A collaborative optimization method and system for an intelligent spinning process path

Abstract

The present invention provides a collaborative optimization method and system for an intelligent spinning process path. The method includes: selecting different spinning process parameters according to orthogonal design, and then obtaining the physical field state and formability quality parameters through finite element simulation and spinning experiments; establishing a neural network model to determine the mapping relationship between the physical field state and the formability quality parameters during part forming, determining the reward function, and training the neural network model; building a deep reinforcement learning framework, embedding the deep reinforcement learning framework into the spinning finite element model, and establishing a spinning forming process parameter control model; the agent in the spinning forming process parameter control model continuously interacts with the environment and continuously adjusts the spinning process parameters according to the return value to maximize the cumulative reward during the forming process and output the optimal spinning process path. The present invention can solve the problems of low accuracy, small applicable range, and limited optimized parameters in the existing spinning process parameter optimization methods.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent manufacturing, and particularly relates to a method and system for collaborative optimization of an intelligent spinning process path. Background Art

[0002] With the continuous improvement of the performance requirements for components in fields such as aerospace, national defense, and transportation, component manufacturing is developing towards high performance, low energy consumption, long life, etc. Among them, complex rotary thin-walled components, such as aircraft engine casings, missile fairings, and automobile wheels, as key components in the above fields, have a wide range of application requirements. Spinning forming, as a continuous local plastic forming process, has been widely used in the manufacturing of complex rotary thin-walled components due to its high process flexibility, simple die structure, low forming load, high production efficiency, etc. However, the shape and quality of the components are affected by various process parameters, such as spinning temperature, roller feed ratio, thinning rate, etc., and the relationship between them is highly non-linear, which poses a great challenge to the manufacturing of high-quality components. Therefore, how to quickly and accurately optimize the above process parameters is of great significance for improving product quality, reducing production costs, and shortening the R & D cycle.

[0003] Regarding the problem of optimizing the forming process parameters, the Chinese patent application with the publication number CN118133615A discloses a method for optimizing the process parameters of fiber metal super-hybrid circular tube power spinning. By conducting a spinning orthogonal experiment, an equation relating the spinning process parameters to the interfacial shear strength in the fiber circular tube is established, and then in combination with finite element simulation, the spinning process parameters corresponding to the best interfacial shear strength of the circular tube are determined. However, the parameters collected by this method have limited dimensions, and the accuracy of the corresponding relationship equation is not high, only being able to give a rough range of process parameters. The Chinese patent application with the publication number CN103182457A discloses a method for optimizing the spinning process parameters of aerospace thin-walled parts based on data mining. Taking the material parameters of the spinning blank and the spinning process parameters as input variables and the part quality index as the output variable, a quality prediction model for spinning thin-walled parts is established through the XGboost algorithm to achieve the optimization of the spinning process parameters. However, the establishment of the model requires a large amount of pre-defined label data and the model is only applicable to such processes, having certain limitations. Moreover, the algorithm itself will occupy a large amount of memory space when dealing with large-scale data, and the calculation cost is relatively high. The Chinese patent application with the publication number CN118607288A discloses a dynamic modeling prediction and online process optimization method for the spinning forming of curved parts. By discretizing the finite element simulation process of spinning forming into a series of continuous stages, the forming results at different stages under different spinning process parameters are obtained. Through the response surface method, a surrogate model between the process parameters and the forming results at different forming stages is established, and then according to the preset optimization objective, the best process parameters for each stage of spinning are obtained to achieve the process optimization of the entire spinning forming process. However, the optimization accuracy of the above method depends on the discretization degree of spinning, and a larger discretization degree will lead to a significant increase in the optimization process, and it can only achieve the optimization of the geometric accuracy of the part, ignoring the influence of process parameters on the mechanical properties of the workpiece. Summary of the Invention

[0004] The object of the present invention is: aiming at the deficiencies existing in the above-mentioned background technology, to provide a solution that can solve the problems of low accuracy, small applicable range, and limited optimized parameters of the existing spinning process parameter optimization methods.

[0005] To achieve the above object, the present invention provides a collaborative optimization method for an intelligent spinning process path, including the following steps:

[0006] S1, Select different spinning process parameters according to the orthogonal design, and then obtain the physical field state during part forming and the formability quality parameters of the part after forming through finite element simulation and spinning experiments;

[0007] S2, Establish a neural network model, determine the mapping relationship between the physical field state during part forming and the formability quality parameters, and determine the reward function through the weighted coefficient method, and train the neural network model;

[0008] S3. Build a deep reinforcement learning framework, and embed the deep reinforcement learning framework into the spinning finite element model through secondary development to establish a spinning process parameter control model.

