Fatigue life prediction method, device, equipment and storage medium for metal materials

Abstract

The present application discloses a fatigue life prediction method, device, equipment and storage medium for metal materials, which relates to the field of materials science and engineering. By constructing a customized reinforcement learning environment, the modified SWT equation is converted into a decision-making process. The state includes the current fatigue life estimate and fatigue life error, and the action is the step factor. Then, a deep deterministic policy gradient algorithm is used to train the agent. The agent selects the step factor in each iteration and updates the current fatigue life estimate in combination with the Newton-Raphson iteration method. The reward function comprehensively considers the fatigue life error value, fatigue life deviation and number of iterations to guide the agent to gradually approach the true solution. By dynamically adjusting the step factor, the solution efficiency and accuracy can be significantly improved. By combining the reinforcement learning agent with the Newton iteration method, the strategy can be continuously optimized according to the environmental feedback to achieve efficient nonlinear equation solving. At the same time, the convergence and stability of the model can be enhanced, and the sensitivity to the initial value can be reduced.

Description

Technical Field

[0001] The present application relates to the field of materials science and engineering, and in particular to a fatigue life prediction method, device, equipment and storage medium for metal materials. Background Art

[0002] Fatigue life prediction is a key issue in the field of materials science and engineering, especially in high-demand applications such as aerospace and automobiles. Accurately predicting the fatigue life of materials under cyclic loads can not only significantly improve the safety and reliability of the structure, but also effectively prevent potential structural failures and reduce maintenance costs.

[0003] In the related art, the SWT (Smith-Watson-Topper, mean stress correction) model is a widely used fatigue life prediction model that estimates the fatigue life of materials by comprehensively considering the elastic and plastic strain components of the material and using the stress and strain amplitude. However, the applicant recognizes that the nonlinear characteristics of the traditional SWT model make its solution process complicated, and traditional numerical methods such as the Newton-Raphson iteration method often face problems such as high computational complexity, slow convergence speed, and high sensitivity to the selection of initial values ​​in practical applications, which not only prolongs the calculation time, but also leads to instability and inaccuracy in the solution process, reducing the calculation efficiency and prediction reliability of the fatigue life of metal materials. Summary of the invention

[0004] In view of this, the present application provides a fatigue life prediction method, device, equipment and storage medium for metal materials. The main purpose is to solve the problem that the use of traditional SWT model and Newton-Raphson iteration method for fatigue life prediction not only prolongs the calculation time, but also causes instability and inaccuracy in the solution process, thereby reducing the calculation efficiency and prediction reliability of the fatigue life of metal materials.

[0005] According to the first aspect of the present application, a fatigue life prediction method for a metal material is provided, the method comprising:

[0006] Acquire a SWT model, and modify the SWT model to obtain a target SWT model;

[0007] Acquire fatigue test data, and construct a reinforcement learning environment using the fatigue test data and the target SWT model;

[0008] Obtaining a deep deterministic policy gradient algorithm, and using the deep deterministic policy gradient algorithm to train the reinforcement learning environment to obtain a fatigue life prediction model;

[0009] The metal material data is acquired, and the metal material data is input into the fatigue life prediction model for prediction to obtain a fatigue life prediction result.

[0010] Optionally, the modifying the SWT model to obtain a target SWT model includes:

[0011] Obtaining a stress concentration factor and a surface roughness coefficient, and using the stress concentration factor and the surface roughness coefficient to modify the SWT model to obtain the target SWT model, wherein:

[0012] ;

[0013] in, represents the comprehensive correction factor, b represents the fatigue strength index, c represents the fatigue ductility index, represents the stress amplitude, represents the maximum stress in the cycle, represents the cyclic strength coefficient, represents the cyclic strain hardening exponent, E represents the elastic modulus, represents the Walker index, represents the fatigue strength coefficient, represents the fatigue ductility coefficient, represents the current fatigue life, where , q represents the notch sensitivity coefficient, represents the stress concentration factor, Represents the surface roughness coefficient.

[0014] Optionally, the acquiring fatigue test data and constructing a reinforcement learning environment using the fatigue test data and the target SWT model includes:

[0015] Preprocessing the fatigue test data, extracting a plurality of fatigue life true values ​​from the preprocessed fatigue test data, and adding the plurality of fatigue life true values ​​to a state space;

[0016] Obtain a step factor adjustment strategy, and add the step factor adjustment strategy to the action space, wherein the step factor adjustment strategy is that the step factor gradually decreases with the training process, and the adjustment range of the step factor is [-0.1, 0.1]. The step factor is used to control the update amplitude of the fatigue life estimation value in each iterative training;

[0017] Obtain a reward function, and construct the reinforcement learning environment using the state space, the action space, the reward function, and the target SWT model, wherein:

[0018] ;

[0019] in, represents the reward function, represents the fatigue life error value, represents the fatigue life estimate, represents the true value of fatigue life, represents the minimum value, Indicates the number of iterations.

[0020] Optionally, the acquiring a deep deterministic policy gradient algorithm, using the deep deterministic policy gradient algorithm to train the reinforcement learning environment to obtain a fatigue life prediction model, includes:

[0021] Obtain the current Actor network, target Actor network, current Critic network, and target Critic network included in the deep deterministic policy gradient algorithm;

[0022] Obtain a first state, determine a first step length factor according to the first state based on the current Actor network, and execute the first step length factor in the first state based on the reinforcement learning environment to obtain a second state and a first reward value;

[0023] Based on the current Actor network, determining a second step factor according to a second state, and performing a next round of training based on the second state and the second step factor, so as to iteratively train to obtain a plurality of states, a plurality of step factors, and a plurality of reward values;

[0024] Storing the multiple states, the step size factor, and the reward value in an experience replay buffer;

[0025] Sample data is obtained from the experience playback buffer, and the current Actor network and the current Critic network are iteratively trained using the sample data to obtain the fatigue life prediction model.

[0026] Optionally, the acquiring the first state, determining a first step length factor according to the first state based on the current Actor network, and executing the first step length factor in the first state based on the reinforcement learning environment to obtain a second state and a first reward value includes:

[0027] Initializing the reinforcement learning environment to obtain a first fatigue life estimate;

[0028] Reading a first fatigue life true value matching the first fatigue life estimate value in the fatigue test data;

[0029] taking the difference between the first fatigue life estimate and the first fatigue life true value as a first fatigue life error value, and taking the first fatigue life estimate and the first fatigue life error value as a first state;

[0030] Based on the current Actor network, determining a first step length factor according to the first state, and sending the first step length factor to the reinforcement learning environment;

[0031] Obtaining a Newton-Raphson iteration method, and calculating the first step length factor using the Newton-Raphson iteration method based on the reinforcement learning environment to obtain a second fatigue life estimate and a first reward;

[0032] A second fatigue life true value matching the second fatigue life estimate is read from the fatigue test data, a difference between the second fatigue life estimate and the second fatigue life true value is taken as a second fatigue life error value, and the second fatigue life estimate and the second fatigue life error value are taken as the second state.

