Vacuum forging global temperature virtual acquisition method and system based on digital twinning

Abstract

This invention relates to the technical field of digital twins and industrial intelligent monitoring, and discloses a method and system for virtual acquisition of full-domain temperature in vacuum forging based on digital twins. First, it reconstructs full-domain temperature field data including unmeasurable points using temperature data from measurable points. Second, it trains a virtual temperature field acquisition model based on full-domain temperature field data from multiple consecutive time periods, enabling the model to learn the temporal evolution law of the temperature field. Then, it embeds this model into a digital twin, driving it to autonomously and continuously deduce the virtual temperature field under real-time input conditions without physical sensors. At preset trigger moments such as forging stage switching, the virtual temperature field is calibrated using spatial registration with measurable temperature data as anchor points through dual-constraint deformation field estimation, and the model state is updated. This invention reduces the independent operation capability that relies on continuous online operation of physical sensors, and improves the robustness of monitoring under extreme conditions in vacuum forging.

Description

Technical Field

[0001] This invention relates to the field of digital twin and industrial intelligent monitoring technology, specifically to a method and system for virtual acquisition of full-domain temperature in vacuum forging based on digital twin. Background Technology

[0002] Vacuum isothermal forging is a core process for manufacturing strategic components such as aero-engine turbine disks and titanium alloy disks. It requires precision forming of difficult-to-deform alloys such as nickel-based and titanium-based materials in a vacuum environment of 1000~1200℃. The temperature distribution in the forging chamber directly affects the microstructure and properties of the workpiece. However, due to the extreme working conditions, physical sensors can only be installed in a few accessible locations such as the furnace wall and the mold support plate. The temperature of critical areas such as the core of the forging and the surface of the mold cavity cannot be directly obtained, forming a monitoring blind spot.

[0003] Existing technologies for acquiring the temperature of unmeasurable regions mainly fall into two categories. The first is finite element multiphysics simulation, which can couple mechanisms such as heat conduction, thermal radiation, contact thermal resistance, and plastic deformation heat to output a complete temperature field. However, a single calculation takes several minutes to several hours, failing to meet real-time monitoring requirements. The second is a data-driven surrogate model method, which uses simulation data to train neural networks to replace finite element calculations. However, it is essentially a static mapping, relying on physical sensors to provide continuous input in each control cycle; once the sensor fails due to high temperature or signal interruption, the model loses its driving input and ceases to function.

[0004] Therefore, existing technologies rely on continuous online input from physical sensors. In the event of sensor failure or signal interruption, the system will lose its ability to update the temperature field, making it difficult to continuously and reliably obtain the full-domain temperature field information of the vacuum forging space. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for virtual acquisition of full-domain temperature in vacuum forging based on digital twins, so as to solve the technical problem that the full-domain temperature field monitoring in the prior art is highly dependent on continuous online input from physical sensors, and the system loses its ability to update the temperature field in the event of sensor failure or signal interruption.

[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A method for virtual acquisition of global temperature in vacuum forging based on digital twins includes the following steps: Points in the vacuum forging furnace where physical sensors can be installed are marked as measurable points, and points where physical sensors cannot be installed are marked as non-measurable points. Temperature data is collected in real time using physical sensors located at measurable points to obtain temperature data for measurable points. Based on the measurable temperature data, the unmeasurable temperature data is reconstructed, and the global temperature field data of the vacuum forging furnace is composed of the measurable temperature data and the unmeasurable temperature data. Under different operating conditions of the vacuum forging furnace, global temperature field data from multiple consecutive moments are selected to form a temperature field evolution dataset. Based on the temperature field evolution dataset, a virtual temperature field acquisition model is trained to predict the global temperature field data of the next moment based on the global temperature field data of the previous moment. The virtual temperature field acquisition model is embedded into the digital twin of the vacuum forging furnace, enabling it to autonomously deduce and generate a virtual temperature field without real-time sensor data. At the moment when the preset triggering condition occurs, the measurable temperature data is used as a local anchor point. The measurable temperature data at the moment of occurrence and the time before the moment of occurrence are used to construct a joint constraint including the first deviation constraint and the second deviation constraint, and solve a global deformation field acting on the virtual temperature field. The virtual temperature field output by the virtual temperature field acquisition model in the digital twin model is spatially registered using the global deformation field, and the spatially registered global virtual temperature field is used as the calibrated global virtual temperature field. The virtual temperature field acquisition model is then updated and optimized based on the calibrated global virtual temperature field. The first deviation constraint is used to ensure that the virtual temperature field after registration is consistent with the temperature data of the measurable point at the current time at the measurable point location. The second deviation constraint is used to ensure that the time evolution trend of the virtual temperature field after registration is consistent with the time evolution trend of the temperature data of measurable points.

[0007] As a preferred embodiment of the present invention, the virtual temperature field acquisition model adopts a time-series prediction model, with the global temperature field data of the previous moment as input and the global temperature field data of the next moment as output. The mean square error between the predicted global temperature field data at the next time step output by the time series prediction model and the true value of the global temperature field data at the next time step in the temperature field evolution dataset is used as the training loss of the time series prediction model.

[0008] As a preferred embodiment of the present invention, under different working conditions, the global temperature field distribution model is used to continuously generate multiple segments of global temperature field time series covering each forging stage, forming the temperature field evolution dataset. The forging stage includes heating, holding, forging, and cooling.

[0009] As a preferred embodiment of the present invention, the method for autonomously generating a virtual temperature field using a digital twin model includes: At the initial moment, the temperature data of measurable points is collected using physical sensors, and the global temperature field data is reconstructed using the temperature data of measurable points. The global temperature field data at the initial moment is synchronously transmitted to the digital twin model. The digital twin model uses the virtual temperature field acquisition model to predict the virtual temperature field at the next moment based on the global temperature field data at the initial moment. The virtual temperature field at the next moment is used as the new input of the virtual temperature field acquisition model to predict the virtual temperature field at the next moment after that. This process is repeated until the preset trigger condition occurs, so as to realize the autonomous and continuous deduction of the virtual temperature field under the condition of real-time input without physical sensors.

[0010] As a preferred embodiment of the present invention, the preset triggering condition is the switching of the forging stage of the vacuum forging furnace.

[0011] As a preferred embodiment of the present invention, the spatial registration method for the virtual temperature field includes: The temperature data of the measurable point collected by the physical sensor at the time of occurrence of the preset trigger condition is obtained and recorded as the true value of the measurable point at the current time. The temperature data of the measurable point collected by the physical sensor at the previous time of occurrence is also obtained and recorded as the true value of the measurable point at the previous time. The virtual temperature field output by the virtual temperature field acquisition model at the previous moment before the occurrence time is obtained and denoted as the virtual temperature field at the previous moment. The virtual temperature field output by the virtual temperature field acquisition model at the occurrence time is also obtained and denoted as the virtual temperature field to be registered at the current moment. Using the measurable point location as the anchor point, a first deviation constraint is constructed between the pre-registered virtual temperature field at the current time and the actual value of the measurable point at the current time, and a second deviation constraint is constructed between the change of the virtual temperature field at the current time relative to the virtual temperature field at the previous time and the change of the actual value of the measurable point at the current time relative to the actual value of the measurable point at the previous time. The first deviation constraint is used to constrain the registered virtual temperature field to be consistent with the actual measured value at the measurable point location, and the second deviation constraint is used to constrain the time evolution trend of the registered virtual temperature field to be consistent with the evolution trend of the actual temperature field. The first deviation constraint and the second deviation constraint are used as joint constraints for deformation field estimation. A global deformation field acting on the virtual temperature field to be registered at the current time is solved, so that the virtual temperature field at the current time after being corrected by the global deformation field satisfies the first deviation constraint and the second deviation constraint at the measurable point. The global deformation field obtained by the solution is used to correct the positions of all unmeasurable points in the virtual temperature field to be registered at the current moment, so as to complete the global spatial registration of the virtual temperature field and obtain the calibrated global virtual temperature field.

[0012] As a preferred embodiment of the present invention, the method for updating and optimizing the virtual temperature field acquisition model based on the calibrated global virtual temperature field includes: Replace the virtual temperature field to be registered at the current time that was originally output by the virtual temperature acquisition model at the time of occurrence with the calibrated global virtual temperature field. The calibrated global virtual temperature field is used as the input to the virtual temperature field acquisition model at the next moment, so that the virtual temperature field acquisition model can continue to extrapolate the virtual temperature field at subsequent moments from the calibrated global virtual temperature field, thereby eliminating the accumulated errors during the autonomous extrapolation process.

[0013] As a preferred embodiment of the present invention, a pre-established global temperature field distribution model for predicting unmeasurable temperature data based on measurable temperature data is used to combine the measurable temperature data and the unmeasurable temperature data output by the global temperature field distribution model into global temperature field data for a vacuum forging furnace.

[0014] To address the aforementioned technical problems, the present invention further provides the following technical solution: A virtual temperature acquisition system for vacuum forging based on digital twins, the system comprising: Physical sensors are placed at measurable points inside the vacuum forging furnace to collect temperature data at those points in real time. The global temperature field reconstruction module is used to reconstruct the non-measurable temperature data based on the measurable temperature data, and the global temperature field data of the vacuum forging furnace is composed of the measurable temperature data and the non-measurable temperature data. The dataset construction module is used to select global temperature field data from multiple consecutive moments under different operating conditions of the vacuum forging furnace to form a temperature field evolution dataset. The model training module is used to train a virtual temperature field acquisition model based on the temperature field evolution dataset. The virtual temperature field acquisition model is used to predict the global temperature field data at the next moment based on the global temperature field data at the previous moment. The digital twin simulation module is used to embed the virtual temperature field acquisition model into the digital twin of the vacuum forging furnace, enabling it to autonomously simulate and generate a virtual temperature field without real-time sensor data. The spatial registration and update module is used to spatially register the virtual temperature field output by the virtual temperature field acquisition model in the digital twin model with the measurable temperature data as the local anchor point at the time of the occurrence of the preset trigger condition, so as to obtain the calibrated global virtual temperature field, and update and optimize the virtual temperature field acquisition model based on the calibrated global virtual temperature field.

