Temperature control method, system and equipment for titanium alloy forging and medium

Abstract

The invention relates to the technical field of control and regulation, in particular to a temperature control method, system, equipment and medium for titanium alloy forging, and the method comprises the following steps: constructing a model library for describing thermodynamic behaviors of forgings, and dynamically generating a double-target temperature track; operating a state observer in real time to estimate the internal temperature field of the forge piece on line, and calculating a feedforward control set value based on a model predictive control framework; meanwhile, a feedback correction control quantity is generated through a feedback correction loop, and online correction is carried out on the model; finally, adaptively fusing the feedforward and feedback control quantities to generate and apply a final driving signal; by combining a physical model and a data driving method, the accurate modeling of deformation heat and the estimation precision of the core temperature are improved, metallurgy and cost constraints are balanced, it is ensured that the forge piece is formed in the optimal temperature window, and the problem that in the prior art, it is difficult to control the microstructure of the forge piece to be formed in the optimal temperature window is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of control adjustment, and particularly relates to a temperature control method, system, device and medium for titanium alloy forging. BACKGROUND

[0002] Titanium alloy is widely used in the field of aerospace and other fields with extremely high requirements for structural reliability due to its excellent specific strength and heat resistance. However, titanium alloy is sensitive to temperature, and the final mechanical properties and fatigue life thereof in the forging process almost completely depend on the microstructure formed in the forging process.

[0003] The prior art usually adopts indirect or macroscopic means in temperature control, such as measuring the furnace temperature or mold temperature of a heating furnace, or directly measuring the surface temperature of a forging, to estimate the workpiece temperature in combination with experience or offline simulation. However, in actual forging operation, especially under a large strain rate and a complex stress state, a large amount of deformation heat is generated inside the workpiece. The internal heat is instantaneous and non-uniformly distributed, which may result in difficulty in real-time acquisition and accurate modeling of the actual temperature of the deformation core region of the forging.

[0004] In summary, the control system in the prior art is difficult to control the core temperature of the forging in real time and accurately, and thus it is difficult to determine whether the core temperature has exceeded or fallen below the critical temperature window required for the formation of microstructure. SUMMARY

[0005] The main purpose of the present application is to provide a temperature control method for titanium alloy forging, which aims to solve the problem that the temperature control method in the prior art is difficult to control the formation of microstructure of the forging within the optimal temperature window.

[0006] To achieve the above-mentioned purpose, the present application provides a temperature control method for titanium alloy forging, which comprises the following steps: acquiring the thermal physical parameters of the forging, and constructing a model library according to the thermal physical parameters, wherein the model library is used to describe the thermodynamic behavior of the forging in the heating and forging process; acquiring the alloy constraint and cost constraint of the forging, and acquiring a double-target temperature control trajectory according to the alloy constraint and cost constraint; acquiring a temperature state correction gain matrix, and acquiring the distribution estimation of the temperature field of the forging according to the temperature state correction gain matrix and the model library; constructing a coupled prediction model according to the distribution estimation of the temperature field of the forging and the target temperature control trajectory, and acquiring a feedforward control set value according to the coupled prediction model; acquiring the measured surface temperature of the forging and the predicted value of the coupled prediction model, acquiring a feedback correction control amount according to the measured surface temperature of the forging, the optimal control sequence and the predicted value of the coupled prediction model, and modifying the coupled prediction model through the feedback correction control amount. According to the feedforward control set value and the feedback correction control amount, an actual control signal is acquired.

[0007] To achieve the above object, the application further provides a control system, which comprises: An offline modeling module is configured to acquire thermal physical parameters of the forging, and to construct a model library according to the thermal physical parameters, wherein the model library is used to describe the thermodynamic behavior of the forging during the heating and forging processes. A trajectory generation module is configured to acquire alloy constraints and cost constraints of the forging, and to acquire a double-target temperature control trajectory according to the alloy constraints and the cost constraints. A state estimation module is configured to acquire a temperature state correction gain matrix, and to acquire distribution estimation of the temperature field of the forging according to the temperature state correction gain matrix and the model library. A feedforward control module is configured to construct a coupled prediction model according to the distribution estimation of the temperature field of the forging and the target temperature control trajectory, and to acquire a feedforward control set value according to the coupled prediction model. A feedback control module is configured to acquire a measured surface temperature of the forging and a predicted value of the coupled prediction model, to acquire a feedback correction control amount according to the measured surface temperature of the forging, the optimal control sequence and the predicted value of the coupled prediction model, and to modify the coupled prediction model through the feedback correction control amount. A temperature control module is configured to acquire an actual control signal according to the feedforward control set value and the feedback correction control amount.

[0008] To achieve the above object, the application further provides a computer device, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to realize the above method.

[0009] To achieve the above object, the application further provides a computer readable storage medium, which stores a computer program, and a processor executes the computer program to realize the above method.

[0010] The application can achieve the following beneficial effects: The present application improves the temperature field prediction accuracy by combining a physical model and a data-driven method, especially accurate modeling of deformation heat under large strain rate conditions; the generation mechanism of the double-target temperature control trajectory balances the metallurgical constraints and cost constraints, ensuring that the forgings are formed within the optimal temperature window; the feedforward-feedback composite control architecture improves the adaptability of the system to process parameter fluctuations and reduces the deviation of core temperature and surface temperature trajectory tracking; the state estimation method based on double correction improves the estimation accuracy of the core temperature, providing a reliable basis for microstructure control; and further solves the problem that the temperature control method in the prior art is difficult to control the microstructure of the forgings to form within the optimal temperature window. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the specific embodiments or the prior art, the drawings needed to be used in the specific embodiments or the prior art description will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale.

[0012] Figure 1 A flowchart of the control method in Embodiment 1 of the present application is shown. Figure 2 A structure block diagram of the control system in Embodiment 10 of the present application is shown.

[0013] The implementation of the present application, functional features and advantages will be further described with reference to the drawings. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0015] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between the components in a certain posture, and if the certain posture changes, the directional indications also change accordingly.

[0016] In the present application, unless otherwise explicitly specified and limited, the terms "connection", "fixation" and the like shall be understood in a broad sense, for example, "connection" can be fixed connection, or detachable connection, or integral; can be mechanical connection, or electrical connection; can be direct connection, or indirect connection through an intermediate medium, can be internal communication of two elements or interaction relationship between two elements, unless otherwise explicitly limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0017] In addition, if the present application involves "first", "second" and the like in the description, the "first", "second" and the like are only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can be explicitly or implicitly included at least one of the features. In addition, the meaning of "and / or" appearing throughout the text includes three parallel schemes. For example, "A and / or B" includes A scheme, or B scheme, or A and B scheme. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of ordinary skilled in the art, when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, nor within the scope of protection required by the present application.

[0018] Embodiment 1 Referring to Figure 1 The present embodiment provides a temperature control method for titanium alloy forging, the control method comprising the following steps: Obtaining the thermal physical parameters of the forging, and constructing a model library according to the thermal physical parameters, the model library being used to describe the thermodynamic behavior of the forging during heating and forging. It can be understood that this is step 1; Obtaining the alloy constraint and cost constraint of the forging, and obtaining a double-target temperature control trajectory according to the alloy constraint and cost constraint. It can be understood that this is step 2; Obtaining a temperature state correction gain matrix, and obtaining a distribution estimate of the forging temperature field according to the temperature state correction gain matrix and the model library. It can be understood that this is step 3; Constructing a coupled prediction model according to the distribution estimate of the forging temperature field and the target temperature control trajectory, and obtaining a feedforward control set value according to the coupled prediction model. It can be understood that this is step 4; Obtaining the measured temperature of the forging surface and the predicted value of the coupled prediction model, obtaining a feedback correction control amount according to the measured temperature of the forging surface, the optimal control sequence and the predicted value of the coupled prediction model, and modifying the coupled prediction model through the feedback correction control amount. It can be understood that this is step 5; According to the feedforward control set value and the feedback correction control amount, an actual control signal is obtained; it can be understood that this is step 6.

[0019] In the conventional existing titanium alloy forging temperature control process, due to the time-varying and spatial heterogeneity of the deformation heat source term, the conventional modeling method cannot effectively identify the internal heat source dynamic characteristics, resulting in systematic deviation between the predicted temperature field value and the actual working condition, which is particularly significant when the forging process parameters change dramatically, directly affecting the synchronization of core temperature and surface temperature trajectory tracking.

[0020] Due to the above problems, non-uniform recrystallization structure will be formed inside the forgings, and abnormal grain growth will occur in local areas. Such microstructure defects cannot be eliminated in the subsequent heat treatment process, resulting in a decrease of two orders of magnitude in the fatigue crack initiation life. More seriously, the temperature control error may cause the imbalance of the competition mechanism between dynamic recrystallization and static recrystallization in the forging process, causing the mechanical property dispersion of batch products to exceed the product certification standard, and thus directly affecting the safe use performance of the products.

