Quenching process parameter optimization method

Abstract

The invention relates to the technical field of heat treatment analysis or optimization, and particularly discloses a quenching process parameter optimization method, which comprises the following steps of: 1, acquiring material performance parameters of a to-be-optimized workpiece; 2, establishing a three-dimensional finite element model of a workpiece to be optimized; step 3, establishing an orthogonal experiment table for finite element simulation; 4, establishing a multi-objective optimization mathematical model and a weight coefficient function; step 5, constructing a back propagation neural network prediction model GA-BPNN optimized based on a genetic algorithm; and step 6, performing process parameter optimization by using a deep reinforcement learning Duelling DQN algorithm. The technical problem that in the prior art, a quenching process parameter optimization method does not consider the characteristic that importance of different workpieces to different performance parameters in different application scenes is different, so that the optimized quenching process parameters are disjointed with actual application is solved.

Description

Technical Field

[0001] The invention relates to the technical field of heat treatment analysis or optimization, and in particular to a method for optimizing quenching process parameters. Background Art

[0002] Quenching is a core process in metal heat treatment, primarily used to improve a material's hardness, wear resistance, and mechanical strength. Its core principle involves heating a metal (such as steel or aluminum alloy) above its critical temperature, where it undergoes austenitization to homogenize the metal's internal structure. Subsequently, the material is rapidly cooled at a rate exceeding the critical cooling rate, undergoing a martensite or bainite phase transformation, improving mechanical properties such as hardness. This process is widely used in aluminum alloys, copper alloys, and steel.

[0003] Currently, there are many heat-treatable materials, such as spring steel. However, during the quenching process, excessively high austenitizing temperatures, short or long holding times, and too slow or too fast cooling rates will affect the final quenching effect of the material. Therefore, each workpiece requires a series of suitable quenching process parameters to achieve optimal mechanical properties through quenching. However, the quenching process involves multiple physical fields and physical quantities such as temperature, phase, and stress / strain, and the various physical fields influence (interact) with each other. It is very difficult to obtain the optimal quenching parameters for a new workpiece using theoretical analysis methods. Therefore, many workpieces need to undergo repeated quenching tests under different quenching process parameters before quenching process production to find the optimal quenching parameters for the workpiece. This method is costly and has a long cycle because it involves a large number of quenching experiments.

[0004] To reduce the cost and cycle time of quenching processes, the industry has adopted the use of finite element technology to model and simulate the quenching process. This technology, through simulation experiments, determines the optimal parameters for the quenching process. For example, patent publication number CN118839547A discloses a method for optimizing spray quenching process parameters. By establishing a finite element model with an embedded quenching factor method and conducting simulation experiments, the obtained simulation data is used to construct a neural network prediction model. This model is then optimized using a non-dominated genetic algorithm and a dung beetle algorithm. This provides a basis for selecting quenching process parameters and avoids the need for extensive quenching experiments.

[0005] Although the above patent solves the technical problem of high cost in determining the optimal process parameters by quenching experiments in traditional technology, the performance parameters of the workpiece will change after quenching, and the importance of these performance parameters on different workpieces is different. For example, for a stabilizer bar, the importance of the core hardness is greater than the maximum axial deformation, and the optimization method in the prior art only wants to bring the core hardness and the maximum axial deformation closer to the ideal value, but the two are intrinsically related. In other words, when the core hardness is as close to its ideal value as possible, the maximum axial deformation may not be as close to its ideal value as possible, and when the maximum axial deformation is as close to its ideal value as possible, the core hardness cannot be as close to its ideal value as possible. The optimization method in the prior art cannot bring both of them as close to the ideal value as possible. The final result is that the two performance parameters of the workpiece after quenching are optimized to the difference between the two and the ideal value and the minimum performance parameter. This results in the two performance parameters of the workpiece being unable to reach the optimal level through the quenching process, and they are all mediocre in different application scenarios. The optimization method does not take into account the different importance of different performance parameters for different workpieces in different application scenarios, and therefore cannot maximize the value of quenching. The optimized quenching process parameters are out of touch with actual applications. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for optimizing quenching process parameters to solve the technical problem that the quenching process parameter optimization method in the above-mentioned prior art does not take into account the different importance of different performance parameters for different workpieces in different application scenarios, and therefore the optimized quenching process parameters are out of touch with actual applications.

