Workpiece cooling method and system based on medium space-time field twinning and generation type control

Abstract

The invention discloses a workpiece cooling method and system based on medium space-time field twinning and generative control, and belongs to the technical field of intelligent manufacturing and metallurgical hot working. The method comprises the steps of collecting physical field data of a cooling medium in real time, performing data fusion and space-time deduction through a fluid dynamic model, and constructing four-dimensional space-time field digital twinning of the cooling medium; geometric data and material attribute data of the workpiece are obtained, four-dimensional space-time field digital twinning is used as a boundary condition, the multi-physics field coupling model is driven to carry out calculation, and internal state evolution information of the workpiece in the cooling process is predicted; by taking preset workpiece target microstructure distribution and / or target residual stress distribution as an optimization target, taking internal state evolution information and four-dimensional space-time field digital twinning as state input, performing online decision making through a deep reinforcement learning model, generating an optimal time sequence control strategy and analyzing the optimal time sequence control strategy into an equipment control instruction, and controlling the equipment according to the optimal time sequence control strategy. And a multi-zone independent execution unit in the active cooling equipment is driven.

Description

Technical Field

[0001] This application relates to the fields of intelligent manufacturing and metallurgical hot processing technology, specifically to a workpiece cooling method and system based on medium spatiotemporal field twinning and generative control. Background Technology

[0002] Heat treatment, especially quenching and cooling, is the core process that determines the final mechanical properties (such as hardness, strength, toughness, and fatigue life) of metal workpieces. When a workpiece is rapidly cooled from a high-temperature austenitic state, a phase transformation occurs, forming the desired microstructure such as martensite. At the same time, residual stress is inevitably generated, leading to deformation.

[0003] However, existing heat treatment cooling process control suffers from fundamental technical bottlenecks: (1) Blind spots in medium field perception: Traditional cooling tanks only place thermocouples at a few points such as the inlet and outlet, which cannot obtain the true and non-uniform three-dimensional spatiotemporal distribution field of the cooling medium (such as water, oil, quenching agent) in the tank. This results in different cooling processes for different parts of the workpiece (such as thick walls and thin walls, front and back surfaces), which is the main reason for uneven cooling, excessive deformation, and even quenching cracks. This reflects the current technology's lack of accurate spatiotemporal understanding of the physical state of the cooling medium.

[0004] (2) Lagging state control: The optimization targets of existing processes are mostly limited to macroscopic performance indicators (such as surface hardness), while the microstructure (type, proportion, morphology and distribution) and residual stress field (such as surface compressive stress which is crucial to fatigue life) that truly determine the service performance of the workpiece are difficult to be directly controlled because they cannot be predicted and quantified online, resulting in a passive situation of "macroscopic inspection and post-event remedy". This reveals that the current technology lacks the ability to effectively predict the evolution of the microstructure and stress state inside the workpiece.

[0005] (3) Rigid process path: Current control relies heavily on static "process cards" and operator experience, lacking dynamic adjustment capabilities. The control system can only switch between limited preset processes (such as selecting tank A or tank B), and cannot generate and execute dynamic, personalized adaptive cooling paths based on information such as real-time water temperature fluctuations and specific geometric features of the workpiece. This reflects the rigidity of the existing process control paradigm, which cannot achieve flexible intelligent control for individual workpieces.

[0006] Therefore, existing heat treatment cooling technologies lack spatiotemporal understanding of the physical state of the cooling medium, lack predictive ability of the microscopic mechanism inside the workpiece, and have rigid process control paradigms, which are problems that this application urgently needs to solve. Summary of the Invention

[0007] The embodiments of this disclosure provide a workpiece cooling method and system based on medium spatiotemporal field twinning and generative control, so as to at least solve the technical problems of existing heat treatment cooling technologies that lack spatiotemporal dimension understanding of the physical state of the cooling medium, lack predictive ability of the microscopic mechanism inside the workpiece, and rigid process control paradigm.

[0008] According to one aspect of the present disclosure, a workpiece cooling method based on media spatiotemporal field twinning and generative control is provided, comprising: real-time acquisition of physical field data of a cooling medium in three-dimensional space; wherein the physical field data includes a temperature field, a flow velocity field, and a concentration field; based on the physical field data, data fusion and spatiotemporal extrapolation are performed through a pre-trained fluid dynamics model to construct a four-dimensional spatiotemporal field digital twin of the cooling medium; wherein the four-dimensional spatiotemporal field digital twin is a digital model used to characterize the temperature field, flow velocity field, and concentration field evolving in three-dimensional space and time dimensions of the cooling medium; acquiring geometric data and material property data of the workpiece to be cooled, and driving a multiphysics coupling model to perform calculations using the dynamic and non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal field digital twin. The system calculates and predicts the internal state evolution information of the workpiece during the cooling process. The multiphysics coupling model is a model that couples heat conduction, phase transformation dynamics, and stress-strain relationships. The internal state evolution information includes microstructure type and distribution, residual stress field, and deformation trend. With the goal of achieving a preset target microstructure distribution and / or target residual stress distribution, the internal state evolution information and the four-dimensional spatiotemporal field digital twin are used as state inputs. A deep reinforcement learning model is used for online decision-making to generate an optimal timing control strategy for dynamically regulating the active cooling equipment in real time. The optimal timing control strategy is then parsed into specific equipment control commands to drive multi-zone independent execution units in the active cooling equipment to perform spatiotemporally differentiated dynamic cooling operations on the workpiece.

[0009] According to another aspect of the present disclosure, a storage medium is also provided, the storage medium including a stored program, wherein the methods described above are executed by a processor when the program is running.

[0010] According to another aspect of the present disclosure, a workpiece cooling system based on media spatiotemporal field twinning and generative control is also provided, comprising: a spatiotemporal field sensing layer, a multiphysics prediction layer, a generative control layer, and an active execution layer; the spatiotemporal field sensing layer is configured with a distributed sensor array and a pre-trained fluid dynamics model, used to collect physical field data of the cooling medium in three-dimensional space in real time, and based on the physical field data, to perform data fusion and spatiotemporal extrapolation through the fluid dynamics model to construct a four-dimensional spatiotemporal field digital twin of the cooling medium; wherein, the physical field data includes a temperature field, a flow velocity field, and a concentration field; the four-dimensional spatiotemporal field digital twin is a digital model used to characterize the temperature field, flow velocity field, and concentration field evolving in three-dimensional space and time dimensions of the cooling medium; the multiphysics prediction layer is connected to the spatiotemporal field sensing layer, used to acquire the geometric data and material property data of the workpiece to be cooled, and to use the dynamic, non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal field digital twin. The system drives a multiphysics coupling model to calculate and predict the internal state evolution information of the workpiece during the cooling process. The multiphysics coupling model is a model that couples heat conduction, phase transformation dynamics, and stress-strain relationships. The internal state evolution information includes microstructure type and distribution, residual stress field, and deformation trend. A generative control layer, connected to the multiphysics prediction layer, is used to optimize the workpiece by achieving a preset target microstructure distribution and / or target residual stress distribution. It takes the internal state evolution information and the four-dimensional spatiotemporal field digital twin as state inputs, and performs online decision-making through a deep reinforcement learning model to generate an optimal timing control strategy for dynamically regulating the active cooling equipment in real time. An active execution layer, connected to the generative control layer and equipped with active cooling equipment, is used to parse the optimal timing control strategy into specific equipment control commands, driving the multi-zone independent execution units in the active cooling equipment to perform spatiotemporally differentiated dynamic cooling operations on the workpiece.