[0021] Based on the above problems, the embodiment explores the feasibility of combining multi-objective optimization and predictive control. Due to the conflict between metallurgical constraints and cost constraints, conventional single-objective optimization cannot balance the two. By establishing a double-target temperature control trajectory generation mechanism, discrete trajectory points that meet the constraints are searched in the simulation environment, and then a continuous trajectory is generated through curve fitting. However, there is still a dynamic error between the simulation model and the actual system, so a feedforward-feedback composite control architecture needs to be designed, in which the feedforward control generates set values based on the prediction model, and the feedback control dynamically corrects the model error according to the measured surface temperature.

[0022] It should be noted that the embodiment realizes the coordinated and accurate regulation of the core temperature and the surface temperature of the forgings by constructing a model library that integrates physical models and data-driven residual compensation, generating a temperature control trajectory based on double-target optimization, and adopting a feedforward-feedback composite control architecture, thereby ensuring that the microstructure is formed within the target temperature window.

[0023] It should also be noted that the embodiment first obtains the thermal physical parameters of the forgings, including thermal conductivity, specific heat capacity, and density, etc. According to these parameters, a model library is constructed to describe the thermodynamic behavior of the forgings during heating and forging, including temperature distribution, heat conduction, etc.

[0024] Next, the alloy constraints and cost constraints of the forgings are obtained, the alloy constraints include temperature window, thermal stress gradient, and soaking time requirements, and the cost constraints consider energy consumption, production efficiency, etc.; according to these constraint conditions, a double-target temperature control trajectory is generated through an optimization algorithm, including the target variation curves of core temperature and surface temperature.

[0025] Then the temperature state correction gain matrix is obtained, and the temperature field distribution of the forging is estimated in combination with the model library. This step improves the accuracy of temperature field estimation by fusing model prediction and measured data.

[0026] According to the temperature field distribution estimation and the target temperature control trajectory, a coupled prediction model is constructed, which considers the influence of temperature, strain and other factors on microstructure evolution. By solving the optimization problem, the feedforward control set value is obtained.

[0027] In the actual control process, the measured temperature of the forging surface is obtained, and the predicted value of the coupled prediction model is compared. According to the deviation, the feedback correction control amount is calculated, and the coupled prediction model is dynamically modified to adapt to the actual working condition changes.

[0028] Finally, the feedforward control set value and the feedback correction control amount are fused to generate the actual control signal. This feedforward-feedback composite control strategy can balance the foresight of model prediction and the robustness of real-time correction.

[0029] Through the above scheme, the embodiment realizes accurate control of the core temperature of the titanium alloy forging. Specifically, by combining physical models and data-driven methods, the temperature field prediction accuracy is improved, especially under large strain rate conditions. The generation mechanism of the dual-target temperature control trajectory balances the metallurgical constraints and cost constraints, ensuring that the forging is formed within the optimal temperature window. The feedforward-feedback composite control architecture improves the system's adaptability to process parameter fluctuations and reduces the deviation between the core temperature and the surface temperature trajectory tracking. The state estimation method based on double correction improves the estimation accuracy of the core temperature, providing a reliable basis for microstructure control. The comprehensive application of these technical means effectively solves the problem of real-time and accurate control of the core temperature of the forging in the prior art, laying a foundation for obtaining uniform and stable microstructure.

[0030] Embodiment 2: In this embodiment, the thermal physical parameters of the forging are obtained, and a model library is constructed according to the thermal physical parameters, which includes the following sub-steps: The thermal physical parameters of the forging are obtained, and a baseline physical model based on the Fourier heat conduction law is established according to the thermal physical parameters. The baseline physical model includes an internal heat source term to be identified; A residual dynamic neural network is constructed according to the thermal physical parameters in the historical data and the baseline physical model. The residual dynamic neural network is used to learn and model the residual error between the baseline physical model and the real data, and the residual error corresponds to the internal heat source term. The parameters of the residual dynamic neural network are globally optimized by the differential evolution algorithm to obtain a refined model. embedding the refined model as the internal heat source term into the benchmark physical model to form a model library; The thermophysical parameters include thermal conductivity, specific heat capacity, and density of the forged piece.

[0031] It should be noted that the benchmark physical model describes the heat conduction process through partial differential equations, but the heat source term in the equation is difficult to express analytically due to material deformation. The residual dynamics neural network receives the temperature gradient and strain rate as input and outputs the residual quantity related to the internal heat source. The differential evolution algorithm generates multiple candidate solutions in the parameter space, each corresponding to a set of neural network weights. The fitting error of the candidate solution on the historical data is calculated through simulation, and the parameter combination with the smallest error is selected. For example, in the forging case of titanium alloy TA15, the optimized neural network reduces the temperature field prediction error from ±25℃ to ±8℃, and the benchmark physical model embedded with the refined model can accurately reflect the coupling effect of deformation heat and external heating, providing a high-precision prediction basis for subsequent temperature control.

[0032] It should also be noted that based on the above, the embodiment realizes accurate modeling of the thermodynamic behavior of the forged piece. By introducing the residual dynamics neural network, the gap between the benchmark physical model and the actual data is bridged, the prediction accuracy of the model is improved, and the application of the differential evolution algorithm ensures the global optimization of the neural network parameters, further enhancing the accuracy and robustness of the model. The model library constructed in this way provides a reliable theoretical basis for subsequent temperature control, which helps to realize accurate control of the microstructure of the forged piece and improve the performance and quality of the titanium alloy forged piece.

[0033] To make the technical solution of step 1 of the present application clearer, the processing process of step 1 is described in detail, which includes the following four sub-steps, specifically including: For sub-step 1.1: The core step is to obtain the thermophysical parameters of the forged piece, and to establish a benchmark physical model based on the Fourier heat conduction law according to the thermophysical parameters, the benchmark physical model containing an internal heat source term to be identified.

[0034] In this embodiment, a three-dimensional transient heat conduction partial differential equation based on the Fourier heat conduction law is established, starting from the first principle of physics. The three-dimensional transient heat conduction partial differential equation describes the physical basis of heat transfer inside the object, and the expression satisfies: Wherein, is a temperature field function, representing the temperature of the internal coordinates (x, y, z) of the forged piece at time t; ρ is the density of titanium alloy, which comes from the material manual or actual measurement; ​cp(T) is the specific heat capacity of the titanium alloy as a function of temperature T, sourced from a material properties database; k(T) is the thermal conductivity of the titanium alloy as a function of temperature T, sourced from a material properties database; t is the time instant; div is the divergence operator to describe the convergence or divergence of heat flux; grad is the gradient operator to describe the rate of change of temperature in space; f is the internal heat source term, which in this embodiment can be represented as volumetric heat generation rate, and in this embodiment is treated as an unknown, to-be-identified function.

[0035] It can be understood that the above expression describes the temperature T at any point inside the forging at any time instant t, which depends on the heat inflow / outflow at that point and whether there is an internal heat source, so as to make the model have physical interpretability and extrapolation capability. The rate of change of temperature T at the coordinate at any time instant t depends on the heat inflow / outflow at that point and whether there is an internal heat source, so as to make the model have physical interpretability and extrapolation capability.

[0036] It can also be understood that, on the basis of the above expression, the partial differential equation is discretized by using finite element analysis (FEA) software (such as ANSYS, ABAQUS) or self-programming to form a simulation model that can be numerically solved; boundary conditions are then set, including thermal convection and thermal radiation between the heating furnace / mold / air; and finally a parameterized reference physical model is output, which can accurately simulate the heat transfer process under the condition of no internal heat source.

[0037] For sub-step 1.2: The core steps are: A residual dynamics neural network is constructed according to the thermal physical parameters in the historical data and the reference physical model, and the residual dynamics neural network is used to learn and model the residual between the reference physical model and the real data, and the residual corresponds to the internal heat source term.

[0038] In this embodiment, a residual dynamics neural network is designed to learn and model the residual between the reference physical model and the real data, and the residual corresponds to the internal heat source term in physics; In this embodiment, instead of directly using a neural network to fit the elusive temperature field, a neural network is used to learn and model the residual between the reference physical model and the real data, which is used to express the internal deformation heat source of the forging during the forging process .

[0039] Specifically, the main idea of the above process is to use historical data sets to perform inverse calculation to take the real, measured temperature field change process (or temperature change of key points) as known and substitute it into the reference physical model to inversely solve the internal heat source term at each time and each position that can make the physical equation hold It can be understood that the inversely solved is the target residual data that needs to be modeled; finally, a residual dynamics neural network (RDNN) is constructed to establish the mapping relationship of the local state of the forging to the local heat generation rate, and in some preferred embodiments, an RBF neural network is selected as the basic structure of the RDNN, and its expression satisfies: ; Wherein, is the internal heat source value predicted by the neural network at time t, and the training target is the residual heat source data inversely solved in step 1.1; is the mapping function of the RBF neural network; , , respectively represent the local strain rate, cumulative strain and current temperature at time t, which are derived from historical data sets or intermediate variables in the simulation process; is the spatial coordinate, which allows the model to learn the non-uniform distribution characteristics of deformation heat in space.

