Thermal Diffusion Digital Twin Model, Temperature Field Real-Time Optimization Control Model and Method for the Molten Casting and Solidification Process of Explosives

Abstract

The present invention discloses a thermal diffusion digital twin model, a temperature field real-time optimization control model and a method for the explosive melting and solidification process. By constructing a data-driven thermal diffusion digital twin model, the real-time simulation of the temperature field of the explosive melting and casting process can be realized, solving the problem that the temperature field distribution of the entire solidification process cannot be quickly obtained in the prior art; by constructing a temperature field optimization control model, the global temperature field is monitored in real time and the future state is predicted based on the thermal diffusion digital twin model, and the process parameters are adjusted to realize the temperature field control of the melting and casting process, solving the problems of many molding defects and poor quality stability caused by the constraints such as the inability to monitor the internal temperature field of the melting and casting process and the inability to optimize the process parameters according to the actual temperature field distribution. The temperature field optimization control method for the explosive melting and casting solidification process is adopted to collect wall temperature data in real time, and the process parameters in the explosive melting and casting process are adjusted in real time to realize the temperature field control of the melting and casting solidification process.

Description

Technical Field

[0001] The present invention belongs to the technical field of data analysis, and specifically relates to a thermal diffusion digital twin model, a real-time optimization control model and method for the temperature field in the process of explosive casting and solidification. Background Art

[0002] As the most important factor affecting the quality of explosive casting and forming, improper control of the temperature field is likely to form defects such as shrinkage cavities, porosity, and cracks. Therefore, it is of great significance to monitor the temperature field distribution in the explosive casting process in real time and dynamically adjust the process parameters based on this to reduce the generation of explosive defects and improve the quality of the casting process.

[0003] Currently, the main methods for studying the temperature field include experimental methods, finite element numerical simulation methods, and machine learning methods, etc.

[0004] Experimental method: Under normal circumstances, experimental research under the same working conditions is the most effective and accurate means to obtain the dynamic change law of the temperature field. However, since physical experiment temperature sensors are only convenient for measuring surface temperature and it is difficult to measure internal temperature, it is difficult to monitor the dynamic change of the temperature field during the casting process through the experimental method.

[0005] Numerical simulation method: With the development of computational fluid dynamics, numerical simulation methods such as the finite element method have been widely used. And with the improvement of relevant theoretical models, the calculation accuracy has basically reached the requirements of engineering practical applications. Currently, it is already possible to obtain the temperature field distribution during the solidification process. However, for the casting process, due to the large model and large number of grids in CFD numerical simulation, the simulation workload is huge, the calculation efficiency is low, the operation is complex and time-consuming. Therefore, it can only be used for pre-simulation and cannot perform real-time simulation and monitoring of the actual temperature field affected by various dynamic uncertain factors during the forming process. Therefore, it cannot support the real-time determination of quality and process adjustment during the process.

[0006] Machine learning: Since the model trained by machine learning technology using a large amount of historical data can discover the laws contained in the data, compared with traditional numerical methods, it does not need to understand the complex physical equations therein, which will reduce time and calculation costs and can achieve the function of rapid prediction. However, currently, the temperature prediction based on machine learning mainly focuses on the temperature change law at a certain position and mainly fits the law through non-linearity. At the same time, machine learning is a black box model without interpretability, and it does not effectively utilize the similarity behind the physical process, requiring a large amount of data, which is difficult to achieve in engineering.

[0007] Temperature field control is to make corresponding adjustments by analyzing the current temperature field data in real time, and apply the decision-making results to the production workshop, ultimately keeping the temperature field in the casting process within a reasonable range. The existing temperature field control and management methods have the following deficiencies: First, they mainly rely on the experience of staff, with poor accuracy, low efficiency and inability to monitor the temperature field status in real time, resulting in no basis for quality control and poor reliability; Second, they are obtained through historical data or finite element model analysis. Although they can improve the casting quality to a certain extent, they lack timeliness and fidelity and are difficult to meet the requirements. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a thermal diffusion digital twin model, a temperature field real-time optimization control model and method for the explosive casting and curing process. Through the constructed data-driven thermal diffusion digital twin model, the real-time simulation of the temperature field in the explosive casting process can be realized, and the problem of being unable to quickly obtain the temperature field distribution in the entire curing process in the existing technology can be solved; Through the temperature field optimization control method, the process parameters are dynamically optimized to realize the temperature field control in the casting process, and the problems such as many forming defects and poor quality stability caused by constraints such as the inability to monitor the internal temperature field in the casting process and the inability to optimize the process parameters according to the actual temperature field distribution are solved.