[0007] In order to solve the above problems, the technical solution adopted by the present invention is as follows: a method for optimizing quenching process parameters, comprising the following steps: Step 1: Obtain material performance parameters of the workpiece to be optimized; Step 2: Using the material performance parameters and workpiece dimensions obtained in step 1, a three-dimensional finite element model of the workpiece to be optimized is established; Step 3: Construct an orthogonal experimental table of quenching process parameters, perform a finite element simulation experiment using the three-dimensional finite element model in step 2, and output the performance parameters of the workpiece after the simulated quenching experiment. The performance parameters and the quenching process parameters together constitute a sample data set; Step 4: Establish a multi-objective optimization mathematical model based on the workpiece performance parameter requirements according to the workpiece application scenario, and rank the target performance parameters of the workpiece in order of importance. Based on the importance ranking, use the weight coefficient conversion method to establish the overall objective optimization function. The specific steps are as follows: S401, constructing a multi-objective optimization function for quenching process parameters based on the requirements of the workpiece application scenario on the workpiece performance parameters; S402, constraining each parameter according to the quenching process parameter range obtained in step 3; S403: Based on the relative importance of the workpiece performance parameters in the application scenario, the workpiece performance parameters are ranked in importance, and the ranking results are assigned different weight coefficients to each sub-objective optimization function. The objective function is minimized to complete the optimization. The overall objective optimization function is expressed as:

[0008] in 、 、 is a performance parameter. When the application scenario requires the maximum performance parameter, the inverse is used. When the requirement is the minimum, the parameter itself is used. 、 、 is the weight coefficient; S404: Establish an importance judgment matrix, calculate weight coefficients based on the importance judgment matrix, bring the weight coefficients into the overall objective function, and output the final overall objective optimization function F; Step 5: Using the sample data in step 3, a GA-BPNN prediction model based on genetic algorithm optimization back propagation neural network is constructed; In step 6, the prediction model GA-BPNN in step 5 is used as a reinforcement learning environment, and the deep reinforcement learning Dueling DQN algorithm is used to optimize the process parameters. The reward function optimized by the Dueling DQN algorithm is: .

[0009] The beneficial effects of this embodiment are: 1. When optimizing quenching parameters, existing quenching process methods typically aim to bring all post-quenching performance parameters closer to their ideal values. However, these parameters are inherently interconnected. When the optimized quenching process parameters maximize the approach to one performance parameter's ideal value, another related performance parameter may not be able to achieve its ideal value. For example, consider two performance parameters of an automotive stabilizer bar (core hardness and maximum axial deformation). Existing optimization methods typically aim to bring both the core hardness and maximum axial deformation closer to their ideal values. However, these two parameters are inherently interconnected. In other words, when the optimized process parameters maximize the approach to the ideal core hardness, the maximum axial deformation may not. Conversely, when the process parameters are adjusted to maximize the approach to the ideal maximum axial deformation, the core hardness may not be able to achieve its ideal value. Optimization methods in the prior art cannot achieve both ideal values. The final result can only optimize the two performance parameters to the difference between the ideal values ​​and the minimum performance parameter. The quenching parameters optimized by this optimization method will make the workpiece relatively balanced in all aspects, but each performance parameter cannot be optimized to its full potential. In actual use, each workpiece has different application scenarios, and the importance of different performance parameters in different application scenarios varies. For example, for a stabilizer bar, the core hardness is more important than the maximum axial deformation. Simply optimizing it to a relatively ideal equilibrium state for both will not bring the core hardness as close to the ideal value as possible, and thus cannot achieve optimal performance in application. However, when establishing a multi-objective optimization mathematical model, the present application first ranks the different performance parameters by relative importance and converts the ranking into a weight coefficient. The inverse of the weight coefficient is used as the reward function for the deep reinforcement learning Dueling DQN algorithm to optimize process parameters. In other words, the present application uses the weight coefficient conversion method to optimize the importance of the multi-objective optimization function and uses it as the final objective constraint to optimize the process parameters using the deep reinforcement learning Dueling DQN algorithm. Therefore, when optimizing the quenching parameters, the weights will be affected, and the performance parameters with the largest weights will be adjusted to the best priority. Therefore, in actual use, the present application can adjust and optimize the multi-objective optimization mathematical model according to the optimal parameters required during actual production, and the optimized quenching parameters are more targeted. Compared with the prior art, the present application can freely adjust a certain performance parameter of the workpiece after quenching to reach the maximum value, making the present application better suitable for quenching process optimization of workpieces in different application scenarios.

