Machine learning prediction method for complex fatigue life of materials based on simple experimental data

Abstract

The present invention belongs to the technical field of material fatigue life prediction, and specifically relates to a machine learning prediction method for material complex fatigue life based on simple test data, including: establishing a simple load life data set of target materials; extracting typical damage mechanisms and corresponding characteristic parameters of materials under complex fatigue loads; encoding physical constraints between fatigue life-damage mechanism-characteristic parameters; designing a non-fully connected physical information-neural network considering physical damage mechanisms; normalizing input features and output features, and training a non-fully connected physical information-neural network; establishing a complex load life data set of target materials; applying physical information-neural networks to perform life extrapolation prediction on the complex load life data set, and identifying and analyzing three types of damage: creep, fatigue, and oxidation. The present invention achieves the purpose of training a model with simple load test data and then directly extrapolating and predicting the complex fatigue load life.

Description

Technical Field

[0001] The present invention belongs to the technical field of material fatigue life prediction, and in particular relates to a machine learning prediction method for material complex fatigue life based on simple test data. Background Art

[0002] The working environment of modern aircraft engines, gas turbines and other equipment is becoming more and more complex, and the structure is subjected to extreme thermal-mechanical coupling loads. Due to factors such as device start-up and shutdown, steady-state operation, and various operating conditions, the service process of key life-limited components such as turbine blades and impellers is accompanied by complex creep-fatigue interaction effects and thermal-mechanical fatigue interaction effects. Therefore, accurately predicting and evaluating fatigue life under complex loads is an important basis for structural design.

[0003] In the traditional paradigm, creep-fatigue interaction and thermomechanical fatigue life assessment often rely on a large amount of test data and parameter fitting, which leads to high time and test costs. A potential idea is to extract information from simple and extensive pure creep and pure fatigue tests, and extrapolate to predict fatigue life under complex loads. In the classic complex fatigue life modeling, the linear damage accumulation model has been widely used because of its simplicity and ease of operation. For example, the time fraction method calculates creep damage and fatigue damage for one cycle, and then uses the creep-fatigue damage envelope to predict the creep-fatigue rupture life. Another example is the Neu-Sehitoglu model, which is widely used in the assessment of thermomechanical fatigue life. By modeling and calculating creep damage, fatigue damage and oxidation damage, fatigue life is predicted based on the time fraction method. However, the increase in data volume and the increase in data dispersion lead to significant errors in traditional life models. It is difficult to consider the creep / fatigue / oxidation damage dominant mechanism and damage interaction at the same time, and it cannot be applied to fatigue life prediction in different temperature ranges and loading types.

[0004] In recent years, the application of machine learning methods under the data-driven paradigm in the fields of mechanical property prediction, material optimization design, and damage identification and detection has attracted widespread attention from researchers. The machine learning model under the neural network architecture shows strong designability and universality. The deep neural network that incorporates physical information and constraints can further analyze the "black box" and enhance the interpretability of the model. Using the neural network model for life modeling under complex loads can make full use of the existing simple load life data set of the material, taking into account the advantages of high precision and high efficiency.

[0005] Chinese patent CN114021481A discloses a creep-fatigue life prediction method based on a physics and data information fusion network. By expanding physical feature engineering and introducing physical loss terms, the prediction accuracy of creep-fatigue life is effectively improved. However, model training still requires a large amount of creep-fatigue interaction test data, and the direct extrapolation of simple loads to predict complex load life is not achieved. Chinese patent CN108170905B discloses a life prediction method for nickel-based high-temperature alloys under thermomechanical fatigue loads. By independently modeling creep damage, fatigue damage, and oxidation damage, reliable prediction of the life of components under thermomechanical fatigue loads can be achieved. However, the above model is only applicable to the life prediction of materials under a single temperature range and loading type, and has great limitations.

