Physical information neural network steel fatigue life prediction method based on meta learning

By adopting a meta-learning physical information neural network in multi-axis fatigue life prediction, combining critical surface model and damage parameters, the problems of large errors and poor interpretation of traditional methods are solved, and high accuracy and physics are achieved under small sample conditions Consistent prediction effect.

CN119069050AActive Publication Date: 2024-12-03TIANJIN UNIV

Patent Information

Application Number
CN202411284904.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-12-03
Estimated Expiration
2044-09-13

AI Technical Summary

Technical Problem

The traditional multi-axis fatigue life prediction method has large errors, high costs, and lacks physical explanatory properties. It is especially poor in small sample conditions and is difficult to widely use.

Method used

The physical information neural network method based on meta-learning is adopted, combining critical surface models and damage parameters to build a physical information neural network, and the weight of physical information item is introduced through the meta-learning framework to optimize model training, and improve prediction accuracy and interpretability.

Benefits of technology

Under small sample conditions, the accuracy and physical consistency of steel multi-axis fatigue life prediction is significantly improved, providing better generalization ability and interpretability, and optimizing the training performance of physical information neural networks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119069050A_ABST
    Figure CN119069050A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of machine learning, and discloses a physical information neural network steel fatigue life prediction method based on meta-learning, which utilizes a meta-learning framework, learns loading path geometric features corresponding to samples by introducing a meta-model taking GRU as a module, updates weights for different samples through model reasoning, and predicts the fatigue life of a steel product. Therefore, the training performance of the physical information neural network is optimized. Different from static optimization, a group of weights are independently updated and optimized for a specific steel material and a sample under a multi-axis loading path. According to the method, path geometric information with non-proportional cyclic hardening degree measurement is introduced through the meta-model, so that the prediction result of the physical information neural network has better physical consistency, and the method has better performance in the aspect of steel multi-axial fatigue life prediction under small samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of machine learning, and in particular relates to a method for predicting fatigue life of steel. Background Art

[0002] Fatigue failure is one of the most common failure forms of metal materials under cyclic loading. Compared with uniaxial fatigue life prediction, multiaxial fatigue makes the prediction more complicated due to the non-proportional additional cyclic hardening effect involved. In order to ensure the reliability and safety of metal structural parts, multiaxial fatigue life prediction has become a research focus. However, although traditional fatigue tests are an important means to study the fatigue properties of materials, they usually consume a lot of time and cost, and due to the complexity of multiaxial loading paths and working conditions, relying solely on experiments to predict the fatigue life of metal materials under complex conditions is obviously limited.

[0003] With the emergence of a large number of empirical methods in recent years, these methods have established various models based on damage parameters such as strain, stress, energy and fracture mechanism. Among them, the critical surface method, as a representative of the energy method, performs well in empirical models because it comprehensively considers multiple factors. However, although the prediction accuracy of the critical surface method is relatively high, its applicability is still limited, and the prediction results can usually only be concentrated within a five-fold error band.

[0004] The progress of modern computer technology and multiaxial fatigue experimental technology has led to the continuous accumulation of multiaxial fatigue data for specific metal materials. Based on this data, training models such as artificial neural networks through data-driven algorithms such as machine learning has become a popular research direction. However, due to the "black box" characteristics of machine learning models and their reliance on a large number of samples, these methods have poor prediction effects in small sample cases and lack sufficient interpretability and physical background, which seriously limits their widespread application.

[0005] Since its introduction in 2019, the physical information neural network, as an innovative model that integrates physical information and data-driven, has achieved remarkable results in many fields. The development of this neural network model that combines physical information and data-driven for metal multi-axial fatigue life prediction can not only introduce physical background and improve the interpretability of the model, but also significantly improve the prediction accuracy under small sample conditions, providing a powerful solution to the limitations of traditional methods and data-driven methods. Summary of the invention

[0006] The purpose of the present invention is to provide a physical information neural network steel fatigue life prediction method based on meta-learning. Based on the critical surface method and its damage parameters, a physical information neural network and a physical information loss function are constructed, and physical information is used as a meta-model input. While introducing physical meaning and background into the data-driven method and increasing interpretability, it also ensures that the prediction ability under small samples is further improved.