[0009] S4. The agent in the spinning process parameter control model continuously interacts with the environment and continuously adjusts the spinning process parameters according to the return value to maximize the cumulative reward during the forming process, and finally outputs the optimal spinning process path.

[0010] Furthermore, the spinning process parameters include the thinning rate, the spindle speed, the roller feed speed, and the heating power, and the spinning process parameters are set as fixed values during the forming process.

[0011] The physical field states include the average strain of the deformed element, the average strain rate, and the average forming temperature. When calculating all physical field states, the stage with an instantaneous strain rate of is removed.

[0012] The formability quality parameters include the part geometric shape parameters and the mechanical properties; the geometric shape parameters include the part roundness, the die fit, and the theoretical wall thickness difference; the mechanical properties include the part hardness, the room temperature tensile properties, and the impact toughness.

[0013] Furthermore, in S2, taking the part forming physical field states obtained by finite element simulation as the input and the formability quality parameters obtained by spinning experiments as the output, calculating the error between the weighted sums of the formability quality parameters, and then updating the weight coefficients of the neural network model by backpropagation until convergence.

[0014] Furthermore, when training the neural network model, K-fold cross-validation is adopted. K-fold cross-validation randomly divides the data into K parts, selects K - 1 parts as the training set each time, and the remaining 1 part as the test set, and performs cross-validation K times. The average value of the K accuracies is taken as the final evaluation index of the neural network model.

[0015] Furthermore, the reward function in S2 is confirmed by the weighted coefficient method, and the expression is:

[0016] ;

[0017] where is the weight coefficient, the size of is allocated according to the importance of each index, and , is the total number of indexes, is the sub-reward item corresponding to the i-th index, is the output result of the mechanical properties in the neural network model, the larger, the greater the reward degree, is the output result of the geometric shape parameters in the neural network model, the smaller, the greater the reward degree.

[0018] Furthermore, the deep reinforcement learning framework in S3 is modeled by a Markov decision process, which is defined as a quadruple: state s, action a, reward r, and state transition probability distribution function p. The state s corresponds to the physical field state of the unit during the spin forming finite element simulation. The action a is the adjustment amount of the spin forming process parameters. The reward r is the calculated value of the reward function defined in S2, which is used to evaluate the immediate impact of the current action on the formability quality. The function p is the probability distribution of transferring to the next physical field state given the state s and action a.

[0019] Furthermore, the spin forming process parameter control model includes an environment and an agent.

[0020] The environment includes a spin forming finite element model and a reward function. The agent includes a policy network and a value network.

[0021] The policy network selects an action according to the current state s and continuously optimizes the policy through the policy gradient method to maximize the cumulative reward. The cumulative reward is the sum of the reward values from the start time to the end time of the simulation, and the expression is:

[0022] ;

[0023] where is the discount rate, which is used to determine the priority of short-term rewards and ranges from 0 to 1. is the time step, represents the end time, represents the time step size, represents at the time step when, according to the current state select and execute the current action the immediate reward value feedback by the environment.

[0024] The value network is used to evaluate the quality of the current policy and provide guidance for updating the policy network.

[0025] Furthermore, the spin forming process parameter control model in S4 controls the interaction process through a preset control strategy: when the environment receives the action a of the agent, the environment obtains the reward r and the next state s′ according to the state s at the current time. The agent optimizes the policy according to the cumulative reward value and determines the next action. During the continuous interaction process, the parameters θ of the policy network and the parameters ω of the value network inside the agent are continuously updated, making the actions of the agent more and more in line with expectations, and finally learning the optimal spin forming process path.

[0026] Furthermore, the parameter update process of the policy network and the value network during the agent interaction is as follows: at time, the agent executes Instantaneous action After that, observe from the environment Instantaneous reward and the new state ; The policy network makes a decision , indicating the policy, but the agent does not immediately execute the decision The value network scores the action :

[0027] ;

[0028] ;

[0029] Among them, is the score of the decision in the state is the score of the decision in the state is the action-value function;

[0030] Based on this, calculate the network target value and the prediction error value :

[0031] ;

[0032] ;

[0033] Update the value network parameter ω according to the gradient descent method:

[0034] ;

[0035] Among them, is the updated value of the value network parameter is the gradient descent learning rate represents the gradient of the parameter ;

[0036] Then update the policy network parameter θ according to the gradient ascent method:

[0037]

[0038] Among them, is the updated value of the policy network parameter is the gradient ascent learning rate represents the gradient of the parameter and represents the output of the policy network, representing that the agent directly selects the action at the state .