[0006] Machine learning has made great progress and breakthroughs in the field of fatigue life prediction of materials, but the established models can only realize life assessment under the same load. There are few reports on extrapolation prediction models and methods from simple loads to complex loads, that is, load level. On the other hand, when facing complex loads such as creep-fatigue interaction and thermomechanical fatigue, damage recognition and physical constraints must be integrated into the model. The powerful designability and universality of deep neural network models provide new ideas for constructing complex fatigue life prediction methods based on simple data. Machine learning models driven by physical information and data information need to be vigorously developed.

[0007] Starting from simple material loads such as pure creep and pure fatigue test data, the present invention extracts characteristic control parameters and encodes physical constraints, builds a physical information-neural network that considers the damage mechanism, and verifies the model's life extrapolation prediction and damage identification analysis capabilities on data sets of complex loads such as creep-fatigue interaction and thermomechanical fatigue tests. Summary of the invention

[0008] In order to solve the above technical problems, the present invention provides a machine learning prediction method for complex fatigue life of materials based on simple test data. By integrating physical knowledge and damage constraints into the neural network model, the purpose of training the model with simple pure creep and pure fatigue test data and then directly extrapolating and predicting complex creep-fatigue interaction and thermomechanical fatigue life is achieved.

[0009] The present invention is implemented by providing a machine learning prediction method for complex fatigue life of materials based on simple test data, comprising the following steps:

[0010] S1: Establish a simple load-life data set for the target material. Simple load means a single pure creep and pure fatigue load. The simple load-life data set includes pure creep test data and pure fatigue test data under various temperatures and loads.

[0011] S2: Extract typical damage mechanisms and corresponding characteristic parameters of materials under complex fatigue loads. The characteristic parameters include creep damage control parameters, fatigue damage control parameters and oxidation damage control parameters.

[0012] S3: Encode the physical constraints between fatigue life, damage mechanism and characteristic parameters. The physical constraints include damage range constraints, negative correlation constraints between life and damage, and positive correlation constraints between damage and parameters.

[0013] S4: Design a non-fully connected physical information-neural network considering the physical damage mechanism. The physical information-neural network consists of a creep damage network block, a fatigue damage network block, an oxidation damage network block, and a life prediction network block.

[0014] S5: The simple load life data set includes the input features and output features of the physical information-neural network. The input features and output features are normalized, and the simple load life data set is divided into a training set and a test set based on stratified sampling. The training set is used to train the non-fully connected physical information-neural network.

[0015] S6: Establish a complex load-life data set for the target material. Complex load means the combined creep-fatigue interaction and thermo-mechanical fatigue load. The complex load-life data set includes creep-fatigue interaction test data of various load-keeping types and thermo-mechanical fatigue test data of various phases.

[0016] S7: Use the trained physical information-neural network to perform life extrapolation prediction on complex load-life data sets, and identify and analyze three types of damage: creep, fatigue, and oxidation.

[0017] Preferably, in step S1, a simple load-life data set is collected from previous tests and material manuals, and the simple load-life data set has twelve-dimensional input features and one-dimensional output features, and the twelve-dimensional input features are:

[0018] Temperature related parameters T in , including: tensile peak temperature T up , compression valley temperature T down ;

[0019] Stress related parameters s in , including: Maximum tensile stress s up1 , Load holding end stress s up2 , Maximum compression stress s down , stress range Δ s ;

[0020] Strain related parameters e in , including: total strain range Δ e , plastic strain range Δ e p ;

[0021] Time-related parameters t in , including: tensile holding time t up , compression holding time t down , tensile loading time t tens , Compress loading time t comp ;

[0022] The one-dimensional output characteristics are: for pure creep test, the creep life time number t , for pure fatigue test is the number of fatigue life cycles N .