[0033] Optionally, the iterative training of the current Actor network and the current Critic network using the sample data to obtain the fatigue life prediction model includes:

[0034] Obtaining a first state, a second state, and a first step length factor from the sample data, inputting the second state into the target Actor network for calculation to obtain a second target step length factor, inputting the second target step length factor and the second state into the target Critic network for calculation to obtain a first target value, inputting the first step length factor and the first state into the current Critic network for calculation to obtain a first current value, and using the first target value and the first current value to train the current Actor network and the current Critic network;

[0035] The third state of the sample data is input into the target Actor network for calculation to obtain a third target step factor, the third target step factor and the third state are input into the target Critic network for calculation to obtain a second target value, the second step factor and the second state are input into the current Critic network for calculation to obtain a second current value, and the second target value and the second current value are used to perform the next round of training on the current Actor network and the current Critic network, so as to iteratively train and obtain the fatigue life prediction model.

[0036] Optionally, the method further comprises:

[0037] Acquire a plurality of fatigue life estimation values, and construct a fatigue life convergence trajectory diagram using the plurality of fatigue life estimation values;

[0038] Acquire a plurality of fatigue life error values, and construct an error convergence diagram using the plurality of fatigue life error values;

[0039] Obtain multiple reward values, and use the multiple reward values ​​to construct a reward value change graph;

[0040] Obtain multiple step factors, and use the multiple step factors to construct a step factor adjustment graph;

[0041] Extracting a plurality of true fatigue life values ​​from the fatigue test data, and constructing a prediction and actual error comparison diagram using the plurality of true fatigue life values ​​and the plurality of fatigue life estimation values;

[0042] Sending the fatigue life convergence trajectory diagram, the error convergence diagram, the reward value change diagram, the step factor adjustment diagram, and the predicted and actual error comparison diagram to a terminal held by a user for model evaluation operation;

[0043] Obtaining a model evaluation result based on feedback from the user's terminal;

[0044] If the model evaluation result indicates that the fatigue life prediction model meets the fatigue life prediction conditions of metal materials, the fatigue life prediction model is stored.

[0045] According to a second aspect of the present application, a fatigue life prediction device for a metal material is provided, the device comprising:

[0046] A correction module is used to obtain a SWT model, correct the SWT model, and obtain a target SWT model;

[0047] A construction module is used to obtain fatigue test data and construct a reinforcement learning environment using the fatigue test data and the target SWT model;

[0048] A training module, used to obtain a deep deterministic policy gradient algorithm, and use the deep deterministic policy gradient algorithm to train the reinforcement learning environment to obtain a fatigue life prediction model;

[0049] The prediction module is used to obtain metal material data, input the metal material data into the fatigue life prediction model for prediction, and obtain a fatigue life prediction result.

[0050] Optionally, the correction module is used to obtain a stress concentration factor and a surface roughness coefficient, and use the stress concentration factor and the surface roughness coefficient to correct the SWT model to obtain the target SWT model, wherein:

[0051] ;

[0052] in, represents the comprehensive correction factor, b represents the fatigue strength index, c represents the fatigue ductility index, represents the stress amplitude, represents the maximum stress in the cycle, represents the cyclic strength coefficient, represents the cyclic strain hardening exponent, E represents the elastic modulus, represents the Walker index, represents the fatigue strength coefficient, represents the fatigue ductility coefficient, represents the current fatigue life, where , q represents the notch sensitivity coefficient, represents the stress concentration factor, Represents the surface roughness coefficient.

[0053] Optionally, the construction module is used to preprocess the fatigue test data, extract multiple fatigue life real values ​​from the preprocessed fatigue test data, and add the multiple fatigue life real values ​​to the state space; obtain a step factor adjustment strategy, add the step factor adjustment strategy to the action space, the step factor adjustment strategy is that the step factor gradually decreases with the training process, the adjustment range of the step factor is [-0.1, 0.1], and the step factor is used to control the update amplitude of the fatigue life estimation value in each iterative training; obtain a reward function, and use the state space, the action space, the reward function, and the target SWT model to construct the reinforcement learning environment, wherein,

[0054] ;

[0055] in, represents the reward function, represents the fatigue life error value, represents the fatigue life estimate, represents the true value of fatigue life, represents the minimum value, Indicates the number of iterations.

[0056] Optionally, the training module is used to obtain the current Actor network, target Actor network, current Critic network, and target Critic network included in the deep deterministic policy gradient algorithm; obtain a first state, determine a first step size factor based on the current Actor network according to the first state, and execute the first step size factor in the first state based on the reinforcement learning environment to obtain a second state and a first reward value; determine a second step size factor based on the current Actor network according to the second state, and perform the next round of training based on the second state and the second step size factor to iteratively train to obtain multiple states, multiple step size factors, and multiple reward values; store the multiple states, the step size factors, and the reward values ​​in an experience replay buffer; obtain sample data from the experience replay buffer, and use the sample data to iteratively train the current Actor network and the current Critic network to obtain the fatigue life prediction model.

[0057] Optionally, the training module is used to initialize the reinforcement learning environment to obtain a first fatigue life estimate; read a first fatigue life true value that matches the first fatigue life estimate in the fatigue test data; use the difference between the first fatigue life estimate and the first fatigue life true value as a first fatigue life error value, and use the first fatigue life estimate and the first fatigue life error value as a first state; based on the current Actor network, determine a first step length factor according to the first state, and send the first step length factor to the reinforcement learning environment; obtain a Newton-Raphson iteration method, and based on the reinforcement learning environment, use the Newton-Raphson iteration method to calculate the first step length factor to obtain a second fatigue life estimate and a first reward; read a second fatigue life true value that matches the second fatigue life estimate in the fatigue test data, use the difference between the second fatigue life estimate and the second fatigue life true value as a second fatigue life error value, and use the second fatigue life estimate and the second fatigue life error value as the second state.

[0058] Optionally, the training module is used to obtain a first state, a second state and a first step factor from the sample data, input the second state into the target Actor network for calculation to obtain a second target step factor, input the second target step factor and the second state into the target Critic network for calculation to obtain a first target value, input the first step factor and the first state into the current Critic network for calculation to obtain a first current value, and use the first target value and the first current value to train the current Actor network and the current Critic network; input the third state of the sample data into the target Actor network for calculation to obtain a third target step factor, input the third target step factor and the third state into the target Critic network for calculation to obtain a second target value, input the second step factor and the second state into the current Critic network for calculation to obtain a second current value, and use the second target value and the second current value to perform the next round of training on the current Actor network and the current Critic network to iteratively train and obtain the fatigue life prediction model.