[0015] Compared with the prior art, the present invention has the following advantages: This invention constructs a virtual temperature field acquisition model that learns the temporal evolution law of the temperature field and embeds it in a digital twin as a self-updating engine. This enables the digital twin model to autonomously deduce the virtual temperature field, thereby improving the independent operation capability of the temperature field monitoring system, which is free from the dependence on physical sensors for continuous online operation.

[0016] This invention employs an intermittent spatial registration mechanism under preset trigger conditions, which uses physical sensor data to globally calibrate the virtual temperature field only at specific moments such as the switching of the forging stage. This transforms the physical sensor from a necessary input for continuously driving the system to an auxiliary reference for periodically calibrating the system's accuracy, thereby reducing the cumulative error of autonomous inference. Attached Figure Description

[0017] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for virtual acquisition of global temperature in vacuum forging based on digital twins, provided in this application embodiment.

[0019] Figure 2 This is a block diagram of a virtual temperature acquisition system for vacuum forging based on digital twins, provided as an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] like Figure 1As shown, this invention provides a virtual temperature acquisition method for the entire vacuum forging process based on digital twins. First, locations within the vacuum forging furnace where sensors can be installed are marked as measurable points, and locations where sensors cannot be installed are marked as unmeasurable points. Local temperature data is acquired using physical sensors at the measurable points, and the entire temperature field, including the temperature at the unmeasurable points, is reconstructed based on this local data. Furthermore, a time-series prediction model, i.e., a virtual temperature acquisition model, is trained using the entire temperature field data from multiple consecutive moments under different operating conditions. This model enables the prediction of the entire temperature field at the next moment based on the temperature field at the previous moment, thereby learning the dynamic evolution of the temperature field.

[0022] After the model is embedded in the digital twin, the digital twin can autonomously and continuously generate a virtual temperature field without real-time input from physical sensors. To suppress the cumulative error of long-term simulations, the system performs spatial registration and calibration of the virtual temperature field at preset trigger times, such as switching between forging stages, using the measurable temperature data collected by physical sensors as local anchor points. The calibration results are then used to update the state of the virtual temperature field acquisition model, allowing subsequent simulations to restart from the calibrated high-precision state. Through this autonomous simulation combined with intermittent calibration working mode, this method can still continuously and reliably output global temperature field information, including unmeasurable critical areas, even with very few physical sensors and the risk of sensor failure.

[0023] The first step is to mark the locations in the vacuum forging furnace where physical sensors can be installed as measurable points, and the locations where physical sensors cannot be installed as non-measurable points. Temperature data is then collected in real time using physical sensors located at the measurable points to obtain the temperature data of the measurable points.

[0024] Construct a three-dimensional geometric model of the forging chamber, forging die, and forged workpiece in the vacuum forging furnace. In the three-dimensional geometric model, the modeling meshes where solid sensors can be installed are marked as points where the temperature can be measured, and the remaining modeling meshes are marked as points where the temperature cannot be measured.

[0025] The geometric model is drawn to scale based on the actual equipment dimensions and structure, clearly distinguishing core components such as the furnace wall, heating elements, mold cavities, and workpiece blanks. After completing the construction of the three-dimensional geometric model, it is meshed to generate a modeling mesh composed of a large number of spatial discrete elements. The modeling mesh serves as the basic unit for finite element calculations, with each mesh node or element carrying the material properties, initial conditions, and temperature data to be solved at that spatial location.

[0026] Based on the modeling mesh, and according to the actual sensor deployment scheme, the modeling mesh where physical sensors can be installed is marked as measurable temperature points. Due to the extreme operating conditions of vacuum, high temperature, and high pressure, physical sensors can only be introduced into the furnace through a small number of vacuum-sealed connectors. In actual deployment, seven thermocouples are typically installed in four locations on the furnace wall of the forging chamber and on a specific surface of the mold. The furnace wall sensors are installed on the inner side of the furnace wall, 50-100mm from the heating element, and their corresponding modeling meshes are located at the corresponding coordinate positions on the inner surface of the furnace wall in the 3D geometric model. The mold sensors are embedded in blind holes on the non-working surface of the mold support plate, and their corresponding modeling meshes are located in the surface units of the non-working surface of the mold support plate. The modeling meshes containing these measurable points are relatively open in space, allowing for reliable temperature values ​​to be output in simulation calculations and directly corresponding to actual sensor readings for subsequent model calibration.

[0027] Accordingly, the remaining modeling meshes where physical sensors cannot be installed are marked as points where temperature cannot be measured, mainly including key locations such as the core of the forging, the surface of the mold cavity, and the flash groove area. The modeling mesh corresponding to the core of the forging is located in the geometric center region of the workpiece blank, which is completely encased in high-temperature metal, making it impossible to drill holes or implant sensors. The modeling mesh corresponding to the surface of the mold cavity is located on the working surface where the mold and the forging are in direct contact. This area bears huge mechanical loads and thermal shocks, and any sensor opening would damage the mold strength and induce stress concentration. The modeling mesh corresponding to the flash groove area is located in the narrow gap near the mold parting surface, which is the channel for the violent flow of material during forging, and there is no space available for installation.

[0028] The temperature history of the modeling mesh in the core of the forging directly determines the degree of dynamic recrystallization and grain refinement effect of difficult-to-deform materials (such as nickel-based and titanium-based alloys). It is a core parameter to ensure the uniformity of the forging structure and the achievement of mechanical properties. The temperature distribution of the modeling mesh on the die cavity surface affects interfacial friction, contact thermal resistance, and die life. Abnormal temperature drops or overheating can directly lead to surface defects in the forging or premature die failure. The temperature evolution of the modeling mesh in the flash groove region can reflect the material filling flow state, providing a direct basis for optimizing forging process parameters and avoiding defects such as folding and incomplete filling. Therefore, mastering the temperature field information at these unmeasurable modeling mesh points is essential to better serve the forging process.

[0029] After completing the construction of the three-dimensional geometric model and mesh generation, this invention then establishes a multiphysics finite element thermal model that can realistically reflect the complex heat transfer process inside a vacuum forging furnace. This finite element thermal model constructs a digital virtual forging environment that covers all key physical mechanisms, thus providing a reliable simulation tool for subsequently generating high-precision simulation datasets.

[0030] The second step is to reconstruct the temperature data of unmeasurable points based on the temperature data of measurable points, and to form the global temperature field data of the vacuum forging furnace by combining the temperature data of measurable points and the temperature data of unmeasurable points.

[0031] Using a pre-established global temperature field distribution model for predicting unmeasurable temperature data based on measurable temperature data, the measurable temperature data and the unmeasurable temperature data output by the global temperature field distribution model are combined to form the global temperature field data of the vacuum forging furnace.

[0032] Specifically, a multi-physics finite element thermal model is established based on a three-dimensional geometric model, which couples nonlinear heat conduction, heat radiation, contact thermal resistance, and plastic deformation heat.

[0033] Furthermore, the methods for constructing multiphysics finite element thermal models include: In the three-dimensional geometric model, the modeling mesh of key areas in the forging chamber, forging die, and forging workpiece is refined to improve the simulation accuracy of key areas. Key areas are locations where temperature changes are drastic and have a significant impact on the microstructure and properties of forgings.

[0034] This invention refines the modeling mesh in key regions of the 3D geometric model. Key regions refer to locations with drastic temperature changes that significantly affect the microstructure and properties of the forging, specifically including the core of the forging, the surface of the mold cavity, the flash groove area, and the vicinity of the mold-workpiece contact surface. Refining the mesh in these regions effectively captures steep local temperature gradients, avoiding the loss of computational accuracy caused by an overly sparse mesh. Furthermore, the solution accuracy of the finite element method is directly related to the mesh size; the finer the mesh, the smaller the discretization approximation error of the spatial derivative of the temperature field. Refining the mesh in regions with large temperature gradients can significantly improve local solution accuracy. Simultaneously, using a coarser mesh in regions with gentler temperature changes, such as the outer side of the furnace wall and the mold support structure, allows for control of the computational scale while maintaining overall solution accuracy, thus balancing computational efficiency.

[0035] Import the thermal conductivity of the material, which varies nonlinearly with temperature, from a pre-established database of thermophysical properties of forging materials. Specific heat capacity and density The parameters, among which, the database of thermal property parameters of forging materials is pre-established through material thermal property testing experiments. Thermal conductivity, specific heat capacity, and density are the material thermal property parameters of the heat conduction control equation in the finite element thermal model, which can directly determine the solution accuracy of the temperature field. The heat conduction control equation is: ; In the formula For temperature, For time, The intensity of the internal heat source.

[0036] Import the thermal conductivity, specific heat capacity, and density parameters of the forging materials, which vary non-linearly with temperature, from a pre-established database of forging material thermophysical parameters. This database was pre-established through material thermophysical testing experiments such as laser scintillation and differential scanning calorimetry, covering thermophysical data of typical difficult-to-deform materials such as nickel-based alloys, titanium-based alloys, and cobalt-based alloys from room temperature to above 1200℃. The reason for using parameters that vary non-linearly with temperature is that in the extreme temperature range of 1000~1200℃ during vacuum forging, the thermal conductivity and specific heat capacity of alloy materials change significantly. Using room-temperature constants would lead to a systematic deviation of tens of degrees Celsius between the calculated and actual temperature field values.