[0007] In order to achieve the above-mentioned object of the invention, the present invention is implemented by the following technical solutions:

[0008] According to one aspect of the present invention, a method for predicting steel fatigue life based on a physical information neural network based on meta-learning is provided, the method comprising:

[0009] S1. Collect fatigue test data set of steel to be predicted;

[0010] S2. After selecting the critical surface model according to the steel to be predicted, the parameters required for the critical surface model are calculated according to the fatigue test data set, and the damage parameters are calculated;

[0011] S3, determining the physical information neural network input characteristics and the physical information neural network output characteristics, and performing data preprocessing on the physical information neural network input characteristics and the physical information neural network output characteristics; wherein the physical information neural network input characteristics adopt the damage parameters obtained in S2, and the physical information neural network output characteristics are fatigue life prediction results;

[0012] S4, build a physical information neural network based on the artificial neural network as the basic model, define the physical information loss function according to the critical surface model selected in S2, and introduce the weight of the physical information item;

[0013] S5, using the normal strain and shear strain information in the fatigue test data set obtained in S1 as the meta-model input feature, using the physical information item weight in S4 as the meta-model output feature, and performing data preprocessing on the meta-model input feature;

[0014] S6. Construct a meta-model and define the mean square error (MSE) of the physical information neural network as the meta-model loss function;

[0015] S7, randomly divide the data sets of metamodel input features, physical information neural network input features, and physical information neural network output features into training sets and test sets, and ensure that the data of the same samples of the metamodel and the physical information neural network are simultaneously divided into the training set or the test set; initialize the metamodel and the physical information neural network, initialize the weights of the physical information items, and set the metamodel training cycle and the physical information neural network training cycle;

[0016] S8. Perform physical information neural network training according to the current physical information item weights, update the output features of the physical information neural network, calculate the physical information loss function and back-propagate it to the physical information neural network;

[0017] S9, in the last cycle of physical neural network training, the mean square error MSE of the physical information neural network is calculated and back-propagated to the meta-model;

[0018] S10, performing meta-model training and updating the weights of physical information items;

[0019] S11. Repeat steps S8-S10 until the set meta-model training cycle is reached, and the fatigue life prediction result is obtained through the physical information neural network.

[0020] Further, in S2, for steel whose fatigue fracture type is shear type, the critical surface model adopts at least two of the FS model, MGS Yu model, MGSE Zhu model, WB model, Liu model, CXH model, Zhu model or GSE model; for steel whose fatigue fracture type is tensile type, the critical surface model adopts at least two of the SWT model, Liu model, CXH model, Zhu model and GSE model.

[0021] Furthermore, in S2, for carbon steel, the critical surface model adopts FS model, SWT model, MGSEZhu model, WB model and CXH(T) model:

[0022] FS Model:

[0023] SWT Model:

[0024] MGSEZhu Model:

[0025] WB Model:

[0026] CXH(T) Model:

[0027] Among them, σ nmax , τ nmax , Δε nmax , Δγ nmax They are the maximum normal stress, maximum shear stress, maximum normal strain range, and maximum normal stress range corresponding to the critical surface when the plane where the maximum shear strain range is located is taken as the critical surface; σ smax , τ smax , Δε smax , Δγ smaxThey are the maximum normal stress, maximum shear stress, maximum normal strain range, and maximum normal stress range corresponding to the critical surface when the plane where the maximum normal strain range is located is taken as the critical surface; Δε n is the positive strain stroke; σ' f is the fatigue strength coefficient, τ' f is the shear fatigue strength coefficient, ε' f is the fatigue ductility coefficient, γ' f is the shear fatigue ductility coefficient, b is the fatigue strength index, c is the fatigue ductility index, b0 is the shear fatigue strength index, c0 is the shear fatigue ductility index, k is the FS model constant, k MGSE is the MGSEZhu model constant, S is the WB model constant; E is Young's modulus, G is the shear modulus, σ y is the yield strength, N f is fatigue life; v e is the Poisson's ratio in the elastic stage, which is taken as 0.3; v p is the Poisson's ratio in the plastic stage, which is taken as 0.5;

[0028] Among them, the critical surface model requires parameters σ' f , τ' f , ε' f , γ' f ,b,c,b0,c0,k,k MGSE and S;

[0029] Wherein, the damage parameter is: nmax , τ nmax , Δε nmax , Δγ nmax , σ smax , τ smax , Δε smax , Δγ smax and Δε n .

[0030] Furthermore, in S3, the physical information neural network input feature is σ nmax , τ nmax , Δε nmax , Δγ nmax , σ smax , τ smax , Δε smax , Δγ smax .