[0059] Optionally, the device further comprises:

[0060] An evaluation module is used to obtain multiple fatigue life estimation values, and use the multiple fatigue life estimation values ​​to construct a fatigue life convergence trajectory diagram; obtain multiple fatigue life error values, and use the multiple fatigue life error values ​​to construct an error convergence diagram; obtain multiple reward values, and use the multiple reward values ​​to construct a reward value change diagram; obtain multiple step factors, and use the multiple step factors to construct a step factor adjustment diagram; extract multiple fatigue life true values ​​from the fatigue test data, and use the multiple fatigue life true values ​​and the multiple fatigue life estimation values ​​to construct a prediction and actual error comparison diagram; send the fatigue life convergence trajectory diagram, the error convergence diagram, the reward value change diagram, the step factor adjustment diagram, and the prediction and actual error comparison diagram to a terminal held by a user for model evaluation operation; obtain the model evaluation result fed back by the user based on the terminal held; if the model evaluation result indicates that the fatigue life prediction model meets the fatigue life prediction conditions of metal materials, the fatigue life prediction model is stored.

[0061] According to a third aspect of the present application, a device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the methods described in the first aspect when executing the computer program.

[0062] According to a fourth aspect of the present application, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of any one of the methods described in the first aspect are implemented.

[0063] By means of the above technical scheme, the present application provides a fatigue life prediction method, device, equipment and storage medium for metal materials. The present application obtains an SWT model, modifies the SWT model, obtains a target SWT model, obtains fatigue test data, uses fatigue test data and the target SWT model to build a reinforcement learning environment, obtains a deep deterministic policy gradient algorithm, uses a deep deterministic policy gradient algorithm to train the reinforcement learning environment, obtains a fatigue life prediction model, obtains metal material data, inputs the metal material data into the fatigue life prediction model for prediction, and obtains fatigue life prediction results. By constructing a custom reinforcement learning environment, the modified SWT equation is converted into a decision-making process, the state includes the current fatigue life estimate and the fatigue life error value, the action is the adjustment of the step size factor, and then the deep deterministic policy gradient algorithm is used to train the intelligent agent. The intelligent agent selects the step size factor in each iteration, and combines the Newton-Raphson iteration method to update the current fatigue life estimate. The reward function comprehensively considers the fatigue life error value, fatigue life deviation and the number of iterations, and guides the intelligent agent to gradually approach the true solution. By dynamically adjusting the step size factor, a balance between rapid convergence and precise adjustment can be achieved, significantly improving the solution efficiency and accuracy. By combining the reinforcement learning agent with the numerical solution advantages of the Newton iteration method, it is possible to continuously optimize the strategy based on environmental feedback and achieve efficient solution of nonlinear equations. At the same time, it can enhance the convergence and stability of the model and reduce sensitivity to initial values.

[0064] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present application. Also, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0066] Figure 1 A schematic diagram of a method flow for predicting fatigue life of a metal material provided in an embodiment of the present application is shown;

[0067] Figure 2A A schematic diagram of a method flow chart for predicting fatigue life of another metal material provided in an embodiment of the present application is shown;

[0068] Figure 2B A schematic diagram of a fatigue life convergence trajectory provided by an embodiment of the present application is shown;

[0069] Figure 2C A schematic diagram of error convergence provided by an embodiment of the present application is shown;

[0070] Figure 2D A schematic diagram of reward value change provided by an embodiment of the present application is shown;

[0071] Figure 2E A schematic diagram of adjusting the step size factor provided in an embodiment of the present application is shown;

[0072] Figure 2F A schematic diagram showing a comparison between a predicted error and an actual error provided by an embodiment of the present application is shown;

[0073] Figure 2G A schematic flow chart of a modified SWT model prediction method based on reinforcement learning provided in an embodiment of the present application is shown;

[0074] Figure 3A A schematic diagram of a fatigue life prediction of a metal material provided in an embodiment of the present application is shown;

[0075] Figure 3B A schematic diagram showing the structure of another metal material fatigue life prediction provided by an embodiment of the present application is shown;

[0076] Figure 4 A schematic diagram of the device structure of a device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0077] The exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0078] The present application embodiment provides a method for predicting fatigue life of a metal material. Figure 1 As shown, the method includes:

[0079] 101. Obtain the SWT model, modify the SWT model, and obtain the target SWT model.

[0080] The nonlinear characteristics of the SWT equation make the solution process complicated, and traditional numerical methods such as the Newton-Raphson iteration method often face problems such as high computational complexity, slow convergence speed, and high sensitivity to the selection of initial values ​​in practical applications. These problems not only prolong the calculation time, but may also lead to instability and inaccuracy in the solution process, affecting the efficiency of fatigue life calculation and prediction reliability.

[0081] To solve this problem, this application proposes a fatigue life prediction method for metal materials. By building a customized reinforcement learning environment, combined with the Newton-Raphson iteration method, the optimization step factor adjustment strategy and the reward function design, the complex nonlinear modified SWT equation can be efficiently solved. The execution subject of this application can be a fatigue life prediction system. The fatigue life prediction system relies on the computing power of the server to provide services to users. The server can be an independent server, or it can provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), as well as big data and artificial intelligence platforms and other basic cloud computing servers, so that the fatigue life prediction system can efficiently and accurately predict fatigue life.

[0082] In the embodiment of the present application, the fatigue life prediction system obtains the SWT model, modifies the SWT model, and obtains the target SWT model. By considering the average stress, using the stress concentration factor and the surface roughness factor to modify the SWT model, a complete modified SWT fatigue life prediction model can be established, which significantly improves the calculation efficiency and accuracy.

[0083] 102. Obtain fatigue test data and use the fatigue test data and the target SWT model to build a reinforcement learning environment.

[0084] In the embodiment of the present application, the fatigue life prediction system obtains fatigue test data, and uses the fatigue test data and the target SWT model to build a reinforcement learning environment. By customizing the reinforcement learning environment, the target SWT equation solving problem is transformed into a decision-making process of interaction between an intelligent agent and the environment, thereby improving the model accuracy.

[0085] 103. Obtain a deep deterministic policy gradient algorithm, use the deep deterministic policy gradient algorithm to train the reinforcement learning environment, and obtain a fatigue life prediction model.

[0086] In the embodiment of the present application, the fatigue life prediction system obtains a deep deterministic policy gradient algorithm, and uses the deep deterministic policy gradient algorithm to train the reinforcement learning environment to obtain a fatigue life prediction model. By combining reinforcement learning with the Newton iteration method, it can be applied to the solution of complex nonlinear target SWT equations, which can not only improve the prediction efficiency and accuracy, but also make the fatigue life prediction model have good robustness and stability.

[0087] 104. Obtain metal material data, input the metal material data into a fatigue life prediction model for prediction, and obtain fatigue life prediction results.

[0088] In the embodiment of the present application, the fatigue life prediction system obtains metal material data, inputs the metal material data into the fatigue life prediction model for prediction, and obtains fatigue life prediction results. The fatigue life prediction model has good convergence and stability, can cope with the complexity and uncertainty in solving nonlinear equations, and has broad engineering application prospects and scientific research value.