[0039] The present invention also provides a collaborative optimization system for an intelligent spinning process path, which adopts a collaborative optimization method for an intelligent spinning process path as described above, and includes a multi-source data acquisition and preprocessing module, a neural network model construction and training module, a deep reinforcement learning framework integration module, and a dynamic process parameter optimization and control module;

[0040] The multi-source data acquisition and preprocessing module is used to generate different combinations of spinning process parameters, perform finite element simulations, and simultaneously obtain physical experiment data;

[0041] The neural network model construction and training module is used to train a neural network model to determine the mapping relationship between the physical field state and the formability quality parameters during part forming;

[0042] The deep reinforcement learning framework integration module is used to deeply integrate deep reinforcement learning with the finite element simulation environment, construct a closed-loop control framework for autonomous optimization of spinning process parameters, and simultaneously establish an intelligent decision-making cycle of state perception → policy generation → action execution → reward feedback;

[0043] The dynamic process parameter optimization and control module is used to dynamically adjust the spinning process parameters through real-time interaction, and finally output the optimal process path.

[0044] The above solution of the present invention has the following beneficial effects:

[0045] The collaborative optimization method and system for an intelligent spinning process path provided by the present invention can dynamically adjust the process parameters during the spinning forming process according to factors such as the ideal deformation temperature, strain rate, and strain of the part, and are applicable to the optimization of process parameters for parts with complex curved generatrices and non-axisymmetric spinning parts, finding the best process path for customizing and optimizing the comprehensive quality of the part, and achieving the characteristics of strong accuracy and high reliability;

[0046] The deep reinforcement learning framework adopted by the present invention is unsupervised learning and does not require a large amount of pre-defined labeled data. The spinning forming process parameter control model can obtain a large amount of data by interacting with the environment, reducing the cost of pre-data acquisition;

[0047] The present invention uses a neural network to determine the reward function in the spinning forming process parameter control model, which can quickly establish the non-linear mapping between factors such as deformation temperature and strain rate during the forming process and the formability quality index, and there are many types of parameters that can be optimized;

[0048] The deep learning framework adopted by the present invention combines the advantages of value function and policy gradient, and uses a policy network and a value network to handle high-dimensional problems with continuous state and action spaces in spinning simulations, avoiding the limitation that the value function method can only handle discrete action space problems, and having higher efficiency compared to pure policy gradient;

[0049] Other beneficial effects of the present invention will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flowchart of the method steps of the present invention;

[0051] Figure 2 It is a schematic diagram of the geometric model assembly in the spinning finite element model in the embodiment of the present invention;

[0052] Figure 3 It is a schematic diagram of the spinning forming process parameter control model in the embodiment of the present invention.

[0053]

DESCRIPTION OF THE REFERENCE NUMERALS

[0054] 1 - blank; 2 - spinning wheel. SPECIFIC IMPLEMENTATION MANNER

[0055] The following describes the implementation manners of the present disclosure through specific specific examples. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. The present disclosure can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts belong to the scope of protection of the present disclosure.

[0056] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is illustrative only. Based on the present disclosure, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and these aspects can be combined in various ways in two or more of them. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, this device and / or this method can be implemented using other structures and / or functions in addition to one or more of the aspects described herein.

[0057] It should also be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present disclosure schematically. The diagrams only show the components related to the present disclosure, rather than being drawn according to the number, shape, and size of the components in actual implementation. The types, quantities, and proportions of the components in actual implementation can be arbitrarily changed, and the component layout type may also be more complex. Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0058] As Figure 1 shown, an embodiment of the present invention provides a collaborative optimization method for an intelligent spinning process path, including the following steps:

[0059] S1, select different spinning process parameters according to the orthogonal design, and then obtain the physical field state during part forming and the formability quality parameters of the part after forming through finite element simulation and spinning experiments.