[0037] The governing equation for heat conduction is a classical parabolic partial differential equation describing the transfer of heat in a solid medium. It means that the increase in internal energy per unit volume of material per unit time equals the net heat inflow through heat conduction plus the heat generated by internal heat sources. Using this equation as the governing equation for the finite element thermal model provides the mathematical foundation for subsequent temperature field solutions.

[0038] Determine the thermal boundary conditions, which consist of the heat flux density boundary of the heating element in the vacuum forging furnace, the heat radiation boundary of the furnace wall in the vacuum forging furnace, and the contact thermal resistance boundary between the forging die and the forging workpiece.

[0039] To ensure a unique solution to the heat conduction governing equation, this invention applies appropriate thermal boundary conditions, which, together with the equation, constitute a boundary value problem. This invention defines three types of thermal boundary conditions: the first is the heat flux density boundary of the heating body, denoted as... Where n is the direction of the outward normal to the boundary, q flux Given the heat flux density, this boundary describes the rate at which the heating element inputs heat to the mold and workpiece surfaces; the second is the thermal radiation boundary of the furnace wall, denoted as... ,in Let T be the surface emissivity, σ be the Stefan-Boltzmann constant, and T be the surface emissivity. amb The first boundary is the ambient temperature, which describes the physical process by which a high-temperature surface dissipates heat to the surrounding furnace wall through thermal radiation in a vacuum environment; the second boundary is the contact thermal resistance boundary between the forging die and the forged workpiece, determined by the contact heat transfer coefficient h. gap The description is as follows: contact heat flow is This boundary condition characterizes the heat transfer resistance effect caused by microscopic irregularities at the interface between the mold and the workpiece. These three types of boundary conditions collectively cover the main heat transfer paths in the vacuum isothermal forging process, coupling thermal radiation and contact thermal resistance effects, enabling the simulation model to realistically simulate the actual heat transfer environment.

[0040] The amount of heat generated by plastic deformation, introduced as the intensity of the internal heat source into the heat conduction control equation, is calculated to establish the dynamic relationship between the deformation amount, deformation rate and temperature of the forged workpiece in the heat conduction control equation. During forging, the enormous mechanical load applied to the workpiece by the press causes severe plastic deformation of the metal, converting some of the mechanical work into heat. This effect must be included in the thermal model. The formula for calculating the heat generated during plastic deformation is: ; in, It is the heat-to-work conversion factor, usually taken as 0.9~0.95, indicating that 90%~95% of the work done in plastic deformation is converted into heat; The flow stress of the material at the current temperature and strain rate is given by the material constitutive model; The equivalent strain rate. The calculated heat generation during plastic deformation. Substituting the internal heat source intensity Q into the heat conduction control equation establishes a dynamic relationship between deformation amount, deformation rate, and temperature. This allows the finite element thermal model to accurately reflect the physical phenomenon of the core temperature rising due to the heat generated by the plastic work of the forging during compression deformation, thus avoiding the problem of underestimating the internal temperature of the forging in the pure heat conduction model.

[0041] The material's thermal properties, thermal boundary conditions, heat generation from plastic deformation, and heat conduction control equations, together with the modeling mesh of the key region after refinement, constitute a boundary value problem for solving the global temperature field. This boundary value problem is then used as a multiphysics finite element thermal model. By integrating the encrypted modeling mesh, the material's thermal properties that vary nonlinearly with temperature, the boundary conditions comprised of three types of thermal boundary conditions and the heat conduction governing equations, and the heat generated by plastic deformation introduced as an internal heat source, a complete boundary value problem for solving the global temperature field is formed. This boundary value problem is the multiphysics finite element thermal model. By performing finite element numerical solutions on this multiphysics finite element thermal model under different forging conditions (different combinations of heating power curves, press speeds, vacuum levels, etc.), the temperature data of each modeling mesh node in the three-dimensional geometric model at each time step can be obtained. These data constitute the global temperature simulation data under the corresponding conditions. This finite element thermal model serves as the core tool for subsequently generating the training dataset, and its fidelity directly determines the upper limit of the prediction accuracy of the final neural network model.

[0042] Among them, by performing finite element solutions on the multiphysics finite element thermal model under different forging conditions, the temperature data of each modeling mesh in the three-dimensional geometric model under each forging condition were obtained. The temperature data of each modeling mesh in the three-dimensional geometric model under each forging condition are used as the global temperature simulation data under each forging condition.

[0043] After constructing the multiphysics finite element thermal model, it is necessary to calibrate and correct its key boundary condition parameters using measured temperature data collected during actual forging processes. Many boundary condition parameters involved in the finite element thermal model, such as the actual distribution coefficient of the heat flux density of the heating element, the equivalent emissivity of the furnace wall thermal radiation, and the thermal conductivity of the die-workpiece interface, are often initially set based on empirical formulas or manual recommendations. However, in actual equipment, these parameters are affected by factors such as installation location, equipment aging, and fluctuations in furnace vacuum, resulting in significant deviations between their actual values ​​and theoretical default values. If an uncalibrated model is used directly for simulation, the output temperature field data will inherit these deviations, which will then be passed to the subsequently trained neural network model, causing a systematic decrease in the overall prediction accuracy. Therefore, by using measured data from physical sensors available on-site, the finite element thermal model is reverse-calibrated to ensure that its simulation output approximates the thermal behavior of the actual vacuum forging scenario as closely as possible.

[0044] Using at least one set of measured temperature data collected by physical sensors during vacuum isothermal forging, the key boundary condition parameters of the multiphysics finite element thermal model are corrected and calibrated so that the error between the simulated temperature data generated by the multiphysics finite element thermal model at the measurable temperature point and the measured temperature data collected by the physical sensors at the measurable temperature point is less than a preset calibration threshold, thus obtaining the corrected and calibrated multiphysics finite element thermal model.

[0045] Furthermore, methods for modifying the key boundary condition parameters of the multiphysics finite element thermal model include: Obtain at least one set of measured temperature data collected at each measurable temperature point.

[0046] This data set should select a complete typical forging process, such as the complete cycle from the heating stage, through heat preservation and homogenization, to the forging stage and cooling, to ensure coverage of stages dominated by different heat transfer mechanisms, such as heating, homogenization, deformation, and cooling. This ensures that subsequent parameter calibration is binding on the boundary conditions of each stage. Measured temperature data are simultaneously collected by seven thermocouples installed at four locations on the furnace wall and on the mold support plate at a preset sampling frequency (e.g., 100Hz), and then led out via a vacuum-sealed connector to an edge computing device for recording and preprocessing.

[0047] Using a multiphysics finite element thermal model, a global simulated temperature field is obtained under the forging conditions corresponding to the measured temperature data. The objective function for parameter correction is to minimize the deviation between the simulated temperature data at each measurable temperature point in the global simulated temperature field and the measured temperature data.

[0048] A multiphysics finite element thermal model was used as the computational tool to simulate the forging conditions corresponding to the measured temperature data. During the simulation, the model's geometric mesh and material thermal properties remained unchanged, and the input process parameters and timing were identical to those in the actual measurement process, including the heating power curve, press displacement and velocity curves, and vacuum setpoint. After running the simulation, the global simulated temperature field under this condition was obtained. From the global simulated temperature field, the temperature values ​​at the modeling mesh nodes corresponding to each measurable temperature point were extracted as the simulated temperature data for each measurable point under this condition.

[0049] The objective function constructed in this invention aims to minimize the deviation between the simulated temperature data and the corresponding measured temperature data at each measurable point. Mathematically, this can be expressed as root mean square error or maximum absolute error. When the simulated temperature and the measured temperature closely match at each measurable point, it can be inferred that the thermal boundary condition parameters used in the model have become consistent with the heat transfer state of the actual equipment. Since the measurable points are distributed in different spatial orientations on the furnace wall and the mold, covering both radiation-dominated and contact-conduction-dominated areas, the deviations at these points can comprehensively reflect the correction requirements for multiple boundary condition parameters such as heat flux density distribution, radiative emissivity, and contact thermal conductivity.

[0050] An optimization algorithm is used to iteratively correct the parameters of the thermal boundary conditions in the multiphysics finite element thermal model until the error between the temperature simulation data and the measured temperature data at each measurable temperature point is less than the preset calibration threshold.

[0051] This invention utilizes an optimization algorithm to iteratively correct key boundary condition parameters in a multiphysics finite element thermal model. Optimization variables can be set as parameters such as the heat flux density ratio coefficient of each zone of the heating element, the equivalent emissivity of the furnace wall, and the thermal conductivity coefficient of the interface between the mold and the workpiece. In each iteration, the finite element simulation is rerun after modifying the optimization variables to obtain the simulation temperature at new measurable points, and the objective function value is calculated. The optimization algorithm automatically searches the parameter space based on the changing trend of the objective function, gradually approximating the optimal parameter combination that minimizes the objective function. The iteration process continues until the error between the simulation temperature data and the measured temperature data at each measurable temperature point is less than a preset calibration threshold, which can be set to a value that matches the measurement accuracy of the physical sensor.

[0052] Substituting the modified boundary conditions into the multiphysics finite element thermal model yields the modified and calibrated multiphysics finite element thermal model.

[0053] This invention substitutes the iteratively corrected boundary condition parameters into a multiphysics finite element thermal model, replacing the initial empirical default values, thus obtaining a corrected and calibrated multiphysics finite element thermal model. This calibrated model has been grounded using physical sensor data, and its simulated temperature field output shows a high degree of consistency with the actual results at measurable points. Therefore, it can be reasonably inferred that the simulation results at unmeasurable points are also closer to the actual physical state. This lays a reliable model foundation for subsequently generating a high-confidence temperature field finite element simulation dataset.

[0054] Using the corrected and calibrated multiphysics finite element thermal model, simulations were performed under multiple sets of vacuum isothermal forging conditions to generate global temperature simulation data covering all forging conditions and including all measurable and non-measurable temperature points. The global temperature simulation data and the corresponding operating parameters were then compiled to form a temperature field finite element simulation dataset.