[0089] The method provided in the embodiment of the present application obtains the SWT model, modifies the SWT model, obtains the target SWT model, obtains fatigue test data, uses the fatigue test data and the target SWT model to build a reinforcement learning environment, obtains a deep deterministic policy gradient algorithm, uses the deep deterministic policy gradient algorithm to train the reinforcement learning environment, obtains a fatigue life prediction model, obtains metal material data, inputs the metal material data into the fatigue life prediction model for prediction, and obtains fatigue life prediction results. By constructing a custom reinforcement learning environment, the modified SWT equation is converted into a decision-making process, the state includes the current fatigue life estimate and the fatigue life error value, the action is the adjustment of the step factor, and then the deep deterministic policy gradient algorithm is used to train the intelligent agent. The intelligent agent selects the step factor in each iteration and updates the current fatigue life estimate in combination with the Newton-Raphson iteration method. The reward function comprehensively considers the fatigue life error value, fatigue life deviation and the number of iterations to guide the intelligent agent to gradually approach the true solution. By dynamically adjusting the step factor, a balance between rapid convergence and precise adjustment can be achieved, significantly improving the solution efficiency and accuracy. By combining the reinforcement learning agent with the numerical solution advantages of the Newton iteration method, it is possible to continuously optimize the strategy based on environmental feedback and achieve efficient solution of nonlinear equations. At the same time, it can enhance the convergence and stability of the model and reduce sensitivity to initial values.

[0090] Further, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, the embodiment of the present application provides another fatigue life prediction method for metal materials, such as Figure 2A As shown, the method includes:

[0091] 201. Obtain the SWT model, modify the SWT model, and obtain the target SWT model.

[0092] In the embodiment of the present application, the fatigue life prediction system obtains the stress concentration factor and the surface roughness coefficient, and uses the stress concentration factor and the surface roughness coefficient to correct the SWT model to obtain the target SWT model as shown in the following formula 1:

[0093] Formula 1: ;

[0094] in, represents the comprehensive correction factor, b represents the fatigue strength index, c represents the fatigue ductility index, represents the stress amplitude, represents the maximum stress in the cycle, represents the cyclic strength coefficient, represents the cyclic strain hardening exponent, E represents the elastic modulus, Represents the Walker index, which characterizes the sensitivity of the material to the mean stress and is used to correct the effect of the mean stress on fatigue life. represents the fatigue strength coefficient, represents the fatigue ductility coefficient, represents the current fatigue life, where , q represents the notch sensitivity coefficient, represents the stress concentration factor, represents the surface roughness coefficient. The target SWT model takes into account the fatigue strength coefficient , fatigue ductility coefficient , Young's modulus E, fatigue strength index b, fatigue ductility index c and other parameters to ensure the accuracy and reliability of the target SWT model in practical applications.

[0095] 202. Obtain fatigue experimental data and use the fatigue experimental data and the target SWT model to build a reinforcement learning environment.

[0096] In the embodiment of the present application, in order to train the solution method of the target SWT model, a large amount of experimental data collection is required to obtain fatigue test data. Specifically, first select a metal material with uniformity and representativeness. Then, a fatigue experiment is carried out, using an extensometer to carry out fatigue experiments under different cyclic stress loading conditions, and the fatigue life (ie, the number of cycles) of each sample is recorded. Subsequently, data collection is carried out to collect data including key parameters such as maximum stress, strain amplitude, fatigue life, etc. to ensure the integrity and accuracy of the data.

[0097] Next, the fatigue life prediction system preprocesses the fatigue test data, cleans and normalizes the collected fatigue test data, removes outliers and noise, and ensures that the data quality is suitable for model training.

[0098] Subsequently, the fatigue life prediction system extracts multiple true fatigue life values ​​from the preprocessed fatigue test data, adds the multiple true fatigue life values ​​to the state space, obtains the step factor adjustment strategy, and adds the step factor adjustment strategy to the action space, wherein the step factor adjustment strategy is that the step factor gradually decreases with the training process. In the early stage of model learning, the step factor is large to accelerate convergence; when the model is close to the true solution, the step factor gradually decreases to improve convergence accuracy and stability. The adjustment range of the step factor is [-0.1, 0.1], and the step factor is used to control the update amplitude of the fatigue life estimate in each iterative training.

[0099] Then, the fatigue life prediction system obtains the reward function and uses the state space, action space, reward function, and target SWT model to build a reinforcement learning environment, where the reward function is the following formula 2:

[0100] Formula 2: ;

[0101] in, represents the reward function, represents the fatigue life error value, represents the fatigue life estimate, represents the true value of fatigue life, Represents the minimum value, which is used to avoid numerical problems of logarithmic functions. represents the number of iterations. The reward function comprehensively considers the fatigue life error, fatigue life deviation and the number of iterations, and can guide the agent to gradually approach the true solution. The fatigue life prediction system can accelerate the solution process of the modified SWT model and significantly improve the computational efficiency by optimizing the step size factor and the reward function.

[0102] 203. Obtain a first state, determine a first step length factor according to the first state based on the current Actor network, and execute the first step length factor in the first state based on the reinforcement learning environment to obtain a second state and a first reward value.

[0103] In an embodiment of the present application, the fatigue life prediction system uses a deep deterministic policy gradient (DDPG) algorithm to train the intelligent agent, so that it can accurately approximate the actual fatigue life value by continuously adjusting the step size factor and combining the Newton-Raphson iteration method.

[0104] Specifically, the fatigue life prediction system obtains the current Actor network, target Actor network, current Critic network, and target Critic network included in the deep deterministic policy gradient algorithm, and initializes the parameters of the current Actor network, target Actor network, current Critic network, and target Critic network. Next, the fatigue life prediction system initializes the reinforcement learning environment to obtain the first fatigue life estimate. It should be noted that the first fatigue life estimate is within a reasonable range to avoid large initial fluctuations. Subsequently, the fatigue life prediction system reads the first fatigue life true value that matches the first fatigue life estimate in the fatigue test data, and uses the difference between the first fatigue life estimate and the first fatigue life true value as the first fatigue life error value, and uses the first fatigue life estimate and the first fatigue life error value as the first state. It should be noted that the fatigue life error value is the difference between the two sides of the target SWT equation, reflecting the deviation between the current estimate and the true value. Then, the fatigue life prediction system determines the first step length factor based on the current Actor network and the first state, sends the first step length factor to the reinforcement learning environment, and calculates the first step length factor based on the reinforcement learning environment using the Newton-Raphson iteration method to obtain the second fatigue life estimate and the first reward. Combined with the numerical solution advantages of the Newton iteration method, the convergence and stability of the model can be enhanced and the sensitivity to the initial value can be reduced. Finally, the fatigue life prediction system reads the second fatigue life true value that matches the second fatigue life estimate in the fatigue test data, takes the difference between the second fatigue life estimate and the second fatigue life true value as the second fatigue life error value, and takes the second fatigue life estimate and the second fatigue life error value as the second state.

[0105] 204. Based on the current Actor network, determine the second step factor according to the second state, and perform the next round of training based on the second state and the second step factor, so as to iteratively train to obtain multiple states, multiple step factors, and multiple reward values.