[0009] To achieve the above purpose, the present invention provides the following technical solutions:

[0010] The present invention first proposes a method for constructing a thermal diffusion digital twin model for the explosive casting and curing process, including the following steps:

[0011] Step 1: Construct a finite element model for the explosive casting and curing process;

[0012] Step 2: Conduct numerical simulation on the explosive casting and curing process. During the simulation process, randomly change the process parameters composed of the temperature of the hot mandrel, the insertion depth of the hot mandrel and the magnitude of the applied pressure, and extract the node data at all time steps during the curing process. The node data includes node coordinates and node temperatures; Synthesize the node data at multiple time steps into a set of temperature field distribution data with process parameters changing over time, and use the temperature field distribution data to construct a data set;

[0013] Step 3: Construct a data-driven thermal diffusion digital twin model

[0014] 31) Randomly extract node data from the temperature field distribution data at the same time step as local information, and encode the node data from the input space into the latent space through an encoder to obtain the initial node state of the function local interpolation;

[0015] 32) Realize spatial discretization through T rounds of message passing in the GNN model, update the node data and globalize the local information;

[0016] 33) Use the decoder to map from the latent space to the output space and predict the temperature field distribution data at the next time step;

[0017] 34) Compare whether the error between the predicted temperature field distribution data at the next time step and the simulated temperature field distribution data at the next time step is less than the set threshold: If so, obtain the data-driven thermal diffusion digital twin model; if not, update the GNN model parameters, use the temperature field distribution data at the next time step as the input, and execute step 31).

[0018] Further, in the first step, the method for constructing the finite element model of the explosive casting and solidification process is as follows:

[0019] 11) Select the mathematical model of the explosive casting and solidification forming process;

[0020] 12) Establish a finite element model of the explosive casting and solidification process according to the actual working conditions, perform preprocessing on the numerical simulation of the explosive casting and solidification process, and complete the mesh division;

[0021] 13) Set the boundary conditions and initial parameters, and randomly change the process parameters for numerical simulation.

[0022] Further, in the second step, normalize the temperature field distribution data.

[0023] Further, in step 32), the node temperature is updated using the temperature transfer function:

[0024]

[0025] where f(r i ) represents the temperature of the spatial node r i ; V j represents the ratio of the mass to density of the surrounding spatial nodes; W(|r i - r j |, h) represents the influence of the distance between the spatial node r i and the spatial node r j on temperature diffusion, h represents the temperature diffusion range; |r i - r j | represents the distance between the spatial node r i and the spatial node r j .

[0026] Further, in step 33), the method for predicting the temperature field distribution data at the next time step is:

[0027] T n+1 = T n + △t dT

[0028] where Tn Indicates the temperature of the current time step; T n+1 represents the temperature of the next time step; △t represents the time interval of a time step; dT represents the temperature change.

[0029] The present invention also discloses a method for constructing a temperature field optimization control model during explosive melting and solidification, comprising the following steps:

[0030] S1: Randomly initialize all parameters w of the Q network, initialize all states and actions of the Q network based on the parameters w, and obtain the corresponding value Q, while clearing the experience replay pool D;

[0031] S2: Collect wall temperature data, use the data constructed by the above method to drive the heat diffusion digital twin model to simulate the temperature field in real time, and obtain the current state of the temperature field s j and global temperature field simulation data, and use the global temperature field simulation data as the characteristic vector φ(s j );

[0032] S3: Use the feature vector φ(s) in the Q network j ) as input, and obtain the Q value output corresponding to all actions of the Q network; use the ∈-greedy method to select the corresponding action a in the current Q value output j , j is the number of iterations;

[0033] S4: In state s j Execute the current action a j , using the data-driven heat diffusion digital twin model to simulate in real time to obtain the new state s j+1 And the corresponding eigenvector φ(s j+1 ) and reward r j+1 , determine the state s j+1 Is it the final state: if yes, terminate the iteration and obtain the temperature field optimization control model; if no, execute step S5);

[0034] S5: The obtained four-tuple (s j , a j , r j+1 ,s j+1 ) is added to the experience replay pool D, and m samples (s) are sampled from the experience replay pool D i , a i , r i+1 ,s i+1 ), where i = j-m+1, j-m+2, ···, j-1, j; j ≥ m; calculate the current target Q value y j , using the target Q value y j Calculate the mean square error loss function and use it to update the Q network parameters;

[0035] Update the status to make s j = s j+1 , r j = r j+1 , j = j + 1;

[0036] Execute step S3.