[0055] This invention employs orthogonal experimental design or Latin hypercube sampling to systematically select representative forging conditions across multiple process parameters, including heating power, vacuum level, press speed, and displacement, covering the entire process of heating, holding, forging, and cooling. For each set of conditions, a calibrated finite element thermal model is simulated, outputting the temperature values ​​of all modeling mesh nodes at each discrete moment, serving as the global temperature simulation data for that condition. After completing the simulations for all conditions, the global temperature simulation data for each set is aggregated with the corresponding condition parameters to form a temperature field finite element simulation dataset. The input to each sample in the dataset consists of the condition parameters and the simulated temperature values ​​at measurable points, while the output is the simulated temperature values ​​at unmeasurable points. This forms an input-output pairing structure, enabling the subsequently trained neural network model to learn the mapping relationship from measurable point temperatures and process parameters to unmeasurable point temperatures.

[0056] After completing the dataset construction, a lightweight neural network model is trained based on the dataset as a global temperature field distribution model. The multi-physics heat transfer law contained in the finite element simulation is distilled into it, so that it can directly predict the temperature of all unmeasurable points from end to end based on the temperature data of measurable points collected in real time by physical sensors and the current operating parameters, thereby realizing real-time control of the global temperature field of the vacuum forging furnace.

[0057] Based on the finite element simulation dataset of the temperature field, a global temperature field distribution model is trained to predict the simulated temperature of each unmeasurable temperature point according to the operating parameters and the simulated temperature of each measurable temperature point. This enables real-time prediction of the unmeasurable temperature point from the measurable temperature point, thereby mastering the global temperature field distribution of the vacuum forging furnace.

[0058] Furthermore, the operating parameters for vacuum isothermal forging include heating power, vacuum level, press load, and press displacement. These four types of parameters correspond to the external energy input of the temperature field, environmental heat transfer conditions, mechanical load, and degree of deformation, respectively, and are the main control variables determining the distribution and evolution of the temperature field. By using these parameters as model inputs, the model can perceive the driving effect of different process operations on the temperature field, thereby maintaining accurate predictive capabilities during operating condition switching.

[0059] The method for constructing the global temperature field distribution model in this invention adopts a teacher-student model architecture. First, a high-precision teacher model with temporal reasoning ability is trained, and then the learned mapping relationship is transferred to a student model with a more concise structure and faster reasoning through knowledge distillation.

[0060] Furthermore, the methods for constructing a global temperature field distribution model include: A teacher model is constructed, which adopts a model architecture consisting of an encoder, a temporal reasoning network, and a decoder.

[0061] The encoder adopts a dual encoder structure. The first encoder is a measurable point encoder, whose input is the temperature data of each measurable point at the current time, and whose output is the potential feature vector of the measurable point.

[0062] The second encoder is the unmeasurable point encoder. During the training phase of the teacher model, its input is the temperature simulation data of each unmeasurable point at the current time, and its output is the potential feature vector of the unmeasurable point.

[0063] The potential feature vectors of measurable points and the potential feature vectors of unmeasurable points are concatenated and fused to obtain the global potential feature vector at the current time.

[0064] The measurable point encoder maps sparse local temperature readings to a high-dimensional latent space, extracting the global temperature field clues contained therein; the unmeasurable point encoder learns the morphological characteristics of the global temperature field from the complete spatial temperature distribution. After the two are spliced ​​and fused, the model can simultaneously learn the correlation between local sensor signals and global temperature morphology, enabling the teacher model to understand the law of temperature response of measurable point temperature changes to unmeasurable regions, laying the foundation for predicting the global temperature field by relying solely on measurable point input during the inference stage.

[0065] It should be noted that the unmeasurable point encoder is only used during the teacher model training phase. Its input comes from the corresponding temperature simulation values ​​in the simulation dataset, and it does not participate in the calculation during the inference phase.

[0066] The temporal reasoning network adopts a long short-term memory network. The input of the long short-term memory network is a concatenated vector of the global latent feature vector at the current time and the operating condition parameter vector at the current time. The output is the global latent feature vector at the current time enhanced by the temporal context. The operating condition parameter vector consists of heating power, vacuum degree, compressor load and compressor displacement.

[0067] Vacuum isothermal forging is a dynamic process. The temperature field at the current moment depends not only on the current heating input and boundary conditions but also on the cumulative effect of historical heat. Long Short-Term Memory (LSTM) networks selectively memorize and forget historical state information through gating mechanisms, enabling them to capture the temporal dependencies of the temperature field across multiple time steps. By jointly encoding the operating parameters and temperature field characteristics and feeding them into the temporal inference network, historical state information is integrated to apply dynamic constraints to the spatial mapping of the current temperature field, maintaining stable and accurate predictions even when switching between different process stages such as heating, holding, forging, and cooling.

[0068] The decoder uses a transposed convolutional decoder to reconstruct the global latent feature vector of the current time step output by the temporal inference network into the global temperature data of the current time step, and extracts the temperature prediction data of each unmeasurable temperature point from the reconstructed global temperature data as the output.

[0069] This invention utilizes the local sensing characteristics of convolution operations to gradually restore the spatial resolution of the temperature field through a learnable upsampling kernel. This can better maintain the spatial continuity and smoothness of the temperature field, avoid the spatial discontinuity temperature jump problem that may occur in fully connected decoders, and make the reconstructed temperature field more physically reasonable.

[0070] The teacher model was trained using a finite element simulation dataset of the temperature field, resulting in a fully trained teacher model. Because the teacher model includes an encoder for unmeasurable points and a long short-term memory network, it can utilize complete global temperature simulation data and historical time-series information during the training phase, thus achieving high prediction accuracy.

[0071] A student model is constructed, consisting of a measurable point encoder and a transposed convolutional decoder. The input of the measurable point encoder in the student model is the temperature data of each measurable point at the current time, and the output is the potential feature vector of the measurable point. The potential feature vector of the measurable point is concatenated with the operating condition parameter vector at the current time and then input into the transposed convolutional decoder. The transposed convolutional decoder outputs the temperature prediction data of each unmeasurable point.

[0072] Compared to the teacher model, the student model in this invention has made two key structural simplifications: First, the encoder for unmeasurable points is removed because the actual temperature data of unmeasurable points cannot be obtained during the inference phase, and this encoder has no actual input source during deployment. Second, the Long Short-Term Memory (LSTM) network is removed because its time-series iterative computation structure introduces inference latency and dependence on historical states, which is not conducive to deployment on edge computing devices. After removing the time-series network, the student model only needs the current measurable point temperature and operating parameters for inference, eliminating the need for time-step iterations. A single forward propagation can complete the prediction, reducing computational complexity and inference latency.

[0073] The student model is trained using the trained teacher model through knowledge distillation, enabling the student model to learn the mapping relationship from measurable temperature and operating parameters to unmeasurable temperature, and the trained student model is used as a global temperature field distribution model.

[0074] In this invention, the teacher model, through a complete dual encoder and temporal network structure, has fully learned the spatial distribution characteristics and temporal evolution of the global temperature field during the training phase. This knowledge is embedded in the output of the teacher model. By having the student model fit the output of the teacher model during training, it is equivalent to transferring the mapping ability learned by the teacher model from locally measurable to globally unmeasurable to the more streamlined student model. Compared to directly training the student model using simulation data, knowledge distillation allows the student model to indirectly utilize the temporal context information extracted from the temporal network in the teacher model even after removing the temporal network. This maintains high prediction accuracy while offering the advantages of lightweight and low-latency deployment.

[0075] In the actual deployment and inference phase, the global temperature field distribution model directly obtains the temperature prediction data for each unmeasurable temperature point based on the temperature data of the measurable temperature points collected in real time by physical sensors and the current operating parameters.

[0076] The student model eliminates the temporal inference network, eliminating the need for historical state information during inference. Prediction can be completed with a single forward computation, reducing inference time per iteration. Furthermore, since the teacher model's training incorporates the residuals of the heat conduction equation as physical constraints into the loss function, the student model indirectly inherits this physical prior through knowledge distillation, maintaining reasonable prediction results even under fluctuating operating conditions. Thus, this step enables real-time and accurate temperature sensing of unmeasurable critical areas such as the core of the forging during vacuum isothermal forging.

[0077] Furthermore, the loss function of the teacher model consists of three parts: a prediction loss term, an unmeasurable point reconstruction loss term, and a residual loss term. The loss function is as follows: ; in, To predict the loss term, For the reconstruction loss of unmeasurable points, For residual loss items, , and All of these are hyperparameters.

[0078] Among them, the prediction loss term provides the main supervision signal, driving the model to learn the mapping relationship from input to output; the unmeasurable point reconstruction loss term strengthens the encoder's ability to represent the spatial features of the temperature field; and the residual loss term embeds the physical laws of heat transfer as soft constraints into the training process, making up for the deficiency of the generalization ability of the pure data-driven model.

[0079] The prediction loss term is the mean square error between the temperature prediction data of each unmeasurable temperature point output by the decoder and the temperature simulation data at the corresponding time in the temperature field finite element simulation dataset. This is the main supervisory signal for the teacher model training, used to directly measure the deviation between the model's predicted values ​​and the simulation true values, driving the model to learn the end-to-end mapping relationship from measurable point temperatures and operating parameters to unmeasurable point temperatures on a large number of simulation samples. The prediction loss term is used to constrain the unmeasurable point encoder to fully learn the spatial distribution characteristics of the global temperature data.