[0106] In an embodiment of the present application, the fatigue life prediction system is based on the current Actor network, determines the second step factor according to the second state, and performs the next round of training based on the second state and the second step factor, so as to obtain multiple states, multiple step factors, and multiple reward values ​​through iterative training. It should be noted that during the training process, the step factor is dynamically adjusted according to the current iteration stage: in the initial stage, a larger step factor is used to accelerate the convergence process; in the convergence stage, the step factor is gradually reduced to improve the solution accuracy and stability and prevent oscillation caused by excessive adjustment.

[0107] 205. Store multiple states, step factors, and reward values ​​in the experience replay buffer.

[0108] In an embodiment of the present application, the fatigue life prediction system stores multiple states, step factors, and reward values ​​in an experience replay buffer for subsequent batch training.

[0109] 206. Obtain sample data from the experience replay buffer, use the sample data to iteratively train the current Actor network and the current Critic network, and obtain a fatigue life prediction model.

[0110] In an embodiment of the present application, a fatigue life prediction system obtains a first state, a second state, and a first step size factor in sample data, inputs the second state into a target Actor network for calculation, and obtains a second target step size factor, inputs the second target step size factor and the second state into a target Critic network for calculation, and obtains a first target value, inputs the first step size factor and the first state into a current Critic network for calculation, and obtains a first current value, and uses the first target value and a current value to train the current Actor network and the current Critic network.

[0111] Then, the fatigue life prediction system inputs the third state of the sample data into the target Actor network for calculation to obtain the third target step factor, inputs the third target step factor and the third state into the target Critic network for calculation to obtain the second target value, inputs the second step factor and the second state into the current Critic network for calculation to obtain the second current value, and uses the second target value and the second current value to perform the next round of training on the current Actor network and the current Critic network to obtain the fatigue life prediction model through iterative training.

[0112] The reinforcement learning agent has adaptive learning capabilities and can continuously optimize strategies based on environmental feedback to achieve efficient nonlinear equation solving, making the fatigue life prediction model adaptable and able to cope with different material parameters and complex nonlinear equations, thereby improving the versatility of the fatigue life prediction method for metal materials.

[0113] 207. The fatigue life prediction model is evaluated using multiple fatigue life estimation values, multiple fatigue life error values, multiple reward values, and multiple step factors obtained during the training process to obtain a model evaluation result; if the model evaluation result indicates that the fatigue life prediction model meets the fatigue life prediction conditions of metal materials, the fatigue life prediction model is stored.

[0114] In the embodiments of the present application, in order to comprehensively evaluate the convergence and stability of the fatigue life prediction model, a variety of scientific and academic standardized evaluation charts are used to evaluate the model training results.

[0115] Specifically, the fatigue life prediction system obtains multiple fatigue life estimates, and uses multiple fatigue life estimates to construct a fatigue life convergence trajectory diagram. The fatigue life convergence trajectory diagram can show the changes of the fatigue life estimates estimated by the intelligent agent with the number of iterations, and gradually approach the true value. Then, the fatigue life prediction system obtains multiple fatigue life error values, and uses multiple fatigue life error values ​​to construct an error convergence diagram. The error convergence diagram can show the changes of fatigue life errors with the number of iterations, and verify the convergence of the fatigue life prediction model. Subsequently, the fatigue life prediction system obtains multiple reward values, and uses multiple reward values ​​to construct a reward value change diagram. The reward value change diagram can show the changes of reward values ​​with the number of iterations, reflecting the model performance. Then, the fatigue life prediction system obtains multiple step factors, and uses multiple step factors to construct a step factor adjustment diagram. The step factor adjustment diagram can show the dynamic adjustment process of the step factor, which is helpful to determine whether the fatigue life prediction model converges. Then, the fatigue life prediction system extracts multiple true fatigue life values ​​from the fatigue test data, and uses multiple true fatigue life values ​​and multiple fatigue life estimates to construct a prediction and actual error comparison chart. The prediction and actual error comparison chart can compare the error predicted by the intelligent agent with the actual error, thereby evaluating the self-assessment ability of the model. Finally, the fatigue life prediction system sends the fatigue life convergence trajectory chart, error convergence chart, reward value change chart, step factor adjustment chart, and prediction and actual error comparison chart to the user's terminal for model evaluation operations, and obtains the model evaluation results based on the user's feedback from the terminal.

[0116] If the model evaluation result indicates that the fatigue life prediction model meets the fatigue life prediction conditions of metal materials, the fatigue life prediction system stores the fatigue life prediction model.

[0117] For example, the fatigue life prediction model of 2024-T3 aluminum alloy is evaluated. FIG. 2B to FIG. 2F Different evaluation charts of the fatigue life prediction model during the prediction process are shown. These charts intuitively show the convergence of the model, error reduction, reward value improvement, step factor adjustment, and comparison between prediction and actual error, which fully reflect the performance and learning effect of the model.

[0118] like Figure 2B As shown in the fatigue life convergence trajectory diagram, the initial estimated value of the agent deviates from the true value, reflecting the initial uncertainty and exploration. As the number of iterations increases, the estimated fatigue life gradually approaches the true value, reflecting the convergence of the model. There are slight fluctuations in the curve, which is caused by the trade-off between exploration and utilization of the agent, which is in line with the characteristics of reinforcement learning. In the end, the estimated value of the agent stabilizes near the true value, indicating that the model has high accuracy. The fatigue life prediction model successfully estimates the fatigue life accurately through iterative learning, showing good convergence and stability.

[0119] like Figure 2C The error convergence diagram shown in the figure shows an exponential decrease in the error, which is a linear decrease in logarithmic coordinates. Although the initial error is large, it decreases rapidly in the first few iterations, showing the efficient learning ability of the fatigue life prediction model. The error reduction rate then slows down, indicating that the fatigue life prediction model becomes more cautious in adjusting its pace when approaching the true value to avoid over-adjustment. It should be noted that slight fluctuations may be caused by exploratory actions or numerical calculation errors, but do not affect the overall convergence trend. It can be seen from the error convergence diagram that the fatigue life prediction model has the ability to quickly reduce the error, and the error convergence effect is ideal, which shows the effectiveness of the fatigue life prediction model in predicting the solution of the modified SWT equation.

[0120] like Figure 2D The reward value change graph shown in the figure shows an overall upward trend, indicating that the performance of the agent is gradually improving. The reward value is low in the initial stage because the error is large and the agent has not yet learned how to adjust. As the error decreases, the reward value gradually increases, which reflects the positive incentive effect of the reward function on error reduction. Moreover, the fluctuations in the reward curve correspond to the fluctuations in the error curve, which reflects the sensitivity of the reward mechanism to the model performance. Therefore, the reward mechanism effectively guides the agent to reduce errors and improves the learning efficiency and convergence speed of the fatigue life prediction model.

[0121] like Figure 2E As shown in the step factor adjustment diagram, the step factor gradually decreases with the number of iterations, from the initial large value to a smaller value. A larger step factor in the initial stage helps the agent to quickly explore the solution space and accelerate convergence. When it is close to the true value, the step factor decreases, which is conducive to fine-tuning and preventing over-correction from causing oscillation. The fluctuation of the step factor reflects the ability of the agent to dynamically adjust the learning rate according to the current state. By adaptively adjusting the step factor, the stability and accuracy of the fatigue life prediction model can be improved, showing a good learning strategy.