[0023] Further preferably, in step S2, typical damage mechanisms of materials under complex fatigue loads include creep damage, fatigue damage and oxidation damage, and creep damage control parameters include tensile peak temperature T up , Maximum tensile stress s up1 , Load holding end stress s up2 , tensile holding time t up and tensile loading time t tens ; Fatigue damage control parameters include tensile peak temperature T up , compression valley temperature T down , Maximum tensile stress s up1 , stress range Δ s , total strain range Δ e and plastic strain range Δ e p ; Oxidative damage control parameters include compression valley temperature T down , Maximum compression stress s down , plastic strain range Δ e p , compression holding time t down and compressed loading time tcomp .

[0024] Further preferably, in step S3, the damage range constraint as one of the physical constraints means that the range of each sub-damage and the total damage is between [0, 1], and the sub-damage is creep damage. D creep , fatigue damage D fatigue , oxidative damage D oxidation , the inequality is as follows:

[0025] (1)

[0026] The second physical constraint, the negative correlation constraint between life and damage, refers to the negative correlation constraint between the output characteristics and each sub-damage and the total damage. That is, the greater the creep damage, fatigue damage, and oxidation damage in physical cognition, the smaller the actual life. N Converted into hours, compared with the creep life time of pure creep test t To form a unified output, the constraint inequality is as follows:

[0027] (2)

[0028] The third physical constraint, the positive correlation constraint between damage and variables, refers to the positive correlation constraint between each sub-damage and the corresponding characteristic parameter, that is, the greater the temperature and load in physical cognition, the greater the creep damage, fatigue damage, and oxidation damage. For creep damage, it is related to the tensile peak temperature. T up , Maximum tensile stress s up1 , tensile holding time t up and tensile loading time t tens Related; for fatigue damage, with stress range Δ s , total strain range Δ e Related; for oxidation damage, with compression valley temperature T down , plastic strain range Δ e p , compression holding time t down and compressed loading time t comp Related, the inequality is as follows:

[0029] (3).

[0030] Further preferably, in the step S4, the constructed non-fully connected physical information-neural network selectively inputs the twelve-dimensional input features into the creep damage network block, the fatigue damage network block and the oxidation damage network block in the network structure, and outputs potential creep damage factors, fatigue damage factors and oxidation damage factors respectively. The three damage factors constitute the damage layer of the physical information-neural network, and the damage layer serves as the input of the life prediction network block to realize the prediction of fatigue life.

[0031] Further preferably, in step S4, the physical information-neural network imposes corresponding physical constraints on the damaged layer: for pure creep test data, the fatigue damage is set to 0, and for pure fatigue test data, the creep damage is set to 0; the physical information-neural network adds the three types of physical constraints as physical loss terms to the loss function of the physical information-neural network, and together with the data loss term, forms a composite loss function L ( i ), ensuring that the neural network has physical consistency:

[0032] (4)

[0033] in, L d ( i ) is the data loss function term, L p ( i ) is the physical loss function term, l is the physical weight coefficient to be optimized, which is used to adjust the relative weight of data loss and physical loss;

[0034] Data loss function L d ( i )Using the mean square error loss function:

[0035] (5)

[0036] in, n represents the number of data points, is the predicted value of the physical information-neural network model, is the actual value of the test;

[0037] Physical loss function term L p ( i ) is obtained by transforming three types of physical constraints with the help of ReLU function.

[0038] According to the inequality 0 ≤ D creep Take ≤ 1 as an example to illustrate the conversion method:

[0039] {L}_{{p}^{*}}(\theta )=\frac {1} {n}\left [ {\sum ^{n}_{i=1} {ReLU(-{D}^{{\theta}_{i}}_{creep})+\sum ^{n}_{i=1} {ReLU({D}^{{\theta}_{I}}_{creep}-1)}}} \right ] (8)

[0040] in, L p* ( i ) is the inequality 0 ≤ D creep ≤ 1 corresponding to the loss function expression, n Represents the number of data points. ReLU is a linear rectifier function. When the input is less than 0, the output is 0. When the output is greater than 0, the output value is equal to the input value. D creep Beyond the range [0, 1], the properties of the ReLU function will make L p* ( i ) produces a positive value, i.e., loss. The remaining inequalities in equations (1) to (3) are converted into corresponding physical loss terms using the above method.

