Material fatigue prediction method and system based on physical guidance neural network

Abstract

The invention relates to the technical field of material fatigue prediction, in particular to a material fatigue prediction method and system based on a physical guidance neural network, and the method comprises the steps: obtaining mechanical response data and material microstructure feature data in a material cyclic loading process; based on the mechanical response data and the material microstructure feature data, constructing a physical guidance neural network model; the method comprises the following steps: a) acquiring fatigue damage parameters through a physical constraint unit by utilizing a physical guidance neural network model; b) incorporating microstructure characteristics of the material into a fatigue damage evolution equation through an attention layer; c) quantizing the uncertainty of the fatigue damage evolution equation through an uncertainty quantization unit; d) obtaining a prediction result through a feature mapping unit; the prediction result is subjected to uncertainty quantification through the prediction unit, a final material fatigue prediction result is obtained, material mechanics knowledge is introduced into the neural network through the physical constraint unit, and it is ensured that the prediction result conforms to the basic physical law.

Description

Technical Field

[0001] The present invention relates to the technical field of material fatigue prediction, and more specifically, to a material fatigue prediction method and system based on a physics-guided neural network. Background Art

[0002] In the field of modern engineering, material fatigue has always been an important and challenging research topic. With the continuous advancement of industrial technology and the increasing complexity of engineering structures, it is becoming increasingly important to accurately predict the fatigue behavior of materials. Traditional fatigue prediction methods mainly rely on experimental data and empirical formulas. Although these methods perform well in certain specific cases, they often have difficulty in dealing with complex loading conditions and new materials.

[0003] In recent years, with the rapid development of artificial intelligence technology, fatigue prediction methods based on machine learning have gradually attracted attention. These methods use a large amount of experimental data to establish the relationship between material properties and fatigue life through algorithms such as neural networks. However, purely data-driven methods also have some inherent limitations. First, these methods usually lack physical interpretability and it is difficult to provide a reasonable physical explanation for the prediction results. Secondly, when faced with situations beyond the range of training data, these methods often have poor extrapolation capabilities and may produce results that contradict physical principles. In addition, existing machine learning methods usually have difficulty in effectively dealing with the influence of material microstructure on fatigue behavior, which is crucial in many cases.

[0004] On the other hand, although fatigue prediction methods based on physical models have good interpretability and generalization capabilities, they often have difficulty capturing the complex nonlinear characteristics of material fatigue behavior. In particular, when considering the influence of microstructure, traditional physical models often appear to be inadequate. In addition, these methods usually require a large number of material parameters, and the acquisition of these parameters often requires complex experiments and a lot of time costs.

[0005] Given the shortcomings of existing technologies, a new fatigue prediction method that can combine the advantages of physical knowledge and data-driven methods is urgently needed. This method should not only maintain the interpretability and generalization capabilities of physical models, but also make full use of the nonlinear fitting capabilities of data-driven methods. At the same time, it should also be able to effectively process material microstructure information and provide reliable uncertainty quantification. Summary of the invention

[0006] The present invention is an innovative solution to the above technical problems. By introducing a physics-guided neural network, the present invention cleverly combines physical knowledge with a data-driven approach to achieve high-precision, explainable material fatigue prediction. This method not only overcomes the problem of pure data-driven methods lacking physical interpretability and extrapolation capabilities, but also solves the difficulty of traditional physical models in handling complex nonlinear relationships.

[0007] The present invention provides a material fatigue prediction method based on a physics-guided neural network, comprising: The acquisition steps include: Obtain the mechanical response data and material microstructure characteristic data during cyclic loading of the material; Processing steps include: Based on the mechanical response data and the material microstructure characteristic data, construct a physical guided neural network model; Using the physical guided neural network model, the following operations are performed: a) Obtain fatigue damage parameters through physical constraint units; b) Incorporating the material's microstructural characteristics into the fatigue damage evolution equation through the attention layer; c) Quantify the uncertainty of fatigue damage evolution equation through uncertainty quantification unit; d) Obtaining prediction results through feature mapping units; Output steps include: The uncertainty of the prediction results is quantified through the prediction unit to obtain the final material fatigue prediction results.