[0011] This application first uses a distributed sensor array to collect real-time physical field data of the cooling medium in three-dimensional space, including temperature, flow velocity, and concentration fields, providing a data foundation for subsequent processing. Then, based on the physical field data, a four-dimensional spatiotemporal digital twin of the cooling medium is constructed using a pre-trained fluid dynamics model. This achieves a leap from "point-like perception" to "full-field cognition" of the cooling medium's physical state, providing a high-dimensional spatiotemporal data foundation for precise control. Secondly, by combining the dynamic, non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal digital twin with the geometric and material property data of the workpiece, a multi-physics coupled model, which couples heat conduction, phase transformation dynamics, and stress-strain relationships, is driven to perform calculations. This enables accurate and quantitative prediction of the evolution of the workpiece's internal microstructure, residual stress distribution, and deformation trends, making the previously invisible "black box" process transparent and computable. Subsequently, targeting the preset microstructure and residual stress distribution, the predicted internal state evolution information and the digital twin of the media's spatiotemporal field are jointly input into a deep reinforcement learning model for online decision-making. This achieves a fundamental shift from "selective" control relying on static process cards to "generative" control that generates optimal timing strategies in real time based on the target. This dynamically generates personalized optimal cooling paths for workpieces with different geometric features and different water immersion states, greatly improving the flexibility and intelligence of the process. Finally, the generated optimal timing control strategy is parsed into specific instructions to drive multi-zone independent execution units in the active cooling equipment, ensuring that the workpiece can withstand spatiotemporally differentiated dynamic cooling operations and actively guide the internal state of the workpiece to the expected target. Therefore, this application, through the deep integration of media spatiotemporal field digital twins, multiphysics field coupled prediction, and generative control, forms a closed loop of "perception-twin-prediction-generation-execution," and decentralizes the control target from macroscopic hardness to microscopic structure and residual stress distribution. This enables refined performance customization such as "surface high-pressure stress" or "core toughness structure," allowing the heat treatment cooling process to leap from "passive experience-based control" to "active precise creation." It establishes a continuous optimization capability driven by both data and mechanism, achieving technical effects such as transparent cooling process mechanism, predictable internal microscopic state of the workpiece, and personalized process control. Furthermore, it solves the technical problems of existing heat treatment cooling technologies lacking spatiotemporal understanding of the physical state of the cooling medium, lacking predictive ability of the internal microscopic mechanism of the workpiece, and having a rigid process control paradigm. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this application, illustrate exemplary embodiments of this disclosure and are used to explain this disclosure, but do not constitute an undue limitation of this disclosure. In the drawings: Figure 1 This is a flowchart of the workpiece cooling method based on medium spatiotemporal field twinning and generative control as described in the embodiments of this application; Figure 2 This is a schematic diagram of the workpiece cooling system based on medium spatiotemporal field twinning and generative control as described in the embodiments of this application. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0014] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] Example According to this embodiment, a workpiece cooling method based on media spatiotemporal field twinning and generative control is provided, implemented in a corresponding application scenario. This application scenario includes a workpiece to be cooled, an active cooling device, and a workpiece cooling system that executes the cooling method. The workpiece is hoisted by a robotic arm and precisely positioned above the active cooling device. Specifically, the active cooling device is a multi-zone spray array, where each nozzle is independently controllable, capable of receiving commands and individually adjusting its media flow rate, pressure, and on / off state. The workpiece cooling system establishes a real-time communication connection with the active cooling device via a control line for issuing control commands. Simultaneously, a distributed sensor array (including but not limited to fiber Bragg grating sensors, infrared thermal imagers, etc.) is deployed inside the active cooling device and in the space surrounding the workpiece to collect multi-dimensional physical field data of the cooling medium and the workpiece in real time. The collected data is continuously transmitted to the workpiece cooling system as input for spatiotemporal field sensing, multi-physics prediction, and generative control.

[0016] Under the aforementioned operating environment, according to the first aspect of this embodiment, a workpiece cooling method based on medium spatiotemporal field twinning and generative control is provided, with reference to... Figure 1 As shown, the method includes: S101: Real-time acquisition of physical field data of the cooling medium in three-dimensional space; wherein, the physical field data includes temperature field, flow velocity field and concentration field; S102: Based on the physical field data, data fusion and spatiotemporal extrapolation are performed through a pre-trained fluid dynamics model to construct a four-dimensional spatiotemporal digital twin of the cooling medium; wherein, the four-dimensional spatiotemporal digital twin is a digital model used to characterize the temperature field, flow velocity field and concentration field that evolve in the three-dimensional space and time dimensions of the cooling medium. S103: Acquire the geometric data and material property data of the workpiece to be cooled, and drive the multiphysics coupling model to perform calculations using the dynamic and non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal digital twin, predicting the internal state evolution information of the workpiece during the cooling process; wherein, the multiphysics coupling model is a model that couples heat conduction, phase transformation dynamics and stress-strain relationship; the internal state evolution information includes microstructure type and distribution, residual stress field and deformation trend; S104: Taking the preset target microstructure distribution and / or target residual stress distribution of the workpiece as the optimization target, the internal state evolution information and the four-dimensional spatiotemporal field digital twin are used as state inputs, and online decision-making is carried out through a deep reinforcement learning model to generate the optimal timing control strategy for dynamic regulation of the active cooling equipment in real time. S105: The optimal timing control strategy is parsed into specific equipment control instructions, which drive the multi-zone independent execution unit in the active cooling device to perform spatiotemporally differentiated dynamic cooling operations on the workpiece.

[0017] In this embodiment of the invention, a distributed sensor array is used to collect physical field data of the cooling medium in three-dimensional space in real time (corresponding to step S101). Specifically, the distributed sensor array can be a heterogeneous sensor network composed of a fiber optic temperature sensor array, an ultrasonic Doppler flow meter, and an electrochemical concentration sensor, which is deployed in the three-dimensional space inside the active cooling device and the space surrounding the workpiece according to a preset spatial topology. The distributed sensor array synchronously collects the three-dimensional temperature distribution, flow velocity vector field, and quenching agent concentration field of the cooling medium at different spatial locations with millisecond-level time resolution, and ensures the spatiotemporal synchronization of multi-source data through a timestamp alignment mechanism, providing a global, real-time data foundation for the subsequent construction of a high-fidelity digital twin.