[0080] Specifically, the latent feature vector of the unmeasurable point output by the unmeasurable point encoder is directly input into the decoder to obtain the temperature data reconstructed from the unmeasurable point at the current moment. Then, the mean square error between the reconstructed temperature data of the unmeasurable point at the current moment and the corresponding unmeasurable point temperature simulation data in the temperature field finite element simulation dataset is calculated, and this mean square error is used as the unmeasurable point reconstruction loss term. The design principle of this loss term is similar to the reconstruction constraint of an autoencoder: the unmeasurable point encoder needs to compress the high-dimensional temperature field data of the unmeasurable point into low-dimensional latent features. If the latent features can be accurately reconstructed back to the original temperature field by the decoder, it means that the encoder has effectively extracted the key features of the spatial distribution of the temperature field; conversely, if the reconstruction error is large, it indicates that the encoder has lost important spatial information.

[0081] By constraining the loss term, this invention enables the unmeasurable point encoder to fully learn the complete spatial morphology of the global temperature field during the training phase. This makes the global latent feature vector formed by splicing the two encoders more expressive and faithful, providing a high-quality feature foundation for the knowledge distillation of the subsequent student model.

[0082] The unmeasurable point reconstruction loss term is the reconstruction error of the unmeasurable point encoder on the temperature data of the unmeasurable points at the current time during the training phase. The unmeasurable point reconstruction loss term is used to constrain the unmeasurable point encoder to fully learn the spatial distribution characteristics of the global temperature data. Specifically, the unmeasurable point latent feature vector output by the unmeasurable point encoder is directly input into the decoder to obtain the temperature data reconstructed by the unmeasurable points at the current time. Then, the mean square error between the temperature data reconstructed by the unmeasurable points at the current time and the corresponding unmeasurable point temperature simulation data in the temperature field finite element simulation dataset is calculated, and the mean square error is used as the unmeasurable point reconstruction loss term. The residual loss term is the residual of the heat conduction control equation in the current global temperature data reconstructed by the decoder; ; In the formula, ρ is the material density, cp is the specific heat capacity at constant pressure of the material, T is the temperature, t is the time, k is the thermal conductivity of the material, and Q is the intensity of the internal heat source.

[0083] If the global temperature field predicted by the model is physically reasonable, then after substituting it into the heat conduction governing equation, the left and right sides of the equation should be balanced, and the residual should be close to zero. Conversely, if the prediction result violates the basic physical laws of heat conduction, the residual will increase significantly. By introducing this physical constraint into the loss function, it is equivalent to applying a regularization term based on physical laws to the model training, so that even in operating conditions where the training data coverage is insufficient, the model's prediction results still follow the physical laws of heat transfer, effectively overcoming the defect of poor generalization ability of pure data-driven models under sparse data or unseen operating conditions.

[0084] In the residual loss term, the partial derivative of temperature with respect to time, ∂T / ∂t, is calculated using the first-order forward difference approximation between the current temperature data and the previous temperature data, and the second-order derivative of temperature with respect to space... The temperature data output by the decoder is discretely calculated using a central difference scheme. During training, each training batch contains temperature data from multiple consecutive time steps to support the discrete calculation of the partial derivatives mentioned above.

[0085] In this invention, data is input in batches at discrete time steps during neural network training, making it impossible to directly obtain the analytical form of partial derivatives. By employing first-order forward difference and central difference schemes, the partial derivatives can be approximated by relying solely on the temperature values ​​of adjacent time steps and adjacent spatial grid nodes within the current batch, without needing to call additional finite element solvers. This allows for seamless integration with the standard neural network training process, ensuring training efficiency while effectively embedding physical constraints.

[0086] In some specific implementations, at any grid point, at time t n The temperature-time partial derivative is approximately: ; In the formula, T n For the grid at the current time t n The temperature value is obtained from the gridded temperature field data of the current moment output by the decoder, T n-1 For the grid at the previous time t n-1 The temperature value is the model output temperature field of the previous control cycle, and Δt is the time step.

[0087] ; In the formula, A x Let A be the heat flux divergence component in the x-direction. y Let A be the heat flux divergence component in the y-direction. z For the heat flux divergence component in the z-direction; ; In the formula, For grid The thermal conductivity of the left interface, For grid temperature, For grid temperature, For grid The thermal conductivity of the right interface, For grid temperature, For grid and grid The equivalent thermal conductivity of the interface between the two is calculated using the harmonic mean. The grid spacing is in the X direction. The y and z directions are completely symmetrical to the X direction, and will not be discussed further here.

[0088] Furthermore, the loss function of the student model consists of two parts: a distillation loss term and a prediction loss term. The distillation loss term is the mean square error between the temperature prediction data of each unmeasurable temperature point output by the student model and the temperature prediction data of each unmeasurable temperature point output by the teacher model.

[0089] The distillation loss term allows the student model to fit the output distribution of the teacher model rather than directly fitting the simulation truth. During the training phase, the teacher model learns the complete spatial characteristics of the global temperature field through an encoder for unmeasurable points and integrates the historical temporal information of the temperature field evolution through a long short-term memory network. This knowledge is already contained in the output of the teacher model. By using the distillation loss term to make the output of the student model approximate the output of the teacher model, it is equivalent to indirectly transferring the temporal context information extracted by the temporal network in the teacher model to the student model. This allows the student model to learn the evolution law of the temperature field in the time dimension even without a temporal inference structure, thus compensating for the accuracy loss caused by removing the temporal network.

[0090] The prediction loss term is the mean square error between the temperature prediction data of each unmeasurable temperature point output by the student model and the temperature simulation data at the corresponding time in the temperature field finite element simulation dataset.

[0091] The prediction loss term serves as an auxiliary supervision signal, directly constraining the student model output to avoid deviating from the simulation truth. Together with the distillation loss term, it enables the student model to maintain the accuracy of fitting the real physical data while inheriting the knowledge of the teacher model.

[0092] In summary, the teacher model, through triple constraints of prediction loss, reconstruction loss for unmeasurable points, and residual loss, fully learns the spatial distribution characteristics, temporal evolution, and heat transfer physics of the temperature field during the training phase. The student model, through dual constraints of distillation loss and prediction loss, effectively inherits the knowledge of the teacher model while maintaining a simplified structure. The design of the two-stage loss functions jointly ensures that the final deployed global temperature field distribution model possesses the comprehensive performance advantages of high accuracy, strong generalization ability, and low inference latency.

[0093] The measurable temperature data collected in real time by physical sensors is used as the real-time measurable temperature data. The unmeasurable temperature data predicted by the global temperature field distribution model based on the real-time measurable temperature data is used as the real-time unmeasurable temperature data. The real-time measurable temperature data and the real-time unmeasurable temperature data are combined to form the real-time global temperature field of the vacuum forging furnace.

[0094] Physical sensors (usually thermocouples arranged at the four positions of the furnace wall and on the mold support plate) collect temperature data of each measurable point in real time according to a preset sampling frequency. At the same time, they obtain the current operating parameters (heating power, vacuum degree, press load, and press displacement) from the equipment control system. After preprocessing, these data are input into the global temperature field distribution model. The model calculates through a single forward propagation and outputs the predicted temperature data of each unmeasurable point. Finally, the measured temperature of the measurable point and the predicted temperature of the unmeasurable point are fused in a unified grid or point cloud data structure to form a complete real-time global temperature field.

[0095] The third step involves selecting global temperature field data from multiple consecutive moments under different operating conditions of the vacuum forging furnace to form a temperature field evolution dataset. Based on this dataset, a virtual temperature field acquisition model is trained to predict the global temperature field data for the next moment from the global temperature field data of the previous moment.

[0096] The virtual temperature field acquisition model adopts a time-series prediction model, taking the global temperature field data of the previous moment as input and the global temperature field data of the next moment as output. The mean square error between the predicted global temperature field data at the next time step output by the time series prediction model and the true value of the global temperature field data at the next time step in the temperature field evolution dataset is used as the training loss of the time series prediction model.

[0097] Under different working conditions, a global temperature field distribution model is used to continuously generate multiple segments of global temperature field time series covering each forging stage, forming a temperature field evolution dataset. The forging stages include heating, holding, forging, and cooling.

[0098] After completing the real-time reconstruction of the global temperature field data, the core task of this step is to endow the digital twin with the ability to autonomously extrapolate, that is, to train a time-series prediction model that can learn the dynamic evolution law of the temperature field, so that it can predict the temperature field distribution at the next moment based solely on the temperature field state at the previous moment without the input of physical sensors, thereby realizing the continuous autonomous generation of the virtual temperature field.

[0099] To train the time-series prediction model, a dataset containing continuous time-series temperature field changes is first required. This approach utilizes the global temperature field distribution model deployed in the second step to continuously generate multiple segments of global temperature field time series covering each forging stage under different operating conditions, forming a temperature field evolution dataset. The forging stage includes four typical stages: heating, holding, forging, and cooling. Each stage has a different dominant heat transfer mechanism: the heating stage is dominated by radiation from the heating element; the holding stage tends towards thermal equilibrium; the forging stage introduces heat from plastic deformation; and the cooling stage is dominated by heat dissipation from the furnace wall. Time-series data covering these four stages enables the time-series prediction model to learn the complete dynamic characteristics of temperature field evolution under different heat transfer mechanisms.

[0100] In selecting operating conditions, the strategy is consistent with that used to generate the full-domain temperature simulation dataset. Orthogonal experimental design or Latin hypercube sampling methods are employed to systematically select representative combinations of operating conditions within the process parameter space, covering variations in multiple dimensions such as different heating power curves, compressor speed, and vacuum levels. The resulting temperature field evolution dataset not only includes the continuous evolution of the temperature field under a single operating condition but also the dynamic response of the temperature field when switching between operating conditions, providing rich training samples for time-series prediction models.