[0031] Further, in S3, data preprocessing of the physical information neural network input features and the physical information neural network output features includes:

[0032] Normalize the physical information neural network input features and the logarithmically processed physical information neural network output features:

[0033]

[0034] Among them, x refers to the input features of each physical information neural network or the logarithmic output features of the physical information neural network, x * represents the normalized value, μ represents the mean of each physical information neural network input feature or the logarithmic physical information neural network output feature, and σ represents the standard deviation of each physical information neural network input feature or the logarithmic physical information neural network output feature.

[0035] Further, in S4, the physical information loss function is:

[0036] L PINN =MSE+ω1L FS +ω2L SWT +ω3L MGSEZhu +ω4L WB +ω5L CXH(T)

[0037] in:

[0038]

[0039]

[0040] ω1, ω2, ω3, ω4 and ω5 are the weights of physical information items, i is the sample number, n is the number of samples, N exp,i is the experimental life of the i-th sample, N pre,i is the predicted life span of the sample with sample number i;

[0041] Among them, σ nmax , τ nmax , Δε nmax , Δγ nmax They are the maximum normal stress, maximum shear stress, maximum normal strain range, and maximum normal stress range corresponding to the critical surface when the plane where the maximum shear strain range is located is taken as the critical surface; σ smax , τ smax , Δε smax , Δγ smax They are the maximum normal stress, maximum shear stress, maximum normal strain range, and maximum normal stress range corresponding to the critical surface when the plane where the maximum normal strain range is located is taken as the critical surface; Δε n is the positive strain stroke; σ' f is the fatigue strength coefficient, τ' f is the shear fatigue strength coefficient, ε' f is the fatigue ductility coefficient, γ' fis the shear fatigue ductility coefficient, b is the fatigue strength index, c is the fatigue ductility index, b0 is the shear fatigue strength index, c0 is the shear fatigue ductility index, k is the FS model constant, k MGSE is the MGSEZhu model constant, S is the WB model constant; E is Young's modulus, G is the shear modulus, σ y is the yield strength, N f is fatigue life; v e is the Poisson's ratio in the elastic stage, which is taken as 0.3; v p is the Poisson's ratio in the plastic stage and is taken as 0.5.

[0042] Further, in S5, performing data preprocessing on the metamodel input features includes: performing normalization processing on the metamodel input features:

[0043]

[0044] Here, y refers to the metamodel input feature, and y' represents the normalized value.

[0045] Furthermore, in S9, the root mean square error MSE of the physical information neural network is:

[0046]

[0047] Among them, i is the sample number, n is the number of samples, and N exp,i is the experimental life of the i-th sample, N pre,i is the predicted life span of the ith sample.

[0048] According to another aspect of the present invention, there is provided a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned meta-learning-based physical information neural network steel fatigue life prediction method.

[0049] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the physical information neural network steel fatigue life prediction method based on meta-learning is implemented.

[0050] The beneficial effects of the present invention are:

[0051] (1) The physical information neural network steel multi-axial fatigue life prediction method under meta-learning of the present invention utilizes the meta-learning framework, introduces the meta-model with GRU as the module to learn the geometric features of the loading path corresponding to the sample, and updates the weights for the loading features corresponding to different samples through model reasoning, thereby optimizing the training performance of the physical information neural network. Different from static optimization, for a specific steel material and a sample under a multi-axial loading path, a set of weights are independently updated and optimized. This method introduces path geometry information with non-proportional cyclic hardening degree measurement through the meta-model so that the prediction results of the physical information neural network have better physical consistency, and have better performance in the prediction of multi-axial fatigue life of steel under small samples.

[0052] (II) The method for predicting multiaxial fatigue life of steel using a physical information neural network under meta-learning of the present invention, wherein the meta-learning framework is a hyperparameter optimization method, which can effectively extract feature differences under different conditions by introducing key physical information as meta-model input features, and dynamically optimize the hyperparameters of the physical information neural network to make the prediction performance of the physical information neural network closer to the true value. At the same time, the search space is effectively reduced, so that the physical information items in the physical information neural network can fully play their role under different conditions.

[0053] (III) The physical information neural network steel multi-axial fatigue life prediction method under meta-learning of the present invention, wherein the physical information neural network loss function has good generalization for fatigue life prediction of different types of steels under different multi-axial loading paths due to the introduction of the weights of the physical information items and the dynamic update of the weights of the physical information items by the meta-learning framework.