[0037] Furthermore, in the said step S5, the current target Q value y j is:

[0038]

[0039] where γ represents the attenuation value; represents the maximum estimated value of the state s j+1 under the action a j+1 .

[0040] Furthermore, in the said step S5, the mean square error loss function is:

[0041]

[0042] where m represents the number of samples; j represents the current iteration number; w represents all the parameters of the Q network.

[0043] A method for optimizing and controlling the temperature field in the process of explosive casting and solidification, which real-time collects the wall temperature data and uses the temperature field optimization control model constructed by the method described above to real-time regulate the process parameters in the explosive casting process, so as to realize the temperature field control in the casting and solidification process.

[0044] The beneficial effects of the present invention are as follows:

[0045] Method for constructing thermal diffusion digital twin model in explosive casting and solidification process of the present invention. First, by constructing a finite element model of the explosive casting and solidification process, through finite element numerical simulation, data for establishing a data-driven thermal diffusion digital twin model is provided. Specifically, during the simulation process, three process parameters, namely the temperature of the hot mandrel, the insertion depth of the hot mandrel, and the applied pressure, are randomly changed, and the nodal data at all time steps is extracted. Sufficient temperature field distribution data is collected and a data set is created. Then, a data-driven thermal diffusion digital twin model is constructed. Since it is difficult to measure the global temperature data during the casting process, the present invention randomly samples and extracts nodal data as local information, and through the GNN model, the local information is propagated to anywhere in the space, and the prediction from local temperature data to the global temperature field is realized through learning the thermal diffusion mechanism, making the local information global. Finally, the data-driven thermal diffusion digital twin model is optimized until the accuracy approaching the output target is obtained, and an optimized data-driven thermal diffusion digital twin model considering the temperature of the hot mandrel, the insertion depth of the hot mandrel, and the applied pressure is obtained, realizing the learning of the temperature field distribution law of the explosive solidification process under different times and different process parameters. In actual use, only by the wall temperature data collected in real time by the thermocouple, the global temperature field distribution can be obtained based on the data-driven thermal diffusion digital twin model, realizing the rapid real-time simulation of the temperature field, and solving the problem in the prior art that the temperature field distribution of the entire explosive casting and solidification process cannot be obtained quickly.

[0046] Method for constructing temperature field optimization control model in explosive casting and solidification process of the present invention. Using the real-time simulation results of the explosive casting and solidification process by the data-driven thermal diffusion digital twin model to quickly identify the temperature field distribution characteristics of the explosive casting and solidification process, according to the determination criteria of the temperature field distribution and the occurrence of casting defects, the process parameters are optimized in real time to realize the real-time regulation of the casting temperature field distribution, thereby improving the quality of explosive casting. The data-driven thermal diffusion digital twin model only needs to monitor the wall temperature in real time through the thermocouple, and the global temperature field distribution can be quickly obtained based on the thermal diffusion digital twin model, realizing the real-time simulation of the temperature field in the explosive casting process, and solving the problem in the prior art that the temperature field distribution of the entire solidification process cannot be obtained quickly. On the basis that the data-driven thermal diffusion digital twin model can perform real-time and rapid simulation on the explosive solidification process, the method of deep reinforcement learning is used to dynamically optimize the process parameters based on the temperature field distribution in the casting process. Through the real-time collection of the wall temperature data by the thermocouple, the global temperature field is monitored in real time and the future state is predicted based on the thermal diffusion digital twin model, and the process parameters are adjusted accordingly to realize the control of the temperature field in the casting process, and solve the problems such as many forming defects and poor quality stability caused by constraints such as the inability to monitor the internal temperature field in the casting process and the inability to optimize the process parameters according to the actual temperature field distribution.