[0018] Then, based on the physical field data, a four-dimensional spatiotemporal digital twin of the cooling medium is constructed through data fusion and spatiotemporal extrapolation using a pre-trained fluid dynamics model (corresponding to step S102). Specifically, the collected sparse point cloud data with spatial coordinate markers is input into a fluid dynamics model (e.g., a U-Net structural proxy model) trained based on computational fluid dynamics forward problem simulation data. The model is guided by a residual loss function constrained by physical conservation laws to perform spatiotemporal interpolation and field evolution extrapolation, outputting a complete four-dimensional data volume of the medium's temperature field, velocity field, and concentration field at a preset spatial resolution (e.g., 1 mm) and a preset temporal resolution (e.g., 0.1 s), serving as the four-dimensional spatiotemporal digital twin of the cooling medium. Thus, a leap from "point-like perception" to "full-field cognition" of the physical state of the cooling medium is achieved, providing a high-dimensional spatiotemporal data foundation for precise control.

[0019] Next, the geometric and material property data of the workpiece to be cooled are acquired, and the dynamic, non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal digital twin are used to drive the multiphysics coupling model to calculate and predict the internal state evolution information of the workpiece during the cooling process (corresponding to step S103). Specifically, by importing the STL format geometric file into the three-dimensional model of the workpiece, and combining parameters such as the CCT curve, thermal expansion coefficient, and yield strength of the material, the convective heat transfer coefficient that varies with spatial location and time provided by the four-dimensional spatiotemporal digital twin is used as the boundary condition. An explicit time integration scheme is used to solve the temperature-microstructure-stress multi-field evolution of the workpiece during the cooling process. Thus, accurate and quantitative prediction of the internal microstructure evolution, residual stress distribution, and deformation trend of the workpiece is achieved, making the originally invisible "black box" process transparent and computable.

[0020] Next, with the preset target microstructure distribution and / or target residual stress distribution of the workpiece as the optimization objective, the internal state evolution information and the four-dimensional spatiotemporal field digital twin are used as state inputs. A deep reinforcement learning model is used for online decision-making, generating in real time the optimal timing control strategy for dynamically regulating the active cooling equipment (corresponding to step S104). Specifically, the Soft Actor-Critic deep reinforcement learning algorithm is used to construct a 256-dimensional state vector from the predicted phase composition fraction, stress tensor, and key features of the medium field at each grid point inside the workpiece. The weighted negative values ​​of the microstructure distribution difference (measured by Wasserstein distance) and the stress distribution difference (measured by root mean square error) are used as the reward function. The optimal control sequence is explored in the action space (e.g., the opening ratio of each nozzle) using a strategy gradient method. This achieves a fundamental shift from "selective" control relying on static process cards to "generative" control that generates optimal timing strategies in real time based on the target. It dynamically generates personalized optimal cooling paths for workpieces with different geometric features and different water entry states, greatly improving the flexibility and intelligence of the process.

[0021] Finally, the optimal timing control strategy is parsed into specific equipment control commands, driving the multi-zone independent execution units in the active cooling equipment to perform spatiotemporally differentiated dynamic cooling operations on the workpiece (corresponding to step S105). Specifically, the normalized action vector output by the reinforcement learning model is converted into a 4-20mA analog signal through a D / A conversion module, and transmitted via PROFINET real-time industrial Ethernet to the proportional control valve and variable frequency pump corresponding to each nozzle. This achieves millisecond-level precise control of the flow rate and pressure of each independent nozzle, forming a dynamic cooling field that matches the geometric characteristics of the workpiece. Thus, it ensures that the workpiece can withstand spatiotemporally differentiated dynamic cooling operations and actively guides the internal state of the workpiece to the expected target.

[0022] As described in the background section, existing heat treatment cooling technologies lack spatiotemporal understanding of the physical state of the cooling medium, lack predictive ability of the microscopic mechanisms inside the workpiece, and have rigid process control paradigms.

[0023] In view of this, this application firstly acquires physical field data such as temperature field, flow velocity field, and concentration field of the cooling medium in three-dimensional space in real time through a distributed sensor array, providing a data foundation for subsequent processing. Then, based on the physical field data, a four-dimensional spatiotemporal digital twin of the cooling medium is constructed through a pre-trained fluid dynamics model, realizing a leap from "point-like perception" to "full-field cognition" of the physical state of the cooling medium, providing a high-dimensional spatiotemporal data foundation for precise control. Secondly, by combining the dynamic and non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal digital twin with the geometric and material property data of the workpiece, a multi-physics coupled model that couples heat conduction, phase transformation dynamics, and stress-strain relationship is driven to perform calculations, realizing accurate and quantitative prediction of the evolution of the internal microstructure, residual stress distribution, and deformation trend of the workpiece, making the originally invisible "black box" process transparent and computable. Subsequently, targeting the preset microstructure and residual stress distribution, the predicted internal state evolution information and the digital twin of the media's spatiotemporal field are jointly input into a deep reinforcement learning model for online decision-making. This achieves a fundamental shift from "selective" control relying on static process cards to "generative" control that generates optimal timing strategies in real time based on the target. This dynamically generates personalized optimal cooling paths for workpieces with different geometric features and different water immersion states, greatly improving the flexibility and intelligence of the process. Finally, the generated optimal timing control strategy is parsed into specific instructions to drive multi-zone independent execution units in the active cooling equipment, ensuring that the workpiece can withstand spatiotemporally differentiated dynamic cooling operations and actively guide the internal state of the workpiece to the expected target. Therefore, this application, through the deep integration of media spatiotemporal field digital twins, multiphysics field coupled prediction, and generative control, forms a closed loop of "perception-twin-prediction-generation-execution," and decentralizes the control target from macroscopic hardness to microscopic structure and residual stress distribution. This enables refined performance customization such as "surface high-pressure stress" or "core toughness structure," allowing the heat treatment cooling process to leap from "passive experience-based control" to "active precise creation." It establishes a continuous optimization capability driven by both data and mechanism, achieving technical effects such as transparent cooling process mechanism, predictable internal microscopic state of the workpiece, and personalized process control. Furthermore, it solves the technical problems of existing heat treatment cooling technologies lacking spatiotemporal understanding of the physical state of the cooling medium, lacking predictive ability of the internal microscopic mechanism of the workpiece, and having a rigid process control paradigm.

[0024] Optionally, based on the physical field data, the operation of constructing a four-dimensional spatiotemporal digital twin of the cooling medium through data fusion and spatiotemporal extrapolation using a pre-trained fluid dynamics model includes: inputting a spatially sparse and timestamped physical field data stream from a distributed sensor array into the pre-trained fluid dynamics model; wherein the fluid dynamics model is a neural network surrogate model for spatiotemporal field reconstruction trained based on computational fluid dynamics simulation data; performing spatiotemporal interpolation and physical field evolution extrapolation on the physical field data stream through the fluid dynamics model, outputting a medium state field that covers the entire three-dimensional space inside the active cooling device and can continuously evolve with time; integrating and encapsulating the currently output medium state field and the historically extrapolated medium state fields to form a four-dimensional spatiotemporal digital twin that can reflect the dynamics of the cooling medium in real time and support future predictions.