[0040] More specifically, the above expression defines a neural network model, the input of which is the real-time physical state (strain, strain rate, temperature, position) of a certain point of the forging, and the output is the instantaneous internal heat source generation rate of the point in that state .

[0041] For sub-step 1.3: The core step is: Optimize the parameters of the residual dynamics neural network globally through the differential evolution algorithm to obtain a refined model.

[0042] In this embodiment, by constructing an adaptive function oriented to the hybrid model, the hyperparameters of the residual dynamics neural network are globally optimized by the differential evolution algorithm to ensure the optimal collaborative work of the reference physical model and the residual dynamics neural network. It can be understood that simply training the RDNN may result in poor coupling effect between the RDNN and the reference physical model, therefore, the step in this embodiment optimizes the whole system to ensure that the final prediction accuracy of the collaborative work of the two reaches the highest.

[0043] The embodiment optimizes the hyperparameters of the residual dynamic neural network globally by designing a fitness function for the hybrid model and using the differential evolution (DE) algorithm, to ensure the optimal collaboration of the benchmark physical model and the residual dynamic neural network, and the expression satisfies: ; In the formula, is the fitness value of the hybrid model, which can also be expressed as the mean square error in some embodiments; N is the total number of sample points in the verification data set; is the specific discrete-time instant of the i-th sample point in the data set; is the real temperature measurement of the forging at time t, which is derived from the data set; is the predicted output of the hybrid model; is a set of RDNN network parameters currently being evaluated by the DE algorithm, which is described as the collection of all centers and widths in some embodiments.

[0044] It can be understood that the above expression calculates the mean square error of the temperature prediction of the entire hybrid model for all real data points under a certain set of RDNN parameters given by the DE algorithm, and takes it as the only evaluation standard for the iterative optimization of the DE algorithm; by evaluating the final output of the hybrid model rather than the intermediate output of the RDNN, it can be ensured that the parameters found by the optimization process can make the physical model and the neural network model work optimally in collaboration, making up for the coupling mismatch problem that may be caused by separate optimization, thereby obtaining a refined RDNN model that is globally optimized and highly coupled with the physical model.

[0045] More specifically, the calculation process of the predicted output of the hybrid model is as follows: RDNN with the to-be-evaluated parameters is used to calculate the heat source term ; is substituted into in the benchmark physical model of sub-step 1.1; ; The equation is numerically solved to obtain the temperature prediction value at time t.

[0046] For sub-step 1.4: The core step is to embed the refined model as the internal heat source term into the benchmark physical model to form a model library.

[0047] In this embodiment, the refined RDNN model is embedded into the benchmark physical model as a heat source term module, forming a final hybrid physical-information model, and the control equation expression is: ; wherein, is the optimal parameter found by the DE algorithm.

[0048] It can be understood that for different deformation behaviors of titanium alloy at different temperature intervals (such as α+β two-phase region and β phase region), sub-steps 1.2 to 1.4 are repeated to establish a dedicated refined RDNN model for each key working condition interval, and finally a hybrid model library is formed.

[0049] Embodiment 3: In this embodiment, the alloy constraint and the cost constraint of the forging are obtained, and the double-target temperature control trajectory is obtained according to the alloy constraint and the cost constraint, including the following sub-steps: Obtain the metallurgical constraint and construct the metallurgical constraint as a constraint condition set; Obtain the cost constraint and construct the cost constraint as a multi-objective optimization function; By using a global optimization algorithm and simulating according to the model library, the optimal discrete trajectory point that minimizes the metallurgical constraint and the cost constraint is searched and determined; The optimal discrete trajectory point obtained is generated into a target temperature control trajectory by a curve fitting technique; The target temperature control trajectory includes a core temperature target trajectory and a surface temperature target trajectory; The alloy constraint includes a temperature window constraint, a thermal stress gradient constraint, and a soaking time constraint.

[0050] It can be understood that the temperature window constraint is realized by setting the upper and lower limits of the temperature at different stages of the forging process, for example, maintaining the core temperature at 950±15°C in the β phase transition region; the thermal stress gradient constraint is realized by limiting the temperature difference between adjacent regions in the temperature field to not more than 50°C / mm. The soaking time constraint requires the workpiece to be kept in the target temperature interval for at least 120 seconds. In the multi-objective optimization function, the heating energy consumption cost is calculated according to the product of heating power and time, and the time cost is converted according to the total forging time. The global optimization algorithm generates a set of candidate trajectory points in each iteration, calls the model library to simulate its temperature evolution process, verifies whether it meets the constraint conditions, and calculates the total cost as the fitness value. The trajectory point that meets all the constraints and has the lowest cost is retained, and finally the core temperature target trajectory and the surface temperature target trajectory are generated by cubic spline interpolation. For example, in the initial heating stage, the surface temperature trajectory rises to the target interval at a rate of 10°C / s, and the core temperature trajectory is synchronized by lag compensation to avoid excessive temperature difference between the inside and outside. In this way, the dual-target temperature control trajectory meets the metallurgical quality requirements while minimizing the heating cost.

[0051] In some embodiments, the above steps can also be described as: The preset metallurgical constraints, including temperature window, thermal stress gradient and soaking time, are constructed into a set of mathematical constraint conditions; A multi-objective optimization function is constructed, which integrates energy consumption cost, equipment wear cost, production time cost and product quality penalty ; A global optimization algorithm is used to search and determine the optimal discrete trajectory points that minimize the multi-objective optimization function in the safe and feasible region defined by the set of mathematical constraint conditions, using the hybrid physical-information model for simulation; The optimal discrete trajectory points are generated into continuous and smooth core temperature target trajectory and surface temperature target trajectory by curve fitting technology.

[0052] Overall, step 2 can be divided into the following 4 sub-steps: For sub-step 2.1: The preset metallurgical constraints, including temperature window, thermal stress gradient and soaking time, are constructed into a set of mathematical constraint conditions.

[0053] The specific process includes: Define temperature window constraint: ensure that the temperature at any time and any position in the forging is within a safe process window, avoiding overburning or entering an unexpected phase region.

[0054] Define the thermal stress gradient constraint: In order to prevent cracks in the internal part of the forging due to excessive temperature difference between the inside and outside, the internal temperature gradient needs to be limited, and the expression satisfies: ; In the formula, is the temperature field gradient norm at time t and position x, that is, the fastest rate of change of the temperature at this point in space; is the maximum allowable temperature gradient, with units of ℃ / mm, which is determined by the material's fracture toughness, thermal expansion coefficient, and other mechanical property parameters through finite element pre-analysis or empirical formula. It can be understood that directly calculating the three-dimensional stress field is very complex, while the temperature gradient is the direct cause of thermal stress and is easier to calculate. By restricting the temperature gradient, the purpose of controlling thermal stress can also be achieved.

[0055] Define the soaking time constraint: In order to ensure the uniformity of the organization and temperature before forging, the forging must stay in the interval close to the target forging temperature for a long enough time, and the expression satisfies:

[0056] In the formula, Total length of the entire heating process; is the Heaviside step function, which is 1 when its argument is greater than or equal to 0, otherwise it is 0; Core temperature of the forging at time t; is the final target forging temperature; is the temperature width of the soaking interval; is the required minimum soaking time, which comes from the forging process specification.

[0057] It can be understood that the above expression accurately calculates the total length of time that the core temperature of the forging is in the soaking temperature interval and ensures that it is not less than a minimum value; by using integration and the Heaviside function, the soaking time can be strictly and accurately defined, which is more robust than a simple timer, especially in scenarios where the temperature has small fluctuations.

[0058] For sub-step 2.2: ​In this embodiment, the vague concept of "optimal heating trajectory" is quantified as a multi-dimensional, computable cost function that integrates energy consumption, equipment wear, production efficiency, and product quality. The input cost parameters include energy cost, equipment wear cost, production time cost, and product quality penalty. Specifically, in some embodiments, the above cost parameters are derived from the factory's cost accounting department, equipment maintenance manual, and quality control standards, and specifically include: Cost coefficient of unit energy (electricity, natural gas) ; Equivalent wear cost coefficient of forging press under unit load ; Production cost per unit time (labor, depreciation, etc.) ; Quality penalty coefficient for deviation from final target temperature .

[0059] The specific processing process is represented as: ; In the formula, is the total cost objective function to be optimized; represents time integration from start to end, accumulating total cost; The heating power applied at time t is one of the optimization control variables; is the forging force function, which represents the forging pressure when the core temperature is and the strain rate is . This function is given by the material's constitutive model, which establishes the key physical relationship between temperature and equipment load; , , , all represent weight coefficients representing cost parameters, which are dimensionless adjustment parameters set according to production priorities; is the terminal quality penalty term, which is used to penalize the deviation of the final core temperature from the target value.

