parameters, including the set of principal axis lengths, the cosine matrix of the principal axis directions, and the coordinate vector of the center point, which serve as the standard covariance ellipsoidal parameter set describing the distribution of uncertainty in the location of the nominal force occurrence point.

[0209]

[0210] in, The original covariance matrix, The eigenvector matrix, It is a diagonal matrix of eigenvalues.

[0211] , , These are the lengths of the three principal axes of the covariance ellipsoid. , , The eigenvalue set obtained through eigenvalue decomposition is {0.27, 0.15, 0.08}, and each column of the eigenvector matrix represents the direction cosine of the corresponding principal axis in the original coordinate system. The square roots of the eigenvalues ​​are calculated to obtain {0.5196, 0.3873, 0.2828} as the length parameters of the ellipsoidal principal axes. Combined with the normal vector of the mold cavity center plane, the principal axis directions are rotated to a plane-aligned state, forming a set of three mutually orthogonal principal axis directions aligned with the plane coordinate system. Finally, the standard covariance ellipsoidal parameter set is generated for confidence interval constraint correction in S7.4, which significantly improves the consistency and geometric interpretability of the positional uncertainty characterization.

[0212] S7.4: The initial three-dimensional coordinate offset vector is corrected by confidence interval constraint using the standard covariance ellipsoid parameter set. Abnormal components that exceed the preset tolerance range are removed and re-weighted and fused to generate corrected three-dimensional coordinates of the nominal force occurrence point with robust characteristics.

[0213] Based on the standard covariance ellipsoid parameter set, confidence interval constraint calculations are performed on each component of the initial 3D coordinate offset vector. The amplitude of each coordinate component is compared with the confidence semi-axis length of the corresponding covariance ellipsoid along the principal axis of that component to obtain an out-of-limit judgment result. Components whose amplitude exceeds the corresponding confidence semi-axis length and whose direction is discontinuous with the historical offset trend are marked as anomalous components, and an anomalous mask vector is generated. The anomalous mask vector is used to perform a rejection operation on the initial 3D coordinate offset vector, replacing the anomalous components with unbiased reference values ​​or historical weighted averages, forming an intermediate coordinate offset vector after anomaly rejection. Based on the covariance matrix of the intermediate coordinate offset vector after anomaly rejection and the standard covariance ellipsoid, weight factors are assigned to each remaining component according to the inverse covariance weighted fusion principle, and the corrected 3D coordinates are calculated using weighted fusion. The consistency of this fusion result with the standard covariance ellipsoid parameter set is checked, ensuring that the position coordinates are within the high confidence interval while correcting them. By using multi-axis constraint elimination and inverse covariance weighted fusion processing, the initial three-dimensional coordinate offset vector from the previous step is transformed into the corrected three-dimensional coordinates of the nominal force generation point with high robustness and high confidence, thereby eliminating the interference of instantaneous anomalies on the final control accuracy.

[0214] For example, in a single forming operation of a GK-5000 large forging press, the initial three-dimensional coordinate offset vector is [0.48mm, -0.35mm, 0.27mm], and the corresponding standard covariance ellipsoidal semi-axis lengths are [0.50mm, 0.40mm, 0.25mm]. The Z-axis value of 0.27mm exceeds the confidence limit of 0.25mm and deviates from the mean Z-axis offset of 0.22mm over the previous 10 cycles by more than 0.03mm, thus being deemed an anomaly. After anomaly masking, the Z-axis value is replaced with 0.22mm, forming an intermediate vector [0.48mm, -0.35mm, 0.22mm]. Combining the diagonal elements of the covariance matrix [0.04, 0.03, 0.02], the inverse covariance weights are calculated to be [25, 33.3, 50]. Substituting these weights into the aforementioned weighted fusion formula yields the corrected results [0.479mm, -0.349mm, 0.221mm]. These corrected coordinates, after undergoing a covariance ellipsoid consistency check, all fall within the 95% confidence interval, effectively suppressing abnormal disturbances. The corrected three-dimensional coordinates of the nominal force generation point are directly available for use by the closed-loop control module, significantly improving the spatial positioning accuracy of the force application point during forming.