[0010] 2. In order to make the model closer to the actual quenching experiment, the existing quenching process parameter optimization method needs to input a large number of different process parameters for repeated finite element modeling and simulation experiments. It is impossible to quickly determine the degree of influence of each process parameter on the quenching result, which wastes computer resources. However, the present application introduces an orthogonal test table for simulation experiments. By utilizing the balanced dispersion of the orthogonal test table, the orthogonal table ensures that the level of each process parameter is evenly distributed in all experiments, avoiding the problem of local concentration of data and the emergence of a large amount of redundant data caused by random input of data for simulation experiments. The orthogonal test table greatly reduces the number of input process parameters required for the simulation experiment of the present application, effectively saving computer resources. Secondly, after using the orthogonal test table in the early stage of the present application, its data dispersion and representativeness are good. Later, when combining the deep reinforcement learning DuelingDQN algorithm to optimize the process parameters, it only needs to optimize between adjacent process parameters in the orthogonal test table, and its calculation amount is greatly reduced.

[0011] 3. The quenching process parameter optimization method of the present invention selects quenching process parameters in a standardized manner without the need for additional experiments when the parameters within the model are known. The process is based on a finite element model. Then, through a large number of simulation experiments, a process parameter and target parameter mapping model GA-BPNN based on a genetic algorithm optimized back propagation neural network is established. The objective function is given according to actual production requirements, and a multi-objective optimization mathematical model is established. Based on the established GA-BPNN, a deep reinforcement learning Dueling DQN algorithm is used to solve the optimization model for the optimal solution.

[0012] Furthermore, the quenching process parameters in step three are austenitizing temperature, holding time, quenching medium temperature, and quenching medium; and the performance data are maximum axial deformation, core hardness value, and surface axial residual tensile stress value.

[0013] Furthermore, the step three includes the following steps: S301 obtaining four quenching process parameter ranges of austenitizing temperature, holding time, quenching medium temperature, and quenching medium type from a database; S302 setting several levels for each quenching process parameter according to the four obtained quenching process parameter ranges to construct an orthogonal test table; S303 performing a simulation experiment using finite element simulation calculation according to the orthogonal test table to obtain the maximum axial deformation, core hardness, and axial residual stress of the workpiece under different parameter combinations.

[0014] Furthermore, the step five includes the following steps: S501, normalizing the simulation test data obtained in step three; S502, selecting an activation function and a loss function to construct a BP neural network; S503, expanding all weights and biases of the BP neural network into a one-dimensional vector to form a chromosome; S504, initializing the population, randomly generating N chromosomes as the parent group, generation=1; S505, performing fitness calculation; S506, selecting, crossover, and mutation to form the generation-th subgroup; S507, merging the generation-th parent group and the subgroup and performing fitness calculation; S508, selecting suitable individuals from the merged group as the generation+1th parent group, generation=generation+1; S509, judging whether generation exceeds the maximum number of population iterations, if so, exiting to obtain the Pareto solution and selecting the best chromosome from it, if not, going to S505; S510, using the best chromosome optimized by the genetic algorithm to initialize the BP neural network. The weights and biases of the network; S511, set learning, select the back propagation algorithm, and use the training set to iteratively update the weights; S512, finally the BP neural network prediction model can be obtained.

[0015] Furthermore, the activation function is selected as: ; The loss function uses the mean square error (MSE): .

[0016] Furthermore, the workpiece material is spring steel, and the specific steps of step 2 are as follows: S201, establishing a geometric model of the spring steel workpiece; S202, meshing the workpiece model; S203, assigning material properties to the workpiece model; S204, setting it to thermal-mechanical coupling analysis and setting the time step; S205, applying symmetry constraints to the symmetric boundaries; S206, applying thermodynamic boundary conditions; S207, setting initial temperature field conditions; S208, setting the heat transfer coefficient boundary conditions of the end face; S209, embedding the quenching factor into the finite element model.