[0060] It can be understood that the above expression expresses the total cost of the entire heating forging process as a single numerical value, and the goal of the optimization algorithm is to find a temperature control path that minimizes this total cost and unifies multiple conflicting production goals into one framework. For example, heating too quickly can save time cost, but will increase energy cost; forging at too low a temperature can save energy, but will cause forging force to increase dramatically, thereby increasing equipment wear cost. The above function provides a mathematical basis for balancing these contradictions.

[0061] For sub-step 2.3: A global optimization algorithm is used to search and determine the optimal discrete trajectory points that minimize the multi-objective optimization function defined in the safe feasible region defined by the set of mathematical constraints, using the hybrid physical-information model for simulation. In this embodiment, by selecting a global optimization algorithm suitable for solving such complex constraint nonlinear optimization problems, such as a genetic algorithm or a particle swarm optimization algorithm, a complete heating-forging-soaking temperature trajectory is represented as a series of temperature set values with key control time points.

[0062] In some preferred embodiments, the iterative optimization process is: Randomly generate a generation of a population containing multiple candidate control trajectories; For each candidate trajectory, call the hybrid physical-information model established in step one to perform a complete simulation to obtain the full-range temperature evolution of the core and surface of the forging under the trajectory and ; Check whether the temperature evolution process obtained by simulation meets all the metallurgical constraints in sub-step 2.1 throughout the process. If any violation occurs, a large penalty value is imposed on the candidate trajectory, or it is directly eliminated; For trajectories that meet all the constraints, calculate their total cost using the optimization function of sub-step 2.2 as their fitness (the lower the cost, the higher the fitness); According to the fitness, generate a better candidate trajectory population in the next generation through selection, crossover, mutation, or velocity-position update operations; Repeat the above process until the algorithm converges or the maximum number of iterations is reached.

[0063] It can also be understood that based on the above process, after iteration to convergence, the result obtained represents the discrete time point-temperature value sequence of the optimal process path, corresponding to the core and surface temperatures respectively.

[0064] For sub-step 2.4: The optimal discrete trajectory points are generated into a continuous and smooth core temperature target trajectory through curve fitting technology and the surface temperature target trajectory .

[0065] In this embodiment, the purpose of implementation is to convert the discrete, original optimization results obtained in the previous step into a smooth, continuous, and easy-to-control mathematical function form for the real-time control system, and then fit the discrete trajectory points into two continuous and smooth mathematical curves through curve fitting techniques such as B-spline or piecewise polynomial interpolation. In this embodiment, the B-spline curve fitting technique is preferred, and its expression satisfies: ; wherein, is the continuous target temperature trajectory function generated by fitting; is the control point of the B-spline curve, whose position is determined by the discrete trajectory points output by sub-step 2.3 through a fitting algorithm; is the B-spline basis function, which is a set of piecewise polynomials determined by the node vector and has local support; n and k are the number of control points and the order of the spline, respectively.

[0066] It can be understood that the above expression will weight average a series of discrete control points through a set of smooth basis functions, and finally obtain two optimal target trajectory functions that can be directly used for subsequent control steps, i.e.: core temperature target trajectory ; surface temperature target trajectory .

[0067] Embodiment 4: In this embodiment, the temperature state correction gain matrix is obtained, and the distribution estimate of the forging temperature field is obtained according to the temperature state correction gain matrix and the model library, including the following sub-steps: According to the state estimate at the previous time on the model library, the prior temperature field state at the current time is obtained; The forging surface temperature measurement value is obtained, and then the temperature measurement innovation is obtained according to the forging surface temperature measurement value and the prior temperature field state; The temperature state correction gain matrix is obtained; Performing a first re-correction: correcting the prior temperature field state using the temperature measurement innovation and the temperature state correction gain matrix; Performing a second re-correction: obtaining a correction factor using the temperature measurement innovation, and obtaining an intermediate temperature state estimate according to the correction factor; The intermediate temperature state estimate is taken as the final posterior state estimate at the current time, and the best estimate value of the core temperature and the distribution estimate of the forging temperature field are obtained.

[0068] It should be noted that in each control cycle, the model library calculates the prior temperature field distribution based on the thermal physical parameters and the deformation heat source, the surface temperature sensor measures the data in real time, and the temperature measurement innovation is generated by comparing the prior prediction value, which is preliminarily corrected by the gain matrix for each node of the temperature field. Subsequently, based on the statistical characteristics of the measurement innovation, a correction factor is calculated to make a secondary adjustment to the intermediate temperature state after the preliminary correction. For example, when the surface temperature measurement value deviates from the predicted value by more than a threshold value, the correction factor dynamically increases the correction amplitude of the internal node to compensate for the modeling error of the deformation heat source. The two-stage correction process improves the accuracy of temperature field estimation from the perspectives of linear error compensation and nonlinear dynamic compensation, respectively, and the final output of the posterior state estimation contains the best estimate value of the core temperature and the full-space temperature distribution, providing high-precision initial conditions for subsequent prediction control.

[0069] Based on the above, the temperature field of the forging is accurately estimated. Due to the adoption of the double correction mechanism, the accuracy of the temperature field estimation is significantly improved. In addition, by introducing the correction factor, the technical solution in the embodiment can better adapt to the temperature variation characteristics under different working conditions.

[0070] Similarly, the above summarizing steps can also be expressed as: The step of observing the real-time running state adopts a double correction physical information observer, specifically including: Mixing the physical-information model and the state estimation of the last moment to predict the prior temperature field state at the current moment; Calculate the temperature measurement innovation between the measured temperature of the forging surface and the predicted value of the corresponding surface temperature in the prior state ; Perform the first heavy correction: use the temperature measurement innovation and a temperature state correction gain matrix to directly correct the prior temperature field state; Perform the second heavy correction: use the temperature measurement innovation to update a heat source model correction factor online for characterizing the uncertainty of the internal heat source model , which is used to dynamically adjust the internal heat source term predicted by the neural network in the mixed physical-information model in subsequent prediction.

[0071] It should be noted that the traditional state observer (such as Kalman filter or Luenberger observer) usually takes the model as a whole and corrects all internal states uniformly based on the output error, while in this embodiment, the model of the dual correction physical information observer is applied not only to correct the temperature state itself, but also to the part with the strongest uncertainty in the model, that is, the RDNN (real-time correction of estimated internal deformation heat source) to real-time feedback and adjust its cognition of internal physical process (especially deformation heat) according to the actual measurement of surface temperature.

[0072] For sub-step 3.1: According to the refined RDNN model in step 1, the distribution of deformation heat source of each node inside the forging at the current time is predicted, and the predicted heat source term is substituted into the discretized physical model (finite element form) in step one to perform a single step of time integration, from the state at time t-1 to predict the prior state estimation at time t.

[0073] The expression satisfies: ; In the formula, is the prior state estimation vector at time t, which includes the predicted temperature values of all discrete nodes of the forging at time t; is the posterior state estimation vector at time t-1, that is, the optimal estimation value output by the last round of observer loop, , is the discretized state transition matrix and input matrix, which is obtained from the reference physical model (finite element discretization) in step one, and is a function of temperature to handle nonlinear problems; is a generalized input vector containing all heat inputs, which not only includes external heating / cooling power, but also includes internal heat source term predicted by RDNN.

[0074] The above expression is the matrix form of the discretized heat conduction equation, which calculates the temperature field distribution at the next time according to the temperature field at the last time and the current heat input. The state space matrix form is used to seamlessly connect with the observer framework in modern control theory, and the matrix , itself can change with temperature, so that this prediction step can adapt to nonlinear systems.

[0075] For sub-step 3.2: From the high-dimensional prior state vector In this case, the predicted temperature value corresponding to the surface measurement point position is extracted to calculate the temperature measurement innovation , the expression satisfies: ; In the formula, is the temperature measurement innovation at time t, which represents a scalar or a small-dimensional vector, which can also be a model-predicted bias signal in this embodiment; is the actual reading of the sensor, such as an infrared thermometer, which is the actual measurement value of the surface temperature of the forging; is a measurement matrix used to filter out the state components corresponding to the measurement points from the complete state vector, which is a row vector for a single surface temperature measurement point, with 1 at the position corresponding to the surface node and 0 at the remaining positions.

[0076] For sub-step 3.3: The traditional observer directly corrects the entire temperature state using the bias signal through a gain matrix, while in this embodiment, the correction process is divided into two parts by a double correction mechanism, which updates the temperature state and the uncertainty of the heat source model, including the first correction and the second correction, specifically: For the first correction: Direct correction of the temperature state vector, similar to the update step of the standard Kalman filter, is used to correct the general prediction bias caused by model simplification, parameter error, etc., and the expression satisfies: ; In the formula, is the intermediate temperature state estimate after the first correction; is the temperature state correction gain matrix, which is obtained by solving the Riccati equation or by offline optimization setting, to balance the response speed and noise resistance.