[0215] S7.5: Based on the corrected three-dimensional coordinates of the nominal force generation point and the standard covariance ellipsoid parameter set, perform a data structure encapsulation operation to bind the two into a unified state feedback object, and generate the corrected three-dimensional coordinates of the nominal force generation point and the covariance ellipsoid.

[0216] Step S8: Based on the corrected three-dimensional coordinates and covariance ellipsoid of the nominal force generation point, dynamically adjust the output ratio of multiple sets of hydraulic cylinders or connecting rods to correct the parallelism between the slider and the worktable, thus completing precise closed-loop control of the nominal force generation point. Specifically, this includes: S8.1: Obtain the corrected three-dimensional coordinates and covariance ellipsoid data of the nominal force generation point. Based on the preset reference position of the mold cavity center, calculate the spatial residual vector and attitude tilt angle deviation of the nominal force generation point, and generate a comprehensive control error signal containing position offset component and angle tilt component.

[0217] S8.2: Receive the comprehensive control error signal, and use the inverse dynamic solution method of the multi-axis linkage hydraulic system to map the position offset component and the angle tilt component into the theoretical compensation displacement of each independent cylinder group, and generate multiple sets of cylinder target displacement command sequences.

[0218] The system receives the comprehensive control error signal output from the previous sub-step. This signal contains spatial position offset and angular tilt components, decomposed from the difference between the corrected nominal force generation point and the mold cavity center. For this error signal, the system calls the inverse dynamics model based on system structural parameters and constraints within the control solution unit of the multi-axis linkage hydraulic system. This transforms the input six-degree-of-freedom error into a set of Cartesian coordinate displacement vectors required at the slider support points. During the inverse solution process, the cylinder arrangement matrix and the slider rigid body kinematic equations are introduced to establish a coupled mapping relationship between input and output. A combination of matrix decomposition and pseudo-inverse solution is used to ensure that the minimum norm solution can still be obtained when the number of degrees of freedom is greater than the number of actuators. A bounded solution space is formed by introducing mechanical stroke and velocity constraints. In the mapping calculation, the position offset component is directly projected onto a translational compensation subspace parallel to the worktable surface. The angular tilt component is transformed into its equivalent displacement component at each cylinder mounting point through a Jacobian matrix transformation from attitude to displacement. The theoretical compensation displacement of the hydraulic cylinders is calculated and constrained saturation is applied to limit its amplitude to within the safe compensation range of each cylinder under the current speed and pressure conditions. The multiple sets of target displacement quantities of the hydraulic cylinders obtained through the above inverse solution are reconstructed into a time-synchronized command sequence according to the cylinder number order, serving as the input control quantity for subsequent output ratio control modules. By mapping and solving the comprehensive error signal within the inverse dynamic domain of the multi-axis hydraulic system, the target displacement command of the hydraulic cylinders that meets the parallelism correction requirements of the working plane is obtained, achieving the precise conversion of error quantities into executable actuation actions.

[0219] S8.3: Based on the target displacement command sequence of the multiple sets of hydraulic cylinders, combined with the real-time collected pressure feedback value and piston speed status of each hydraulic cylinder, a proportional-integral-derivative adjustment strategy with feedforward compensation is executed to calculate the opening correction coefficient of each set of servo proportional valves and generate real-time output ratio control parameters for multiple sets of hydraulic cylinders.

[0220] Based on multiple sets of target displacement command sequences for hydraulic cylinders, real-time sampling data from pressure sensors in each hydraulic cylinder cavity, and speed status information from piston displacement encoders, a multi-channel dynamic control data matrix is ​​constructed, which includes three types of elements: displacement setpoint, pressure feedback value, and speed status.