[0017] Furthermore, the material performance parameters of the spring steel to be optimized obtained in step 1 include flow stress, Poisson's ratio, TTT curve, CCT curve, thermal conductivity, phase composition, specific heat capacity, latent heat, Young's modulus and volume expansion coefficient.

[0018] Furthermore, step six includes the following steps: S601, problem description and goal; S602, initialization of Dueling DQN algorithm model; S603, interaction with the environment and sample collection; S604, experience replay; S605, training the main network; S606, updating the target network; S607, action selection strategy update; S608, training termination condition; S609, using the model for process optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart representing the optimization method of the present application. DETAILED DESCRIPTION

[0020] The following is further described in detail through specific implementation methods: The present invention provides a method for optimizing parameters in a quenching process based on deep reinforcement learning, comprising the following steps: Step 1: Obtain material performance parameters of the workpiece to be optimized; Step 2: Using the material performance parameters and workpiece dimensions obtained in step 1, a three-dimensional finite element model of the workpiece to be optimized is established; Step 3: Construct an orthogonal experimental table of quenching process parameters, perform a finite element simulation experiment using the three-dimensional finite element model in step 2, and output the performance parameters of the workpiece after the simulated quenching experiment. The performance parameters and the quenching process parameters together constitute a sample data set; Step 4: Establish a multi-objective optimization mathematical model based on the workpiece performance parameter requirements according to the workpiece application scenario, and rank the target performance parameters of the workpiece in order of importance. Based on the importance ranking, use the weight coefficient conversion method to establish the overall objective optimization function. The specific steps are as follows: S401, constructing a multi-objective optimization function for quenching process parameters based on the requirements of the workpiece application scenario on the workpiece performance parameters; S402, constraining each parameter according to the quenching process parameter range obtained in step 3; S403: Based on the relative importance of the workpiece performance parameters in the application scenario, the workpiece performance parameters are ranked in importance, and the ranking results are assigned different weight coefficients to each sub-objective optimization function. The objective function is minimized to complete the optimization. The overall objective optimization function is expressed as:

[0021] in 、 、 is a performance parameter. When the application scenario requires the maximum performance parameter, the inverse is used. When the requirement is the minimum, the parameter itself is used. 、 、 is the weight coefficient; S404: Establish an importance judgment matrix, calculate weight coefficients based on the importance judgment matrix, bring the weight coefficients into the overall objective function, and output the final overall objective optimization function F; Step 5: Using the sample data in step 3, a GA-BPNN prediction model based on genetic algorithm optimization back propagation neural network is constructed; In step 6, the prediction model GA-BPNN in step 5 is used as a reinforcement learning environment, and the deep reinforcement learning Dueling DQN algorithm is used to optimize the process parameters. The reward function optimized by the Dueling DQN algorithm is: 。

[0022] In this embodiment, in step 1, the material performance parameters of the spring steel to be optimized are obtained, including flow stress, Poisson's ratio, TTT curve, CCT curve, thermal conductivity, phase composition, specific heat capacity, latent heat, Young's modulus and volume expansion coefficient.

[0023] In this embodiment, the workpiece is made of spring steel. In step 2, a three-dimensional finite element model of the spring steel to be optimized is established, including the following steps: S201, establishing a geometric model of the spring steel workpiece; S202, meshing the workpiece model; S203, assigning material properties to the workpiece; S204, set to thermal-mechanical coupling analysis and set the time step; S205, applying symmetry constraints to the symmetric boundaries; S206, applying thermodynamic boundary conditions; S207, setting initial temperature field conditions; S208, setting the heat transfer coefficient boundary condition of the end surface; S209, embedding the quenching factor into the finite element model.