[0101] The virtual temperature field acquisition model adopts a time-series prediction model architecture. Its input is the global temperature field data of the previous time step, and its output is the global temperature field data of the next time step. This autoregressive input-output design allows the model to use its own output as the input for the next round during the inference phase, forming a closed-loop rolling inference. This is the key to the digital twin's ability to operate autonomously without external input.

[0102] In the model training phase of this invention, the mean squared error (MSE) between the predicted global temperature field at the next time step output by the time-series prediction model and the corresponding actual global temperature field data at the next time step in the temperature field evolution dataset is used as the training loss. This loss function directly measures the model's fitting accuracy to the time-series evolution of the temperature field, driving the model to learn the mapping relationship from historical states to future states. The role of the MSE as a loss function is to apply a higher penalty weight to larger deviations, thereby enabling the model to prioritize prediction accuracy during critical periods such as the forging stage with large and drastic temperature gradients and the initial cooling stage.

[0103] This invention trains a virtual temperature field acquisition model by learning the dynamic evolution of the temperature field from continuous time-series data generated by a global temperature field distribution model. This model possesses the ability to predict the future, unlike static mapping models such as the global temperature field distribution model which passively respond to input. Instead, it actively predicts the subsequent evolution of the temperature field. Once this model is embedded in a digital twin, the digital twin possesses a core engine capable of autonomous operation even without external input, laying the foundation for the subsequent autonomous prediction in the fourth step and the intermittent registration in the fifth step.

[0104] The fourth step is to embed the virtual temperature field acquisition model into the digital twin of the vacuum forging furnace, so that it can autonomously deduce and generate a virtual temperature field without real-time sensor data.

[0105] Methods for autonomously extrapolating and generating virtual temperature fields using digital twin models include: At the initial moment, the temperature data of measurable points is collected using physical sensors, and the global temperature field data is reconstructed using the temperature data of measurable points. The global temperature field data at the initial moment is synchronously transmitted to the digital twin model. The digital twin model uses the virtual temperature field acquisition model to predict the virtual temperature field at the next moment based on the global temperature field data at the initial moment. The virtual temperature field at the next moment is used as the new input of the virtual temperature field acquisition model to predict the virtual temperature field at the next moment after that. This process is repeated until the preset trigger condition occurs, so as to realize the autonomous and continuous deduction of the virtual temperature field under the condition of real-time input without physical sensors.

[0106] After training the virtual temperature field acquisition model, this invention embeds it into a digital twin model built on a vacuum forging furnace, making it the virtual engine of the digital twin. This drives the digital twin model to autonomously and continuously generate a virtual temperature field under real-time input without physical sensors, thereby reducing dependence on physical sensors.

[0107] Traditional temperature field monitoring methods based on data assimilation require physical sensors to provide new observation data in each control cycle to correct the model state. Once the sensor signal is interrupted, the data assimilation process is forced to terminate due to the lack of observation input, and the system immediately loses its ability to update the temperature field. The autonomous extrapolation mechanism proposed in this invention allows physical sensors to be needed only at the initial moment of extrapolation and during subsequent intermittent calibration. During the long operation phase between two calibrations, the digital twin can autonomously advance entirely based on its own time-series prediction capabilities.

[0108] Initially, the digital twin does not start from scratch. Instead, it uses temperature data from measurable points collected by physical sensors to reconstruct the global temperature field using a global temperature field distribution model. This global temperature field data serves as the initial state for the simulation. This initial state is generated driven by real measurement data, ensuring the accuracy and reliability of the simulation's starting point. After synchronously transmitting the initial global temperature field data to the digital twin model, the model invokes its built-in virtual temperature field acquisition model to predict the virtual temperature field at the next moment based on the initial global temperature field data.

[0109] Afterward, it enters a closed-loop autonomous simulation mode: the virtual temperature field predicted for the next moment is used as the new input for the virtual temperature field acquisition model, and then the virtual temperature field for the next moment after that is predicted. This process is repeated iteratively to form a self-sufficient rolling chain of prediction, feedback, and re-prediction. This iterative simulation process continues until the preset trigger condition (such as the forging stage switching moment) occurs, without the need for any physical sensor signals.

[0110] The virtual temperature field acquisition model has learned the dynamic evolution of the temperature field at different forging stages during the training phase. These patterns are objective manifestations of heat transfer physics in specific equipment and possess a certain degree of determinism. As long as the initial state is accurate and the operating parameters (heating power, vacuum degree, etc.) are executed according to the preset process curve, the model can make reasonable predictions about future states based on these patterns. Physically speaking, the temperature field is a continuous field constrained by the heat conduction equation, and its temporal evolution will not undergo arbitrary jumps, which provides a predictable physical basis for the time series model.

[0111] Through a working mode combining initial state injection and autonomous rolling simulation, the digital twin model has, for the first time, achieved offline operation capability independent of physical sensors. Even when sensors fail due to high temperatures, signal interference, or active hibernation, the system can still maintain continuous output of the virtual temperature field, providing operators with uninterrupted global temperature information reference and improving the robustness and survivability of the monitoring system under extreme vacuum forging conditions. Simultaneously, this autonomous simulation mechanism also provides the operational basis for the intermittent spatial registration in the fifth step. Because the digital twin can autonomously simulate between registrations, the physical sensors can shift from a necessary role of continuous operation to an auxiliary role of periodic calibration.

[0112] The fifth step is to use the measurable temperature data as a local anchor point at the time of the occurrence of the preset triggering condition, and construct a joint constraint including the first deviation constraint and the second deviation constraint using the measurable temperature data at the time of occurrence and the time before the occurrence, and solve a global deformation field acting on the virtual temperature field. The virtual temperature field output by the virtual temperature field acquisition model in the digital twin model is spatially registered using the global deformation field. The spatially registered global virtual temperature field is then used as the calibrated global virtual temperature field. Based on the calibrated global virtual temperature field, the virtual temperature field acquisition model is updated and optimized. Among them, the first deviation constraint is used to ensure that the virtual temperature field after registration is consistent with the temperature data of the measurable point at the current time at the measurable point location; The second deviation constraint is used to ensure that the temporal evolution trend of the virtual temperature field after registration is consistent with the temporal evolution trend of the measurable temperature data.

[0113] The preset trigger condition is the switching of the forging stage in the vacuum forging furnace.

[0114] Spatial registration methods for virtual temperature fields include: The system acquires the temperature data of the measurable point collected by the physical sensor at the moment of occurrence of the preset trigger condition, and records it as the true value of the measurable point at the current moment. It also acquires the temperature data of the measurable point collected by the physical sensor at the moment before the occurrence, and records it as the true value of the measurable point at the previous moment. The virtual temperature field output by the virtual temperature field acquisition model at the time of occurrence is obtained and denoted as the virtual temperature field at the previous time. The virtual temperature field output by the virtual temperature field acquisition model at the time of occurrence is also obtained and denoted as the virtual temperature field to be registered at the current time. This virtual temperature field to be registered is a direct product of the model's autonomous inference and has not yet been calibrated and corrected, and contains accumulated errors. Using the measurable point location as the anchor point, a first deviation constraint is constructed between the virtual temperature field to be registered at the current time and the actual value of the measurable point at the current time, and a second deviation constraint is constructed between the change of the virtual temperature field at the current time relative to the virtual temperature field at the previous time and the change of the actual value of the measurable point at the current time relative to the actual value of the measurable point at the previous time. The first deviation constraint is used to ensure that the virtual temperature field after registration is consistent with the actual measured value at the measurable point location, and the second deviation constraint is used to ensure that the time evolution trend of the virtual temperature field after registration is consistent with the evolution trend of the actual temperature field. The first and second deviation constraints are used as joint constraints for deformation field estimation. A global deformation field acting on the virtual temperature field to be registered at the current time is solved, so that the virtual temperature field at the current time after global deformation field correction satisfies the first and second deviation constraints at the measurable point. The global deformation field obtained by the solution is used to correct the positions of all unmeasurable points in the virtual temperature field to be registered at the current moment, so as to complete the global spatial registration of the virtual temperature field and obtain the calibrated global virtual temperature field.

[0115] Although the autonomous simulation mechanism built in the fourth step enables the digital twin to operate independently of physical sensors, any time-series prediction model inevitably accumulates errors during long-term rolling simulations. These errors primarily originate from: residuals in the model's learning of temperature field evolution; subtle fluctuations in operating conditions not covered by simulation data during actual forging; and the gradual amplification of single-step prediction errors in closed-loop iterations as the number of simulation steps increases. Therefore, it is necessary to introduce real measurement data from physical sensors at appropriate times to calibrate the virtual temperature field, thereby "pulling back" accumulated deviations and ensuring the long-term accuracy of the digital twin.

[0116] This invention sets the preset trigger condition as the switching of the forging stage in a vacuum forging furnace. The complete process of vacuum isothermal forging can be divided into four stages: heating, holding, forging, and cooling. Each stage has a clear process node (such as forging starting after holding ends, cooling starting after forging ends, etc.). The selection of the forging stage switching moment as the trigger condition is based on the following advantages: First, the stage switching moment is a natural breakpoint in process control. At this time, the system state changes significantly, requiring high accuracy in the temperature field, and the marginal value of registration is maximized. Second, the stage switching moment has a clear identifiable marker in the process sequence, and can be accurately triggered without additional anomaly detection mechanisms. Third, the duration of the forging stage is usually on the order of several minutes to tens of minutes. Within this interval, the cumulative error of the time series prediction model is still within a controllable range, and registration can effectively prevent the continuous amplification of errors.

[0117] When the preset trigger condition occurs, the system performs spatial registration. The core task of registration is to globally correct the regions in the virtual temperature field that have deviated from the real state due to accumulated errors, based on the actual measurement data at the measurable points.