[0221] The multi-channel dynamic control data matrix is ​​input into the feedforward channel of the proportional-integral-derivative controller with a feedforward compensation branch. Based on the difference between the rate of change of the target displacement of each cylinder and the current actual speed, an estimated opening correction amount is generated to offset the response lag caused by execution delay and load inertia in advance.

[0222] The feedforward-compensated adjustment signal, together with the difference between the target displacement and the actual displacement, is input into the proportional term branch of the PID main loop. The instantaneous position error is amplified by the proportional coefficient to form a fast response component, which is then weighted and summed with the cumulative historical position error integral in the integral term and the predicted trend component calculated based on the velocity state in the derivative term within the controller.

[0223] The above weighted summation result is mapped to a standardized servo proportional valve opening correction coefficient vector using normalized adjustment coefficients. Each opening correction coefficient is calculated using the following formula:

[0224] Among them, K adj,i K is the opening correction coefficient for the i-th group of servo proportional valves. p,i K I,i K D,i These are the proportional, integral, and differential gain coefficients, respectively, and e(t) is the instantaneous position error function of the cylinder group.

[0225] The target opening degree of each group of hydraulic cylinders is multiplied element by element with the above opening degree correction coefficient to generate a vector of real-time output ratio control parameters for multiple groups of hydraulic cylinders under the current working conditions. This vector serves as the direct input reference for subsequent differential adjustment of hydraulic flow and pressure.

[0226] By using the above-mentioned feedforward compensation and PID control method, the target displacement command sequence and real-time monitoring status generated in the previous step are transformed into multiple sets of cylinder output ratio parameters that are dynamically stable and respond quickly, thereby achieving the real-time control effect of precise correction of slider posture.

[0227] For example, in the four-cylinder independent drive structure of the GK-5000 large forging press, the target displacement command sequence of the hydraulic cylinder in a certain working cycle is set as [0.25mm, 0.18mm, 0.22mm, 0.20mm], and the collected real-time pressure feedback values ​​are [12.5MPa, 11.8MPa, 12.2MPa, 12.0MPa], and the piston speed is [8.5mm / s, 8.1mm / s, 8.3mm / s, 8.4mm / s]. The above data matrix is ​​input into a PID controller with a feedforward branch. The feedforward path uses the difference between the target displacement change rate and the speed state to generate a pre-compensation signal to guide the fast response. The proportional gain coefficient K... p,i The configuration is [0.8, 0.82, 0.81, 0.80], with an integral gain coefficient K. I,i The configuration is [0.15, 0.14, 0.15, 0.14], and the differential gain coefficient K... D,i All values ​​are set to 0.05. K is calculated according to the formula. adj,iThe values ​​[0.972, 0.968, 0.970, 0.969] are multiplied by their respective target opening values ​​to generate multiple sets of real-time cylinder output ratio control parameters [0.243mm, 0.174mm, 0.213mm, 0.194mm]. Applying these control parameters to execute hydraulic output, monitoring showed that the flatness deviation of the lower surface of the slider decreased from the initial 0.08mm to 0.015mm, achieving a significant improvement in parallelism maintenance capability and the stability of the nominal force application point accuracy.

[0228] S8.4: By applying the real-time output ratio control parameters of the multiple sets of hydraulic cylinders, the high-pressure servo hydraulic system is driven to perform differentiated output adjustment of the flow and pressure of each set of hydraulic cylinders, thereby changing the distribution of support reaction force in different areas of the slider and generating dynamic correction action of the slider posture.

[0229] S8.5: Monitor the change in the flatness of the lower surface of the slider after the execution of the dynamic correction action of the slider posture, verify whether the nominal force generation point returns to the tolerance range of the mold cavity center, and if the accuracy convergence condition is met, lock the current real-time output ratio control parameters of multiple sets of hydraulic cylinders to complete the precise closed-loop control of the nominal force generation point.

[0230] This application also provides a precise control system for the nominal force generation point of a large forging press, which uses the above-mentioned precise control method for the nominal force generation point of a large forging press to correct the position of the nominal force generation point.