[0024] Furthermore, in step three, an orthogonal experimental table is established for finite element simulation, including the following steps: S301 obtains four quenching process parameter ranges: austenitizing temperature, holding time, quenching medium temperature, and quenching medium type; S302 uses the ranges of the above four quenching process parameters to set several levels for each quenching process parameter to construct an orthogonal experimental table; S303 performs simulation experiments using finite element simulation calculations based on the orthogonal experimental table, and outputs the performance data of the spring steel workpiece after quenching under different parameter combinations: maximum axial deformation , core hardness value , surface axial residual tensile stress value The performance data of the spring steel workpiece after quenching and the four quenching process parameters together constitute the sample data set.

[0025] In this embodiment, in step 5, a GA-BPNN prediction model based on genetic algorithm optimization back propagation neural network is constructed, which includes the following steps: S501, obtain sufficient sample data from finite element analysis and divide it into training set, validation set and test set. Use Min-Max normalization to limit the data value range to [0, 1], continuous type: , discrete type: Where 0: water, oil: 1; S502, set the network structure. According to the number of neurons in the input layer of 4 and the number of neurons in the output layer of 3, set the hidden layers 1, 2 and 3. The number of neurons in each layer uses the empirical formula: , ; Choose an activation function:

[0026] The loss function uses the mean square error (MSE):

[0027] S503, all weights and biases of the BP neural network are expanded into one-dimensional vectors to form chromosomes.

[0028] S504, initialize the population and randomly generate N chromosomes as the parent group. generation=1 ; S505, perform fitness calculation; the fitness function evaluates the BP neural network performance corresponding to each chromosome, and the fitness function is the model on the training set. MSE Reverse value, where ϵ is a small positive number to prevent division by zero errors; S506, selection, crossover, and mutation form the generation Daiziqun; S507, p. generation The parent group and the child group are merged and the fitness is calculated; S508, select suitable individuals from the merged group as the first generation+1 Father group, generation= generation+1 ; S509, judgment generationWhether the maximum population iteration number is exceeded, if so, jump out to obtain the Pareto solution and select the best chromosome from it, if not, go to S08; S510, using genetic algorithm to optimize the best chromosome, initialize BP The weights and biases of the network; S511, set learning rate η=0.01 , using the back propagation algorithm Adam , iteratively update the weights using the training set; S512, can finally be obtained BP Neural network prediction models.

[0029] In this embodiment, in step 4, establishing a multi-objective optimization mathematical model includes the following steps: S401, constructing a multi-objective optimization function for quenching process parameters based on the requirements of the workpiece application scenario on the workpiece performance parameters; Establish a mathematical model for the optimization problem. Determine the target quantity of output according to the actual production requirements and the application scenario of the workpiece and the target performance parameters of the workpiece after quenching. For example, the maximum axial deformation of the spring steel stabilizer bar workpiece needs to be output due to the actual application scenario. , core hardness value Maximum and surface axial residual tensile stress values There are three performance parameters, so the number of output objectives of the multi-objective optimization function is 3.

[0030] According to the application scenario, the ideal value of the performance parameter is constrained. For example, the application scenario of the spring steel stabilizer bar workpiece expects the maximum axial deformation to be Minimum, core hardness value Maximum and surface axial residual tensile stress values Minimum, so the multi-objective optimization function of quenching process parameters is constructed as follows:

[0031] in is the austenitizing temperature, t For the holding time, Tc is the quenching medium temperature, h It is the quenching medium type.

[0032] S402: Constrain each parameter based on the quenching process parameter range obtained in step 3. For example, if the workpiece is a stabilizer bar made of spring steel, first crawl the experimental data related to quenching of spring steel from the Internet to build a database. Based on the upper and lower limits of the process parameters in the database, set the constraints of the multi-objective optimization function as follows:

[0033] S403: Based on the relative importance of the workpiece performance parameters in the application scenario, the workpiece performance parameters are ranked in importance, and the ranking results are assigned different weight coefficients to each sub-objective optimization function. The objective function is minimized to complete the optimization. The overall objective optimization function is expressed as:

[0034] in 、 、 is a performance parameter. When the application scenario requires the maximum performance parameter, the inverse is used. When the requirement is the minimum, the parameter itself is used. 、 、 is the weight coefficient; For example: The stabilizer bar workpiece of spring steel needs to output the maximum axial deformation , core hardness value Maximum and surface axial residual tensile stress values Three performance parameters, and the ideal value is expected to be: Maximum axial deformation Minimum, core hardness value Maximum and surface axial residual tensile stress values smallest; The weight coefficient conversion method is used to give each sub-goal optimization function according to the relative importance of the goal. Different weight coefficients The optimization is completed by minimizing the objective function. The overall objective optimization function is expressed as:

[0035] The weight coefficient is determined based on the performance requirements of the workpiece according to the application scenario of the workpiece and the actual production and manufacturing experience. For example: In terms of the importance of the automobile stabilizer bar to the quenching effect, the core hardness is greater than the maximum axial deformation and greater than the surface axial residual tensile stress. Minimizing quenching deformation and reducing residual stress as much as possible while ensuring that the hardness meets the use requirements is an inherent requirement for process parameter optimization. That is, the importance of the hardness of the stabilizer bar core is slightly greater than the maximum axial deformation, and is obviously more important than the surface axial residual tensile stress. The maximum axial deformation is slightly more important than the surface axial residual tensile stress. The number 3 is used to represent the comparison result of "slightly greater importance", the number 5 is used to represent the comparison result of "obviously important", and the number 1 is used to represent the comparison result of "equal importance". Therefore, the optimization importance judgment matrix between the sub-goals of the stabilizer bar quenching process is shown in Table 1:

[0036] The maximum eigenvalue of the matrix can be obtained by calculation ， for The corresponding eigenvector is calculated to obtain the eigenvector The three values ​​are the weight coefficients of each sub-optimization objective, and we get:

[0037] In this embodiment, in step six, deep reinforcement learning is used Dueling DQN The algorithm performs process parameter optimization, including the following steps: S601 uses the prediction model established in step 3 as an environment for reinforcement learning; S602 transforms the process parameter optimization task into a reinforcement learning problem. ,action , the reward function ; S603 Initialize discount factor , learning rate ， —Exploration rate in greedy strategy and decay rules, experience replay pool size; S604 Build Estimate A neural network with an input layer input state , the hidden layer is shared by the fully connected layer, which is used to extract state features, and the activation function is used ReLU , branch layer (a branch estimates the state value , and the other branch estimates the action advantage ) and the output layer uses the formula ; S605 Randomly initialize the main network , copy the main network parameters to the target network ; S606 Experience replay pool is initialized to empty and used to store samples ; S607 from the current state Start and select actions according to the ε-greedy strategy ; S608 Execution Action , calculate the new state through the prediction model of process parameters and corresponding performance indicators, as well as rewards ; S609 will sample Deposit into experience replay pool ; S610 When the number of samples in the experience replay pool is sufficient, randomly sample a small batch , for each sample, calculate the target Q value , if the status Is the terminal state: ,otherwise: ), The target network Output value; S611 training The neural network of the value is used to calculate the current state action Q value , using mean square error (MSE) as the loss function: , use gradient descent to update the parameters of the main network ; S612 updates the target network and synchronizes the parameters of the main network to the target network at regular intervals: ; S613 action selection strategy update, using the exploration rate decay mechanism to gradually reduce , moving from more exploration in the early stages to development in the later stages: ; S614 When the preset number of training rounds is reached or the reward value is stabilized at a high level, the training is terminated; S615 uses the trained main network to perform process optimization, given the initial process parameters , predicting each action through the network Q Value. Select Q The action with the largest value , adjust the process parameters according to the action, and iterate the optimization until the target is met or convergence is achieved.