[0041] Further preferably, in step S5, for inputting characteristic temperature related parameters T in , stress related parameters s in , strain related parameters e in , time-related parameters t in The following methods are used for normalization:

[0042] (6)

[0043] Among them, the symbols without a horizontal bar are the parameters before normalization, and the symbols with a horizontal bar are the parameters after normalization; is the melting point temperature of the material under standard atmospheric pressure, is the yield strength of the material at the test temperature, is the elastic modulus of the material at the test temperature, It is the time period of a single cycle;

[0044] For output features, the number of fatigue life cycles of pure fatigue test N Converted into hours, compared with the creep life time of pure creep test tTo form a unified output, the output characteristics of pure fatigue test and pure creep test are normalized by the following method:

[0045] (7)

[0046] in, and are the creep life time before and after normalization, Essentially the logarithm of the lifetime in seconds.

[0047] Further preferably, in step S5, the simple load-life data set is randomly disrupted and then divided into subsets according to the types of pure creep tests and pure fatigue tests, and each subset is divided into a training set and a test set in proportion using a stratified sampling method.

[0048] Further preferably, in step S5, the data stream is forward propagated in a single training of the physical information-neural network, and then the composite loss function of formula (4) is L ( i ) for back propagation, and the gradient descent algorithm is used to iteratively update the network parameters. The number of hidden layers and neurons in each layer of the creep damage network block, fatigue damage network block, oxidation damage network block and life prediction network block are determined by trial and error. The remaining hyperparameters include the initial learning rate α , weight decay coefficient W d and the physical weight coefficient l ,The grid search algorithm is used to optimize the hyperparameters of the physical information-neural network on the training set, and the maximum number of iterations is fixed to 2000.

[0049] Further preferably, in step S6, the complex load life data set has the same twelve-dimensional input features as in step S1, and the output feature is the number of fatigue life cycles. N Creep-fatigue interaction tests include tensile load-holding-CP, compression load-holding-PC and tension-compression load-holding-CC tests, and thermomechanical fatigue tests include in-phase-IP and reverse-phase-OP tests;

[0050] The input characteristics and output characteristics of creep-fatigue interaction test data of various loading types are normalized according to formula (6) and formula (7); the temperature-related parameters and time-related parameters of thermomechanical fatigue test data of various phases are normalized according to formula (6), the maximum tensile stress, loading end stress, and maximum compressive stress in the stress-related parameters are normalized according to formula (6) using the yield strength at the corresponding peak and valley temperatures, and the stress range and strain-related parameters in the stress-related parameters are normalized according to formula (6) using the corresponding yield strength and elastic modulus at the median temperature, and the output characteristics are normalized according to formula (7).

[0051] Further preferably, steps S4-S7 are implemented by programming including but not limited to Python language or Java language or commercial computing software Matlab.

[0052] Compared with the prior art, the advantages of the present invention are:

[0053] The present invention makes full use of the existing extensive simple load test data set, extracts characteristic control parameters and encodes physical constraints, introduces damage recognition and physical constraints into a carefully designed neural network architecture, and establishes a complex fatigue life prediction method based on simple data. The model is trained only with known simple pure creep and pure fatigue test data, and can then directly perform high-precision life extrapolation predictions on unknown complex creep-fatigue interactions and thermomechanical fatigue tests; at the same time, the model can identify three types of creep, fatigue and oxidation damage factors, and has the ability to evaluate damage mechanisms and characteristics under complex fatigue loads. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A flow chart of the method provided by the present invention on a certain metal material;

[0055] Figure 2 This is an architecture diagram of the non-fully connected physical information-neural network constructed in step S4 of the present invention;

[0056] Figure 3 The life prediction results of the neural network model trained by the present invention in simple load data sets, namely pure creep test data (a) and pure fatigue test data (b);