[0231] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0232] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0233] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for precisely controlling the nominal force generation point of a large forging press, specifically including: S1: Generate the parameter space error sensitivity gradient matrix based on the structural dimension parameters and typical working condition sequences of a large forging press; S2: Collect the measured slider posture data of the current forging cycle, compare it with the predicted value of the digital twin model, and calculate the spatial residual vector of the nominal force occurrence point; S3: Map the spatial residual vector of the nominal force occurrence point to the parameter space error sensitivity gradient matrix, filter out the physical parameter channels with sensitivity weights higher than the preset threshold, and generate a high-sensitivity parameter calibration set. S4: Based on the calibration set of highly sensitive parameters, locate the local geometric domain that is strongly coupled with the digital twin model, and logically separate the local geometric domain from the global reference finite element mesh model to generate the local sub-model region to be reconstructed. S5: Based on the local sub-model region to be reconstructed, a simplified sub-model template is matched from the fast response template library to generate a local physical behavior alternative model; S6: If a step change is detected in the sensor data, freeze all current parameter updates and input the spatial residual vector of the nominal force occurrence point into the abnormal pattern recognizer to generate a local physical behavior instantaneous reset command. S7: Executes the output result projection operation of the local physical behavior replacement model or the local physical behavior instantaneous reset command, directly projecting the output result of all local sub-models onto the nominal force generation point definition plane, generating the corrected three-dimensional coordinates and covariance ellipsoid of the nominal force generation point. S8: Adjust the output ratio of multiple sets of hydraulic cylinders or connecting rods according to the three-dimensional coordinates of the nominal force generation point and the covariance ellipsoid.

2. The method for precisely controlling the nominal force generation point of a large forging press according to claim 1, characterized in that, The generation of the parameter spatial error sensitivity gradient matrix based on the structural size parameters and typical working condition sequence of the large forging press is specifically achieved by sequentially disturbing the physical parameters of the machine body Young's modulus, the equivalent compressive stiffness of the die, and the support stiffness of the worktable based on the structural size parameters and typical working condition sequence of the large forging press, recording the amplitude and direction of the influence of each parameter on the three-dimensional coordinate offset of the center point of the lower surface of the slider, and generating the parameter spatial error sensitivity gradient matrix.

3. The method for precisely controlling the nominal force generation point of a large forging press according to claim 1, characterized in that, The digital twin model includes a material constitutive parameter library, a boundary condition mapping table, a multiphysics coupling solution engine, and a parameter-geometric domain mapping table.

4. The method for precisely controlling the nominal force generation point of a large forging press according to claim 1, characterized in that, The global benchmark finite element mesh model is obtained by discretizing the machine frame columns, mold mounting surfaces, and workbench support areas in the structural dimensional parameters and typical working condition sequence data of a large forging press based on finite element mesh generation technology.

5. The method for precisely controlling the nominal force generation point of a large forging press according to claim 1, characterized in that, The fast response template library stores multiple simplified sub-model templates that have been verified by real forging load spectrum excitation. Each template contains a reduced-order basis function matrix of the local geometric domain, a set of constitutive relation parameters, and boundary condition mapping rules.

6. The method for precisely controlling the nominal force generation point of a large forging press according to claim 1, characterized in that, The anomaly pattern recognizer is a classification model trained and constructed based on historical fault samples, offline simulation data, and experimental verification data.