[0054] (IV) The physical information neural network under meta-learning of the present invention is used to predict the multi-axial fatigue life of steel. The physical information neural network optimized under the meta-learning framework is applied to the field of multi-axial fatigue life prediction of steel. By introducing the physical information term in the loss function and the loading path geometry information as the meta-model input, the empirical model in the field is used as the physical meaning guidance for model training, thereby obtaining a machine learning life prediction method with both interpretability and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0056] Figure 1 A schematic diagram of the framework of the physical information neural network steel fatigue life prediction method based on meta-learning provided by the present invention;

[0057] Figure 2 A schematic diagram of the implementation process of the physical information neural network in the method provided by the present invention;

[0058] Figure 3 The ablation experimental performance of the method provided by the present invention on three metal materials: 316LN, 316L and 304;

[0059] Figure 4 The figure shows a comparison between the predicted value and the experimental value of fatigue life of three metal material data sets: 316LN, 316L and 304. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] In order to solve the problems in the background technology, the present invention provides the following solution ideas in predicting the multiaxial fatigue life of steel with both accuracy and interpretability:

[0062] Traditional empirical models have good physical meaning, but due to experimental deviations and the theoretical nature of empirical models, the accuracy of multi-axial fatigue life prediction of steel is average. Data-driven methods have good accuracy in multi-axial fatigue life prediction of steel, but due to the "black box" characteristics of data-driven models such as machine learning, the physical meaning and interpretability of such prediction models are low. At the same time, since fatigue data of complex paths are relatively scarce, improving the accuracy of fatigue life prediction under small samples has also become a direction for the development of data-driven methods.

[0063] Meta-learning refers to a branch of machine learning that aims to design algorithms that can learn how to learn. It accelerates the hyperparameter search process for new tasks by learning optimization strategies on related tasks.

[0064] In order to propose a method that has both physical meaning, interpretability and accuracy, the present invention constructs a loss function that uses the critical surface model in the traditional empirical method as the source of physical information, and constructs a physical information neural network on this basis. The weights of each sub-item in the physical information sub-item of the loss function of the physical information neural network are adjusted with the material and loading path to adapt to different multi-axial fatigue characteristics. The weight adjustment is carried out through the training of the meta-learning model. Under the meta-learning framework, the meta-model learns the geometric features of each loading path, which is used to update the weights of the physical information sub-items of the physical information neural network loss function, so as to better predict the multi-axial fatigue life of steel.

[0065] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] Example 1

[0067] like Figure 1 As shown, this embodiment provides a method for predicting multi-axial fatigue life of steel based on a physical neural network based on meta-learning, which specifically includes:

[0068] S1. Collect the fatigue test data set of the steel to be predicted, and divide the data of the fatigue test data set into a training set and a validation set.

[0069] The fatigue experimental data set of the steel to be predicted can be obtained from published literature or fatigue experiments.

[0070] The fatigue test data set specifically includes multi-axial loading hysteresis loops, real-time normal stress, normal strain, shear stress, shear strain signal values, and corresponding life cycles.

[0071] Steel refers to carbon steel and alloy steel, etc.; fatigue test data is obtained based on experiments on steel specimens of certain shape, size and performance made by pressure processing in the form of ingots, billets, etc.

[0072] S2. Select the critical surface model according to the type of steel to be predicted, and then calculate the parameters required for the critical surface model according to the fatigue test data set, and calculate the damage parameters.

[0073] Generally, for steels with shear fatigue fracture type, the selectable critical surface models include FS model, MGS Yu model, MGSE Zhu model, WB model, Liu model, CXH model, Zhu model, and GSE model; for steels with tensile fatigue fracture type, the selectable critical surface models include SWT model, Liu model, CXH model, Zhu model, and GSE model; for each type, the critical surface model can be a combination of any two or more of the above models.

[0074] In this embodiment, for 316LN, 316L, and 304 steels, the selected critical surface model is:

[0075] FS Model:

[0076] SWT Model:

[0077] MGSEZhu Model:

[0078] WB Model:

[0079] CXH(T) Model:

[0080] Among them, σ nmax , τ nmax , Δε nmax , Δγ nmax The plane with the maximum shear strain range is taken as the critical surface, the maximum normal stress, maximum shear stress, maximum normal strain range, maximum normal stress range corresponding to the critical surface, σ smax , τ smax , Δε smax , Δγ smax The plane with the maximum normal strain range is taken as the critical surface, the maximum normal stress, maximum shear stress, maximum normal strain range, maximum normal stress range corresponding to the critical surface, Δε n is the positive strain stroke; σ' f is the fatigue strength coefficient, τ' f is the shear fatigue strength coefficient, ε' f is the fatigue ductility coefficient, γ' f is the shear fatigue ductility coefficient, b is the fatigue strength index, c is the fatigue ductility index, b0 is the shear fatigue strength index, c0 is the shear fatigue ductility index, k is the FS model constant, k MGSE is the MGSEZhu model constant, S is the WB model constant; E is Young's modulus, G is the shear modulus, σ y is the yield strength, N f is the fatigue life.