[0008] Preferably, the mechanical response data include load, displacement, strain, stress, strain rate, strain energy density and fatigue life.

[0009] Preferably, the fatigue damage parameters include stress, strain, strain energy density, hardening exponent and plastic strain.

[0010] Preferably, the fatigue damage evolution equation is: , in, Indicates fatigue damage, represents the equivalent plastic strain amplitude, Indicates the current cycle number.

[0011] Preferably, the expression model of the feature mapping unit is: , in, Indicates Inputs, Indicates The output of each physical layer node, and denote the weight matrix and the offset vector respectively, represents a non-linear activation function.

[0012] Preferably, the expression model of the attention layer is: , in, represents the output of the attention layer, Indicates The attention coefficient of the input, Indicates input.

[0013] Preferably, the expression model of the uncertainty quantification unit is: mean ,std , in, Represents the model output results, represents the model prediction value, Represents normal distribution, mean() represents the mean function, and std() represents the standard deviation function.

[0014] As a preference, an adaptive learning step is also included: Based on the initial fatigue parameters, the initial fatigue damage evolution equation is established; Through iterative optimization and continuous adjustment of fatigue parameters, the final fatigue damage evolution equation is obtained.

[0015] Preferably, the expression of the initial fatigue parameter is: , in, represents the material constant, represents the equivalent plastic strain amplitude, represents the starting point of fatigue plastic strain, Indicates the hardening index.

[0016] A material fatigue prediction system based on physics-guided neural networks, including: A data acquisition module is used to obtain the mechanical response data and material microstructure characteristic data during the material cyclic loading process; A physical constraint module, used for obtaining fatigue damage parameters based on the mechanical response data and material microstructure characteristic data; an attention module to incorporate the material’s microstructural characteristics into the fatigue damage evolution equation; Uncertainty quantification module, used to quantify the uncertainty of fatigue damage evolution equations; Feature mapping module, used to obtain prediction results; The prediction module is used to quantify the uncertainty of the prediction results and obtain the final material fatigue prediction results; Among them, the physical constraint module, attention module, uncertainty quantification module, feature mapping module and prediction module constitute a physical guided neural network model.

[0017] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0018] Specifically, the present invention introduces material mechanics knowledge into the neural network through the physical constraint unit, ensuring that the prediction results conform to the basic physical laws. The introduction of the attention layer enables the model to adaptively process the material microstructure information, which is difficult to achieve with traditional methods. The design of the uncertainty quantification unit provides a reliability assessment for the prediction results, which is crucial for engineering decision-making.

[0019] More importantly, the method of the present invention achieves the synergistic effect of physical models and data-driven methods. The prior knowledge provided by the physical model provides effective constraints and guidance for the training of the neural network, reduces the risk of overfitting, and improves the generalization ability of the model. At the same time, the nonlinear fitting ability of the neural network complements the shortcomings of the physical model and can capture complex material behaviors. This complementary and synergistic effect significantly improves the accuracy and reliability of fatigue prediction.

[0020] In addition, the adaptive learning mechanism of the present invention enables the model to continuously optimize the fatigue damage evolution equation, which makes the prediction method highly adaptable and able to cope with various materials and loading conditions. This adaptive capability is not available in traditional methods, which enables the method of the present invention to maintain high accuracy when facing new materials or complex working conditions.

[0021] In general, the material fatigue prediction method and system based on physics-guided neural network provided by the present invention successfully solves the problem that physical models and data-driven methods in the prior art are difficult to balance, and achieves high-precision, explainable and reliable fatigue prediction. This innovation not only promotes the development of material fatigue prediction technology, but also provides a powerful tool for material design, structural optimization and life assessment, and is expected to have an important impact in many fields such as aerospace, automotive industry, energy engineering, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The present invention is a flow chart of the method.