[0060] Among them, the spinning process parameters include the thinning rate, spindle speed, feed speed of the spinning wheel 2, heating power, etc. The spinning process parameters are set as fixed values during the forming process to reduce the simulation calculation amount and experimental operation difficulty. In this embodiment, appropriate spinning process parameters within a suitable range are selected through the orthogonal design, and then spinning finite element simulation and spinning experiments are carried out under the process conditions selected in the orthogonal table to obtain the average physical field state during part forming and the formability quality parameters of the part after forming.

[0061] In the spinning finite element simulation, the physical field state includes the average strain of the deformed element, average strain rate, average forming temperature, etc. When calculating all the physical field states, the stage with an instantaneous strain rate of 0 should be removed. Specifically, the history curves of the temperature, strain, strain rate, etc. of the elements in the part bus direction can be selected for extraction, and then the data points with an instantaneous strain rate greater than 0 and the values of other physical field state variables corresponding to the elements at that moment are screened out, and their respective averages are calculated.

[0062] The spinning finite element model can be constructed using existing commercial software such as ABAQUS, ANSYS, etc. For example, constructing a spinning finite element model through ABAQUS includes: confirming the material properties of the components; establishing the assembly relationship of the components; setting geometric boundary conditions, including the die motion trajectory, contact relationship of each component, application of load conditions, and application of thermal boundary conditions, etc.; ensuring uniform element size during mesh division, with at least 3 or more elements in the wall thickness direction of the blank 1, selecting the reduced integration element for the element type, and adopting the hourglass enhancement mode; selecting the Explicit method for model solution to avoid convergence problems.

[0063] After the spinning experiment, the formability quality of the part is tested. The quality parameters include the part's geometric shape parameters and mechanical properties. The geometric shape parameters include the part's roundness, die fitting degree, and theoretical wall thickness difference, etc.; the mechanical properties include the hardness of the workpiece, room temperature tensile properties, and impact toughness, etc.

[0064] S2. Establish a neural network model, determine the mapping relationship between the physical field state during part forming and the formability quality parameters, and determine the reward function by the weighted coefficient method to train the neural network model.

[0065] In this embodiment, each structural parameter in the neural network model includes the initial weight and threshold, the number of hidden layers, the number of neurons in each layer, and the selection of activation functions, etc., which are determined by an optimization algorithm. The optimization algorithm can adopt the Bayesian algorithm, genetic algorithm, particle swarm optimization algorithm, etc. The training process of the neural network model is specifically as follows: taking the physical field state of part forming obtained by finite element simulation as the input and the formability quality parameters obtained by the spinning experiment as the output, calculating the error between the weighted sums of the quality parameters, and then updating the weight coefficients of the neural network model by backpropagation until convergence. Among them, the physical field state and the formability quality parameters are normalized and normalized to between 0 and 1 according to the method, and then input, output, etc. are performed.

[0066] It should be noted that in this embodiment, K-fold cross-validation is used during the training of the neural network model to improve the reliability and stability of model evaluation. K-fold cross-validation maximizes the use of data and reduces evaluation bias by repeatedly dividing the dataset. Generally, the smaller the amount of training data, the larger the K value. Specifically, the above data is randomly divided into K parts. Each time, K - 1 parts are selected as the training set, and the remaining 1 part is the test set. Cross-validation is performed K times, and the average value of the K accuracies is taken as the final evaluation index of the neural network model. When the amount of data is less than 1000 groups, generally 5-fold or 10-fold cross-validation is selected.

[0067] In this embodiment, the reward function is confirmed by the weighted coefficient method, and the specific expression is:

[0068] ;

[0069] Among them, is the weight coefficient, The size of is allocated according to the importance of each index, and , is the total number of indexes, is the sub-reward item corresponding to the i-th index, is the output result of the mechanical properties in the neural network model, and the larger its value, the greater the degree of reward, It is the output result of the geometric shape parameters in the neural network model. The smaller the value, the greater the degree of reward. Generally speaking, the mechanical properties and geometric shape accuracy of parts are equally important. It can be set to 1 / n.

[0070] S3. Build a deep reinforcement learning framework, and embed the deep reinforcement learning framework into the spinning finite element model through secondary development to establish a spinning forming process parameter control model.