[0081] In addition, the above five critical surface models can also be applied to other carbon steels except 316LN, 316L, and 304 steels.

[0082] The fatigue parameter τ' on the right side of the above critical surface model equation is f , γ' f , b0 and c0 are obtained by fitting the torsion data; σ' f , ε' f , b and c are obtained by fitting the uniaxial data. Due to the lack of torsion data for 316L, the Mises equivalent stress criterion is used to obtain:

[0083]

[0084] b o =b

[0085] c o =c

[0086] E, G, σ y Please check the data sheet of the corresponding steel material to obtain it.

[0087] Table 1 lists the parameters required for the above critical surface model for 316LN, 316L, and 304 stainless steel:

[0088] Table 1 316LN, 316L, 304 stainless steel material parameters

[0089]

[0090]

[0091] And the corresponding damage parameters are calculated according to the following formula:

[0092]

[0093] Among them, θ is the angle between the calculation plane and the axis of the steel sample, ε θ is the stress value on the calculation plane, ε x is the axial normal strain value, γ θ is the calculated plane shear strain value, τ xy is the axial shear stress value, σ θ is the calculated plane normal stress value, σ x is the axial normal stress value.

[0094] The normal strain and shear strain range are obtained as follows: Δε θ =ε θmax -ε θmin

[0095] Δγ θ =γ θmax -γ θmin

[0096] And normal stress, shear stress range:

[0097] Δσ θ =σ θmax -σ θmin

[0098] Δτ θ =τ θmax -τ θmin

[0099] According to the critical surfaces defined by different critical surface methods, the damage function σ is obtained. nmax , τ nmax , Δε nmax , Δγ nmax , σ smax , τ smax , Δε smax , Δγ smax and Δε n .

[0100] S3. Determine the input characteristics and output characteristics of the physical information neural network, and perform data preprocessing on the input characteristics and output characteristics of the physical information neural network.

[0101] Among them, the input feature of the physical information neural network adopts the damage parameter obtained by S2, and the output feature of the physical information neural network is the fatigue life prediction result.

[0102] The physical information neural network input feature selected in this embodiment is: nmax , τ nmax , Δε nmax , Δγ nmax , σ smax , τ smax , Δε smax and Δγ smax .

[0103] The physical information neural network input features and the physical information neural network output features are preprocessed, specifically, the physical information neural network input features and the physical information neural network output features after logarithmic processing are normalized:

[0104]

[0105] Among them, x refers to the input features of each physical information neural network or the logarithmic output features of the physical information neural network, x * represents the normalized value, μ represents the mean of each physical information neural network input feature or the logarithmic physical information neural network output feature, and σ represents the standard deviation of each physical information neural network input feature or the logarithmic physical information neural network output feature.

[0106] S4. Construct a physical information neural network based on the artificial neural network as the basic model, define the physical information loss function according to the critical surface model selected in S2, and introduce the weight of the physical information item.

[0107] σ nmax , τ nmax , Δε nmax , Δγ nmax , σ smax , τ smax , Δε smax and Δγ smax As 8 input features, an artificial neural network is constructed. Among them, an artificial neural network refers to a mathematical model in the field of machine learning and cognitive science that imitates the structure and function of a biological neural network (the central nervous system of animals, especially the brain).

[0108] The structure of the artificial neural network in this embodiment is 5 hidden layers, and the number of neurons in each layer is 16, 16, 8, 8, and 4. The output is the predicted life span after logarithmic normalization.

[0109] On this basis, the physical information loss function is defined to construct a physical information neural network, which is expressed as:

[0110] L PINN =MSE+ω1L FS +ω2L SWT +ω3L MGSEZhu +ω4L WB +ω5L CXH(T)

[0111] in:

[0112]

[0113]

[0114] ω1, ω2, ω3, ω4 and ω5 are the physical information weights introduced by the above physical information loss function, i is the sample number, n is the number of samples, N exp,i is the experimental life of the i-th sample, N pre,i is the predicted life span of the ith sample.

[0115] S5. The normal strain and shear strain information in the fatigue test data set in S1 are used as the input features of the meta-model, and the weights of the physical information items in S4 are used as the output features of the meta-model.

[0116] The input features of the meta-model are: ε and γ, namely normal strain and shear strain.