[0023] Figure 2 This is a diagram of the internal structure of the physical guided neural network model of the present invention.

[0024] Figure 3 The figure is a flow chart of the adaptive learning method of the present invention.

[0025] Figure 4 Flow chart of the uncertainty quantification method of the present invention. DETAILED DESCRIPTION

[0026] Please refer to Figure 1-4 The present invention provides a material fatigue prediction method and system based on a physics-guided neural network. The method achieves high-precision and explainable material fatigue prediction by combining physical models and data-driven methods.

[0027] First, the method includes an acquisition step, a processing step and an output step. In the acquisition step, the mechanical response data and the material microstructure characteristic data during the material cyclic loading process are acquired. These data are the basis for material fatigue prediction.

[0028] Specifically, mechanical response data include load, displacement, strain, stress, strain rate, strain energy density and fatigue life. For example, load data can be the stress amplitude during cyclic loading, usually in MPa; displacement data can be the deformation of the material during loading, usually in mm; strain data reflects the relative deformation degree of the material, usually expressed in percentage; stress data represents the stress state inside the material, also in MPa; strain rate data reflects the speed of material deformation, which can be in s -1 ; Strain energy density data represents the energy absorbed per unit volume of material, and the unit can be J / m 3 ; Fatigue life data are usually expressed in terms of number of cycles.

[0029] In the processing step, a physics-guided neural network model is constructed based on the acquired mechanical response data and material microstructure characteristic data. This is one of the core innovations of the present invention, which improves the interpretability and generalization ability of the model by introducing physical knowledge into the neural network.

[0030] Next, a series of operations are performed using the constructed physics-guided neural network model. First, fatigue damage parameters are obtained through physical constraint units. Fatigue damage parameters include stress, strain, strain energy density, hardening index, and plastic strain. These parameters describe the fatigue state of the material from a physical perspective and provide an important basis for subsequent predictions.

[0031] For example, the stress parameter can be selected as equivalent stress, such as Von Mises stress, in MPa; the strain parameter can be selected as equivalent plastic strain, dimensionless; the strain energy density parameter reflects the energy absorbed by the material, in J / m 3 ; The hardening exponent reflects the strain hardening characteristics of the material and is usually a dimensionless positive number; the plastic strain parameter describes the irreversible deformation of the material and is also dimensionless.

[0032] Then, the microstructural characteristics of the material are incorporated into the fatigue damage evolution equation through the attention layer. This step takes into account the influence of the material microstructure on the fatigue behavior and improves the accuracy of the prediction. Microstructural characteristics may include grain size (usually in μm), crystal orientation distribution (which can be expressed in Euler angles), volume fraction of second phase particles (expressed in percentage), etc.

[0033] Next, the uncertainty of the fatigue damage evolution equation is quantified by the uncertainty quantification unit. This step takes into account the uncertainty in the prediction process and improves the reliability of the prediction results. Uncertainty quantification can be expressed using probability distributions, such as normal distribution or beta distribution.

[0034] Subsequently, the prediction results are obtained through the feature mapping unit, which combines the physical model with the neural network model to achieve nonlinear mapping from input features to prediction results.

[0035] Finally, in the output step, the prediction unit quantifies the uncertainty of the prediction results to obtain the final material fatigue prediction results. This step not only gives the prediction value, but also provides the confidence interval of the prediction results, providing more comprehensive information for engineering decision-making.

[0036] One advantage of the method of the present invention is that it combines the interpretability of the physical model with the nonlinear fitting ability of the neural network. By introducing the physical constraint unit, the method can ensure that the prediction results conform to the basic principles of material mechanics and avoid the physically unreasonable results that may occur in pure data-driven methods. At the same time, the introduction of the attention layer enables the model to adaptively focus on the impact of different microscopic features on fatigue behavior, improving the accuracy of the prediction and the adaptability of the model.