[0057] Figure 4 The life extrapolation prediction results of the neural network model trained by the present invention on complex load data, namely creep-fatigue interaction test data (a) and thermomechanical fatigue test data (b);

[0058] Figure 5 The figure is a relationship diagram between the predicted life and each partial damage and total damage factor on the creep-fatigue interaction test data set by the neural network model of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the specific implementation of the present invention is further described in detail below with reference to the accompanying drawings, taking the complex fatigue life assessment of a nickel-based high-temperature alloy as an example. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0060] The present invention provides a machine learning prediction method for complex fatigue life of materials based on simple test data, see Figure 1 , is a simplified flow chart of a complex fatigue life assessment method for a nickel-based high-temperature alloy provided in an embodiment of the present invention. In this embodiment, the alloy is a nickel-based high-temperature alloy. The specific steps are:

[0061] S1: Establish a simple load-life data set for the target material. Simple load means a single pure creep and pure fatigue load. The simple load-life data set includes pure creep test data and pure fatigue test data under various temperatures and loads.

[0062] S2: Extract typical damage mechanisms and corresponding characteristic parameters of materials under complex fatigue loads, including creep damage control parameters, fatigue damage control parameters and oxidation damage control parameters;

[0063] S3: Encode the physical constraints between fatigue life, damage mechanism and control parameters, including damage range constraints, negative correlation constraints between life and damage, and positive correlation constraints between damage and parameters;

[0064] S4: Design a non-fully connected physical information-neural network considering the physical damage mechanism. The network consists of a creep damage network block, a fatigue damage network block, an oxidation damage network block, and a life prediction network block.

[0065] S5: normalize the model input parameters and model output parameters, and divide the simple load data set into a training set and a test set based on stratified sampling, and use the training set to train the neural network;

[0066] S6: Establish a complex load-life data set for the target material. Complex load means the combined creep-fatigue interaction and thermo-mechanical fatigue loads. The data set includes creep-fatigue interaction test data of various load types and thermo-mechanical fatigue test data of various phases.

[0067] S7: Use the trained neural network to perform life extrapolation prediction on complex load data sets, and identify and analyze three types of damage: creep, fatigue, and oxidation.

[0068] In a specific embodiment, the metal material described in step S1 is selected as a nickel-based high-temperature alloy.

[0069] In a specific embodiment, the fatigue life data of the relevant materials described in steps S1 and S6 are collected from publicly published literature and material manuals at home and abroad.

[0070] In a specific embodiment, the ratio of the training set to the test set when the data set is divided in step S5 is 7:3.

[0071] In a specific embodiment, steps S4-S6 are implemented by programming in Python.

[0072] In order to demonstrate the application of the technical means of the present invention more clearly, the present invention has made a practical application of material data. Taking a selected nickel-based high-temperature alloy as an example, the above technical means are used to realize the life extrapolation prediction and damage identification analysis under complex loads.

[0073] S1. The pure creep and pure fatigue test data of the target materials from the publicly published literature and material manuals at home and abroad are collected to form a simple load-life data set. There are 87 groups of pure creep data at different temperatures and stress levels, and 90 groups of pure fatigue data at different temperatures and strain levels.

[0074] S2. Extract typical damage mechanisms and corresponding characteristic parameters of materials under complex fatigue loads, including creep damage, fatigue damage and oxidation damage. Creep damage control parameters include tensile peak temperature T up , Maximum tensile stress s up1 , Load holding end stress s up2 , tensile holding time t up and tensile loading time t tens ; Fatigue damage parameters include tensile peak temperature T up , compression valley temperature T down , Maximum tensile stress s up1 , stress range Δ s , total strain range Δ e and plastic strain range Δ e p ; Oxidative damage control parameters include compression valley temperature T down , Maximum compression stress s down , plastic strain range Δ e p , compression holding time t down and compressed loading timet comp .