[0117] Normalize the meta-model input features:

[0118]

[0119] Here, y refers to the metamodel input feature, and y' represents the normalized value.

[0120] The metamodel output features are: the weights of the physical information items ω1, ω2, ω3, ω4 and ω5 in S4.

[0121] S6. Construct a meta-model using the gated recurrent unit as the meta-model module and define the meta-model loss function.

[0122] Using the gated recurrent unit as the module of the meta-model, a total of 3 identical modules are constructed, with the number of hidden layers being 32. The first gated recurrent unit is pre-connected to the input features, and the last gated recurrent unit (GRU) is followed by a fully connected layer, which then outputs the weights required by the physical neural network.

[0123] GRU mainly includes update gate, reset gate, candidate hidden state, and hidden state.

[0124] The update gate is used to determine the proportion of information from the current time step that is passed to the next time step:

[0125] z t =σ(W z ·[h t-1 ,x t ])

[0126] Among them, z t is the output of the update gate, σ is the Sigmoid activation function, W z is the weight matrix, h t-1 is the hidden state of the previous time step, x t is the input for the current time step.

[0127] The reset gate determines the forgetting ratio of the previous time step information:

[0128] r t =σ(W r ·[h t-1 ,x t ])

[0129] Among them, r t is the output of the reset gate, σ is the Sigmoid activation function, W r is the weight matrix, h t-1 is the hidden state of the previous time step, x t is the input for the current time step.

[0130] Under the action of the reset gate, the candidate hidden state of the current time step is calculated, and under the action of the update gate, the weighted sum with the previous hidden state is calculated to obtain the final hidden state.

[0131] The meta-model loss function is defined as:

[0132]

[0133] Among them, i is the sample number of the physical information neural network, n is the number of samples of the physical information neural network, and N exp,i is the experimental life of the i-th sample, N pre,i is the predicted life span of the ith sample.

[0134] S7. Randomly divide the data sets of metamodel input features, physical information neural network input features, and physical information neural network output features into training sets and test sets, and ensure that the sample division of the metamodel and the physical information neural network is consistent, that is, the data of the same samples in the metamodel and the physical information neural network are divided into the training set or test set at the same time.

[0135] Initialize the metamodel and physical information neural network, initialize the weights of physical information items, and set the metamodel training cycle and physical information neural network training cycle.

[0136] After creating a database for the physical neural network input features and physical neural network output features in S3 and the metamodel input features and metamodel output features in S5, a random seed is assigned to randomly divide the data corresponding to the samples in the database into training sets and test sets.

[0137] In this embodiment, 80% of the data is used for the training set and 20% is used for the validation set.

[0138] Initially, the metamodel and physical information network hyperparameters and physical information item weights are randomly assigned. The number of iterations of the metamodel is set to 50, and the number of cycles in a single training of the physical neural network is set to 1500 cycles.

[0139] S8. Perform physical information neural network training according to the current physical information item weights, update the output features of the physical information neural network, calculate the physical information loss function and back-propagate it to the physical information neural network.

[0140] According to the weights of the physical information items, the physical information neural network training is performed once every 1500 cycles. While predicting the fatigue life in each cycle, the corresponding physical information loss function is calculated, and the value of the physical information loss function is back-propagated to update the hyperparameters for fatigue life prediction in the next cycle.

[0141] S9. In the last cycle of a training of the physical neural network, calculate the MSE and back-propagate it to the meta-model as its loss function value.

[0142] In the 1500th cycle of a single training of the physical information neural network, the root mean square error (MSE) of the test set is calculated and back-propagated to the meta-model for the meta-model to update the hyperparameters.

[0143] S10. The meta-model is trained once and the weights of the physical information items are updated.

[0144] After the meta-model updates the hyperparameters, a training is performed to update a set of weights required for the physical information terms of the physical information neural network loss function.

[0145] S11. Repeat steps S8-S10 until the set meta-model training cycle is reached, and obtain fatigue life prediction results through the physical information neural network.

[0146] Figure 2 The specific process of meta-learning in the physical information neural network steel fatigue life prediction method is demonstrated.

[0147] Figure 3 The ablation experiment performance of the method proposed in this invention on three metal materials: 316LN, 316L and 304. Figure 3 It can be seen that the gated recurrent unit-physical information neural network performs better than the artificial neural network and the physical information neural network with unoptimized weights (weight is 1) in repeated experiments on the multi-axial fatigue data set of three metal materials.