[0037] In addition, the application of uncertainty quantification unit enables this method to not only give point estimates but also provide reliability assessment of prediction results, which is of great significance for risk assessment and decision making in engineering applications. The design of feature mapping unit and prediction unit ensures that the model has sufficient complexity to capture the nonlinear characteristics of material fatigue behavior.

[0038] In general, the method provided by the present invention achieves high-precision prediction of material fatigue behavior through physical guided neural networks, providing a powerful tool for material design, structural optimization and life assessment. The application of this method is expected to significantly improve the efficiency and accuracy of material fatigue performance assessment, which is of great significance for improving the safety and reliability of engineering structures. In a preferred embodiment of the present invention, the fatigue damage evolution equation is designed as a dynamic, cyclically dependent expression. Specifically, the equation can be expressed as: , in, Indicates fatigue damage, represents the equivalent plastic strain amplitude, represents the current cycle number. This equation cleverly captures the essence of fatigue damage accumulation, that is, the damage state of the current cycle depends on the damage state of the previous cycle and the degree of plastic deformation of the current cycle. Function Here, it plays a key role and can be designed as a nonlinear function to reflect the complex behavior of the material during fatigue. For example, Can be defined as: , in, , and is a material-related parameter that can be obtained by fitting experimental data. Usually, The possible value range is between 0.1 and 10. and It may be between 0.5 and 2. This form of function can reflect the nonlinear growth of fatigue damage with the number of cycles, as well as the relationship between the damage growth rate and the degree of plastic deformation.

[0039] Another innovative point of the method of the present invention is the design of the feature mapping unit. The expression model of the feature mapping unit is defined as: , In this expression, Indicates Inputs, Indicates The output of each physical layer node, and denote the weight matrix and the offset vector respectively, represents a nonlinear activation function. This design allows the model to learn complex nonlinear relationships, thereby better capturing the complexity of material fatigue behavior. Preferably, the nonlinear activation function You can choose the ReLU (rectified linear unit) function, that is: , The advantage of the ReLU function is that it can effectively alleviate the gradient vanishing problem, and at the same time has high computational efficiency, which is conducive to the training of deep networks.

[0040] In one embodiment of the present invention, the expression model of the attention layer is designed as follows: , here, represents the output of the attention layer, Indicates The attention coefficient of the input, Indicates Inputs. Attention coefficient It can be calculated by the softmax function: , in, It is an attention score calculated by another small neural network. This design enables the model to adaptively adjust the degree of attention to different input features, thereby improving the accuracy of prediction and the interpretability of the model.

[0041] Another important innovation of the method of the present invention is the design of the uncertainty quantification unit. The expression model of the uncertainty quantification unit is defined as: , In this expression, Represents the model output results, represents the model prediction value, represents normal distribution, mean() represents mean function, and std() represents standard deviation function. This design not only gives point estimates, but also provides uncertainty estimates of the prediction results, which is of great significance for engineering decision-making and risk assessment.

[0042] For example, if the model predicts a fatigue life with a mean of 100,000 cycles and a standard deviation of 10,000 cycles, we can get a confidence interval. Assuming we want a 95% confidence level, the 95% confidence interval for fatigue life is approximately [80,400, 119,600] cycles. This information is valuable for engineering design and maintenance decisions.

[0043] In general, the present invention realizes a material fatigue prediction model that has both a physical basis and a strong learning ability through the innovative design of fatigue damage evolution equation, feature mapping unit, attention layer and uncertainty quantification unit. This method can not only provide accurate prediction results, but also give a reliability estimate of the prediction, providing a powerful tool for the research and engineering application of material fatigue behavior. In a preferred embodiment of the present invention, an adaptive learning step is introduced to further improve the prediction accuracy and adaptability of the model. This adaptive learning step mainly includes two key links: the establishment of the initial fatigue damage evolution equation and the iterative optimization process.