[0075] S3. Based on the current knowledge and understanding of a large number of material damage mechanisms, the physical constraints between fatigue life, damage type and control parameters are encoded. The damage range constraint of physical constraint one is shown in formula (1), the negative correlation constraint between life and damage of physical constraint two is shown in formula (2), and the positive correlation constraint between damage and variable of physical constraint three is shown in formula (3).

[0076] S4. Design and build a non-fully connected physical information-neural network considering the physical damage mechanism. The network is generally composed of four major network blocks: creep damage, fatigue damage, oxidation damage and life prediction. The specific network architecture is as follows: Figure 2 As shown. The neural network structure selectively inputs 12-dimensional input features to the front-end creep, fatigue, and oxidation damage network blocks, and outputs potential creep, fatigue, and oxidation damage factors. The three major damage factors constitute the damage layer as the input of the back-end life prediction network block to achieve the prediction of creep life or fatigue life; on the carefully designed damage layer, for pure creep test data, the fatigue damage is set to 0, that is, D fatigue = 0, for pure fatigue test data, the creep damage is set to 0, that is, D creep = 0; In the back-propagation loss function, the three types of physical constraint inequalities in equations (1)-(3) are transformed into corresponding physical loss terms according to the example of equation (8): L p ( i ), and the data loss term in formula (5) L d ( i ) together form a composite loss function L ( i ), as shown in formula (4).

[0077] S5. A neural network model was built based on the PyTorch library, and the temperature-related parameters, stress-related parameters, strain-related parameters, and time-related parameters of the simple load data set were normalized, as shown in formula (6); the output life was normalized using logarithmic life, as shown in formula (7). The normalized data set was segmented and divided into pure creep and pure fatigue subsets according to the test type; a stratified sampling method was further used to sample each subset according to the training: test = 7:3 to obtain the training set and data set for training and testing of the neural network model. The Adam optimization algorithm was used to iteratively update the network parameters, and the maximum number of iterations was fixed at 2000; the grid search algorithm was used to optimize the hyperparameters selected by the neural network model. Initial learning rate αThe range is [0.001, 0.005, 0.008, 0.01, 0.1], weight decay coefficient W d The range is [0.0001, 0.0005, 0.001, 0.01, 0.1], the physical weight coefficient l The range is [0.01, 0.1, 0.2, 0.5, 1, 10]. The final hyperparameter optimization results are: α = 0.008, W d = 0.0005, l = 0.2.

[0078] S6. Collect the creep-fatigue interaction and thermomechanical test data of the target material from the publicly published literature and material manuals at home and abroad to form a complex load-life data set. There are 124 sets of creep-fatigue interaction data of different load types and 14 sets of thermomechanical fatigue data of different phase types. The input characteristics and output characteristics of creep-fatigue interaction data are still normalized according to equations (6) and (7). The temperature-related parameters and time-related parameters of thermomechanical fatigue test data of multiple phases are normalized according to equation (6). The maximum tensile stress, load-holding end stress, and maximum compressive stress in the stress-related parameters are normalized according to equation (6) using the yield strength at the corresponding peak and valley temperature. The stress range and strain-related parameters in the stress-related parameters are normalized according to equation (6) using the corresponding yield strength and elastic modulus at the median temperature. The output characteristics are normalized according to equation (7).

[0079] S7. A simple load data set for a nickel-based superalloy. Figure 3 (a) and (b) show the prediction results of pure creep and pure fatigue life by the trained physical information-neural network model. It can be seen that the prediction results of almost all data are within the ±2 times error band. The trained model is used to directly perform life extrapolation prediction on complex load data sets. The results are shown in Figure 2. Figure 4 As shown in (a) and (b), most of the data prediction results are within the ±3 times error band. The model realizes the unified life prediction of creep-fatigue interaction tests of various loading types and thermomechanical fatigue tests of different phase types. Taking the creep-fatigue interaction data set as an example, Figure 5 The relationship between the predicted life and the creep / fatigue / oxidation damage factor and the total damage factor is given, which meets the physical constraints encoded in step S3. It should be noted that it is precisely because of the integration of damage recognition and physical constraints that the neural network model under the data-physics bidirectional driving paradigm successfully realizes the life extrapolation prediction from simple loads to complex loads, that is, at the load level.