[0122] like Figure 2F The comparison chart of predicted and actual errors shows that the trends of predicted error and actual error are basically consistent, and both show a trend of decreasing with the increase of iteration number. The closeness of the two curves indicates that the agent's estimation of its own error is relatively accurate. It should be noted that the fluctuations in the curves may be caused by numerical calculation errors or the exploration behavior of the agent. In the end, both the predicted error and the actual error reached a low level, which can verify the accuracy of the fatigue life prediction model. In summary, through a variety of scientific and academic standardized evaluation charts, the learning process and performance of the model can be fully demonstrated, which is convenient for verification and application in academic research and engineering applications, and improves the reliability and stability of the fatigue life prediction model.

[0123] 208. Obtain metal material data, input the metal material data into a fatigue life prediction model for prediction, and obtain fatigue life prediction results.

[0124] In an embodiment of the present application, a fatigue life prediction system acquires metal material data, inputs the metal material data into a fatigue life prediction model for prediction, and obtains a fatigue life prediction result.

[0125] like Figure 2G The process of the modified SWT equation prediction method based on reinforcement learning is shown. First, a fatigue life prediction model is established by modifying the SWT equation. Then, fatigue data collection and preprocessing are performed. Subsequently, a reinforcement learning environment is constructed, the state space and action space are defined, the fatigue life error, fatigue life deviation, number of iterations and reward function are designed, and the modified SWT equation solution process is transformed into a decision-making process of multi-agent interaction with the environment. Then, the DDPG algorithm is used to train the agent, the step size factor is optimized through multiple iterative learning, and the fatigue life estimate is updated in combination with the Newton iteration method. Finally, the model performance is comprehensively evaluated by generating a variety of evaluation charts. This application combines reinforcement learning with the Newton iteration method and applies it to the solution of complex nonlinear modified SWT equations, demonstrating its potential and advantages in material fatigue life prediction, which can not only improve the prediction efficiency and accuracy, but also has good robustness and stability, and has broad engineering application prospects and scientific research value. Moreover, by constructing a customized reinforcement learning environment, optimizing the reward function and the step size adjustment strategy, it can cope with the complexity and uncertainty in the solution of nonlinear equations, thereby meeting the high standard requirements of modern engineering for material performance evaluation.

[0126] The method provided in the embodiment of the present application obtains the SWT model, modifies the SWT model, obtains the target SWT model, obtains fatigue test data, uses the fatigue test data and the target SWT model to build a reinforcement learning environment, obtains a deep deterministic policy gradient algorithm, uses the deep deterministic policy gradient algorithm to train the reinforcement learning environment, obtains a fatigue life prediction model, obtains metal material data, inputs the metal material data into the fatigue life prediction model for prediction, and obtains fatigue life prediction results. By constructing a custom reinforcement learning environment, the modified SWT equation is converted into a decision-making process, the state includes the current fatigue life estimate and the fatigue life error value, the action is the adjustment of the step factor, and then the deep deterministic policy gradient algorithm is used to train the intelligent agent. The intelligent agent selects the step factor in each iteration and updates the current fatigue life estimate in combination with the Newton-Raphson iteration method. The reward function comprehensively considers the fatigue life error value, fatigue life deviation and the number of iterations to guide the intelligent agent to gradually approach the true solution. By dynamically adjusting the step factor, a balance between rapid convergence and precise adjustment can be achieved, significantly improving the solution efficiency and accuracy. By combining the reinforcement learning agent with the numerical solution advantages of the Newton iteration method, it is possible to continuously optimize the strategy based on environmental feedback and achieve efficient solution of nonlinear equations. At the same time, it can enhance the convergence and stability of the model and reduce sensitivity to initial values.

[0127] Further, as Figure 1 In a specific implementation of the method, the present application provides a fatigue life prediction device for metal materials, such as Figure 3A As shown, the device includes: a correction module 301, a construction module 302, a training module 303 and a prediction module 304.

[0128] A correction module 301 is used to obtain a SWT model, correct the SWT model, and obtain a target SWT model;

[0129] A construction module 302 is used to obtain fatigue test data and construct a reinforcement learning environment using the fatigue test data and the target SWT model;

[0130] A training module 303 is used to obtain a deep deterministic policy gradient algorithm, and use the deep deterministic policy gradient algorithm to train the reinforcement learning environment to obtain a fatigue life prediction model;

[0131] The prediction module 304 is used to obtain metal material data, input the metal material data into the fatigue life prediction model for prediction, and obtain a fatigue life prediction result.

[0132] In a specific application scenario, the correction module 301 is used to obtain a stress concentration factor and a surface roughness factor, and use the stress concentration factor and the surface roughness factor to correct the SWT model to obtain the target SWT model, wherein:

[0133] ;

[0134] in, represents the comprehensive correction factor, b represents the fatigue strength index, c represents the fatigue ductility index, represents the stress amplitude, represents the maximum stress in the cycle, represents the cyclic strength coefficient, represents the cyclic strain hardening exponent, E represents the elastic modulus, represents the Walker index, represents the fatigue strength coefficient, represents the fatigue ductility coefficient, represents the current fatigue life, where , q represents the notch sensitivity coefficient, represents the stress concentration factor, Represents the surface roughness coefficient.

[0135] In a specific application scenario, the construction module 302 is used to preprocess the fatigue test data, extract multiple fatigue life real values ​​from the preprocessed fatigue test data, and add the multiple fatigue life real values ​​to the state space; obtain a step factor adjustment strategy, add the step factor adjustment strategy to the action space, the step factor adjustment strategy is that the step factor gradually decreases with the training process, the adjustment range of the step factor is [-0.1, 0.1], and the step factor is used to control the update amplitude of the fatigue life estimation value in each iterative training; obtain a reward function, and use the state space, the action space, the reward function, and the target SWT model to construct the reinforcement learning environment, wherein,

[0136] ;

[0137] in, represents the reward function, represents the fatigue life error value, represents the fatigue life estimate, represents the true value of fatigue life, represents the minimum value, Indicates the number of iterations.

[0138] In a specific application scenario, the training module 303 is used to obtain the current Actor network, target Actor network, current Critic network, and target Critic network included in the deep deterministic policy gradient algorithm; obtain a first state, determine a first step size factor based on the current Actor network according to the first state, and execute the first step size factor in the first state based on the reinforcement learning environment to obtain a second state and a first reward value; determine a second step size factor based on the current Actor network according to the second state, and perform the next round of training based on the second state and the second step size factor to iteratively train to obtain multiple states, multiple step size factors, and multiple reward values; store the multiple states, the step size factors, and the reward values ​​in an experience replay buffer; obtain sample data from the experience replay buffer, and use the sample data to iteratively train the current Actor network and the current Critic network to obtain the fatigue life prediction model.