[0044] First, based on the initial fatigue parameters, the initial fatigue damage evolution equation is established. This initial equation provides a good starting point for subsequent optimization. Preferably, the expression of the initial fatigue parameters can be defined as: , in, represents the material constant, represents the equivalent plastic strain amplitude, represents the starting point of fatigue plastic strain, represents the hardening exponent. The initial values ​​of these parameters can be set based on the basic properties of the material and prior knowledge.

[0045] For example, for most metal materials, The initial value of can be set between 0.5 and 2. The initial value of can be set between 1 and 3. It can be determined based on the monotonic tensile test results of the material, and is usually taken as 1.5 to 2 times the yield strain.

[0046] Next, through the iterative optimization process, the fatigue parameters are continuously adjusted to obtain the final fatigue damage evolution equation. This process can use optimization algorithms such as gradient descent to minimize the prediction error.

[0047] Preferably, an optimization algorithm with adaptive learning rate, such as Adam optimizer, can be used to accelerate convergence speed and improve optimization effect.

[0048] At each iteration, the model adjusts the parameters based on the difference between the predicted and observed values. For example, if the model tends to underestimate high cycle fatigue life, you might need to reduce Value or increase Conversely, if the model tends to overestimate low cycle fatigue life, it may be necessary to increase value or decrease This adaptive adjustment process allows the model to better adapt to the fatigue behavior of different materials and loading conditions.

[0049] Preferably, a convergence threshold can be set, for example, when the prediction error change of 5 consecutive iterations is less than 1%, the optimization process is considered to have converged and the iteration is stopped. This can avoid overfitting and waste of computing resources while ensuring the optimization effect.

[0050] Another important feature of the method of the present invention is its system implementation. The material fatigue prediction system based on physical guided neural network includes multiple functional modules, each module undertakes a specific task, and together achieves high-precision fatigue prediction.

[0051] The data acquisition module 1 is responsible for acquiring the mechanical response data and material microstructure characteristic data during the material cyclic loading process. This module may include various sensors and data acquisition equipment, such as strain gauges, force sensors, displacement sensors, etc. Preferably, the module may also include a data preprocessing unit for filtering, denoising and data standardization to improve the accuracy of subsequent analysis.

[0052] Physical constraint module 2 obtains fatigue damage parameters based on the acquired data. The core of this module is to integrate physical knowledge into the data-driven model. For example, it may contain a stress-strain relationship calculation unit based on material science theory and a cumulative damage calculation unit based on damage mechanics.

[0053] Attention module 3 is responsible for incorporating the microstructural characteristics of the material into the fatigue damage evolution equation. This module may contain multiple attention heads, each focusing on different microscopic features, such as grain size, second phase distribution, etc. In this way, the model can automatically learn which microscopic features have a greater impact on fatigue behavior.

[0054] The uncertainty quantification module 4 is used to quantify the uncertainty of the fatigue damage evolution equation. This module may use techniques such as Bayesian neural network or ensemble learning to obtain the probability distribution of the prediction results. Preferably, the module may also include a sensitivity analysis unit to evaluate the influence of different input parameters on the prediction results.

[0055] The feature mapping module 5 is responsible for obtaining the prediction results. This module may contain multiple layers of neural networks, each layer corresponding to a different level of abstraction of the material fatigue behavior. For example, the lower layers may focus on local stress concentrations, while the higher layers may focus on the overall damage evolution trend.

[0056] Finally, the prediction module 6 quantifies the uncertainty of the prediction results to obtain the final material fatigue prediction results. This module not only outputs the predicted median value, but also gives the predicted confidence interval. Preferably, the module can also include a visualization unit to display the prediction results in a graphical manner for easy understanding and use by engineers.