[0071] For example, the deep reinforcement learning framework can be written using the VUMAT subroutine in the ABAQUS software. In this embodiment, the deep reinforcement learning framework is modeled by a Markov decision process and is defined as a quadruple: state s, action a, reward r, and state transition probability distribution function p. Among them, the state s corresponds to the physical field state of the unit during the spinning finite element simulation. In this embodiment, it is specifically defined as the physical field state of the deformed unit that is undergoing plastic deformation during the spinning forming process, so as to avoid reading other undeformed units in the model and causing a significant increase in the calculation amount; the action a is the adjustment amount of the spinning process parameters; the reward r is the calculated value of the reward function defined in S2, which is used to evaluate the immediate impact of the current action on the formability quality; the function p is the probability distribution of transferring to the next physical field state under the given state s and action a, that is, the probability distribution of transferring to the next state under the given state s and action a.

[0072] It should be noted that the spinning forming process parameter control model established in this embodiment consists of an environment and an agent. Among them, the environment part consists of a spinning finite element model and a reward function; the agent part consists of a policy network (actor) and a value network (critic).

[0073] In this embodiment, the policy network can select an action according to the current state s and continuously optimize the policy through the policy gradient method (such as deterministic policy gradient, DPG) to maximize the cumulative reward. The specific expression is:

[0074] ;

[0075] Among them, the objective function the gradient of the policy network parameters of, represents the expectation of the state s under the distribution of, represents the policy the probability of the state s appearing under, is the action value function, which represents the expected cumulative return value after taking the action a in the state s under the policy of, is the gradient of the action value function with respect to the action a, is the output of the policy network, indicating the gradient of the output of the policy network with respect to the parameter . Therefore, the policy network can select an action according to the current state s. The cumulative reward is the sum of the reward values from the start to the end of the simulation, and the expression is as follows:

[0076] ;

[0077] where is the discount rate, which is used to determine the priority of short-term rewards, ranging from 0 to 1, is the time step, represents the end time, represents the time step size, represents at the time step , according to the current state select and execute the current action , and then the immediate reward value feedback by the environment. When tends to 0, the model pays more attention to short-term rewards, and vice versa, treats the rewards of all time steps equally. During spin forming, the properties of each part of the part are equally important, so should be close to 1.

[0078] In this embodiment, the value network can evaluate the quality of the current policy, provide guidance for the update of the policy network, and improve the learning stability and efficiency.

[0079] S4. The agent in the spin forming process parameter control model continuously interacts with the environment and continuously adjusts the spin forming process parameters according to the return value, thereby maximizing the cumulative reward in the forming process and finally outputting the optimal spin forming process path.

[0080] In this embodiment, the spin forming process parameter control model controls the interaction process through a preset control strategy: when the environment receives the action a of the agent (such as heating power, feed speed of the spinning wheel 2, etc.), the environment obtains the reward r and the next state s′ according to the state s (deformation unit temperature, strain, etc.) at the current time. The agent optimizes the policy according to the cumulative reward value and determines the next action, that is, starts the next interaction. During the continuous interaction process, the parameters θ of the policy network and the parameters ω of the value network inside the agent are continuously updated, making the actions of the agent more and more in line with expectations, and finally learning the optimal spin forming process path.

[0081] The parameter update process of the policy network and the value network during agent interaction is specifically as follows: at time, the agent executes the action at time , and then observes from the environment the reward at time and the new state ; The policy network makes a decision , but the agent does not immediately execute the decision ; The value network scores the actions:

[0082] ;

[0083] ;

[0084] Among them, is the score of the decision in the state, is the score of the decision in the

[0085] state, and is the action value function; Based on this, calculate the network target value

[0086] (TD target) and the prediction error value

[0087] (TD error):

[0088] Then update the value network parameter ω according to the gradient descent method:

[0089] ;

[0090] Among them, is the updated value of the value network parameter, is the gradient descent learning rate, generally in the range of 10 -5 ~10 -2 The value is related to the task complexity, network structure, etc. represents the gradient of the parameter . Then update the policy network parameter θ according to the gradient ascent method:

[0091] ;

[0092] Among them, is the updated value of the policy network parameter, is the gradient ascent learning rate, and its range is similar to , represents the gradient of the parameter , is the output of the policy network, representing the action directly selected by the agent in the state when .

[0093] It should be noted that in this embodiment, when updating the network parameters inside the agent, experience replay is used to store the interaction records between the agent and the environment for reusing the collected data, overcoming the problems of correlation and non-stationary distribution of network data, and improving the training efficiency and stability of network data. Among them, the interaction records are stored in the form of quadruples and stored. Each time the network is updated, multiple groups of records (mini-batch) are randomly selected and the corresponding TD target and TD error value are calculated, and then the average gradient is obtained to update the network parameters. It should be noted that the size of the mini-batch is usually between 32 and 512, and a smaller network can choose 32 or 64.