[0077] For the second correction: Correction of the internal heat source model uncertainty. It can be understood that the bias signal not only reflects the error of temperature prediction, but also represents the possible systematic bias of the internal heat source prediction. Therefore, an internal variable representing this uncertainty, i.e., the heat source model correction factor, is updated through the temperature innovation (bias signal), and the expression satisfies: ; In the formula, is the optimal estimate of the heat source model correction factor updated at time t; The heat source correction gain is a scalar parameter that needs to be tuned, and determines the speed of adjusting the heat source model according to the measurement error.

[0078] In the above expression, an integral feedback loop is established. If the measured surface temperature is consistently higher than the predicted temperature, it means that the internal deformation heat source may be underestimated, and the formula will gradually increase the correction factor (greater than 1), and vice versa, and finally obtain the first corrected intermediate temperature state estimation and the updated heat source model correction factor.

[0079] For sub-step 3.4: The first corrected intermediate temperature state estimation is taken as the final posterior state estimation at the current time t, that is, , In other embodiments, the core physical quantities of interest in the subsequent control step can also be extracted from the high-dimensional final state vector , including: The best estimate of the core temperature ; The complete internal temperature field distribution estimate ; The current heat source model correction factor .

[0080] Embodiment 5: In this embodiment, the optimal control sequence according to the distribution estimation of the forging temperature field and the target temperature control trajectory includes the following sub-steps: Define a multi-time domain framework; Obtain the microstructure evolution model of the forging grain, and construct a coupled prediction model according to the microstructure evolution model, the distribution estimation of the forging temperature field, and the target temperature control trajectory after dimensionality lifting in the multi-time domain framework; Obtain the tactical cost and the strategic cost, and construct a composite cost function according to the tactical cost and the strategic cost; At each control time, obtain the optimal control sequence in the multi-time domain framework with the objective of minimizing the composite cost function, and take the first element of the optimal control sequence as the feedforward control set value.

[0081] It should be noted that the multi-time domain framework is divided into three nested optimization levels: the first level covers the heating control sequence for the next five minutes, the second level predicts the grain size evolution for the next thirty minutes, and the third level evaluates the microstructure uniformity throughout the forging cycle. When building the coupled prediction model, the core temperature and surface temperature in the temperature field distribution estimation are input as boundary conditions into the microstructure evolution model, while the grain size output by the model is fed back to the temperature control module to form a closed loop. Through dimensionality expansion, the state variables of the prediction model are expanded from a single temperature field to a multi-physical field set containing strain energy, phase fraction, and grain size. In the compound cost function, the weight of the tactical cost decays linearly with the forging stage, and the weight of the strategic cost increases exponentially in the phase transition sensitive temperature interval. At each control time, a sequence quadratic programming algorithm is used to solve the multi-time domain optimization problem to generate an optimal sequence containing the next ten control steps, where the first control quantity directly acts on the heating system, and the subsequent control quantities are used as prediction benchmarks for rolling optimization.

[0082] It should also be noted that the above scheme can simultaneously consider temperature control and microstructure evolution at multiple time scales to achieve precise control of the forging process. Thus, the short-term temperature control accuracy can be ensured while optimizing the long-term microstructure evolution process, thereby improving the final mechanical properties and fatigue life of the forged part. Further, by introducing a compound cost function of tactical cost and strategic cost, the short-term control effect and long-term microstructure evolution target can be balanced during control to achieve more comprehensive and efficient temperature control. Specifically, this method can effectively reduce temperature fluctuations, reduce thermal stress gradients, and ensure that the forged part forms an ideal microstructure structure within the optimal temperature window.

[0083] In some embodiments, the above summarizing step can be represented as: The state vector of the system is dimensionally expanded to contain both the temperature field state and the key microstructure parameter state; A compound cost function composed of tactical cost and strategic cost is constructed , wherein the tactical cost is used to evaluate the tracking error of the dual-objective temperature trajectory in the short-term tactical prediction time domain , and the strategic cost is used to evaluate the influence on the final microstructure and comprehensive economic benefits in the long-term strategic prediction time domain that extends to the end of the process; At each control time, an optimal control sequence is obtained by solving a constrained optimization problem that aims to minimize the compound cost function , and the first element of the sequence is used as the feedforward control set value.

[0084] It can be understood that each control moment in this step is no longer simply calculating an optimal control power or an optimal control temperature, but through a hierarchical and far-sighted optimization process, a control instruction is calculated which can not only meet the short-term temperature accurate tracking, but also ensure the long-term final organizational performance optimization. Specifically, this step includes the following four interrelated sub-steps: For sub-step 4.1: The state of the system is extended from a purely thermodynamic state (temperature) to a thermodynamic-metallurgical coupled state, and the new state vector contains not only temperature but also key microstructure parameters, and the expression satisfies: ; In the formula, is the extended state vector at time t; is the temperature estimation vector of all discrete nodes inside the forging, derived from the output of step three; is the average grain size vector of all discrete nodes inside the forging, whose initial value is known and the subsequent value is calculated online through an evolution model coupled with temperature; is the alpha phase volume fraction vector of all discrete nodes inside the forging.

[0085] In the above expression, a new and more comprehensive system state description is defined, and the real-time updated extended state vector is finally obtained.

[0086] The definition of the multi-time domain framework includes: Tactical prediction time domain : A short time window (such as tens of seconds) for fine and fast temperature trajectory tracking and disturbance suppression; Strategic prediction time domain : A longer time window that runs through the end of the forging process, used to evaluate the far-reaching impact of current decisions on the final product quality (such as final grain size), usually .

[0087] For sub-step 4.2: Through the hybrid model and metallurgical model constructed in step one, the future temperature and organizational evolution panorama in the two different time domains is simultaneously deduced according to an assumed future control sequence. Specifically, by performing the following steps at each time step in a loop from t to t+ : Calculate heat source: According to the temperature, strain, etc. in the current state, call the RDNN model to calculate the internal deformation heat source; Solve temperature field: Substitute heat source term and external heating power into the reference physical model (partial differential equation), numerically solve the temperature field of the next time step ; Solve microstructure field: Substitute the updated temperature field as input into the microstructure evolution model (e.g. grain growth equation), solve the grain size and phase fraction of the next time step ; .

[0088] Update the extended state for the next time step calculation.

[0089] It can be understood that the above process will continue until the end of the strategic time domain , so as to obtain a complete extended state prediction trajectory covering the entire remaining process flow, i.e. the future extended state prediction trajectory ; It can be understood that the input of step 4.2 in the above process is the current extended state vector output by sub-step 4.1 ; The mixed physical-information model library constructed in step one; The microstructure evolution model introduced in step 4.1; A candidate control sequence given by the optimization algorithm: , where is the control time domain.

[0090] For sub-step 4.3: By obtaining the future extended state prediction trajectory output by sub-step 4.2, the optimal double-target temperature trajectory dynamically generated in step two, the final desired target microstructure parameters, and the weight coefficients of various costs (the same as defined in step 2), a composite cost function composed of tactical cost and strategic cost is constructed, the expression satisfies: ; In the formula, is the tactical cost function, a standard quadratic tracking error function; is the predicted future temperature vector; is the double-target trajectory vector obtained from step two; is the weighted norm square, representing the weighted sum of tracking errors, where is the tactical tracking error weight matrix, used to adjust the relative importance of core and surface temperature tracking; The effort weight for tactical control, used to penalize excessively large short-term control actions.

[0091] The strategic cost function, evaluating the impact on the final outcome; The strategic weight coefficient, a key tuning parameter, used to balance the short-term tracking accuracy and the long-term strategic goal; The predicted final average grain size, derived from the endpoint of the extended-state prediction trajectory; The final target grain size required by the process; The total economic cost prediction from the current time to the end of the process, calculated in a similar way as in Step 2, but only for the future part.

[0092] In the above expression, a weighted sum is taken, simultaneously evaluating both the short-term performance (whether the recent temperature target is closely tracked) and the long-term impact (whether the final grain size and economic optimality are guaranteed) of a control decision; the short-sightedness problem of traditional MPC is solved.

[0093] For sub-step 4.4: At each control time, an optimal control sequence is obtained by solving a constrained optimization problem with the objective of minimizing the composite cost function , and the first element of the sequence is taken as the feedforward control setpoint.

[0094] This embodiment solves the following constrained optimization problem at each control time t: ; Specifically, a nonlinear programming (NLP) solver or metaheuristic optimization algorithm is used to find the optimal control sequence According to the receding horizon principle of MPC, although we have calculated a whole sequence , at the current time t, we only take the first element of the sequence as the final feedforward control quantity; and at the next time t+1, a new state measurement and estimation are obtained, and then the whole Step Four is repeated to perform a complete optimization calculation again to obtain the optimal feedforward control setpoint at the current time .