[0148] Figure 4 The comparison between the predicted value and experimental value of multiaxial fatigue life of three metal material data sets of 316LN, 316L and 304 by the method proposed in this invention. Figure 4 It can be seen that the predicted values ​​of the gated recurrent unit-physical information neural network on both the test set and the training set are within the 2-fold error band, and most of them are concentrated within the 1.5-fold error band.

[0149] Example 2

[0150] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the physical information neural network steel fatigue life prediction method based on meta-learning in Example 1.

[0151] Example 3

[0152] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the meta-learning-based physical information neural network steel fatigue life prediction method in Example 1.

[0153] 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.

[0154] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above description of the embodiments is only used to help understand the method and core ideas of the present invention; at the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A physical information neural network steel fatigue life prediction method based on meta-learning, It is characterized in that The method includes: S1. Collect fatigue test data set of steel to be predicted; S2. After selecting the critical surface model according to the steel to be predicted, the parameters required for the critical surface model are calculated according to the fatigue test data set, and the damage parameters are calculated; S3, determining the physical information neural network input characteristics and the physical information neural network output characteristics, and performing data preprocessing on the physical information neural network input characteristics and the physical information neural network output characteristics; wherein the physical information neural network input characteristics adopt the damage parameters obtained in S2, and the physical information neural network output characteristics are fatigue life prediction results; S4, build a physical information neural network based on the artificial neural network as the basic model, define the physical information loss function according to the critical surface model selected in S2, and introduce the weight of the physical information item; S5, using the normal strain and shear strain information in the fatigue test data set obtained in S1 as the meta-model input feature, using the physical information item weight in S4 as the meta-model output feature, and performing data preprocessing on the meta-model input feature; S6. Construct a meta-model and define the mean square error (MSE) of the physical information neural network as the meta-model loss function; S7, randomly divide the data sets of metamodel input features, physical information neural network input features, and physical information neural network output features into training sets and test sets, and ensure that the data of the same samples of the metamodel and the physical information neural network are simultaneously divided into the training set or the test set; initialize the metamodel and the physical information neural network, initialize the weights of the physical information items, and set the metamodel training cycle and the physical information neural network training cycle; S8. Perform physical information neural network training according to the current physical information item weights, update the output features of the physical information neural network, calculate the physical information loss function and back-propagate it to the physical information neural network; S9, in the last cycle of physical neural network training, the mean square error MSE of the physical information neural network is calculated and back-propagated to the meta-model; S10, performing meta-model training and updating the weights of physical information items; S11. Repeat steps S8-S10 until the set meta-model training cycle is reached, and the fatigue life prediction result is obtained through the physical information neural network.

2. According to claim 1, a physical information neural network steel fatigue life prediction method based on meta-learning, It is characterized in that In S2, for steels whose fatigue fracture type is shear type, the critical surface model adopts at least two of the FS model, MGS Yu model, MGSE Zhu model, WB model, Liu model, CXH model, Zhu model or GSE model; for steels whose fatigue fracture type is tensile type, the critical surface model adopts at least two of the SWT model, Liu model, CXH model, Zhu model and GSE model.

3. According to claim 2, a physical information neural network steel fatigue life prediction method based on meta-learning, It is characterized in that In S2, for carbon steel, the critical surface model adopts FS model, SWT model, MGSEZhu model, WB model and CXH(T) model: FS Model: SWT Model: MGSEZhu Model: WB Model: CXH(T) Model: Among them, σ nmax , τ nmax , Δε nmax , Δγ nmax They are the maximum normal stress, maximum shear stress, maximum normal strain range, and maximum normal stress range corresponding to the critical surface when the plane where the maximum shear strain range is located is taken as the critical surface; σ smax , τ smax , Δε smax , Δγ smax They are the maximum normal stress, maximum shear stress, maximum normal strain range, and maximum normal stress range corresponding to the critical surface when the plane where the maximum normal strain range is located is taken as the critical surface; Δε n is the positive strain stroke; σ' f is the fatigue strength coefficient, τ' f is the shear fatigue strength coefficient, ε' f is the fatigue ductility coefficient, γ' f is the shear fatigue ductility coefficient, b is the fatigue strength index, c is the fatigue ductility index, b 0 is the shear fatigue strength index, c 0 is the shear fatigue ductility index, k is the FS model constant, k MGSE is the MGSEZhu model constant, S is the WB model constant; E is Young's modulus, G is the shear modulus, σ y is the yield strength, N f is fatigue life; v e is the Poisson's ratio in the elastic stage, which is taken as 0.3; v p is the Poisson's ratio in the plastic stage, which is taken as 0.5; Among them, the critical surface model requires parameters σ' f , τ' f , ε' f , γ' f , b, c, b 0 , c 0 , k, k MGSE and S; Among them, the damage parameters are: p nmax ,t nmax ,No nmax ,Dg nmax ,s smax ,t smax ,No smax ,Dg smax No n 。 4. According to the physical information neural network steel fatigue life prediction method based on meta-learning in claim 1, It is characterized in that In S3, the physical information neural network input feature is σ nmax , τ nmax , Δε nmax , Δγ nmax , σ smax , τ smax , Δε smax , Δγ smax .