[0038] The above is only an embodiment of the present invention, and the common knowledge such as the specific structure and characteristics of the scheme is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for optimizing quenching process parameters, characterized in that: The following steps are involved: Step 1: Obtain material performance parameters of the workpiece to be optimized; Step 2: Using the material performance parameters and workpiece dimensions obtained in step 1, a three-dimensional finite element model of the workpiece to be optimized is established; Step 3: Construct an orthogonal experimental table of quenching process parameters, perform a finite element simulation experiment using the three-dimensional finite element model in step 2, and output the performance parameters of the workpiece after the simulated quenching experiment. The performance parameters and the quenching process parameters together constitute a sample data set; Step 4: Establish a multi-objective optimization mathematical model based on the workpiece performance parameter requirements of the workpiece application scenario, including the following steps: S401, constructing a multi-objective optimization model for quenching process parameters based on the requirements of the workpiece application scenario for the workpiece performance parameters; S402, constraining each parameter according to the quenching process parameter range obtained in step 3; S403: The performance parameters of the workpieces are ranked according to their importance based on the application scenario, and the ranking results are assigned different weight coefficients to the sub-objective optimization functions. The optimization is completed by minimizing the objective function. The overall objective optimization function is expressed as: in 、 、 is a performance parameter. When the application scenario requires the maximum performance parameter, the inverse is used. When the requirement is the minimum, the parameter itself is used. 、 、 is the weight coefficient; S404: Establish an importance judgment matrix, calculate the weight coefficient through the importance judgment matrix, bring the weight coefficient into the overall objective function, and output the final overall objective optimization function. F ; Step 5: Use the sample data in step 3 to build a prediction model based on genetic algorithm optimized back propagation neural network GA-BPNN ; Step 6: The prediction model in step 5 GA-BPNN Using deep reinforcement learning as a reinforcement learning environment Dueling DQN The algorithm is used to optimize the process parameters, where Dueling DQN The reward function optimized by the algorithm is: .

2. The method for optimizing quenching process parameters according to claim 1, characterized in that: The quenching process parameters in step three are austenitizing temperature, holding time, quenching medium temperature, and quenching medium; the performance parameters are maximum axial deformation, core hardness value, and surface axial residual tensile stress value.

3. The method for optimizing quenching process parameters according to claim 2, characterized in that: The step three includes the following steps: S301 obtains four quenching process parameter ranges of austenitizing temperature, holding time, quenching medium temperature, and quenching medium type from the database; S302 sets several levels for each quenching process parameter based on the obtained four quenching process parameter ranges to construct an orthogonal test table; S303 performs simulation experiments using finite element simulation calculations based on the orthogonal experimental table to obtain the maximum axial deformation, core hardness, and axial residual stress of the workpiece to be optimized under different parameter combinations.

4. The method for optimizing quenching process parameters according to claim 1, wherein: The step five includes the following steps: S501, normalizing the sample data obtained in step 3; S502, selecting an activation function and a loss function to construct a BP neural network; S503, expanding all weights and biases of the BP neural network into a one-dimensional vector to form a chromosome; S504, population initialization, randomly generate N chromosomes as parent groups, generation=1 ; S505, performing fitness calculation; S506, selection, crossover, and mutation form the generation Daiziqun; S507, the generation-generation parent group and child group are merged and fitness is calculated; S508, select suitable individuals from the merged group as the first generation+1 Father group, generation=generation +1 ; S509, judgment generation Whether the maximum population iteration number is exceeded, if so, jump out to obtain the Pareto solution and select the best chromosome from it, if not, go to S505; S510, using genetic algorithm to optimize the best chromosome, initialize BP The weights and biases of the network; S511, set up learning, select the back propagation algorithm, and use the training set to iteratively update the weights; S512, can finally be obtained BP Neural network prediction models.

5. The method for optimizing quenching process parameters according to claim 4, characterized in that: The activation function selection: ; The loss function uses the mean square error (MSE): .

6. The method for optimizing quenching process parameters according to claim 1, wherein: The workpiece material to be optimized is a spring steel workpiece, and the specific steps of step 2 are as follows: S201, establishing a geometric model of the spring steel workpiece; S202, meshing the workpiece model; S203, assigning material properties to the workpiece model; S204, set to thermal-mechanical coupling analysis and set the time step; S205, applying symmetry constraints to the symmetric boundaries; S206, applying thermodynamic boundary conditions; S207, setting initial temperature field conditions; S208, setting the heat transfer coefficient boundary condition of the end surface; S209, embedding the quenching factor into the finite element model.

7. The method for optimizing quenching process parameters according to claim 6, characterized in that: The material performance parameters of the spring steel to be optimized obtained in the step 1 include flow stress, Poisson's ratio, TTT curve, CCT curve, thermal conductivity, phase composition, specific heat capacity, latent heat, Young's modulus and volume expansion coefficient.

Images