[0094] It should be noted that in this embodiment, noise can be added when the agent interacts with the environment to improve the exploration ability of the agent, avoid the strategy falling into local optimum, and improve the learning efficiency. Among them, noise can be added in the output action of the policy network, the state input of the value network, and the experience replay stage. For example, before the agent executes the action selected by the policy, a Gaussian distribution form can be added, such as , where is the standard deviation, c is the noise range, is used to be restricted to the interval .

[0095] The effect of this method will be further described through specific cases below. Taking an AZ31 magnesium alloy disc with a diameter of 200 mm and a thickness of 4 mm as blank 1 to manufacture a magnesium alloy fairing (thin-walled conical part), it is required to optimize the laser heating power, the feed speed of the spinning wheel 2, and the spindle speed during shear spinning to achieve the highest possible room temperature tensile properties and impact toughness of the part under the condition of meeting the geometric dimension accuracy (roundness tolerance of 0.01 mm and die fitting degree within 0.1 mm). The specific steps are as follows:

[0096] Determine the forming parameter range of the magnesium alloy conical part according to the size and material of the spinning blank 1, that is, the heating power is 1000 W, 1500 W, 2000 W, 2500 W, the feed speed of the spinning wheel 2 is 2 mm / s, 4 mm / s, 6 mm / s, and the spindle speed is 3 rps, 6 rps, 9 rps. Use the L16(4^5) orthogonal table to determine 16 groups of orthogonal experiments with different factor levels.

[0097] Based on the Abaqus finite element software, establish a laser-assisted magnesium alloy shear spinning model, such as Figure 2As shown in the figure. The boundary conditions of the model are set according to the orthogonal table. The laser heat source and the movement trajectory of the spinning wheel 2 in the model can be embedded through the VDEFLUX and VDISP subroutines. When meshing, ensure that there are 3 or more elements in the thickness direction, and the element type is C3D8RT enhanced by hourglass. The solver is selected as Explicit. After all simulations are completed, extract the temperature, strain, and strain rate history curves of all elements in the generatrix direction of the conical part, and calculate the average values of the corresponding variables when the instantaneous strain rate is not 0. Based on the spinning experimental platform, carry out a spinning orthogonal experiment, measure the geometric accuracy of the spun conical part, and record the roundness and die fitting degree of the part; then test the room temperature tensile properties and impact toughness of the part.

[0098] Establish a BP neural network model. The input layer is the average temperature, average strain, and average strain rate of the elements, and the output layer is the roundness, die fitting degree, room temperature tensile properties, and impact toughness; after normalizing the training data, use 10-fold cross-validation to train the neural network model; select the mean absolute error to evaluate the quality of the model; the number of hidden layers and neurons in the model are optimized using the Bayesian algorithm to ensure that the model has the best prediction ability. Use the weighted coefficient method to determine the reward function of the spinning forming process parameter control model:

[0099] ;

[0100] Among them, when the roundness and die fitting degree tolerances are less than 0.01 mm and 0.1 mm respectively, a = 1, otherwise it is 0; and are the tensile properties and impact toughness respectively.

[0101] Based on the Abaqus secondary development platform, use the VUMAT subroutine to embed the actor-critic deep learning framework into the finite element model to establish a spinning forming process parameter control model. As Figure 3 shown, it is divided into two parts: the environment and the agent. The environment part consists of the shear spinning finite element model and the reward function, and the agent consists of the value network and the policy network, belonging to the deep deterministic policy gradient algorithm (DDPG). The model uses the temperature, strain, and strain rate of the deformed elements during spinning forming as the system state s , and the laser heating power, spindle speed, and feeding speed of the spinning wheel 2 within each control step as the system actions a . Due to the local forming characteristics of spinning, only the elements near the contact point between the spinning wheel 2 and the blank 1 are selected each time the system state is read, reducing the state space of the system.