[0025] In this embodiment of the invention, before inputting the physical field data stream into the fluid dynamics model, the physical field data stream collected by the distributed sensor array, which includes three-dimensional spatial coordinate markers and timestamps, can be preprocessed by data cleaning and outlier removal to organize it into a spatiotemporal graph structure before being input into the fluid dynamics model. The fluid dynamics model employs a neural network architecture based on Fourier operators. This model, pre-trained on CFD simulation data covering different flow conditions, possesses the ability to learn complete flow field physical laws from sparse observation data.

[0026] Then, through the multi-scale convolution and spatiotemporal attention mechanism of the fluid dynamics model, three-dimensional spatial interpolation and time-series prediction are performed on the input sparse sensor data under physical constraints, outputting a medium state field that covers the entire three-dimensional space inside the active cooling device and can continuously evolve with time. Subsequently, the high-resolution medium state field output at the current time step is integrated with the extrapolation results from multiple historical time steps through a spatiotemporal database to construct a four-dimensional data cube containing a time-dimensional index. This data cube maintains its latest state through a rolling time window update mechanism and exposes a RESTful API interface to provide real-time query and prediction services to upstream systems, forming a true digital twin capable of reflecting the dynamics of the cooling medium in real time and supporting short-term (e.g., 5-second) predictions.

[0027] Thus, it achieves accurate reconstruction of the medium state from low spatial sampling rate to full-basin coverage, upgrading traditional single-point monitoring to four-dimensional dynamic field perception with physical consistency. It not only accurately characterizes the real-time state of the medium at any location, but also provides forward-looking decision support for process control through time-series prediction capabilities, and provides accurate dynamic boundary conditions for subsequent multi-physics coupling analysis.

[0028] Optionally, the operation of acquiring the geometric data and material property data of the workpiece to be cooled, and using the dynamic, non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal field digital twin to drive the multiphysics coupling model to calculate and predict the internal state evolution information of the workpiece during the cooling process, includes: mapping the four-dimensional spatiotemporal field digital twin onto the geometric surface of the workpiece to obtain the dynamic heat transfer coefficients at different locations on the workpiece surface at different times, constituting the dynamic, non-uniform heat transfer boundary conditions; and inputting the heat transfer boundary conditions, the geometric data and material property data of the workpiece into the multiphysics field. A coupled model is used; wherein the material property data includes at least the phase transformation kinetic curve, thermophysical parameters, and mechanical property parameters of the material; coupled calculations are performed through the multiphysics coupling model to sequentially solve the three-dimensional transient temperature field inside the workpiece, the phase transformation evolution dependent on the change of the three-dimensional transient temperature field, and the stress-strain field caused by the temperature gradient and the change of phase transformation volume; based on the solution results of the phase transformation evolution, the microstructure type and distribution inside the workpiece during the cooling process are determined; based on the solution results of the stress-strain field, the residual stress field and deformation trend inside the workpiece during the cooling process are determined.

[0029] In this embodiment of the invention, the medium temperature and flow velocity data in the four-dimensional spatiotemporal digital twin are first matched with the surface mesh of the workpiece's three-dimensional model. Through correlational calculations based on the principle of convective heat transfer, the medium state at each surface location is converted into a real-time changing heat transfer coefficient, establishing boundary conditions that dynamically update with spatial location and time. Next, these boundary conditions, along with the workpiece material's thermophysical parameters, phase transition characteristic curves, and mechanical performance parameters, are input into a multiphysics coupling model for calculation. The calculation process sequentially solves for: the temperature change history of each point inside the workpiece over time, the phase transition type and transformation variable evolution based on the temperature history, and the stress-strain development process caused by temperature inhomogeneity and phase transition volume effects. Then, based on the phase transition calculation results, the microstructure types and proportions formed at different cooling stages at various locations inside the workpiece are determined, generating a microstructure distribution map that accurately displays the spatial distribution of each phase. Finally, based on the stress field calculation results, the magnitude and distribution of residual stress after workpiece cooling are extracted, and the three-dimensional deformation and deformation trend caused by uneven cooling are calculated.

[0030] Thus, a complete prediction chain from the medium field to the internal state of the workpiece was established, realizing accurate quantitative prediction of microstructure evolution and residual stress. This transformed the invisible "black box" process in traditional heat treatment into a completely transparent and calculable process, providing key input data for intelligent control.

[0031] Optionally, the internal state evolution information and the four-dimensional spatiotemporal field digital twin are used together as state inputs. Online decision-making is performed through a deep reinforcement learning model to generate, in real time, the optimal timing control strategy for dynamically regulating the active cooling equipment. This includes: aligning and splicing the microstructure type and distribution, residual stress field, and deformation trend included in the internal state evolution information with the temperature field, flow velocity field, and concentration field represented by the four-dimensional spatiotemporal field digital twin to form a unified state vector, which serves as the input data for the deep reinforcement learning model; defining the adjustable parameters of each independent execution unit in the active cooling equipment as the deep reinforcement learning model's parameters. The learning model's action space; wherein the adjustable parameters include at least pressure, flow rate, and on / off state; a reward function is constructed based on the preset microstructure distribution of the workpiece target and / or the target residual stress distribution; the reward function is used to quantify the difference between the predicted final internal state and the target state, and guide the deep reinforcement learning model to optimize decisions in the direction of maximizing the cumulative reward; through the deep reinforcement learning model, based on the state vector at the current moment, the instantaneous action that maximizes the reward function is explored and decided in the action space, and the optimal timing control strategy is constituted by a series of instantaneous actions from the current moment to the end of cooling.

[0032] In this embodiment of the invention, multi-source state data fusion is first performed. Data representing the microstructure distribution, residual stress field, and deformation trend of the workpiece's internal state are combined with data representing the temperature field, flow velocity field, and concentration field of the external medium environment. These data are then processed using a coordinate system unification, and feature stitching forms a unified state vector with 512 dimensions. This vector fully describes the real-time state of the workpiece-medium coupling system. Next, the pressure adjustment range (0-1MPa), flow control range (0-100L / min), and on / off state of each independent nozzle in the active cooling device are defined as the basic units of the action space, forming action vectors with multiple continuous control dimensions. This ensures that each execution unit can be independently and precisely controlled.

[0033] Then, based on the preset target microstructure distribution of the workpiece (e.g., surface martensite content > 90%) and target residual stress distribution (e.g., surface compressive stress > 300 MPa), a reward function is constructed by calculating the similarity score between the predicted state and the target state. This function simultaneously considers the degree of microstructure transformation completion and the uniformity of stress distribution. Finally, the deep reinforcement learning model receives the latest state vector and searches in real time in the action space for control actions that maximize the cumulative reward. A control strategy covering the entire cooling cycle is generated through temporal action sequence planning, which specifies in detail the optimal operating parameters of each nozzle at different times.

[0034] Thus, a real-time decision-making closed loop was established from multi-source state perception to precise action generation, realizing dynamic process adjustment based on the internal state feedback of the workpiece. This broke through the limitations of traditional fixed process routes and ensured that each workpiece could obtain the optimal cooling path matching its specific state through online optimization.