[0057] In general, the method and system of the present invention achieves high-precision and explainable material fatigue prediction by combining physical knowledge and data-driven methods. Its adaptive learning ability and uncertainty quantification function enable it to cope with various complex materials and loading conditions, providing a powerful tool for material design and structural optimization.

[0058] In order to verify the superiority of the material fatigue prediction method based on physics-guided neural network of the present invention, we designed a set of comparative experiments. The commonly used aviation aluminum alloy 7075-T6 was selected as the research object, and fatigue life prediction was performed using the method of the present invention (Example 1), the traditional physical model method (Comparative Example 1) and the pure data-driven method (Comparative Example 2).

[0059] Example 1 adopts the physical guided neural network method of the present invention, combining the physical constraint unit, the attention layer and the uncertainty quantification unit. Comparative Example 1 adopts the traditional fatigue prediction model based on the Morrow equation. Comparative Example 2 uses a multi-layer perceptron neural network without physical constraints.

[0060] We collected 100 sets of fatigue test data of 7075-T6 aluminum alloy, including stress amplitude, strain amplitude, loading frequency, microstructure characteristics (such as average grain size) and corresponding fatigue life. These data were randomly divided into training set (70%) and test set (30%). In order to comprehensively evaluate the performance of each method, we selected the following key indicators:

[0061] 1. Average relative error (ARE): reflects the overall accuracy of the prediction.

[0062] 2. Coefficient of determination (R²): reflects the degree of correlation between the predicted value and the actual value.

[0063] 3. Extrapolation capability: Use data outside the training set to evaluate the generalization performance of the model.

[0064] 4. Computation time: reflects the computational efficiency of the method.

[0065] 5. Uncertainty quantification: Assess the reliability of prediction results.

[0066] The detection methods of each indicator are as follows: 1.ARE = ​​(1 / n) * Σ|(predicted value - actual value) / actual value| * 100%; 2. R² is calculated through regression analysis; 3. Extrapolation capability is evaluated by prediction error at high stress levels; 4. The computation time is measured using the running time under the same hardware configuration; 5. Uncertainty quantification through coverage assessment of prediction intervals;

[0067] The experimental results are shown in the following table: Evaluation Metrics Example 1 Comparative Example 1 Comparative Example 2 ARE (%) 5.2 12.8 8.7 R² 0.96 0.85 0.92 Extrapolation error (%) 9.5 25.3 18.6 Calculation time(s) 8.5 2.1 6.3 95% confidence interval coverage (%) 94.3 N / A N / A

[0068] It can be seen from the experimental results that the method of the present invention (Example 1) shows obvious advantages in multiple key indicators:

[0069] 1. Prediction accuracy: The average relative error of Example 1 is only 5.2%, which is much lower than the traditional physical model (12.8%) and the pure data-driven method (8.7%). This shows that the method of the present invention successfully combines the advantages of physical knowledge and data-driven to achieve more accurate fatigue life prediction.

[0070] 2. Goodness of fit: The R² value of Example 1 reached 0.96, which is significantly higher than that of the other two methods. This shows that the method of the present invention can not only accurately predict the average trend, but also well capture the discrete characteristics of the data.

[0071] 3. Extrapolation capability: When predicting fatigue life at high stress levels outside the training range, the error of Example 1 (9.5%) is much lower than that of the comparison method. This demonstrates the generalization advantage of physics-guided neural networks and reflects the importance of physical constraints.

[0072] 4. Computational efficiency: Although the computation time of Example 1 is slightly longer than that of the simple physical model, this computational cost is completely acceptable considering its significantly improved prediction performance. Compared with the pure data-driven method, the method of the present invention provides higher accuracy while also improving computational efficiency.

[0073] 5. Uncertainty quantification: Example 1 can provide reliable uncertainty estimates, and its actual coverage rate (94.3%) of the 95% confidence interval is very close to the theoretical value. This feature is of great value in engineering decision-making, while traditional methods cannot provide this information.