7. The method for precisely controlling the nominal force generation point of a large forging press according to claim 1, characterized in that, S3 specifically includes: Obtain the spatial residual vector of the nominal force occurrence point and the parameter space error sensitivity gradient matrix. Perform weighted projection operation on each component of the spatial residual vector of the nominal force occurrence point and the corresponding row vector of the parameter space error sensitivity gradient matrix to generate a parameter sensitivity response scalar sequence that characterizes the theoretical contribution of each physical parameter to the current spatial deviation. Based on the parameter sensitivity response scalar sequence, absolute value normalization is performed, and the normalized values ​​are mapped to the dimensionless interval from zero to one to generate a standardized parameter sensitivity weight distribution vector. Receive the parameter sensitivity weight distribution vector, call the threshold dynamic judgment logic, combine the load level of the current forging cycle with the historical noise baseline to calculate the dynamic cutoff threshold, and generate the sensitivity screening critical value used to distinguish between significant and non-significant parameters. By performing an element-by-element comparison operation on the standardized parameter sensitivity weight distribution vector using the sensitivity screening threshold, the index identifiers of all physical parameters whose weight values ​​are greater than the sensitivity screening threshold are extracted, and a preliminary candidate list of highly sensitive physical parameters is generated. Based on the preliminary candidate list of highly sensitive physical parameters, parameter type deduplication and priority sorting are performed. The selected physical parameters are arranged in descending order of sensitivity weight and encapsulated into structured data objects to generate the final set of highly sensitive parameter calibrations as the input basis for local sub-model reconstruction.

8. The method for precisely controlling the nominal force generation point of a large forging press according to claim 1, characterized in that, S4 specifically includes: Based on the physical parameter type identifier in the high-sensitivity parameter calibration set, the parameter-geometric domain mapping relationship table is retrieved to obtain the local geometric domain topological feature data that is strongly coupled with each high-sensitivity physical parameter, and a local geometric domain feature index set to be located is generated. Using the feature index set of the local geometric domain to be located, spatial matching query is performed on the node coordinate data of the global benchmark finite element mesh model. The Euclidean distance weight between each node and the feature topology center is calculated to generate a set of local geometric domain spatial coordinates containing the influence range of highly sensitive parameters. Based on the set of spatial coordinates of the local geometric domain, extract the corresponding element connection relationship and material property data in the global reference finite element mesh model, and construct the initial sub-mesh model of the local geometric domain containing complete physical information; Separate internal free nodes and external constraint nodes from the initial sub-mesh model of the local geometric domain, and generate a list of boundary condition nodes that define the interaction interface between the local geometric domain and the global environment; Based on the boundary condition node list, a logical cutting operation is performed on the initial sub-mesh model of the local geometric domain, removing redundant connections with the global mesh and encapsulating independent data structures to generate a local sub-model region to be reconstructed with independent solution capabilities.

9. The method for precisely controlling the nominal force generation point of a large forging press according to claim 1, characterized in that, S5 specifically includes: The geometric topology and current stress-strain state of the local sub-model region to be reconstructed are obtained. The geometric topology is mapped to mesh nodes. The displacement constraint vectors and force load distribution characteristics of key boundary nodes are extracted to generate local region boundary condition descriptors. Based on the local region boundary condition descriptor, the fast response template library is searched and traversed, the Euclidean distance between the current boundary feature and the historical load spectrum fingerprint of each simplified sub-model template in the library is calculated, and a template matching score sequence is generated. Based on the template matching score sequence, threshold filtering and optimal index positioning operations are performed to select the simplified sub-model template with the highest score that has been verified by real forging load spectrum excitation as the target candidate and generate the target simplified sub-model template identifier. Using the target simplified sub-model template identifier, the corresponding constitutive relation parameter set and reduced basis function matrix are called. The local region boundary condition descriptor is substituted into the reduced basis function matrix for projection transformation to reconstruct a local dynamic equation set that adapts to the current working condition and generate an instantiated local physical behavior alternative model. The instantiated local physical behavior substitution model is subjected to millisecond-level response performance verification, and single-step time integration derivation is performed to output the displacement field increment and reaction torque in the local domain, forming the local physical behavior substitution model output result that can be directly projected onto the plane defining the nominal force occurrence point.

10. A precise control system for the nominal force generation point of a large forging press, characterized in that, The position of the nominal force generation point is corrected by using the precise control method of the nominal force generation point of the large forging press according to claim 1.

Images