[0035] Optionally, parsing the optimal timing control strategy into specific device control commands to drive the operation of the multi-zone independent execution units in the active cooling device includes: parsing the instantaneous actions corresponding to each time step in the timing control strategy; wherein each instantaneous action contains a control vector, and each dimension of the control vector corresponds to a target control parameter of an independent execution unit; converting the parsed target control parameters into underlying drive signals that can be received by the corresponding actuators in the active cooling device via an industrial fieldbus or real-time Ethernet protocol; wherein the underlying drive signals include at least analog voltage signals, pulse width modulation signals, or servo positioning commands; and synchronously sending the underlying drive signals to the multi-zone independent execution units according to the time sequence specified by the timing control strategy.

[0036] In this embodiment of the invention, control commands are first parsed, decoding the real-time actions of each time step in the timing control strategy to extract control vectors containing multiple dimensions. Each dimension corresponds to the target pressure value (0-1MPa), target flow rate value (0-100L / min), and on / off state of an independent nozzle, forming a set of control parameters recognizable by the device. Next, signal conversion and transmission are performed. The parsed target control parameters are converted into underlying drive signals receivable by the actuator via the PROFINET industrial Ethernet protocol. Pressure regulation commands are converted into 0-10V analog voltage signals, flow control commands into PWM pulse width modulation signals, and on / off states into digital signals, ensuring control accuracy of ±0.02MPa for pressure and ±1L / min for flow rate. Finally, multi-unit synchronous control is implemented. According to the millisecond-level time sequence specified by the timing strategy, drive signals are synchronously sent to all execution units via a distributed I / O module, ensuring that each nozzle can accurately perform opening / closing, pressure regulation, and flow regulation operations according to the predetermined timing sequence, achieving independent and precise control of the cooling intensity of each area.

[0037] Thus, a precise conversion channel from digital control strategy to physical execution action was established, realizing millisecond-level synchronous control of multiple execution units, ensuring the accurate reproduction of spatiotemporally differentiated cooling processes, and providing reliable execution guarantee for the precise control of the internal structure of the workpiece.

[0038] Optionally, the method further includes: after the workpiece has cooled, obtaining measured data of the actual microstructure, actual residual stress, and actual deformation of the workpiece through at least one of metallographic analysis, X-ray diffraction stress measurement, and three-dimensional scanning deformation measurement; calculating the difference between the measured data and the prediction results output by the multiphysics coupling model to generate a model error signal; and using the model error signal, iteratively updating the parameters of the multiphysics coupling model and / or the deep reinforcement learning model through a backpropagation algorithm or a gradient descent optimization algorithm.

[0039] In this embodiment of the invention, firstly, measured data are collected and processed. After the workpiece has cooled, metallographic samples are prepared from representative key areas of the workpiece (such as thick-walled areas, thin-walled areas, and stress concentration areas) using wire cutting. The samples are then mechanically ground and polished, sequentially using metallographic abrasive paper of different grits for coarse grinding to fine grinding, followed by mirror polishing with diamond polishing agent. The polished samples are then chemically etched with a 4% nitric acid alcohol solution for 10-30 seconds until the microstructure is clearly visible.

[0040] Then, tissue images were acquired at magnifications ranging from 100× to 1000× using a metallographic microscope. The acquired images were processed using a deep learning-based image segmentation algorithm: first, Gaussian filtering was used to remove noise; then, a U-Net convolutional neural network was used to perform semantic segmentation of different phases such as martensite, bainite, and ferrite in the images; finally, pixel statistics were used to calculate the area percentage of each phase, completing the quantitative analysis of the tissue.

[0041] Residual stress was scanned at preset measurement points using an X-ray diffractometer, employing sin... 2 The diffraction angle shift of the {211} crystal plane family at different ψ angles was measured using the ψ method, and the residual stress value was calculated using elasticity formulas. Point cloud data of the workpiece surface was acquired using a 3D laser scanner, and after best fitting and alignment with the original CAD model, the normal deviation of each point was calculated as the deformation.

[0042] Next, the measured microstructure distribution, residual stress data, and deformation are compared with the prediction results corresponding to the multiphysics coupling model in a gridded manner. The absolute and relative errors of each grid point are calculated, and an error signal containing the spatial error distribution is generated to identify the systematic deviation of the model in a specific temperature range or geometric structure.

[0043] Finally, based on the distribution characteristics of the error signal, the material thermal properties and phase transition kinetics parameters in the multiphysics coupling model are calibrated, and the value function network parameters in the deep reinforcement learning model are updated. The optimization process adopts a periodic batch processing mode, with a concentrated model update performed after the cooling process of a certain number (e.g., 50 pieces) of workpieces is completed, ensuring that the model continuously approximates the actual physical process.

[0044] Thus, a continuous optimization closed loop from physical measurement to model self-calibration was established. Standardized metallographic sample preparation procedures and automated image analysis algorithms ensured the accuracy and repeatability of the measured data, providing reliable experimental data support for model optimization. This measured data-driven parameter calibration effectively improved the prediction accuracy of the multiphysics model and the decision reliability of the reinforcement learning model, enabling the system to continuously learn and self-improve.

[0045] Optionally, the fluid dynamics model is a spatiotemporal graph neural network model; and the operation of performing spatiotemporal interpolation and physical field evolution deduction on the physical field data stream through the fluid dynamics model to output a medium state field covering the entire internal three-dimensional space of the active cooling device and capable of continuous evolution with time sequence includes: abstracting the physical location and connection relationship of the distributed sensor array into nodes and edges of a graph structure, and using the physical field data as the dynamic features of the nodes; using the spatiotemporal graph neural network model, based on the graph structure of the current time and historical time, performing iterative reasoning in a discrete time step manner to predict the full field medium state at the next time step; and combining the prediction results of multiple consecutive time steps in chronological order to form a medium state field covering the entire internal three-dimensional space of the active cooling device and capable of continuous evolution with time sequence.

[0046] In this embodiment of the invention, the three-dimensional spatial coordinates of each sensor in the distributed sensor array are first used as the location of graph nodes. Edge connections are established based on the spatial proximity between sensors (distance threshold set to 50mm, for example), forming a static graph topology with multiple nodes and edges. Then, the temperature, flow rate, and concentration data collected at each time step are used as dynamic features of the nodes. A spatiotemporal graph convolution module simultaneously captures spatial neighborhood relationships and temporal series patterns. This module includes spatial graph convolution layers and temporal convolution layers. Spatial convolution is implemented based on Chebyshev multinomial approximation, while temporal convolution uses one-dimensional causal convolution to ensure temporal dependence. Next, with a time step of 0.1 seconds, based on the graph structure data of the current and previous historical moments, a gated recurrent unit is used to achieve temporal memory and iteratively predict the overall medium state at future preset time steps. After each prediction, the node features are updated, and the prediction result is used as the input for the next time step, forming a closed-loop inference mechanism. Finally, the predicted node states are mapped onto a regular three-dimensional mesh using an inverse distance weighted interpolation algorithm, generating a medium state field covering the entire internal space of the device. The predicted fields of continuous time steps are integrated in chronological order to form a medium state field that covers the entire three-dimensional space inside the active cooling device and can evolve continuously with the time sequence.