[0118] First, the temperature data of the measurable point collected by the physical sensor at the moment the preset trigger condition occurs is acquired and recorded as the true value of the measurable point at the current moment. Simultaneously, the temperature data of the measurable point collected by the physical sensor at the moment before this occurrence is acquired and recorded as the true value of the measurable point at the previous moment. These two sets of actual measurement data not only provide static spatial information at the current moment but also contain dynamic information about the temperature field changes between adjacent moments, providing a data foundation for the subsequent construction of dual constraints.

[0119] Secondly, virtual temperature field data for the corresponding time moment is obtained from the digital twin model: the virtual temperature field output by the virtual temperature field acquisition model at the time of occurrence is obtained and denoted as the virtual temperature field at the previous time moment; the virtual temperature field output by the virtual temperature field acquisition model at the time of occurrence is obtained and denoted as the virtual temperature field to be registered at the current time moment. This virtual temperature field is a direct product of the model's autonomous inference and has not yet undergone calibration correction, thus containing accumulated errors.

[0120] Next, using the measurable point location as anchor points, two deviation constraints are constructed. The first deviation constraint is the deviation between the virtual temperature field to be registered at the current moment and the actual value of the measurable point at the current moment. This constraint ensures that the virtual temperature field after registration is consistent with the actual measured value at the measurable point location, addressing the spatial accuracy problem. The second deviation constraint is the deviation between the change in the virtual temperature field at the current moment relative to the virtual temperature field at the previous moment and the change in the actual value of the measurable point at the current moment relative to the actual value of the measurable point at the previous moment. This constraint ensures that the temporal evolution trend of the virtual temperature field after registration is consistent with the evolution trend of the actual temperature field, addressing the temporal consistency problem.

[0121] The introduction of a second bias constraint is a key feature of the spatial registration method in this invention. Existing technologies, when calibrating sensor data and model predictions, typically use only the current measurement value as a single constraint for correction. While this method achieves alignment at the measurable point location at the current moment, the corrected temperature field may exhibit discontinuities in the time dimension compared to historical states, and its evolution trend may not match the actual physical process, leading to physically unreasonable initial states in subsequent inferences. The second bias constraint, through forced registration, ensures that the virtual temperature field not only conforms to the current state but also to the changing trend, allowing the correction result to be better embedded in the continuous temporal evolution, consistent with the physical nature of the temperature field as a spatiotemporal continuum.

[0122] Compared to a single calibration method that only uses the first deviation constraint, the dual-constraint deformation field estimation adopted in this invention can effectively avoid problems such as discontinuity with historical states and distortion of evolution trends in the time dimension of the registered temperature field. This improves the physical rationality of the initial state of subsequent inferences, slows down the rate of error accumulation, and ensures the stability of the virtual temperature field during long-term autonomous inferences.

[0123] By using the first and second bias constraints as joint constraints for deformation field estimation, a global deformation field acting on the virtual temperature field to be registered at the current moment is solved. The deformation field is a vector field defined on the global space, describing the mapping relationship from the original value to the corrected value at each spatial location. At measurable points, the deformation field is strictly constrained by the first bias constraint, aligning the corrected temperature value with the measured value. In the region near the measurable point, the deformation field is guided by the second bias constraint, smoothly propagating in a direction consistent with the evolution trend of the true temperature field. In regions far from the measurable point, the deformation field is reasonably extrapolated based on the spatial continuity of the temperature field. By solving this deformation field, the local true information at the measurable point is effectively diffused to the entire global space, completing the temperature correction for all unmeasurable points. Finally, the solved global deformation field is used to correct all unmeasurable points in the virtual temperature field to be registered at the current moment, resulting in the calibrated global virtual temperature field.

[0124] It should be noted that although the number of nodes in a finite element mesh can reach thousands to tens of thousands, the temperature concerns that truly require precise control in the process are concentrated in a limited number of critical areas such as the core of the forging, the surface of the die cavity, and the flash groove. The remaining numerous nodes are non-critical areas with gentle temperature gradients and contribute little to the forging quality assessment. The temperature field itself possesses significant spatial continuity and smoothness. Seven measurable points, acting as spatial anchors, are sufficient to reasonably propagate local corrections to all critical nodes under double bias constraints using existing radial basis function interpolation, thin-plate splines, or regularization optimization methods based on finite element shape functions. During the solution process, minimizing the bending energy or gradient of the deformation field is the objective. Using the double bias constraints at the measurable points as hard constraints or high-weight soft constraints, a unique and smooth global deformation field can be stably obtained, achieving reliable extrapolation from sparse anchors to a limited number of critical concerns.

[0125] In some specific implementations... Step 1: Assume the virtual temperature field to be registered at the current moment is defined in the finite element mesh node set {x}. j |j=1,2,...,N}, where N is the total number of grid nodes. The global deformation field d(x) is defined as the temperature correction at each node, i.e., the temperature T after registration. corrected (x)=T pred (x)+d(x), where T pred(x) represents the virtual temperature field to be registered.

[0126] The deformation field is represented by radial basis function expansion: d(x)=∑ i w i φ(‖xx i ||)+p(x), where {x i |i = 1, 2, ..., M} represents the locations of M measurable points (M = 7), φ(r) is the radial basis function, and the Gaussian function φ(r) = exp(-r) is chosen. 2 / σ 2 ) or thin plate spline φ(r)=r 2 log(r), w i The weighting coefficients are to be determined, and p(x) is a low-order polynomial term to ensure the overall stability of the deformation field.

[0127] The second step is to construct a dual-bias constraint: obtain the true value T of the measurable point at the current time. real (t) and the true value T of the measurable point at the previous time. real (t-1), and the virtual temperature field T at the previous moment. pred (t-1) and the virtual temperature field T to be registered at the current time pred (t).

[0128] First deviation constraint: After registration, the temperature at measurable points must match the true value, i.e., T pred (x i )+d(x i ) = T real (x i ,t), i=1,2,...,M; Second deviation constraint: The time change of the registered temperature field is consistent with the actual change, i.e., [T] pred (x i ,t)+d(x i ,t)]-T pred (x i ,t-1)=T real (x i ,t)-T real (x i ,t-1), i=1,2,...,M; The virtual temperature field T at the previous moment pred (t-1) has been calibrated in the previous registration period and is considered an accurate value, so no correction is needed.

[0129] The third step is to construct the optimization objective function: transforming the double constraint into a regularized least squares problem. minJ(w)=λ1∑ i [T] pred (x i)+d(x i )-T real (x i ,t)] 2 +λ2∑ i [(T] pred (x i ,t)+d(x i ,t)-T pred (x i ,t-1))-(T real (x i ,t)-T real (x i ,t-1))] 2 +λ reg ·R(d); Where λ1 is the weight of the first deviation constraint, λ2 is the weight of the second deviation constraint, and λ reg For regularization weights. The regularization term R(d) is selected from the bending energy or gradient norm of the deformation field: or ; The regularization term constrains the deformation field to change smoothly in space, avoiding overfitting and non-physical temperature jumps, while reasonably propagating local correction information at measurable points to unmeasurable regions.

[0130] Step 4: Substitute the radial basis function expansion of the deformation field into the objective function to obtain the weighting coefficient w. i The least squares problem with polynomial coefficients. For the regularization term, a finite difference or finite element discrete approximation is used at the grid nodes.

[0131] This least squares problem can be written in standard matrix form: min‖Aw-b‖ 2 +λ reg w T Kw, where A is the constraint matrix, b is the deviation vector, and K is the regularization matrix. Its analytical solution is: w* = (A T A+λ reg K) -1 A T b. Since there are only 7 measurable points, the constraint matrix is ​​14 × |w| (14 equations in total due to double constraints). The system of equations is small and can be solved directly. Weights λ1, λ2, λ... reg Determined by cross-validation or L-curve method.

[0132] Step 5: After obtaining the weight coefficient w*, calculate x for each unmeasurable point in the entire domain. j Deformation field correction d(x) at the location j The calibrated global temperature field is obtained as follows: T corrected (x j) = T pred (x j )+d(x j ), j = 1, 2, ..., N.

[0133] This process propagates the dual deviation constraint information at the seven measurable points to all finite element mesh nodes through spatial smoothing interpolation of radial basis functions, thus completing the spatial registration from sparse anchor points to the global temperature field.

[0134] Methods for updating and optimizing the virtual temperature field acquisition model based on the calibrated global virtual temperature field include: Replace the virtual temperature field to be registered at the current moment that was originally output by the virtual temperature field acquisition model at the moment of occurrence with the calibrated global virtual temperature field. The calibrated global virtual temperature field is used as the input to the virtual temperature field acquisition model at the next moment, so that the virtual temperature field acquisition model can continue to extrapolate the virtual temperature field at subsequent moments from the calibrated global virtual temperature field, thereby eliminating the accumulated errors during the autonomous extrapolation process.

[0135] After spatial registration, the virtual temperature field acquisition model needs to be updated and optimized based on the calibrated global virtual temperature field to eliminate accumulated errors and allow subsequent simulations to start from the calibrated high-precision state. The calibrated global virtual temperature field replaces the virtual temperature field to be registered at the current moment, which was originally output by the virtual temperature field acquisition model at the time of occurrence. The calibrated global virtual temperature field serves as the starting state input for the next round of simulations, allowing the model to continue simulating the virtual temperature field at subsequent moments from the calibrated state. This operation is equivalent to resetting the state of the rolling simulation chain, clearing the errors accumulated before the registration moment, and allowing subsequent simulations to restart from a high-precision starting point corrected by real data, effectively suppressing the continuous accumulation of errors across stages.