[0047] In addition, due to the poor accuracy, low efficiency and inability to monitor the temperature field status in real time of traditional temperature field control methods, there is no basis for quality control, and the reliability is poor, making it difficult to meet production requirements. Combining the powerful decision-making ability of reinforcement learning, in the process of constructing a data-driven thermal diffusion digital twin model, by learning the influence of different process parameters on the temperature field during the casting process and how to automatically adjust the process parameters at the next moment based on the current temperature field distribution to improve the temperature field distribution, and then on the basis of real-time acquisition of temperature field data, according to the defect occurrence judgment criterion based on the temperature field during the casting process, the optimal process parameters are decided in real time to regulate the subsequent temperature field distribution, ultimately realizing real-time and dynamic control of the temperature field during the casting process, thereby reducing the occurrence of various defects during the casting process and improving the quality of explosive casting. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to make the objectives, technical solutions and beneficial effects of the present invention clearer, the present invention provides the following drawings for description:

[0049] Figure 1 It is the schematic diagram of the embodiment of the method for constructing the thermal diffusion digital twin model of the explosive casting and solidification process of the present invention;

[0050] Figure 2 It is the schematic diagram of the constructed finite element model;

[0051] Figure 3 It is the mesh division diagram of the finite element model;

[0052] Figure 4 It is the result diagram of the numerical simulation temperature field distribution of the finite element model;

[0053] Figure 5 It is the schematic diagram of synthesizing the node data into the temperature field distribution data;

[0054] Figure 6 It is the corresponding relationship diagram between the process parameters and the temperature field distribution data;

[0055] Figure 7 It is the schematic diagram of the embodiment of the method for constructing the temperature field optimization control model of the explosive casting and solidification process of the present invention;

[0056] Figure 8 It is the schematic diagram of the modeling principle of the temperature field optimization control model;

[0057] Figure 9 It is the schematic diagram of the embodiment of the temperature field optimization control method of the explosive casting and solidification process of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited do not limit the present invention.

[0059] As Figure 1 shown, the method for constructing a thermal diffusion digital twin model for the explosive casting and solidification process of this embodiment includes the following steps.

[0060] Step 1: Construct a finite element model for the explosive casting and solidification process, the method is as follows:

[0061] 11) Select the mathematical model for the explosive casting and solidification forming process, that is, the heat transfer model for the cooling process;

[0062] 12) Establish a finite element model for the explosive casting and solidification process according to the actual working conditions. As Figure 2 shown, perform preprocessing on the numerical simulation of the explosive casting and solidification process, complete the mesh division, as Figure 3 shown;

[0063] 13) Set the boundary conditions and initialization parameters, and randomly change the process parameters for numerical simulation, as Figure 4 shown.

[0064] Step 2: Perform numerical simulation on the explosive casting and solidification process. During the simulation, randomly change the process parameters composed of the temperature of the hot mandrel, the insertion depth of the hot mandrel, and the magnitude of the applied pressure, and extract the node data at all time steps during the solidification process. The node data includes node coordinates and node temperatures; synthesize the node data at multiple time steps into a set of temperature field distribution data of the process parameters changing with time, and use the temperature field distribution data to construct a data set. Specifically, the method for creating the data set is:

[0065] 21): Synthesize the node data at each time step of the solidification process into a set of temperature field distribution data changing with the real time series under the process conditions, as Figure 5 shown;

[0066] 22) In order to establish the relationship between different process parameters and the solidification time and the temperature field distribution, make the process parameters and the solidification time correspond one-to-one with the temperature field distribution data, as Figure 6 shown;

[0067] 23) In order to eliminate the difference between the data orders of magnitude, better reflect the mutual relationship between various factors, and reduce the error in the data processing process, perform normalization processing on the temperature field distribution data.

[0068] Step 3: Construct a data-driven thermal diffusion digital twin model

[0069] 31) Randomly extract node data from the temperature field distribution data at the same time step as local information, and encode the node data from the input space into the latent space through the encoder to obtain the initial node state for local interpolation of the function.

[0070] 32) Achieve spatial discretization through T rounds of message passing in the GNN model, update the node data, and globalize the local information.

[0071] 33) Use the decoder to map from the latent space to the output space and predict the temperature field distribution data at the next time step.

[0072] 34) Compare whether the error between the predicted temperature field distribution data at the next time step and the simulated temperature field distribution data at the next time step is less than the set threshold: If so, obtain the data-driven thermal diffusion digital twin model; if not, update the GNN model parameters, use the temperature field distribution data at the next time step as the input, and execute step 31).