[0047] Thus, it achieves accurate reconstruction and prediction of high-resolution medium field from sparse sensor data to the entire flow domain. By leveraging the unique spatial relationship modeling capabilities of graph structures, it effectively preserves the physical characteristics of the flow field, solves the interpolation problem of traditional methods at complex geometric boundaries, and provides high-quality boundary condition input for multiphysics coupling analysis.

[0048] Optionally, the method further includes: while driving the active cooling device to execute the device control command, acquiring the actual temperature distribution of the workpiece surface in real time using an infrared thermal imager or thermocouple array deployed on the workpiece surface; comparing the actual temperature distribution with the workpiece surface temperature field in the three-dimensional transient temperature field calculated by the multiphysics coupling model in real time, and determining that the cooling process is abnormal when the deviation between the two exceeds a preset threshold; based on the deviation, triggering online replanning of the deep reinforcement learning model, or switching to a preset safe cooling strategy, to dynamically intervene in the execution parameters of the active cooling device.

[0049] In this embodiment of the invention, the steps of real-time monitoring and dynamic intervention include: First, deploying an infrared thermal imager array and a thermocouple array in key areas of the workpiece surface to synchronously collect temperature distribution data of the workpiece surface at a high sampling frequency, and ensuring the accuracy and coverage integrity of temperature monitoring through data fusion processing. Then, comparing the spatial distribution consistency of the real-time collected workpiece surface temperature distribution with the surface temperature field calculated by the multiphysics coupling model, and calculating the temperature deviation statistics for each region. When the temperature deviation exceeds a preset threshold for multiple consecutive time steps, it is determined that an anomaly has occurred in the cooling process, and a diagnostic signal containing the location of the abnormal area and the deviation level is generated. Finally, corresponding control adjustments are initiated according to the anomaly level: for minor anomalies, online replanning of the deep reinforcement learning model is triggered, and the control strategy is locally optimized based on real-time temperature feedback; for severe anomalies, the system immediately switches to a preset safe cooling strategy, adopting more conservative cooling parameters to ensure process safety, and achieving seamless switching of the strategy through a rapid control loop refresh mechanism.

[0050] Thus, a rapid response closed loop from real-time monitoring to intelligent decision-making was established. Through multi-source temperature monitoring and online diagnosis, abnormal detection and graded handling of the cooling process were realized, effectively preventing workpiece quality problems caused by sudden changes in medium state or model prediction deviations. This significantly improved the system's reliability and adaptability, providing important safety assurance for intelligent cooling control.

[0051] In summary, the beneficial effects of the embodiments of this application are as follows: 1. Achieve mechanism transparency: Through 4D spatiotemporal twinning and multiphysics prediction, the "black box" quenching process is transformed into a "white box" computable and predictable physical process.

[0052] 2. Achieve microscopic control: For the first time, the control target is shifted from macroscopic hardness to microscopic structure and residual stress distribution, enabling refined performance customization such as "surface high pressure stress" or "core toughness structure".

[0053] 3. Achieving process flexibility: Through generative control, the system is freed from dependence on static process cards and can dynamically generate personalized optimal cooling paths for workpieces with different geometric features and different water immersion states, greatly improving the flexibility and robustness of the process.

[0054] It should be noted that the training processes of the fluid dynamics model, multiphysics coupling model, and deep reinforcement learning model described in this application are all implemented using mature training methods commonly used in the relevant technical fields. For example, by acquiring a labeled sample set composed of computational fluid dynamics simulation or historical experimental data, a gradient descent-based backpropagation algorithm is used to supervise the learning of model parameters, and the model accuracy is optimized by minimizing the loss function (such as mean square error, physical information residual, etc.). Model training is a fundamental preparatory step for achieving its expected function, and the specific training algorithm and hyperparameter selection are not improvements in this application. The core of this invention lies in integrating the pre-trained model into the closed-loop system of "perception-twin-prediction-generation-execution" in a specific way, and solving the specific technical problem of workpiece cooling through their synergistic effect.

[0055] According to a second aspect of this embodiment, a workpiece cooling system 200 based on media spatiotemporal field twinning and generative control is provided, comprising: a spatiotemporal field sensing layer 210, a multiphysics prediction layer 220, a generative control layer 230, and an active execution layer 240; the spatiotemporal field sensing layer 210 is configured with a distributed sensor array and a pre-trained fluid dynamics model, used to collect physical field data of the cooling medium in three-dimensional space in real time, and based on the physical field data, to perform data fusion and spatiotemporal extrapolation through the fluid dynamics model to construct a four-dimensional spatiotemporal field digital twin of the cooling medium; wherein, the physical field data includes a temperature field, a flow velocity field, and a concentration field; the four-dimensional spatiotemporal field digital twin is a digital model used to characterize the temperature field, flow velocity field, and concentration field evolving in three-dimensional space and time dimensions of the cooling medium; the multiphysics prediction layer 220 is connected to the spatiotemporal field sensing layer 210, used to acquire the geometric data and material property data of the workpiece to be cooled, and to utilize the dynamic, non-uniform heat transfer provided by the four-dimensional spatiotemporal field digital twin. Boundary conditions drive a multiphysics coupling model to calculate and predict the internal state evolution information of the workpiece during the cooling process. The multiphysics coupling model couples heat conduction, phase transformation dynamics, and stress-strain relationships. The internal state evolution information includes microstructure type and distribution, residual stress field, and deformation trend. The generative control layer 230, connected to the multiphysics prediction layer 220, is used to optimize the workpiece by achieving a preset target microstructure distribution and / or target residual stress distribution. It takes the internal state evolution information and the four-dimensional spatiotemporal digital twin as state inputs, and uses a deep reinforcement learning model for online decision-making to generate an optimal timing control strategy for dynamically regulating the active cooling equipment. The active execution layer 240, connected to the generative control layer 230, is equipped with an active cooling equipment. It parses the optimal timing control strategy into specific equipment control commands, driving the multi-zone independent execution units in the active cooling equipment to perform spatiotemporally differentiated dynamic cooling operations on the workpiece.

[0056] In this embodiment of the invention, the workpiece cooling system 200 based on media spatiotemporal field twinning and generative control includes the following main modules: a spatiotemporal field sensing layer 210, a multiphysics prediction layer 220, a generative control layer 230, and an active execution layer 240. The data flow relationship between the modules is as follows: the spatiotemporal field sensing layer 210 provides the "4D media field" to the multiphysics prediction layer 220; the multiphysics prediction layer 220 provides the "predicted structure and stress" to the generative control layer 230; and the generative control layer 230 provides the "timing control strategy" to the active execution module of the active execution layer 240. In addition, the active execution layer 240 also has a closed-loop correction module, which obtains "measured data" from the active execution layer 240 as feedback and updates the models of the multiphysics prediction layer 220 and the generative control layer 230 respectively.

[0057] The spatiotemporal field sensing layer 210 serves as the system's sensing unit. In specific implementations, a distributed sensor array can be deployed within cooling equipment (such as a cooling tank or spray box). For example, a high-temperature resistant fiber grating (FBG) array can be arranged along the X, Y, and Z dimensions to acquire the three-dimensional temperature field of the cooling medium (such as a quenching agent) in real time with a time resolution of 0.1 seconds. Simultaneously, an ultrasonic Doppler flowmeter array is used to acquire the three-dimensional flow velocity field of key sections.