[0102] During the simulation run, the agent in the spinning forming process parameter control model continuously interacts with the environment, that is, the environment receives the action , and obtains the reward according to the previous moment's environment state With the system state , this part of the information is stored in the replay experience in the form of a tuple; then a batch of 32 experiences (min-batch) is sampled from the replay experience, and then the new actions under each group of experiences are obtained through the decision-making of the policy network in the agent To enhance the exploration ability of the model, noise is applied to the actions For the action noise of the heating power, it can be set = 0.5, c = 25; the value network in the agent evaluates the previous action and the action generated by the decision-making to obtain and , and then calculate the TD target value and the TD error ; update the parameters in the value network according to the error value , and update the parameters in the decision network . The learning rate during parameter update is set to 0.001. After the network update is completed, the agent makes a decision to output an action , and starts the next interaction.

[0103] During the spin forming process, the agent calculates the cumulative reward sum from the start time to the end time of the simulation , Take 0.9. By selecting the process path with the maximum cumulative reward, the change information of the spin forming process parameters, that is, the optimal spin forming process parameters, can be obtained.

[0104] In summary, the collaborative optimization method for the intelligent spin forming process path provided in this embodiment can realize the dynamic adjustment of process parameters during the spin forming process according to factors such as the ideal deformation temperature, strain rate, and strain of the part, and is applicable to the process parameter optimization of parts with complex curved busbars and non-axisymmetric spin parts, finding the best process path for customizing and optimizing the comprehensive quality of parts, and realizing the characteristics of strong accuracy and high reliability.

[0105] Based on the same inventive concept, this embodiment also provides a collaborative optimization system for the intelligent spinning process path, including a multi-source data acquisition and preprocessing module, a neural network model construction and training module, a deep reinforcement learning framework integration module, and a dynamic process parameter optimization and control module. Among them, the multi-source data acquisition and preprocessing module is used to generate different combinations of spinning process parameters, perform finite element simulations, and simultaneously obtain physical experiment data; the neural network model construction and training module is used to train the neural network model to determine the mapping relationship between the physical field state and the formability quality parameters during part forming; the deep reinforcement learning framework integration module is used to deeply integrate deep reinforcement learning with the finite element simulation environment to construct a closed-loop control framework for autonomous optimization of spinning process parameters, and simultaneously establish an intelligent decision-making cycle of state perception → policy generation → action execution → reward feedback; the dynamic process parameter optimization and control module is used to dynamically adjust the spinning process parameters through real-time interaction and finally output the optimal process path.

[0106] The collaborative optimization system for the intelligent spinning process path provided in this embodiment has the same inventive concept and beneficial effects as the foregoing method, which will not be elaborated here.

[0107] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0108] The above embodiments only represent several implementation manners of the present application, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A collaborative optimization method for the intelligent spinning process path, characterized in that It includes the following steps: S1. Select different spinning process parameters according to orthogonal design, and then obtain the physical field state during part forming and the formability quality parameters of the part after forming through finite element simulation and spinning experiments; S2. Establish a neural network model, determine the mapping relationship between the physical field state during part forming and the formability quality parameters, and determine the reward function through the weighted coefficient method to train the neural network model; S3. Build a deep reinforcement learning framework, embed the deep reinforcement learning framework into the spinning finite element model through secondary development, and establish a spinning forming process parameter control model; S4. The agent in the spinning forming process parameter control model continuously interacts with the environment and continuously adjusts the spinning process parameters according to the return value to maximize the cumulative reward during the forming process, and finally outputs the optimal spinning process path.

2. The collaborative optimization method for an intelligent spinning process path according to claim 1, characterized in that The spinning process parameters include the thinning rate, spindle speed, roller feed speed, and heating power, and the spinning process parameters are set as fixed values during the forming process; The physical field state includes the average strain of the deformation unit, average strain rate, and average forming temperature. The stage with an instantaneous strain rate of is removed when calculating all physical field states; The formability quality parameters include part geometric shape parameters and mechanical properties; the geometric shape parameters include part roundness, die fitting degree, and theoretical wall thickness difference; the mechanical properties include part hardness, room temperature tensile properties, and impact toughness.

3. The collaborative optimization method for an intelligent spinning process path according to claim 1, characterized in that In S2, the physical field state of part forming obtained by finite element simulation is used as the input, and the formability quality parameters obtained by spinning experiments are used as the output. The error between the weighted sums of the formability quality parameters is calculated, and then the weight coefficients of the neural network model are updated by backpropagation until convergence.

4. The collaborative optimization method for an intelligent spinning process path according to claim 1, characterized in that When training the neural network model, K-fold cross-validation is adopted. K-fold cross-validation randomly divides the data into K parts. Each time, K - 1 parts are selected as the training set, and the remaining 1 part is the test set. Cross-validation is performed K times, and the average value of the K accuracies is taken as the final evaluation index of the neural network model.