[0073] When the liquid explosive fills the projectile mold, the solidification of the explosive mainly considers unsteady heat conduction, which is usually described by the unsteady heat conduction partial differential equation, and this is also the traditional numerical simulation method. Based on this, the data-driven thermal diffusion digital twin model considers representing the fluid domain as a dense set of particles (nodes), and each node will contain temperature information and position information. Since heat diffuses to the surrounding area when the temperature is different, the temperature of each node will affect the temperature of the surrounding nodes. Normally, the law of node temperature diffusion is related to the position, temperature, and diffusion distance of the node itself, and its heat diffusion property can be represented by a continuous function. Therefore, if the positions and temperature values of several nodes are known, the temperature can be diffusively transmitted through the temperature propagation function of the nodes. Finally, the output value can be read from any node in the system by interpolating and weighted summing between the node values, and the value of any new query point in space can be predicted. Therefore, in this embodiment, the data-driven thermal diffusion digital twin model is to establish a temperature propagation model through the mechanism of heat transfer to predict the temperature. Obviously, the data-driven thermal diffusion digital twin model in this embodiment does not learn to directly map the input to the global solution, but learns simpler local models and then uses message passing to derive the global solution. It can have far fewer nodes than the training examples and sparse connectivity, making its computational efficiency higher.

[0074] When the liquid explosive fills the projectile mold, it satisfies Fourier's law and is described by the unsteady heat conduction partial differential equation:

[0075]

[0076] The three-dimensional heat conduction formula after simplification and transformation is:

[0077]

[0078] In the formula:

[0079]

[0080] where ρ represents the average density of the liquid phase and the solid phase, ρ L represents the liquid phase density, and ρ s represents the solid phase density, f L represents the liquid phase volume fraction, and f s represents the solid phase volume fraction, c represents the isobaric specific heat capacity, T represents the temperature of the liquid medicine, t represents time, k represents the thermal conductivity, and L represents the latent heat of crystallization; k x represents the thermal conductivity in the X direction; k y represents the thermal conductivity in the Y direction; k z represents the thermal conductivity in the Z direction.

[0081] Perform time discretization on it, and derive the temperature T n at the next moment from the temperature T n+1 at the current moment:

[0082]

[0083] where T n represents the temperature at the current time step; T n+1 represents the temperature at the next time step; △t represents the time interval of one time step; dT represents the temperature change.

[0084] After time discretization, spatial discretization needs to be performed on it. Because the temperature of any particle not only changes with time but also diffuses heat to the surrounding particles. Normally, it is a continuous function, and through it, the influence degree of heat diffusion to the surrounding particles is characterized, which is mainly related to the temperature of the particle itself and the distance to other particles. Since the particle diffusion has a certain range, and the temperature of any particle can also be obtained by weighted summation of the surrounding particles. Thus, the node temperature is updated using the temperature transfer function:

[0085]

[0086] where f(r i ) represents the temperature of the spatial node r i ; V j represents the ratio of the mass to the density of the surrounding spatial nodes; W(|r i - r j |, h) represents the influence of the distance between the spatial node r i and the spatial node r j on temperature diffusion, h represents the temperature diffusion range; |r i - rj | represents the spatial node r i The distance between j and the spatial node r

[0087] For the method for constructing a thermal diffusion digital twin model in the explosive casting and solidification process of this embodiment, first, by constructing a finite element model of the explosive casting and solidification process, through finite element numerical simulation, data for establishing a data-driven thermal diffusion digital twin model is provided; specifically, during the simulation process, three process parameters, namely the temperature of the hot mandrel, the insertion depth of the hot mandrel, and the applied pressure, are randomly changed, the node data at all time steps is extracted, sufficient temperature field distribution data is collected and a data set is created; then, a data-driven thermal diffusion digital twin model is constructed. Since it is difficult to measure the global temperature data during the casting process, the present invention randomly samples and extracts node data as local information, and propagates the local information to anywhere in the space through the GNN model, and realizes the prediction from local temperature data to the global temperature field through learning the thermal diffusion mechanism, making the local information global; finally, the data-driven thermal diffusion digital twin model is optimized until the accuracy approaching the output target is obtained, and an optimized data-driven thermal diffusion digital twin model considering the temperature of the hot mandrel, the insertion depth of the hot mandrel, and the applied pressure is obtained, realizing the learning of the temperature field distribution law during the explosive solidification process under different moments and different process parameters; in actual use, only by the wall temperature data collected in real time by the thermocouple, the global temperature field distribution can be obtained based on the data-driven thermal diffusion digital twin model, realizing the rapid real-time simulation of the temperature field, and solving the problem in the prior art that the temperature field distribution of the entire explosive casting and solidification process cannot be obtained quickly.