[0139] In a specific application scenario, the training module 303 is used to initialize the reinforcement learning environment to obtain a first fatigue life estimate; read a first fatigue life true value matching the first fatigue life estimate in the fatigue test data; use the difference between the first fatigue life estimate and the first fatigue life true value as a first fatigue life error value, and use the first fatigue life estimate and the first fatigue life error value as a first state; based on the current Actor network, determine a first step length factor according to the first state, and send the first step length factor to the reinforcement learning environment; obtain the Newton-Raphson iteration method, and based on the reinforcement learning environment, use the Newton-Raphson iteration method to calculate the first step length factor to obtain a second fatigue life estimate and a first reward; read a second fatigue life true value matching the second fatigue life estimate in the fatigue test data, use the difference between the second fatigue life estimate and the second fatigue life true value as a second fatigue life error value, and use the second fatigue life estimate and the second fatigue life error value as the second state.

[0140] In a specific application scenario, the training module 303 is used to obtain a first state, a second state and a first step factor from the sample data, input the second state into the target Actor network for calculation to obtain a second target step factor, input the second target step factor and the second state into the target Critic network for calculation to obtain a first target value, input the first step factor and the first state into the current Critic network for calculation to obtain a first current value, and use the first target value and the first current value to train the current Actor network and the current Critic network; input the third state of the sample data into the target Actor network for calculation to obtain a third target step factor, input the third target step factor and the third state into the target Critic network for calculation to obtain a second target value, input the second step factor and the second state into the current Critic network for calculation to obtain a second current value, and use the second target value and the second current value to perform the next round of training on the current Actor network and the current Critic network to iteratively train and obtain the fatigue life prediction model.

[0141] In specific application scenarios, such as Figure 3B As shown, the device further includes: an evaluation module 305 .

[0142] The evaluation module 305 is used to obtain multiple fatigue life estimation values, and use the multiple fatigue life estimation values ​​to construct a fatigue life convergence trajectory diagram; obtain multiple fatigue life error values, and use the multiple fatigue life error values ​​to construct an error convergence diagram; obtain multiple reward values, and use the multiple reward values ​​to construct a reward value change diagram; obtain multiple step factors, and use the multiple step factors to construct a step factor adjustment diagram; extract multiple fatigue life true values ​​from the fatigue test data, and use the multiple fatigue life true values ​​and the multiple fatigue life estimation values ​​to construct a prediction and actual error comparison diagram; send the fatigue life convergence trajectory diagram, the error convergence diagram, the reward value change diagram, the step factor adjustment diagram, and the prediction and actual error comparison diagram to the terminal held by the user for model evaluation operation; obtain the model evaluation result fed back by the user based on the terminal held; if the model evaluation result indicates that the fatigue life prediction model meets the fatigue life prediction conditions of metal materials, the fatigue life prediction model is stored.

[0143] The device provided in the embodiment of the present application obtains the SWT model, modifies the SWT model, obtains the target SWT model, obtains fatigue test data, uses the fatigue test data and the target SWT model to build a reinforcement learning environment, obtains a deep deterministic policy gradient algorithm, uses the deep deterministic policy gradient algorithm to train the reinforcement learning environment, obtains a fatigue life prediction model, obtains metal material data, inputs the metal material data into the fatigue life prediction model for prediction, and obtains fatigue life prediction results. By constructing a custom reinforcement learning environment, the modified SWT equation is converted into a decision-making process, the state includes the current fatigue life estimate and the fatigue life error value, the action is the adjustment of the step factor, and then the deep deterministic policy gradient algorithm is used to train the intelligent agent. The intelligent agent selects the step factor in each iteration and updates the current fatigue life estimate in combination with the Newton-Raphson iteration method. The reward function comprehensively considers the fatigue life error value, fatigue life deviation and the number of iterations, and guides the intelligent agent to gradually approach the true solution. By dynamically adjusting the step factor, a balance between rapid convergence and precise adjustment can be achieved, significantly improving the solution efficiency and accuracy. By combining the reinforcement learning agent with the numerical solution advantages of the Newton iteration method, it is possible to continuously optimize the strategy based on environmental feedback and achieve efficient solution of nonlinear equations. At the same time, it can enhance the convergence and stability of the model and reduce sensitivity to initial values.

[0144] It should be noted that for other corresponding descriptions of the functional units involved in the fatigue life prediction device for metal materials provided in the embodiment of the present application, reference can be made to Figure 1 and Figures 2A to 2G The corresponding description in will not be repeated here.

[0145] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0146] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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, they should be considered to be within the scope of this specification.

[0147] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

[0148] In an exemplary embodiment, see Figure 4 , and also provides a device, which includes a bus, a processor, a memory and a communication interface, and may also include an input and output interface and a display device, wherein each functional unit can communicate with each other through the bus. The memory stores a computer program, and the processor is used to execute the program stored in the memory and execute the fatigue life prediction method of the metal material in the above embodiment.

[0149] A storage medium stores a computer program, which, when executed by a processor, implements the steps of the fatigue life prediction method for metal materials.

[0150] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by hardware, or by means of software plus necessary general hardware platforms. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0151] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily required for implementing the present application.

[0152] Those skilled in the art will appreciate that the modules in the device in the implementation scenario can be distributed in the device in the implementation scenario according to the implementation scenario description, or can be changed accordingly and located in one or more devices different from the implementation scenario. The modules in the above implementation scenario can be combined into one module, or can be further split into multiple sub-modules.

[0153] The above application serial numbers are for description only and do not represent the advantages or disadvantages of the implementation scenarios.

[0154] The above disclosure only discloses several specific implementation scenarios of the present application. However, the present application is not limited thereto, and any changes that can be conceived by technicians in this field should fall within the scope of protection of the present application.

Claims

1. A fatigue life prediction method for metal materials, characterized in that: include: Acquiring a SWT model, and modifying the SWT model to obtain a target SWT model, including: acquiring a stress concentration factor and a surface roughness factor, and modifying the SWT model using the stress concentration factor and the surface roughness factor to obtain the target SWT model, wherein: Among them, K f ′ represents the comprehensive correction factor, b represents the fatigue strength index, c represents the fatigue ductility index, σ a represents the stress amplitude, σ max represents the maximum stress in the cycle, K ′ represents the cyclic strength coefficient, n ′ represents the cyclic strain hardening index, E represents the elastic modulus, γ represents the Walker index, σ f ′ represents the fatigue strength coefficient, ε f ′ represents the fatigue ductility coefficient, N f represents the current fatigue life, where K f ′ =[1+q(K t -1)]×C s , q represents the notch sensitivity coefficient, K t represents the stress concentration factor, C s It represents the surface roughness coefficient; Fatigue test data are obtained, and a reinforcement learning environment is constructed using the fatigue test data and the target SWT model, including: preprocessing the fatigue test data, extracting multiple fatigue life real values ​​from the preprocessed fatigue test data, and adding the multiple fatigue life real values ​​to the state space; obtaining a step factor adjustment strategy, adding the step factor adjustment strategy to the action space, wherein the step factor adjustment strategy is that the step factor gradually decreases with the training process, and the adjustment range of the step factor is [-0.1, 0.1]. The step factor is used to control the update amplitude of the fatigue life estimation value in each iterative training; obtaining a reward function, and using the state space, the action space, the reward function, and the target SWT model to construct the reinforcement learning environment, wherein, R=-log 10 (|F(N f )|+∈)-0.5log 10 (|N f -N f,true |+∈)-0.001n; Among them, R represents the reward function, F(N f ) represents the fatigue life error value, N f represents the estimated fatigue life, N f,true represents the true value of fatigue life, ∈ represents the minimum value, and n represents the number of iterations; Obtaining a deep deterministic policy gradient algorithm, and using the deep deterministic policy gradient algorithm to train the reinforcement learning environment to obtain a fatigue life prediction model; The metal material data is acquired, and the metal material data is input into the fatigue life prediction model for prediction to obtain a fatigue life prediction result.