[0136] like Figure 2 As shown, a virtual temperature acquisition system for vacuum forging based on digital twins is disclosed. The system includes: Physical sensors are placed at measurable points inside the vacuum forging furnace to collect temperature data at those points in real time. The global temperature field reconstruction module is used to reconstruct the temperature data of unmeasurable points based on the temperature data of measurable points. The global temperature field data of the vacuum forging furnace is composed of the temperature data of measurable points and the temperature data of unmeasurable points. The dataset construction module is used to select global temperature field data from multiple consecutive moments under different operating conditions of the vacuum forging furnace to form a temperature field evolution dataset. The model training module is used to train a virtual temperature field acquisition model based on the temperature field evolution dataset. The virtual temperature field acquisition model is used to predict the global temperature field data at the next moment based on the global temperature field data at the previous moment. The digital twin simulation module is used to embed the virtual temperature field acquisition model into the digital twin of the vacuum forging furnace, enabling it to autonomously simulate and generate a virtual temperature field without real-time sensor data. The spatial registration and update module is used to spatially register the virtual temperature field output by the virtual temperature field acquisition model in the digital twin model with the measurable temperature data as the local anchor point when the preset trigger condition occurs, so as to obtain the calibrated global virtual temperature field, and update and optimize the virtual temperature field acquisition model based on the calibrated global virtual temperature field.

[0137] This invention constructs a virtual temperature field acquisition model that learns the temporal evolution law of the temperature field and embeds it in a digital twin as a self-updating engine. This enables the digital twin model to autonomously deduce the virtual temperature field, thereby improving the independent operation capability of the temperature field monitoring system, which is free from the dependence on physical sensors for continuous online operation.

[0138] This invention employs an intermittent spatial registration mechanism under preset trigger conditions, which uses physical sensor data to globally calibrate the virtual temperature field only at specific moments such as the switching of the forging stage. This transforms the physical sensor from a necessary input for continuously driving the system to an auxiliary reference for periodically calibrating the system's accuracy, thereby reducing the cumulative error of autonomous inference.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. All such modifications or substitutions should be covered within the protection scope of this application, and should not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A digital-twin-based global temperature virtual acquisition method for vacuum forging, characterized in that, Includes the following steps: Points in the vacuum forging furnace where physical sensors can be installed are marked as measurable points, and points where physical sensors cannot be installed are marked as non-measurable points. Temperature data is collected in real time using physical sensors located at measurable points to obtain temperature data for measurable points. Based on the measurable temperature data, the unmeasurable temperature data is reconstructed, and the global temperature field data of the vacuum forging furnace is composed of the measurable temperature data and the unmeasurable temperature data. Under different operating conditions of the vacuum forging furnace, global temperature field data from multiple consecutive moments are selected to form a temperature field evolution dataset. Based on the temperature field evolution dataset, a virtual temperature field acquisition model is trained to predict the global temperature field data of the next moment based on the global temperature field data of the previous moment. The virtual temperature field acquisition model is embedded into the digital twin of the vacuum forging furnace, enabling it to autonomously deduce and generate a virtual temperature field without real-time sensor data. At the moment when the preset triggering condition occurs, the measurable temperature data is used as a local anchor point. The measurable temperature data at the moment of occurrence and the time before the moment of occurrence are used to construct a joint constraint including the first deviation constraint and the second deviation constraint, and solve a global deformation field acting on the virtual temperature field. The virtual temperature field output by the virtual temperature field acquisition model in the digital twin model is spatially registered using the global deformation field, and the spatially registered global virtual temperature field is used as the calibrated global virtual temperature field. The virtual temperature field acquisition model is then updated and optimized based on the calibrated global virtual temperature field. The first deviation constraint is used to ensure that the virtual temperature field after registration is consistent with the temperature data of the measurable point at the current time at the measurable point location. The second deviation constraint is used to ensure that the time evolution trend of the virtual temperature field after registration is consistent with the time evolution trend of the temperature data of measurable points.

2. The method for virtual acquisition of full-domain temperature in vacuum forging based on digital twin as described in claim 1, characterized in that, The virtual temperature field acquisition model adopts a time-series prediction model, taking the global temperature field data of the previous moment as input and the global temperature field data of the next moment as output. The mean square error between the predicted global temperature field data at the next time step output by the time series prediction model and the true value of the global temperature field data at the next time step in the temperature field evolution dataset is used as the training loss of the time series prediction model.

3. The method for virtual acquisition of full-domain temperature in vacuum forging based on digital twin according to claim 2, characterized in that, Under different working conditions, the global temperature field distribution model is used to continuously generate multiple segments of global temperature field time series covering each forging stage, forming the temperature field evolution dataset. The forging stage includes heating, holding, forging, and cooling.

4. The method for virtual acquisition of full-domain temperature in vacuum forging based on digital twin as described in claim 3, characterized in that, Methods for digital twin models to autonomously extrapolate and generate virtual temperature fields include: At the initial moment, the temperature data of measurable points is collected using physical sensors, and the global temperature field data is reconstructed using the temperature data of measurable points. The global temperature field data at the initial moment is synchronously transmitted to the digital twin model. The digital twin model uses the virtual temperature field acquisition model to predict the virtual temperature field at the next moment based on the global temperature field data at the initial moment. The virtual temperature field at the next moment is used as the new input of the virtual temperature field acquisition model to predict the virtual temperature field at the next moment after that. This process is repeated until the preset trigger condition occurs, so as to realize the autonomous and continuous deduction of the virtual temperature field under the condition of real-time input without physical sensors.

5. The method for virtual acquisition of full-domain temperature in vacuum forging based on digital twin according to claim 4, characterized in that, The preset trigger condition is the switching of the forging stage in the vacuum forging furnace.

6. The method for virtual acquisition of full-domain temperature in vacuum forging based on digital twin as described in claim 5, characterized in that, The spatial registration method for the virtual temperature field includes: The temperature data of the measurable point collected by the physical sensor at the time of occurrence of the preset trigger condition is obtained and recorded as the true value of the measurable point at the current time. The temperature data of the measurable point collected by the physical sensor at the previous time of occurrence is also obtained and recorded as the true value of the measurable point at the previous time. The virtual temperature field output by the virtual temperature field acquisition model at the previous moment before the occurrence time is obtained and denoted as the virtual temperature field at the previous moment. The virtual temperature field output by the virtual temperature field acquisition model at the occurrence time is also obtained and denoted as the virtual temperature field to be registered at the current moment. Using the measurable point location as the anchor point, a first deviation constraint is constructed between the virtual temperature field to be registered at the current time and the actual value of the measurable point at the current time, and a second deviation constraint is constructed between the change of the virtual temperature field at the current time relative to the virtual temperature field at the previous time and the change of the actual value of the measurable point at the current time relative to the actual value of the measurable point at the previous time. The first deviation constraint is used to constrain the virtual temperature field after registration to be consistent with the actual measured value at the measurable point location, and the second deviation constraint is used to constrain the time evolution trend of the virtual temperature field after registration to be consistent with the evolution trend of the actual temperature field. The first deviation constraint and the second deviation constraint are used as joint constraints for deformation field estimation. A global deformation field acting on the virtual temperature field to be registered at the current time is solved, so that the virtual temperature field at the current time after being corrected by the global deformation field satisfies the first deviation constraint and the second deviation constraint at the measurable point. The positions of all unmeasurable points in the virtual temperature field to be registered at the current moment are corrected using the global deformation field obtained by the solution, and the global spatial registration of the virtual temperature field is completed, thus obtaining the calibrated global virtual temperature field.

7. The method for virtual acquisition of full-domain temperature in vacuum forging based on digital twin as described in claim 6, characterized in that, The method for updating and optimizing the virtual temperature field acquisition model based on the calibrated global virtual temperature field includes: Replace the virtual temperature field to be registered at the current time that was originally output by the virtual temperature acquisition model at the time of occurrence with the calibrated global virtual temperature field. The calibrated global virtual temperature field is used as the input to the virtual temperature field acquisition model at the next moment, so that the virtual temperature field acquisition model can continue to extrapolate the virtual temperature field at subsequent moments from the calibrated global virtual temperature field, thereby eliminating the accumulated errors during the autonomous extrapolation process.

8. The method for virtual acquisition of full-domain temperature in vacuum forging based on digital twin according to claim 7, characterized in that, Using a pre-established global temperature field distribution model for predicting unmeasurable temperature data based on measurable temperature data, the measurable temperature data and the unmeasurable temperature data output by the global temperature field distribution model are combined to form the global temperature field data of the vacuum forging furnace.

9. A virtual temperature acquisition system for vacuum forging based on digital twins, characterized in that, The system, applicable to the virtual acquisition method for full-domain temperature acquisition in vacuum forging based on digital twins as described in any one of claims 1-8, comprises: Physical sensors are placed at measurable points inside the vacuum forging furnace to collect temperature data at those points in real time. The global temperature field reconstruction module is used to reconstruct the non-measurable temperature data based on the measurable temperature data, and the global temperature field data of the vacuum forging furnace is composed of the measurable temperature data and the non-measurable temperature data. The dataset construction module is used to select global temperature field data from multiple consecutive moments under different operating conditions of the vacuum forging furnace to form a temperature field evolution dataset. The model training module is used to train a virtual temperature field acquisition model based on the temperature field evolution dataset. The virtual temperature field acquisition model is used to predict the global temperature field data at the next moment based on the global temperature field data at the previous moment. The digital twin simulation module is used to embed the virtual temperature field acquisition model into the digital twin of the vacuum forging furnace, enabling it to autonomously simulate and generate a virtual temperature field without real-time sensor data. The spatial registration and update module is used to construct a joint constraint including a first deviation constraint and a second deviation constraint using the measurable temperature data as a local anchor point at the time of occurrence of a preset trigger condition, and to solve a global deformation field acting on the virtual temperature field. The module then uses the global deformation field to spatially register the virtual temperature field output by the virtual temperature field acquisition model in the digital twin model, and uses the spatially registered global virtual temperature field as the calibrated global virtual temperature field. Based on the calibrated global virtual temperature field, the module updates and optimizes the virtual temperature field acquisition model.

Images