[0088] As Figure 7 shown, this embodiment also proposes a method for constructing a temperature field optimization control model in the explosive casting and solidification process, including the following steps:

[0089] S1: Randomly initialize all parameters w of the Q network, initialize all states and actions of the Q network based on the parameters w, and obtain the corresponding value Q, and at the same time empty the experience replay pool D. Specifically, the parameters w include the number of iteration rounds T, the state feature dimension n, the action set A, the step size α, the decay factor γ, the exploration rate ∈, the Q network structure, and the number of samples m for batch gradient descent;

[0090] S2: Collect wall temperature data, use the data-driven thermal diffusion digital twin model constructed by the method as described above in this embodiment to perform real-time simulation on the temperature field, obtain the current state s of the temperature field j and the global temperature field simulation data, and use the global temperature field simulation data as the feature vector φ(s j ) of the Q network;

[0091] S3: Use the feature vector φ(s in the Q networkj ) As the input, obtain the Q-value outputs corresponding to all actions of the Q-network; select the corresponding action a using the ∈-greedy method from the current Q-value outputs j , where j is the number of iterations;

[0092] S4: In state s j Execute the current action a j , and use the data-driven thermal diffusion digital twin model to obtain the new state s through real-time simulation j+1 along with the corresponding feature vector φ(s j+1 ) and the reward r j+1 , and determine whether the state s j+1 is the final state: if so, terminate the iteration to obtain the temperature field optimization control model; if not, execute step S5);

[0093] S5: Add the obtained quadruple (s j , a j , r j+1 , s j+1 ) to the experience replay pool D, sample m samples (s i , a i , r i+1 , s i+1 ) from the experience replay pool D, where i = j - m + 1, j - m + 2, ···, j - 1, j; j ≥ m; calculate the current target Q-value y j , and use the target Q-value y j to calculate the mean squared error loss function, and update the Q-network parameters using the mean squared error loss function;

[0094] Update the state, let s j = s j+1 , r j = r j+1 , and j = j + 1;

[0095] Execute step S3.

[0096] Among them, the current target Q-value y j is:

[0097]

[0098] Among them, γ represents the attenuation value; represents the maximum estimated value of the state s j+1 under the action a j+1 .

[0099] The mean squared error loss function is:

[0100]

[0101] Among them, m represents the number of samples; j represents the current iteration number; w represents all the parameters of the Q network.

[0102] After establishing the thermal diffusion digital twin model, based on the wall temperature data collected in real time by the thermocouple, the global temperature field distribution can be obtained based on the thermal diffusion digital twin model, realizing the real-time simulation of the temperature field. Since the generation of defects in the casting process is closely related to the temperature field distribution, in order to reduce the generation of defects, it is necessary to evaluate the temperature field at each moment, and use the defect occurrence judgment criterion based on the casting temperature field as the temperature field optimization standard to optimize the process parameters. To achieve the temperature field control of the casting process, a deep reinforcement learning method is used to optimize the process parameters of the casting process in real time. A temperature field control model for the casting process is established based on deep reinforcement learning, which mainly consists of two parts: the environment and the agent. The environment consists of the thermal diffusion digital twin model, which is used to provide the temperature field changes at different times under different process parameters, used to construct the history, and provide training data for the temperature field control model. During the training process, first, the environment receives the action and the temperature state at the current moment, obtains the current reward and observation value according to the previous environment state, and inputs them into the agent. The agent outputs the next-step action and starts the next interaction. During the interaction between the agent and the environment (each time the agent interacts with the environment, the environment returns the reward, telling the agent whether the just action is good or bad), the network parameters are continuously updated using the deep reinforcement learning algorithm according to the target value, realizing the dynamic control of the temperature field in the casting process. In actual application, first, initialize the working state and set the initial process parameters. Then, based on the wall temperature collected in real time, obtain the global temperature field distribution at the current moment based on the thermal diffusion digital twin model. Finally, compare the temperature field data with the target value to check whether the quality requirements are met, and adjust the action (process parameters) at the next moment based on this. Repeat the above process until the entire casting process is completed.