[0058] These sparse, multimodal sensing data are then fed into a pre-trained fluid dynamics model (e.g., a CFD surrogate model built on U-Net or Fourier neural operators), which outputs a high-resolution four-dimensional spatiotemporal field F(x, y, z, t) covering the entire cooling region through real-time interpolation and extrapolation. This field contains temperature and velocity information at each coordinate point at each time step.

[0059] The multiphysics prediction layer 220 serves as the core of the system's analysis, receiving the workpiece's three-dimensional geometric model (such as a STEP file) and material data (such as CCT curves and thermophysical parameters). This layer constructs a "thermal-phase transition-force" multiphysics coupled prediction model, preferably implemented using a physical information neural network (PINN).

[0060] In terms of heat conduction, the boundary condition of the multiphysics coupling prediction model is no longer a simple fixed temperature, but a non-uniform dynamic 4D spatiotemporal field F(x, y, z, t) provided by the spatiotemporal field sensing layer 210, thereby accurately capturing the heat transfer differences of the workpiece surface (such as the front and back surfaces).

[0061] In phase transformation simulation, multiphysics coupled prediction models are based on temperature history and calculate the phase transformation process at each node using the JMAK equation or the Koistinen-Marburger equation, outputting the volume fraction of structures such as martensite, bainite, and austenite.

[0062] In stress and deformation prediction, the multiphysics coupled prediction model comprehensively considers thermal strain, phase transformation strain, and plastic strain to simulate the evolution of the stress field inside the workpiece and predict the final residual stress distribution and deformation. The output of this layer is a high-dimensional tensor, characterizing the internal microstate of the workpiece after cooling.

[0063] The generative control layer 230, serving as the system's decision-making core, is one of the key innovations of this invention. In specific implementations, it can be implemented using a deep reinforcement learning model (such as the PPO or SAC algorithm). The state inputs of the deep reinforcement learning model include the "microscopic state map of the workpiece's interior" output by the multiphysics prediction layer 220 and the "4D spatiotemporal field of the medium" provided by the spatiotemporal field perception layer 210. Its action output is a high-dimensional continuous control vector A. t For example, for a spray array with 16 individual nozzles, A t Given a 16-dimensional vector, A t = [P1, P2, ..., P 16 ], where P i This represents the pressure or flow rate of the i-th nozzle at time t.

[0064] The reward function R is designed based on the "target state" set by the operator (such as surface compressive stress of -300MPa, martensite content in the core greater than 90%, and total deformation less than 0.1mm). It is usually a negative value of the difference between the predicted state and the target state (such as negative Wasserstein distance or root mean square error).

[0065] The deep reinforcement learning agent learns how to adjust control actions A through extensive trial-and-error training in a simulated environment (i.e., the multiphysics prediction layer 220). t This guides the internal state of the workpiece to approach the target reward R. Ultimately, a novel and optimal timing control strategy A = [A1, A2, ..., A...] is generated for this workpiece. T ].

[0066] The active execution layer 240, acting as the system's execution unit, receives the timing control strategy A generated by the generative control layer 230 and transmits the strategy vector A through a PLC or servo controller. t The commands are converted into control instructions for physical devices in real time. For example, at t=1 second, A1 = [1.0, 0.8, 0.1, ...] is executed (i.e., nozzle 1 is fully open, nozzle 2 is 80% open, nozzle 3 is 10% open, etc.), and at t=2 seconds, A2 = [0.9, 0.9, 0.3, ...] is executed, thereby achieving differentiated, dynamic, and active cooling control for different parts of the workpiece (such as thick-walled and thin-walled parts).

[0067] In the closed-loop correction step, after the workpiece has cooled, the residual stress in key areas is measured by X-ray diffraction, or the deformation is measured by three-dimensional scanning. The obtained "real data" is compared with the "predicted data" of the multiphysics prediction layer 220. The model parameters of the multiphysics prediction layer 220 and the generative control layer 230 are iteratively updated using error signals through backpropagation and other methods, thereby endowing the system with the ability to learn and continuously optimize.

[0068] It should be noted that the workpiece cooling system based on medium spatiotemporal field twinning and generative control provided in this embodiment can realize all the functions and steps in the above method embodiments, solve the same technical problems, and achieve the same technical effects. The similarities will not be repeated here.

[0069] According to a third aspect of this embodiment, a storage medium is provided, the storage medium including a stored program, wherein the method described above is executed by a processor when the program is running.

[0070] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0071] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A workpiece cooling method based on medium spatiotemporal field twinning and generative control, characterized in that, include: Real-time acquisition of physical field data of the cooling medium in three-dimensional space; wherein, the physical field data includes temperature field, flow velocity field and concentration field; Based on the physical field data, a four-dimensional spatiotemporal digital twin of the cooling medium is constructed by performing data fusion and spatiotemporal extrapolation through a pre-trained fluid dynamics model; wherein, the four-dimensional spatiotemporal digital twin is a digital model used to characterize the temperature field, flow velocity field and concentration field of the cooling medium in three-dimensional space and time dimensions. The geometric and material property data of the workpiece to be cooled are acquired, and the dynamic and non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal digital twin are used to drive a multiphysics coupling model to calculate and predict the internal state evolution information of the workpiece during the cooling process. The multiphysics coupling model is a model that couples heat conduction, phase transformation dynamics, and stress-strain relationship. The internal state evolution information includes microstructure type and distribution, residual stress field, and deformation trend. With the goal of achieving the preset target microstructure distribution and / or target residual stress distribution of the workpiece, the internal state evolution information and the four-dimensional spatiotemporal field digital twin are used as state inputs. Online decision-making is carried out through a deep reinforcement learning model to generate the optimal timing control strategy for dynamically regulating the active cooling equipment in real time. The optimal timing control strategy is parsed into specific equipment control instructions, which drive the multi-zone independent execution units in the active cooling equipment to perform spatiotemporally differentiated dynamic cooling operations on the workpiece.

2. The method according to claim 1, characterized in that, Based on the physical field data, the operation of constructing a four-dimensional spatiotemporal digital twin of the cooling medium through data fusion and spatiotemporal extrapolation using a pre-trained fluid dynamics model includes: A spatially sparse and timestamped physical field data stream from a distributed sensor array is input into the pre-trained fluid dynamics model; wherein the fluid dynamics model is a neural network surrogate model for spatiotemporal field reconstruction trained based on computational fluid dynamics simulation data. The physical field data stream is spatiotemporally interpolated and the physical field evolution is deduced using the fluid dynamics model, resulting in a medium state field that covers the entire three-dimensional space inside the active cooling device and can continuously evolve with time. The current output medium state field and the medium state field derived from historical simulations are integrated and encapsulated to form a four-dimensional spatiotemporal digital twin that can reflect the dynamics of the cooling medium in real time and support future predictions.