[0095] Embodiment 6: In this embodiment, the construction process of the coupled prediction model includes: Obtaining the external heating power; ​calculating the deformation heat source of the refining model according to the temperature and strain in the current state; substituting the deformation heat source and the external heating power into the reference physical model to obtain a temperature field of a next time step; substituting the temperature field of the next time step as input into the organization evolution model to obtain grain size and phase fraction of the next time step; updating the extended state again until the end of the multi-time domain framework is reached.

[0096] It should be noted that, in the process of iteration at each time step, the external heating power is collected in real time by the actual power sensor of the heating device, and the deformation heat source is obtained by calculating the quadratic invariant of the current temperature field and the strain rate tensor. Both are input into the reference physical model to obtain the temperature field distribution at the next time by solving the three-dimensional unsteady heat conduction equation. The temperature field is transmitted to the organization evolution model, in which the grain size can be calculated by associating the Zener-Hollomon parameter with the dynamic recrystallization model, and the phase fraction can be predicted by the Avrami equation combined with the temperature history. The four-dimensional state variables of the extended state update include temperature field, strain field, grain size and phase fraction, forming a closed loop iteration until the prediction range of the multi-time domain framework is covered. Through the simultaneous fusion of the dual effects of external heat input and internal heat generation, the temperature field and microstructure are cooperatively predicted.

[0097] It should also be noted that, based on the above process, the embodiment realizes accurate prediction of the temperature field and microstructure evolution of the forging during forging, thereby better controlling the core temperature of the forging and ensuring that it is always within the critical temperature window required for the formation of the target microstructure. Further, the method takes into account the effect of deformation heat, improving the accuracy of temperature control and helping to obtain an ideal microstructure, thereby improving the final mechanical properties and fatigue life of the forging.

[0098] In some preferred embodiments, the above steps can also be represented as: the deviation between the measured temperature of the forging surface and the model predicted value is decomposed into a bias error component for characterizing the slowly changing model mismatch and a dynamic error component for characterizing the transient dynamic mismatch through a low-pass filter; the bias error component is used to update the physical boundary parameters in the hybrid physical-information model online through a slow loop integral type update rule; the dynamic error component is used to update the heat source model correction factor in step four in real time through a fast loop correction loop ; ​​a low gain controller is used to handle the residual error to generate the final feedback correction control amount .

[0099] It should be noted that in the conventional feedback control process, the prediction error is input into a simple controller (such as PID) to generate a compensation value of the control amount, and in the present embodiment, a hierarchical model backtracking correction framework is constructed, which no longer only corrects the final control output, but regards the prediction error as a kind of deviation, analyzes its most possible deviation amount by analyzing the dynamic characteristics of the error, and then directly corrects the most uncertain parameters or modules inside the model in real time.

[0100] It should also be noted that the present embodiment establishes an intelligent online feedback mechanism to analyze the deviation between the model prediction and the physical reality in real time, and traces the deviation to the most likely error link inside the model, whether it is the inaccurate estimation of slowly changing physical boundary parameters or the prediction failure of rapidly changing internal deformation heat source model, and then dynamically corrects the model itself, and finally uses a low gain controller to handle the residual random error that is not absorbed by the model. The specific process includes the following four logically progressive sub-steps: For sub-step 5.1: In the present embodiment, it is assumed that the total error is superimposed by two types of errors with different properties: One is the low-frequency / bias error caused by the static or slowly changing deviation of the model to the steady-state physical parameters (such as heat exchange coefficient); The other is the high-frequency / dynamic error caused by the prediction failure of the model to the transient dynamics (mainly deformation heat).

[0101] In order to separate these two types of errors, a digital low-pass filter is designed to extract the bias error component, and its expression satisfies: ; In the formula, is the bias error component estimated at time t; is the bias error estimation value at the last time; is the original total error at the current time, which comes from sub-step 3.2; is the filter coefficient.

[0102] Finally, the bias component is subtracted from the total error to obtain the dynamic error component, and then the final bias error component and the dynamic error component are obtained.

[0103] For sub-step 5.2: In this embodiment, the slowly varying error from the previous step is used to continuously fine-tune the parameters in the baseline physical model that are most likely to have static bias, such as the heat transfer coefficients. Specifically, an integral-type parameter updating law is established to correct the boundary parameters online. If the model-predicted surface temperature consistently goes up, it means that the model might underestimate the heat dissipation, and the equivalent heat transfer coefficient of the whole system needs to be increased. The expression is: ; where, is the updated estimate of the convective heat transfer coefficient at time t, which will be immediately used in the model calculation of the next step three (state observation) and step four (MPC prediction); is the estimate of the convective heat transfer coefficient at the previous time, whose initial value comes from the offline identification in step one; is the adaptive gain, which is a small positive number that determines the speed of parameter correction. This value needs to be tuned based on experience or simulation to ensure the stability of parameter adjustment.

[0104] For sub-step 5.3: It should be noted that, compared to the integrator in step three, the I-D controller designed in this embodiment can respond more quickly and stably to the changes in dynamic error , which directly translates the rapid fluctuations in surface temperature into rapid adjustments to the prediction of internal heat sources. At the same time, it can be understood as an online learning mechanism: when a sudden impact occurs, if the measured surface temperature rises faster than the RDNN prediction, the above expression will immediately increase the heat source model correction factor, so that the model generates more deformation heat than originally expected in the next instant, and eventually obtains a heat source model correction factor whose expression satisfies: ; where, is the updated heat source model correction factor at time t, which will be used in the model prediction of the next step three and step four, i.e. ; , are the integral gain and differential gain of the heat source correction loop, respectively. These two parameters need to be tuned through simulation to match the typical generation and dissipation rate of deformation heat.

[0105] For sub-step 5.4: It should be noted that after the compensation and correction of the internal parameters of the model in the previous two steps, most of the systematic errors have been absorbed by the model itself in theory, and this step is responsible for handling those most random and unpredictable residual errors to generate a traditional control compensation signal.

[0106] In some specific embodiments, the remaining error is preferably processed by an existing proportional (P) controller, i.e. satisfying: ; In the formula, is the final feedback correction control quantity, with the unit of power (W) or equivalent units; is the residual error proportional gain, which is a small positive number that needs to be set.

[0107] It should be noted that through the above four sub-steps, an innovative feedback correction system is constructed. It not only satisfies the simple correction of the output of the controller, but also realizes the layered and online adaptive correction of the physical parameters (slow loop) and the data-driven module (fast loop) through error decomposition and attribution, and finally only uses a traditional low-gain controller as the tail processing, so as to greatly improve the online prediction accuracy of the entire hybrid physical-information model while ensuring the robustness and stability of the system.

[0108] Example 7: In this embodiment, the feedback correction control quantity is obtained according to the measured temperature of the forging surface, the optimal control sequence and the predicted value of the coupled prediction model, which includes the following sub-steps: The deviation between the measured temperature of the forging surface and the predicted value of the coupled prediction model is decomposed into a bias error component and a dynamic error component through a filter; According to the bias error component, the physical boundary parameters in the coupled prediction model are updated; According to the dynamic error component, the correction factor is updated; The residual error is processed by a controller to obtain the feedback correction control quantity.

[0109] It should be noted that the original deviation between the measured surface temperature and the predicted value is first separated by a filter: a Butterworth low-pass filter extracts the bias error component below the set cutoff frequency, and a Kalman filter extracts the high-frequency dynamic error component. The bias error component is used to correct the mold contact thermal resistance parameters, such as updating the contact thermal resistance value by least squares method, to eliminate the steady-state deviation caused by boundary condition modeling error. The dynamic error component is input to the state estimation module to adjust the weight coefficients in the temperature state correction gain matrix, enhancing the tracking ability of transient disturbances. The residual error is processed by a proportional-integral controller to generate a feedback correction control quantity, which is superimposed on the feedforward control set value to form a closed-loop correction. This process processes error components by frequency bands, avoiding high-frequency noise interference with model parameter identification, and improving the response speed of the system to dynamic deviations.

[0110] Further, the precise estimation and control of the forging temperature field are realized, the temperature deviation is decomposed into bias and dynamic components, the model parameters and correction factors are updated respectively, which can effectively improve the accuracy of temperature field estimation; at the same time, the feedback correction mechanism is introduced, which can compensate for the model prediction error in time, ensuring the real-time and robustness of temperature control. This method can more accurately control the core temperature of the forging, keeping it within the optimal temperature window, which is conducive to obtaining ideal microstructure and improving the mechanical properties and service life of the forging.

[0111] In some preferred embodiments, the above steps can also be represented as: The step of fusing control and generating the final actual driving signal adopts an adaptive control authority fusion and direct driving framework based on uncertainty evaluation, specifically including: According to the heat source model correction factor The degree of deviation from its ideal value, a feedforward control confidence score is calculated in real time to quantify the reliability of the feedforward control set value ; Through a nonlinear authority allocation function, the feedforward control confidence score , the authority weight of the feedforward control set value and the feedback correction control quantity is dynamically calculated, and weighted fusion is performed to obtain an ideal comprehensive control instruction; The ideal comprehensive control instruction is subjected to amplitude constraint and rate constraint processing to ensure that it meets the physical constraints of the actuator, generating the final actual driving signal.