[0103] As Figure 8 shown, the agent is the executing entity, which can operate the executing entity to make different choices (that is, actions), such as adjusting the process parameters and changing the temperature field distribution; the environment has one state after another (temperature field distributions at different times and under different process parameters), which is used to feedback the state to the agent; the action is to adjust the process parameters to change the environment, that is, the state; the reward is used to evaluate whether the action provided by the agent changes the environment state for better or worse. Generally speaking, during the casting process, the agent continuously adjusts the action according to the comparison between the temperature field and the target temperature field, changing the state, and finally realizing the real-time control of the temperature field.

[0104] The method for constructing an optimized control model of the temperature field in the explosive casting and solidification process of this embodiment uses a data-driven thermal diffusion digital twin model to quickly identify the distribution characteristics of the temperature field in the explosive casting and solidification process based on the real-time simulation results of the explosive casting and solidification process. According to the distribution of the temperature field and the judgment criteria for the occurrence of casting defects, the process parameters are optimized in real time to achieve real-time regulation of the distribution of the casting temperature field, thereby improving the quality of explosive casting. The data-driven thermal diffusion digital twin model only needs to monitor the wall temperature in real time through a thermocouple, and then can quickly obtain the global temperature field distribution based on the thermal diffusion digital twin model, realizing real-time simulation of the temperature field in the explosive casting process, and solving the problem in the prior art that the temperature field distribution of the entire solidification process cannot be obtained quickly. On the basis that the data-driven thermal diffusion digital twin model can perform real-time and rapid simulation of the explosive solidification process, the method of deep reinforcement learning is used to dynamically optimize the process parameters based on the temperature field distribution in the casting process. Through the real-time acquisition of the wall temperature data by the thermocouple, the global temperature field is monitored in real time and the future state is predicted based on the thermal diffusion digital twin model, and the process parameters are adjusted accordingly to achieve temperature field control in the casting process, solving problems such as many forming defects and poor quality stability caused by constraints such as the inability to monitor the internal temperature field in the casting process and the inability to optimize the process parameters according to the actual temperature field distribution.

[0105] In addition, due to the poor accuracy, low efficiency and inability to monitor the temperature field state in real time of the traditional temperature field control method, there is no basis and poor reliability for quality control, making it difficult to meet the production requirements. Combining the powerful decision-making ability of reinforcement learning, in the process of constructing the data-driven thermal diffusion digital twin model, by learning the influence of different process parameters on the temperature field in the casting process and how to automatically adjust the process parameters at the next moment to improve the temperature field distribution according to the current temperature field distribution, and then, on the basis of real-time acquisition of temperature field data, according to the defect occurrence judgment criteria based on the temperature field in the casting process, the optimal process parameters in the process are decided in real time to regulate the subsequent temperature field distribution, and finally, real-time and dynamic control of the temperature field in the casting process is realized, thereby reducing the occurrence of various defects in the casting process and improving the quality of explosive casting.

[0106] As Figure 9 shown, this embodiment also proposes an optimized control method for the temperature field in the explosive casting and solidification process, which collects wall temperature data in real time and uses the temperature field optimized control model constructed by the method described above in this embodiment to regulate the process parameters in the explosive casting process in real time to achieve temperature field control in the casting and solidification process.

[0107] The above-described embodiments are only preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for constructing a thermal diffusion digital twin model of the explosive casting and solidification process, characterized in that: It includes the following steps: Step 1: Construct a finite element model of the explosive casting and solidification process; Step 2: Conduct numerical simulation on the explosive casting and solidification process. During the simulation, randomly change the process parameters composed of the temperature of the hot mandrel, the insertion depth of the hot mandrel, and the applied pressure magnitude, and extract the node data at all time steps during the solidification process. The node data includes node coordinates and node temperature; Synthesize the node data at multiple time steps into a set of temperature field distribution data with process parameters varying over time, and use the temperature field distribution data to construct a data set; Step 3: Construct a data-driven thermal diffusion digital twin model 31) Randomly extract node data from the temperature field distribution data at the same time step as local information, and encode the node data from the input space into the latent space through an encoder to obtain the initial node state of function local interpolation; 32) Achieve spatial discretization through T rounds of message passing in the GNN model, update the node data and globalize the local information; 33) Use the decoder to map from the latent space to the output space and predict the temperature field distribution data at the next time step; 34) Compare whether the error between the predicted temperature field distribution data at the next time step and the simulated temperature field distribution data at the next time step is less than the set threshold: If so, obtain the data-driven thermal diffusion digital twin model; If not, update the GNN model parameters, use the temperature field distribution data at the next time step as the input, and execute step 31).