5. According to claim 1, a physical information neural network steel fatigue life prediction method based on meta-learning, It is characterized in that In S3, data preprocessing of the physical information neural network input features and the physical information neural network output features includes: Normalize the physical information neural network input features and the logarithmically processed physical information neural network output features: Among them, x refers to the input features of each physical information neural network or the logarithmic output features of the physical information neural network, x * represents the normalized value, μ represents the mean of each physical information neural network input feature or the logarithmic physical information neural network output feature, and σ represents the standard deviation of each physical information neural network input feature or the logarithmic physical information neural network output feature.

6. According to claim 1, a physical information neural network steel fatigue life prediction method based on meta-learning, It is characterized in that In S4, the physical information loss function is: L PINN =MSE+ω 1 L FS +oh 2 L SWT +oh 3 L MGSEZhu +oh 4 L WB +oh 5 L CXH(T) in: ω 1 ,ω 2 ,ω 3 ,ω 4 and ω 5 is the weight of the physical information item, i is the sample number, n is the number of samples, N exp,i is the experimental life of the i-th sample, N pre,i is the predicted life span of the sample with sample number i; Among them, σ nmax , τ nmax , Δε nmax , Δγ nmax They are the maximum normal stress, maximum shear stress, maximum normal strain range, and maximum normal stress range corresponding to the critical surface when the plane where the maximum shear strain range is located is taken as the critical surface; σ smax , τ smax , Δε smax , Δγ smax They are the maximum normal stress, maximum shear stress, maximum normal strain range, and maximum normal stress range corresponding to the critical surface when the plane where the maximum normal strain range is located is taken as the critical surface; Δε n is the positive strain stroke; σ' f is the fatigue strength coefficient, τ' f is the shear fatigue strength coefficient, ε' f is the fatigue ductility coefficient, γ' f is the shear fatigue ductility coefficient, b is the fatigue strength index, c is the fatigue ductility index, b 0 is the shear fatigue strength index, c 0 is the shear fatigue ductility index, k is the FS model constant, k MGSE is the MGSEZhu model constant, S is the WB model constant; E is Young's modulus, G is the shear modulus, σ y is the yield strength, N f is fatigue life; v e is the Poisson's ratio in the elastic stage, which is taken as 0.3; v p is the Poisson's ratio in the plastic stage and is taken as 0.

5.

7. According to claim 1, a physical information neural network steel fatigue life prediction method based on meta-learning, It is characterized in that In S5, performing data preprocessing on the metamodel input features includes: performing normalization processing on the metamodel input features: Here, y refers to the metamodel input feature, and y' represents the normalized value.

8. According to claim 1, a physical information neural network steel fatigue life prediction method based on meta-learning, It is characterized in that In S9, the root mean square error MSE of the physical information neural network is: Among them, i is the sample number, n is the number of samples, and N exp,i is the experimental life of the i-th sample, N pre,i is the predicted life span of the ith sample.

9. A computer device, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the meta-learning-based physical information neural network steel fatigue life prediction method described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the physical information neural network steel fatigue life prediction method based on meta-learning described in any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Creep fatigue life prediction method based on fused physical neural network

    CN114021481A

  • Physical information driven metal material notch fatigue life machine learning prediction method

    CN117993304A

  • Stiffness degradation prediction model training method, application method and related device

    CN118332798A

  • Explosive near-earth explosion shock wave overpressure peak prediction method based on physical information neural network

    CN118536075A

  • Systems and methods to predict fatigue lives of aluminum alloys under multiaxial loading

    US20100235110A1

Cited By

  • Multi-scale multi-deep neural network driven metal material fatigue life prediction method

    CN119694462A

  • Multi-fidelity physical information neural network fatigue life prediction method under defect-induced failure

    CN119885876A

  • Broken line abnormity prediction method and device and computer equipment

    CN119917844A

  • Multi-axial fatigue life prediction method based on physical information

    CN120063906A

  • A multiaxial fatigue life prediction method based on physical information

    CN120063906B