[0112] It can be understood that in the embodiment, an adaptive control authority fusion and direct driving framework based on uncertainty evaluation is constructed, which combines fusion and execution into one step, no longer blindly superimposes control signals, but evaluates the reliability of two decision sources of feedforward (from MPC) and feedback (from HMCF) in real time, dynamically allocates control weights, and directly synthesizes the driving signal finally applied to the physical actuator (such as power supply, valve), and specifically includes the following four sub-steps: For sub-step 6.1: According to the heat source model correction factor The degree of deviation from its ideal value, a feedforward control confidence score for quantifying the reliability of the feedforward control set value is calculated in real time , the expression satisfies: ; In the formula, The feedforward control confidence score at time t, which is dimensionless, ranges from 0 to 1; is a confidence decay coefficient, which is a positive adjustment constant for adjusting the sensitivity of confidence to model deviation, and the value is set by offline simulation; is the real-time heat source model correction factor from step three.

[0113] It can be understood that in the above expression, an exponential decay function is used, which can make the confidence in the model (i.e. the confidence in the feedforward control) very high (close to 1) when the model correction factor is close to the ideal value 1; and once significantly deviates from 1, indicating that the model is undergoing a dramatic adaptive correction, its reliability will quickly decrease. This nonlinear relationship better reflects the degree of trust when the model is stable than linear mapping, and finally outputs a feedforward control confidence score.

[0114] For sub-step 6.2: In the embodiment, according to the feedforward control confidence score , the authority weight of the feedforward control set value and the feedback correction control amount is dynamically calculated by a nonlinear authority allocation function, and weighted fusion is performed to obtain an ideal comprehensive control instruction. More specifically, the function is preferably a Sigmoid type authority allocation function, and the expression satisfies: ; In the formula, is the dynamic authority weight of the feedforward control, ranging from 0 to 1; is the authority switching gain; For the confidence threshold, a constant between 0 and 1 (e.g. 0.8) that represents the critical point at which we start to trust the feedforward model.

[0115] Finally, the final ideal control command is synthesized by the following expression : ; In the above expression, the feedforward and feedback control quantities are weighted and averaged according to the dynamically calculated authority weight, to obtain the most reasonable integrated control intention at the current time. This weighting method ensures that when the model is highly reliable, the system behavior is mainly dominated by step four, achieving optimal performance; when the model is inaccurate, the system can smoothly switch to the more conservative and robust feedback control mode dominated by step five, ensuring safety.

[0116] For sub-step 6.3: After two-step processing of the ideal command (amplitude constraint and rate constraint), the final actual driving signal is output .

[0117] For amplitude constraint satisfaction: ; For rate constraint satisfaction: ; In the formula, is the maximum allowed output amplitude of the actuator; is the maximum maximum power rate of change of the actuator; is the actual driving signal issued at the previous time t-1.

[0118] In the above expression, it is mainly used to ensure that the difference between the command issued at the current time and the command at the previous time does not exceed the maximum amount that the actuator can safely change within a single control period.

[0119] For sub-step 6.4: It should be noted that this step is used to form a closed loop at the lowest level to verify whether the control command is accurately executed, and to feed back the execution deviation information to the upper layer, constituting a monitoring of the health of the entire control system, which specifically includes the following steps: Obtain the actual state value read back from the actuator ; Calculate the execution deviation, the expression satisfies: ; In the formula, is the execution error.

[0120] In some embodiments, the method further comprises the following steps: monitoring the execution error and triggering an actuator fault alarm if the absolute value of the error continuously exceeds a preset threshold.

[0121] In some embodiments, the method further comprises the following steps: calculating the execution error as a new information flow, feeding back to the HMCF framework of step five, in which the execution error can be regarded as another form of uncertainty bias and used to further correct the physical model parameters or the output of the RDNN; for example, if the instruction continuously exceeds the actual output, it may indicate that the heating efficiency is declining, and the HMCF can accordingly fine-tune the heat exchange coefficient in the model.

[0122] Embodiment 8: In this embodiment, the step of obtaining the actual control signal according to the feedforward control set value and the feedback correction control amount comprises the following sub-steps: obtaining a feedforward control confidence score according to the degree of deviation of the correction factor from the ideal value; obtaining the authority weight of the feedforward control set value and the feedback correction control amount according to the feedforward control confidence score through a nonlinear authority allocation function, and performing weighted fusion to obtain the ideal integrated control instruction after fusion.

[0123] The step of obtaining the ideal integrated control instruction after fusion further comprises the following steps: performing amplitude constraint and rate of change constraint processing on the ideal integrated control instruction to obtain the actual control signal.

[0124] In combination with the foregoing embodiments, it can be understood that the control confidence score calculation module monitors the deviation of the correction factor output by the state estimator from the theoretical value in real time, and automatically reduces the feedforward control weight when the deviation exceeds the preset threshold; the nonlinear authority allocation function generates a dynamic fusion coefficient according to the confidence score, giving priority to the feedforward set value when the model prediction accuracy is high, and enhancing the feedback correction effect when the measurement noise is large. The control instruction after dynamic fusion enters the constraint processing link, the amplitude constraint module clamps the instruction value exceeding the rated power of the heater to the safe range, and the rate of change constraint module prevents temperature mutation by setting the maximum temperature rising rate.

[0125] Embodiment 9: In the global optimization process of the residual dynamic neural network parameters through the differential evolution algorithm, the process comprises: randomly generating a population containing multiple candidate control trajectories; For each candidate trajectory within the population, call the refinement model to simulate the temperature evolution of the core and surface of the forging under the candidate trajectory; Check whether the temperature evolution process obtained by simulation meets the metallurgical constraints throughout the process. If not, impose a penalty value on the candidate trajectory, or directly eliminate it; For the trajectory that meets all the constraints, calculate the total cost using the cost constraint, and use the total cost as the fitness; According to the fitness, generate the next generation of candidate trajectories through selection, crossover, mutation or speed-position update operations; Repeat the above process until the algorithm converges or reaches the maximum number of iterations.

[0126] In combination with the foregoing embodiments, it can be understood that the differential evolution algorithm first encodes the candidate trajectories in each iteration, converting the temperature control parameters into real number vectors. The refinement model performs full-process simulation on the control strategy corresponding to each vector, outputting time series data of the core and surface temperatures. The constraint checking module simultaneously verifies whether the temperature sequence is within the target phase transition temperature range throughout the process. For individuals that violate the constraints, a penalty term is added in the fitness calculation stage to reduce their survival probability. The top 20% of elite individuals in terms of fitness are directly retained for the next generation, and the remaining individuals generate new solutions through differential mutation operations. During the evolution process, the Cauchy mutation operator is used to enhance global search capability, and a simulated annealing mechanism is introduced to prevent premature convergence. After a predetermined number of iterations, the algorithm outputs the individual with the optimal fitness as the final temperature control trajectory, which achieves dual optimization of energy cost and time cost while meeting the metallurgical constraints.

[0127] Example 10: As shown in the accompanying drawings, Figure 2 a control system is provided in this embodiment, which comprises: an offline modeling module, configured to obtain the thermal physical parameters of the forging, and construct a model library based on the thermal physical parameters, the model library being used to describe the thermodynamic behavior of the forging during the heating and forging processes; a trajectory generation module, configured to obtain the alloy constraints and cost constraints of the forging, and obtain a double-target temperature control trajectory based on the alloy constraints and cost constraints; a state estimation module, configured to obtain a temperature state correction gain matrix, and obtain a distribution estimation of the temperature field of the forging based on the temperature state correction gain matrix and the model library; a feedforward control module, configured to construct a coupled prediction model based on the distribution estimation of the temperature field of the forging and the target temperature control trajectory, and obtain a feedforward control set value based on the coupled prediction model; A feedback control module is configured to acquire a measured temperature of the forging surface and a predicted value of the coupled prediction model, acquire a feedback correction control amount according to the measured temperature of the forging surface, the optimal control sequence and the predicted value of the coupled prediction model, and correct the coupled prediction model by using the feedback correction control amount. A temperature control module is configured to acquire an actual control signal according to the feedforward control set value and the feedback correction control amount.

[0128] It should be noted that by deeply fusing physical laws and data intelligence, a residual dynamics neural network is innovatively used to learn and quantify internal heat source terms dominated by complex deformation heat and difficult to analyze. By constructing a mathematical metallurgical constraint and a multi-objective cost function, and using a global optimization algorithm, a dedicated two-objective temperature trajectory is dynamically generated for each forging task, thereby ensuring that the control target itself has adaptability and optimality for specific working conditions.