2. The method for constructing a thermal diffusion digital twin model of the explosive casting and solidification process according to claim 1, characterized in that: In the said step 1, the method for constructing a finite element model of the explosive casting and solidification process is as follows: 11) Select the mathematical model of the explosive casting and solidification forming process; 12) Establish a finite element model of the explosive casting and solidification process according to the actual working conditions, perform preprocessing on the numerical simulation of the explosive casting and solidification process, and complete the mesh generation; 13) Set boundary conditions and initialization parameters, and randomly change the process parameters for numerical simulation.

3. The method for constructing a thermal diffusion digital twin model of the explosive casting and solidification process according to claim 1, characterized in that: In the said step 2, the temperature field distribution data is normalized.

4. The method for constructing a thermal diffusion digital twin model of the explosive casting and solidification process according to claim 1, characterized in that: In the said step 32), the node temperature is updated using the temperature transfer function: where, f(r i ) represents the temperature of spatial node r i ; V j represents the ratio of the mass to density of surrounding spatial nodes; W(|r i - r j |, h) represents the influence of the distance between spatial node r i and spatial node r j on temperature diffusion, where h represents the temperature diffusion range; |r i - r j | represents the distance between spatial node r i and spatial node r j .

5. The method for constructing a thermal diffusion digital twin model of the explosive casting and solidification process according to claim 1, characterized in that: In the said step 33), the method for predicting the temperature field distribution data at the next time step is: T n+1 = T n + △tdT where, T n represents the temperature at the current time step; T n+1 represents the temperature at the next time step; △t represents the time interval of one time step; dT represents the temperature change.

6. A method for constructing a temperature field optimization control model of the explosive casting and solidification process, characterized in that: It includes the following steps: S1: Randomly initialize all parameters w of the Q network, initialize all states and actions of the Q network based on the parameters w, and obtain the corresponding value Q. At the same time, clear the experience replay pool D; S2: Collect wall temperature data, and use the data-driven thermal diffusion digital twin model constructed by the method described in any one of claims 1-5 to perform real-time simulation on the temperature field to obtain the current state s of the temperature field j and the global temperature field simulation data. Take the global temperature field simulation data as the feature vector φ(s j ) of the Q network; S3: Use the feature vector φ(s j ) as the input in the Q-network to obtain the Q-value outputs corresponding to all actions of the Q-network; Select the corresponding action a in the current Q-value output using the ε-greedy method j , where j is the number of iterations; S4: At state s j Execute the current action a j , and use the data-driven thermal diffusion digital twin model to perform real-time simulation to obtain a new state s j+1 and the corresponding feature vector φ(s j+1 ) and reward r j+1 , and determine whether the state s j+1 is the final state: if so, terminate the iteration to obtain the temperature field optimization control model; if not, execute step S5); S5: Add the obtained quadruple (s j , a j , r j+1 , s j+1 ) to the experience replay pool D, and sample m samples (s i , a i , r i+1 , s i+1 ) from the experience replay pool D, where i = j - m + 1, j - m + 2, ···, j - 1, j; j≥m; Calculate the current target Q-value y j , and use the target Q-value y j to calculate the mean squared error loss function and update the Q-network parameters using the mean squared error loss function; Update the status to make s j = s j+1 , r j = r j+1 , and j = j + 1; Execute step S3.

7. The method for constructing an optimized control model of the temperature field in the explosive casting and solidification process according to claim 6, characterized in that: In the step S5, the current target Q value y j is: where γ represents the attenuation value; represents the maximum estimated value of j+1 the state s j+1 under the action a.

8. The method for constructing an optimized control model of the temperature field in the explosive casting and solidification process according to claim 6 or 7, characterized in that: In the step S5, the mean square error loss function is: where, m represents the number of samples; j represents the current iteration number; w represents all the parameters of the Q network.

9. An optimized control method for the temperature field in the explosive casting and solidification process, characterized in that: The wall temperature data is collected in real time, and the process parameters in the explosive casting process are regulated in real time by using the optimized control model of the temperature field constructed by the method according to any one of claims 6-8, so as to realize the temperature field control of the casting and solidification process.

Images