3. The method according to claim 1, characterized in that, The operation of acquiring the geometric and material property data of the workpiece to be cooled, and using the dynamic and non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal digital twin to drive a multiphysics coupling model to calculate and predict the internal state evolution information of the workpiece during the cooling process, includes: The four-dimensional spatiotemporal field digital twin is mapped onto the geometric surface of the workpiece to obtain the dynamic heat transfer coefficients at different locations on the workpiece surface at different times, thus constituting the dynamic, non-uniform heat transfer boundary conditions. The heat transfer boundary conditions, the geometric data of the workpiece, and the material property data are input into the multiphysics coupling model; wherein, the material property data includes at least the phase transformation dynamics curve, thermophysical parameters, and mechanical property parameters of the material; The coupling calculation is performed using the multiphysics coupling model to sequentially solve the three-dimensional transient temperature field inside the workpiece, the phase transformation evolution dependent on the change of the three-dimensional transient temperature field, and the stress-strain field caused by the temperature gradient and the change of phase transformation volume. Based on the solution results of the phase transformation evolution, the microstructure type and distribution inside the workpiece during the cooling process are determined; based on the solution results of the stress-strain field, the residual stress field and deformation trend inside the workpiece during the cooling process are determined.

4. The method according to claim 1, characterized in that, Using the internal state evolution information and the four-dimensional spatiotemporal field digital twin as state inputs, an online decision-making process is performed through a deep reinforcement learning model to generate, in real time, the optimal timing control strategy for dynamically regulating the active cooling equipment, including: The microstructure type and distribution, residual stress field and deformation trend included in the internal state evolution information are aligned and spliced ​​with the temperature field, flow velocity field and concentration field of the four-dimensional spatiotemporal digital twin representation to form a unified state vector, which is used as the input data of the deep reinforcement learning model. The adjustable parameters of each independent execution unit in the active cooling device are defined as the action space of the deep reinforcement learning model; wherein, the adjustable parameters include at least pressure, flow rate, and on / off state; Based on the preset microstructure distribution and / or residual stress distribution of the workpiece target, a reward function is constructed; the reward function is used to quantify the difference between the predicted final internal state and the target state, and guide the deep reinforcement learning model to optimize decision-making in the direction of maximizing the cumulative reward; Using the deep reinforcement learning model, based on the state vector at the current moment, the model explores and decides on the instantaneous action in the action space that maximizes the reward function, and the series of instantaneous actions from the current moment to the end of the cooldown constitutes the optimal timing control strategy.

5. The method according to claim 4, characterized in that, The optimal timing control strategy is parsed into specific device control commands to drive the operation of the multi-zone independent execution units in the active cooling device, including: The instantaneous actions corresponding to each time step in the timing control strategy are parsed; wherein each instantaneous action contains a control vector, and each dimension of the control vector corresponds to the target control parameters of an independent execution unit. The parsed target control parameters are converted into underlying drive signals that can be received by the corresponding actuators in the active cooling device via industrial fieldbus or real-time Ethernet protocol; wherein, the underlying drive signals include at least analog voltage signals, pulse width modulation signals or servo positioning commands; According to the time sequence specified by the timing control strategy, the underlying drive signal is synchronously sent to the multi-zone independent execution unit.

6. The method according to claim 1, characterized in that, Also includes: After the workpiece has cooled down, the actual microstructure, actual residual stress and actual deformation of the workpiece are obtained by at least one of metallographic analysis, X-ray diffraction stress measurement and three-dimensional scanning deformation measurement. The difference between the measured data and the prediction results output by the multiphysics coupling model is calculated to generate a model error signal; Using the model error signal, the parameters of the multiphysics coupling model and / or the deep reinforcement learning model are iteratively updated through backpropagation or gradient descent optimization algorithms.

7. The method according to claim 2, characterized in that, The fluid dynamics model is a spatiotemporal graph neural network model; and, through the fluid dynamics model, spatiotemporal interpolation and physical field evolution deduction are performed on the physical field data stream to output a medium state field that covers the entire three-dimensional space inside the active cooling device and can continuously evolve with time, including: The physical locations and connections of the distributed sensor array are abstracted as nodes and edges in a graph structure, and the physical field data is used as the dynamic features of the nodes. Using the spatiotemporal graph neural network model, based on the graph structure of the current moment and historical moments, iterative reasoning is performed in a discrete time step manner to predict the state of the entire medium at the next moment. The prediction results of multiple consecutive time steps are combined in chronological order to form a medium state field that covers the entire three-dimensional space inside the active cooling device and can evolve continuously with the time series.

8. The method according to claim 3, characterized in that, Also includes: While driving the active cooling device to execute the device control commands, the actual temperature distribution of the workpiece surface is collected in real time by an infrared thermal imager or thermocouple array deployed on the workpiece surface. The actual temperature distribution is compared in real time with the workpiece surface temperature field in the three-dimensional transient temperature field calculated by the multiphysics coupling model. When the deviation between the two exceeds a preset threshold, it is determined that the cooling process is abnormal. Based on the deviation, online replanning of the deep reinforcement learning model is triggered, or a preset safe cooling strategy is switched to dynamically intervene in the execution parameters of the active cooling device.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, the method described in any one of claims 1 to 8 is performed by a processor.

10. A workpiece cooling system based on medium spatiotemporal field twinning and generative control, characterized in that, include: The spatiotemporal field perception layer, multiphysics prediction layer, generative control layer, and active execution layer are all included. The spatiotemporal field perception layer is configured with a distributed sensor array and a pre-trained fluid dynamics model. It is used to collect physical field data of the cooling medium in three-dimensional space in real time, and to perform data fusion and spatiotemporal extrapolation based on the physical field data through the fluid dynamics model to construct a four-dimensional spatiotemporal field digital twin of the cooling medium. The physical field data includes temperature field, flow velocity field and concentration field. The four-dimensional spatiotemporal field digital twin is a digital model used to characterize the temperature field, flow velocity field and concentration field that evolve in three-dimensional space and time. The multiphysics prediction layer, connected to the spatiotemporal field sensing layer, is used to acquire the geometric and material property data of the workpiece to be cooled. Using the dynamic, non-uniform heat transfer boundary conditions provided by the four-dimensional spatiotemporal field digital twin, it drives the multiphysics coupling model to calculate and predict the internal state evolution information of the workpiece during the cooling process. The multiphysics coupling model is a model that couples heat conduction, phase transformation dynamics, and stress-strain relationships. The internal state evolution information includes microstructure type and distribution, residual stress field, and deformation trend. The generative control layer, connected to the multiphysics prediction layer, is used to optimize the workpiece target microstructure distribution and / or target residual stress distribution. It takes the internal state evolution information and the four-dimensional spatiotemporal field digital twin as state inputs, and makes online decisions through a deep reinforcement learning model to generate the optimal timing control strategy for dynamically regulating the active cooling equipment in real time. The active execution layer, connected to the generative control layer, is equipped with an active cooling device. This device is used to parse the optimal timing control strategy into specific equipment control commands, driving the multi-zone independent execution units in the active cooling device to perform spatiotemporally differentiated dynamic cooling operations on the workpiece.

Images