[0074] These results fully demonstrate the superiority of the method of the present invention. Through the innovative design of the physics-guided neural network, this method successfully combines physical knowledge with data-driven learning, which not only improves the prediction accuracy, but also enhances the interpretability and generalization ability of the model. In particular, this method shows obvious advantages when dealing with complex material behaviors and loading conditions.

[0075] In addition, the uncertainty quantification capability provided by the method of the present invention provides important support for engineering design and risk assessment, which is difficult to achieve with traditional methods.

[0076] In general, this set of comparative experiments strongly verified the innovation and practical value of the present invention in the field of material fatigue prediction. It not only significantly improved the prediction accuracy, but also showed unique advantages in model generalization, interpretability and uncertainty quantification, providing a more advanced and reliable method for the research and engineering application of material fatigue behavior.

[0077] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, replacement, and improvement made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A material fatigue prediction method based on a physics-guided neural network, characterized in that: include: The acquisition steps include: Obtain the mechanical response data and material microstructure characteristic data during cyclic loading of the material; Processing steps include: Based on the mechanical response data and the material microstructure characteristic data, construct a physical guided neural network model; Using the physical guided neural network model, the following operations are performed: a) Obtain fatigue damage parameters through physical constraint units; b) Incorporating the material's microstructural characteristics into the fatigue damage evolution equation through the attention layer; c) Quantify the uncertainty of fatigue damage evolution equation through uncertainty quantification unit; d) Obtaining prediction results through feature mapping units; Output steps include: The uncertainty of the prediction results is quantified through the prediction unit to obtain the final material fatigue prediction results.

2. The method according to claim 1, characterized in that The mechanical response data include load, displacement, strain, stress, strain rate, strain energy density and fatigue life.

3. The method according to claim 1, characterized in that The fatigue damage parameters include stress, strain, strain energy density, hardening index and plastic strain.

4. The method according to claim 1, characterized in that: The fatigue damage evolution equation is: ， in, Indicates fatigue damage, represents the equivalent plastic strain amplitude, Indicates the current cycle number.

5. The method according to claim 1, characterized in that The expression model of the feature mapping unit is: ， in, Indicates Inputs, Indicates The output of each physical layer node, and Represent the weight matrix and bias vector respectively, represents a non-linear activation function.

6. The method according to claim 1, characterized in that The expression model of the attention layer is: ， in, represents the output of the attention layer, Indicates The attention coefficient of the input, Indicates input.

7. The method according to claim 1, characterized in that The expression model of the uncertainty quantification unit is: mean ,std , in, Represents the model output results, represents the model prediction value, Represents normal distribution, mean() represents the mean function, and std() represents the standard deviation function.

8. The method according to claim 1, characterized in that It also includes an adaptive learning step: Based on the initial fatigue parameters, the initial fatigue damage evolution equation is established; Through iterative optimization and continuous adjustment of fatigue parameters, the final fatigue damage evolution equation is obtained.

9. The method according to claim 8, characterized in that The expression of the initial fatigue parameter is: ， in, represents the material constant, represents the equivalent plastic strain amplitude, represents the starting point of fatigue plastic strain, Indicates the hardening index.

10. Material fatigue prediction system based on physical guided neural network, characterized in that: include: A data acquisition module is used to obtain the mechanical response data and material microstructure characteristic data during the material cyclic loading process; A physical constraint module, used for obtaining fatigue damage parameters based on the mechanical response data and material microstructure characteristic data; an attention module to incorporate the material’s microstructural characteristics into the fatigue damage evolution equation; Uncertainty quantification module, used to quantify the uncertainty of fatigue damage evolution equations; Feature mapping module, used to obtain prediction results; The prediction module is used to quantify the uncertainty of the prediction results and obtain the final material fatigue prediction results; Among them, the physical constraint module, attention module, uncertainty quantification module, feature mapping module and prediction module constitute a physical guided neural network model.

Images