5. The collaborative optimization method of an intelligent spinning process path according to claim 1, characterized in that Reward function in S2 Confirmed by the weighted coefficient method, the expression is: ； Among them, is the weight coefficient, The size of is allocated according to the importance of each index, and , is the total number of indexes, is the sub-reward item corresponding to the i-th index, is the output result of mechanical properties in the neural network model, The larger it is, the greater the reward degree, is the output result of geometric shape parameters in the neural network model, The smaller it is, the greater the reward degree.

6. The collaborative optimization method for an intelligent spinning process path according to claim 1, characterized in that In S3, the deep reinforcement learning framework is modeled by a Markov decision process, which is defined as a quadruple: state s, action a, reward r, and state transition probability distribution function p; the state s corresponds to the physical field state of the unit during spinning finite element simulation; the action a is the adjustment amount of the spinning process parameters; the reward r is the calculated value of the reward function defined in S2, which is used to evaluate the immediate impact of the current action on the formability quality; the function p is the probability distribution of transferring to the next physical field state given the state s and action a.

7. The collaborative optimization method for an intelligent spinning process path according to claim 6, characterized in that, The spinning forming process parameter control model includes an environment and an agent; The environment includes the spinning finite element model and the reward function; the agent includes a policy network and a value network; The policy network selects actions based on the current state s and continuously optimizes the policy through the policy gradient method to maximize the cumulative reward, where the cumulative reward is the sum of the reward values from the start to the end of the simulation, and the expression is: ； wherein, is the discount rate, which is used to determine the priority of short-term returns and ranges from 0 to 1, represents the time step, is the moment, represents the end moment, represents at the time step when, according to the current state select and execute the current action after, the immediate reward value feedback by the environment; The value network is used to evaluate the quality of the current policy and provide guidance for updating the policy network.

8. The collaborative optimization method for an intelligent spinning process path according to claim 7, characterized in that, The spin forming process parameter control model in S4 controls the interaction process through a preset control strategy: when the environment receives the action a of the agent, the environment obtains the reward r and the state s' at the next moment according to the state s at the current moment. The agent optimizes the strategy based on the cumulative reward value and determines the next action. During the continuous interaction process, the parameters θ of the policy network and the parameters ω of the value network inside the agent are continuously updated, making the actions of the agent more and more in line with expectations, and finally learning the optimal spin forming process path.

9. The collaborative optimization method for an intelligent spinning process path according to claim 8, characterized in that The parameter update process of the policy network and the value network during agent interaction is as follows: At time, the agent executes the action at time, and then observes from the environment the reward at and the new state ; the policy network makes a decision , denotes the policy, but the agent does not immediately execute the decision the value network scores the action : ； ； Among them, is the score for the decision made in the state, is the score for the decision made in the state, is the action value function; Based on this, calculate the network target value and the prediction error value : ； ； Update the value network parameter ω according to the gradient descent method: ； Among them, is the updated value of the value network parameter, is the gradient descent learning rate, represents the parameter gradient; Then update the policy network parameter θ according to the gradient ascent method: ； Among them, is the updated value of the policy network parameters, is the gradient ascent learning rate, represents the parameter gradient, represents the output of the policy network, representing that the agent directly selects the action at the state .

10. A collaborative optimization system for an intelligent spinning process path, which adopts a collaborative optimization method for an intelligent spinning process path as described in any one of claims 1-9, characterized in that, It includes a multi-source data acquisition and preprocessing module, a neural network model construction and training module, a deep reinforcement learning framework integration module, and a dynamic process parameter optimization and control module; The multi-source data acquisition and preprocessing module is used to generate different combinations of spin forming process parameters, perform finite element simulations, and obtain physical experiment data at the same time; The neural network model construction and training module is used to train the neural network model and determine the mapping relationship between the physical field state and the formability quality parameters during part forming; The deep reinforcement learning framework integration module is used to deeply integrate deep reinforcement learning with the finite element simulation environment, construct a closed-loop control framework for autonomous optimization of spin forming process parameters, and establish an intelligent decision-making cycle of state perception → policy generation → action execution → reward feedback at the same time; The dynamic process parameter optimization and control module is used to dynamically adjust the spin forming process parameters through real-time interaction, and finally output the optimal process path.

Images