[0129] Embodiment 11: Based on the same inventive idea as the foregoing embodiments, the present embodiment provides a computer device, which comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0130] Embodiment 12: Based on the same inventive idea as the foregoing embodiments, the present embodiment provides a computer readable storage medium, which stores a computer program, and a processor executes the computer program to implement the method described above.

[0131] It should be noted that in this document, the term "comprising" or "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or system. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or system including the element.

[0132] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0133] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned example method can be realized by means of software and a necessary general hardware platform, and of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk) and includes a plurality of instructions for causing a multimedia terminal device (which can be a mobile phone, a computer, a television receiver, or a network device, etc.) to execute the method described in each embodiment of the present application.

[0134] The above is only a preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent flow transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied to other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for temperature control in titanium alloy forging, characterized in that, The control method includes the following steps: The thermophysical parameters of the forging are obtained, and a model library is constructed based on the thermophysical parameters. The model library is used to describe the thermodynamic behavior of the forging during the heating and forging process. Obtain the alloy constraints and cost constraints of the forging, and obtain the dual-objective temperature control trajectory based on the alloy constraints and cost constraints; Obtain the temperature state correction gain matrix, and obtain the distribution estimate of the temperature field of the forging based on the temperature state correction gain matrix and the model library; A coupled prediction model is constructed based on the distribution estimation of the temperature field of the forging and the target temperature control trajectory, and the feedforward control setpoint is obtained based on the coupled prediction model. The measured surface temperature of the forging and the predicted value of the coupled prediction model are obtained. Based on the measured surface temperature of the forging, the optimal control sequence and the predicted value of the coupled prediction model, the feedback correction control quantity is obtained, and the coupled prediction model is corrected by the feedback correction control quantity. The actual control signal is obtained based on the feedforward control setpoint and the feedback correction control quantity.

2. The temperature control method for titanium alloy forging as described in claim 1, characterized in that, The process of obtaining the thermophysical parameters of the forging and constructing a model library based on these parameters includes the following sub-steps: The thermal properties of the forging are obtained, and a reference physical model based on Fourier's law of heat conduction is established based on the thermal properties. The reference physical model includes an internal heat source term to be identified. A residual dynamics neural network is constructed based on the thermal property parameters in historical data and the benchmark physical model. The residual dynamics neural network is used to learn and model the residual between the benchmark physical model and the real data. The residual corresponds to the internal heat source term. The parameters of the residual dynamic neural network are globally optimized using the differential evolution algorithm to obtain a refined model. The refined model is embedded as the internal heat source item into the baseline physical model to form a model library; The thermophysical parameters include the thermal conductivity, specific heat capacity, and density of the forging.

3. The temperature control method for titanium alloy forging as described in claim 1, characterized in that, The process of obtaining the alloy constraints and cost constraints of the forging, and obtaining the dual-objective temperature control trajectory based on the alloy constraints and cost constraints, includes the following sub-steps: Obtain the metallurgical constraints and construct them into a set of constraint conditions; Obtain cost constraints and construct them into a multi-objective optimization function; By employing a global optimization algorithm and performing simulations based on a model library, the optimal discrete trajectory points that minimize the metallurgical constraints and cost constraints are searched and determined. The optimal discrete trajectory points are obtained through curve fitting technology to generate the target temperature control trajectory. The target temperature control trajectory includes a core temperature target trajectory and a surface temperature target trajectory; The alloy constraints include temperature window constraints, thermal stress gradient constraints, and soaking time constraints.

4. The temperature control method for titanium alloy forging as described in claim 1, characterized in that, Obtaining the temperature state correction gain matrix and estimating the distribution of the forging temperature field based on the temperature state correction gain matrix and the model library includes the following sub-steps: Based on the state estimate of the previous moment in the model library, obtain the prior temperature field state at the current moment; Obtain the surface temperature measurement value of the forging, and then obtain the temperature measurement innovation based on the surface temperature measurement value of the forging and the prior temperature field state; Obtain the temperature state correction gain matrix; Perform the first correction: correct the prior temperature field state using the temperature measurement innovation and temperature state correction gain matrix; Perform a second correction: use the temperature measurement innovation to obtain a correction factor, and obtain an intermediate temperature state estimate based on the correction factor; The intermediate temperature state estimate is used as the final posterior state estimate at the current moment, and the best estimate of the core temperature and the distribution estimate of the forging temperature field are obtained.

5. The temperature control method for titanium alloy forging as described in claim 1, characterized in that, The process of obtaining the optimal control sequence based on the distribution estimation of the forging temperature field and the target temperature control trajectory includes the following sub-steps: Define a multi-temporal framework; Obtain the microstructure evolution model of the forging grains, and construct a coupled prediction model in a multi-time domain framework based on the microstructure evolution model, the distribution estimation of the forging temperature field, and the target temperature control trajectory after dimensionality increase. Obtain tactical and strategic costs, and construct a composite cost function based on tactical and strategic costs; At each control time, within a multi-time-domain framework, the optimal control sequence is obtained with the goal of minimizing the composite cost function, and the first element of the optimal control sequence is used as the feedforward control setpoint.

6. The temperature control method for titanium alloy forging as described in claim 5, characterized in that, The construction process of the coupled prediction model includes: Obtain external heating power; Calculate the deformation heat source of the refining model based on the current temperature and strain. Substitute the deformable heat source and external heating power into the baseline physical model to obtain the temperature field at the next time step; The temperature field of the next time step is used as input and substituted into the microstructure evolution model to obtain the grain size and phase fraction of the next time step; Then update the dimensionality expansion state until the end of the multi-temporal framework is reached.

7. The temperature control method for titanium alloy forging as described in claim 4, characterized in that, The process of obtaining the measured surface temperature of the forging and the predicted value of the coupled prediction model, and obtaining the feedback correction control quantity based on the measured surface temperature of the forging, the optimal control sequence, and the predicted value of the coupled prediction model, includes the following sub-steps: The deviation between the measured surface temperature of the forging and the predicted value of the coupled prediction model is decomposed into bias error component and dynamic error component by a filter. Update the physical boundary parameters in the coupled prediction model based on the bias error components; Update the correction factor based on the dynamic error components; The residual error is then processed by the controller to obtain the feedback correction control quantity.

8. The temperature control method for titanium alloy forging as described in claim 1, characterized in that, The step of obtaining the actual control signal based on the feedforward control setpoint and the feedback correction control quantity includes the following sub-steps: The confidence score of feedforward control is obtained based on the degree to which the correction factor deviates from the ideal value; Using a nonlinear permission allocation function, the permission weights of the feedforward control setpoint and the feedback correction control quantity are obtained based on the feedforward control confidence score, and then weighted and fused to obtain the ideal integrated control command after fusion.

9. The temperature control method for titanium alloy forging as described in claim 8, characterized in that, After obtaining the ideal integrated control command after fusion, the following steps are also included: The ideal integrated control command is subjected to amplitude and rate of change constraints to obtain the actual control signal.

10. The temperature control method for titanium alloy forging as described in claim 2, characterized in that, In the process of globally optimizing the parameters of the residual dynamic neural network using the differential evolution algorithm, the global optimization process includes: A population containing multiple candidate control trajectories is randomly generated for one generation; For each candidate trajectory within the population, the refining model is invoked for simulation to obtain the full temperature evolution of the core and surface of the forging under the candidate trajectory. Check whether the temperature evolution process obtained from the simulation satisfies the metallurgical constraints throughout. If it does not, apply a penalty value to the candidate trajectory or eliminate it directly. For a trajectory that satisfies all constraints, calculate the total cost using cost constraints and use the total cost as the fitness. Based on fitness, next-generation candidate trajectories are generated through selection, crossover, mutation, or velocity-position update operations; Repeat the above process until the algorithm converges or reaches the maximum number of iterations.

11. A control system, characterized in that, The control system includes: An offline modeling module is used to acquire the thermophysical parameters of forgings and construct a model library based on the thermophysical parameters. The model library is used to describe the thermodynamic behavior of forgings during heating and forging processes. The trajectory generation module is used to obtain the alloy constraints and cost constraints of the forging, and to obtain a dual-objective temperature control trajectory based on the alloy constraints and cost constraints. The state estimation module is used to obtain the temperature state correction gain matrix and obtain the distribution estimate of the temperature field of the forging based on the temperature state correction gain matrix and the model library. The feedforward control module is used to construct a coupled prediction model based on the distribution estimation of the temperature field of the forging and the target temperature control trajectory, and to obtain the feedforward control setpoint based on the coupled prediction model. The feedback control module is used to obtain the measured temperature of the forging surface and the predicted value of the coupled prediction model, obtain the feedback correction control quantity based on the measured temperature of the forging surface, the optimal control sequence and the predicted value of the coupled prediction model, and correct the coupled prediction model through the feedback correction control quantity. A temperature control module is used to obtain the actual control signal based on the feedforward control setpoint and the feedback correction control quantity.

12. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-10.

Images