2. The method according to claim 1, characterized in that The obtaining of a deep deterministic policy gradient algorithm, using the deep deterministic policy gradient algorithm to train the reinforcement learning environment to obtain a fatigue life prediction model, includes: Obtain the current Actor network, target Actor network, current Critic network, and target Critic network included in the deep deterministic policy gradient algorithm; Obtain a first state, determine a first step length factor according to the first state based on the current Actor network, and execute the first step length factor in the first state based on the reinforcement learning environment to obtain a second state and a first reward value; Based on the current Actor network, determining a second step factor according to a second state, and performing a next round of training based on the second state and the second step factor, so as to iteratively train to obtain a plurality of states, a plurality of step factors, and a plurality of reward values; Storing the multiple states, the step size factor, and the reward value in an experience replay buffer; Sample data is obtained from the experience playback buffer, and the current Actor network and the current Critic network are iteratively trained using the sample data to obtain the fatigue life prediction model.

3. The method according to claim 2, characterized in that The obtaining of the first state, determining a first step length factor according to the first state based on the current Actor network, and executing the first step length factor in the first state based on the reinforcement learning environment to obtain a second state and a first reward value, includes: Initializing the reinforcement learning environment to obtain a first fatigue life estimate; Reading a first fatigue life true value matching the first fatigue life estimate value in the fatigue test data; taking the difference between the first fatigue life estimate and the first fatigue life true value as a first fatigue life error value, and taking the first fatigue life estimate and the first fatigue life error value as a first state; Based on the current Actor network, determining a first step length factor according to the first state, and sending the first step length factor to the reinforcement learning environment; Obtaining a Newton-Raphson iteration method, and calculating the first step length factor using the Newton-Raphson iteration method based on the reinforcement learning environment to obtain a second fatigue life estimate and a first reward; A second fatigue life true value matching the second fatigue life estimate is read from the fatigue test data, a difference between the second fatigue life estimate and the second fatigue life true value is taken as a second fatigue life error value, and the second fatigue life estimate and the second fatigue life error value are taken as the second state.

4. The method according to claim 2, characterized in that: The iterative training of the current Actor network and the current Critic network using the sample data to obtain the fatigue life prediction model includes: Obtaining a first state, a second state, and a first step length factor from the sample data, inputting the second state into the target Actor network for calculation to obtain a second target step length factor, inputting the second target step length factor and the second state into the target Critic network for calculation to obtain a first target value, inputting the first step length factor and the first state into the current Critic network for calculation to obtain a first current value, and using the first target value and the first current value to train the current Actor network and the current Critic network; The third state of the sample data is input into the target Actor network for calculation to obtain a third target step factor, the third target step factor and the third state are input into the target Critic network for calculation to obtain a second target value, the second step factor and the second state are input into the current Critic network for calculation to obtain a second current value, and the second target value and the second current value are used to perform the next round of training on the current Actor network and the current Critic network, so as to iteratively train and obtain the fatigue life prediction model.

5. The method according to claim 1, characterized in that The method further comprises: Acquire a plurality of fatigue life estimation values, and construct a fatigue life convergence trajectory diagram using the plurality of fatigue life estimation values; Acquire a plurality of fatigue life error values, and construct an error convergence diagram using the plurality of fatigue life error values; Obtain multiple reward values, and use the multiple reward values ​​to construct a reward value change graph; Obtain multiple step factors, and use the multiple step factors to construct a step factor adjustment graph; Extracting a plurality of true fatigue life values ​​from the fatigue test data, and constructing a prediction and actual error comparison diagram using the plurality of true fatigue life values ​​and the plurality of fatigue life estimation values; Sending the fatigue life convergence trajectory diagram, the error convergence diagram, the reward value change diagram, the step factor adjustment diagram, and the predicted and actual error comparison diagram to a terminal held by a user for model evaluation operation; Obtaining a model evaluation result based on feedback from the user's terminal; If the model evaluation result indicates that the fatigue life prediction model meets the fatigue life prediction conditions of metal materials, the fatigue life prediction model is stored.

6. A fatigue life prediction device for metal materials, characterized in that: include: The correction module is used to obtain the SWT model, correct the SWT model, and obtain the target SWT model, including: obtaining a stress concentration factor and a surface roughness coefficient, and correcting the SWT model using the stress concentration factor and the surface roughness coefficient to obtain the target SWT model, wherein: Among them, K f ′ represents the comprehensive correction factor, b represents the fatigue strength index, c represents the fatigue ductility index, σ a represents the stress amplitude, σ max represents the maximum stress in the cycle, K ′ represents the cyclic strength coefficient, n ′ represents the cyclic strain hardening index, E represents the elastic modulus, γ represents the Walker index, σ f ′ represents the fatigue strength coefficient, ε f ′ represents the fatigue ductility coefficient, N f represents the current fatigue life, where K f ′ =[1+q(K t -1)]×C s , q represents the notch sensitivity coefficient, K t represents the stress concentration factor, C s It represents the surface roughness coefficient; A construction module is used to obtain fatigue test data, and use the fatigue test data and the target SWT model to construct a reinforcement learning environment, including: preprocessing the fatigue test data, extracting multiple fatigue life real values ​​from the preprocessed fatigue test data, and adding the multiple fatigue life real values ​​to the state space; obtaining a step factor adjustment strategy, adding the step factor adjustment strategy to the action space, the step factor adjustment strategy is that the step factor gradually decreases with the training process, the adjustment range of the step factor is [-0.1, 0.1], and the step factor is used to control the update amplitude of the fatigue life estimation value in each iterative training; obtaining a reward function, using the state space, the action space, the reward function, and the target SWT model to construct the reinforcement learning environment, wherein, R=-log 10 (|F(N f )|+∈)-0.5log 10 (|N f -N f,true |+∈)-0.001n; Among them, R represents the reward function, F(N f ) represents the fatigue life error value, N f represents the estimated fatigue life, N f,true represents the true value of fatigue life, ∈ represents the minimum value, and n represents the number of iterations; A training module, used to obtain a deep deterministic policy gradient algorithm, and use the deep deterministic policy gradient algorithm to train the reinforcement learning environment to obtain a fatigue life prediction model; The prediction module is used to obtain metal material data, input the metal material data into the fatigue life prediction model for prediction, and obtain a fatigue life prediction result.

7. A device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images