[0080] The above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. The above embodiment of the present invention can also be modified in various ways. That is, all simple, equivalent changes and modifications made according to the claims and the description of the present invention application fall within the protection scope of the claims of the present invention.

Claims

1. A machine learning prediction method for complex fatigue life of materials based on simple test data, characterized in that: The following steps are involved: S1: Establish a simple load-life data set for the target material, including pure creep test data and pure fatigue test data under various temperatures and loads; S2: Extract typical damage mechanisms and corresponding characteristic parameters of materials under complex fatigue loads. The characteristic parameters include creep damage control parameters, fatigue damage control parameters and oxidation damage control parameters. S3: Encode the physical constraints between fatigue life, damage mechanism and characteristic parameters. The physical constraints include damage range constraints, negative correlation constraints between life and damage, and positive correlation constraints between damage and parameters. S4: Design a non-fully connected physical information-neural network considering the physical damage mechanism. The physical information-neural network consists of a creep damage network block, a fatigue damage network block, an oxidation damage network block, and a life prediction network block. S5: The simple load life data set includes the input features and output features of the physical information-neural network. The input features and output features are normalized, and the simple load life data set is divided into a training set and a test set based on stratified sampling. The training set is used to train the non-fully connected physical information-neural network. S6: Establish a complex load-life data set for the target material, including creep-fatigue interaction test data of various load-keeping types and thermomechanical fatigue test data of various phases; S7: Apply the trained physical information-neural network to perform life extrapolation prediction on complex load-life data sets, and identify and analyze three types of damage: creep, fatigue, and oxidation; In step S1, the simple load life data set has twelve-dimensional input features and one-dimensional output features, and the twelve-dimensional input features are: Temperature related parameters T in , including: tensile peak temperature T up , compression valley temperature T down ; Stress related parameters σ in , including: Maximum tensile stress σ up1 , Load holding end stress σ up2 , Maximum compression stress σ down , stress range Δ σ ; Strain related parameters ε in , including: total strain range Δ ε , plastic strain range Δ ε p ; Time-related parameters t in , including: tensile holding time t up , compression holding time t down , tensile loading time t tens , Compress loading time t comp ; The one-dimensional output characteristics are: for pure creep test, the creep life time number t , for pure fatigue test is the number of fatigue life cycles N ; In step S2, typical damage mechanisms of materials under complex fatigue loads include creep damage, fatigue damage and oxidation damage, and creep damage control parameters include tensile peak temperature T up , Maximum tensile stress σ up1 , Load holding end stress σ up2 , tensile holding time t up and tensile loading time t tens ; Fatigue damage control parameters include tensile peak temperature T up , compression valley temperature T down , Maximum tensile stress σ up1 , stress range Δ σ , total strain range Δ ε and plastic strain range Δ ε p ; Oxidative damage control parameters include compression valley temperature T down , Maximum compression stress σ down , plastic strain range Δ ε p , compression holding time t down and compressed loading time t comp； In step S3, the damage range constraint means that the range of each sub-damage and the total damage is between [0, 1], and the sub-damage is creep damage. D creep , fatigue damage D fatigue , oxidative damage D oxidation , the inequality is as follows: （1） The negative correlation constraint between life and damage refers to the negative correlation constraint between the output characteristics and each sub-damage and the total damage. That is, the greater the creep damage, fatigue damage, and oxidation damage in physical cognition, the smaller the actual life. N Converted into hours, compared with the creep life time of pure creep test t To form a unified output, the constraint inequality is as follows: （2） The positive correlation constraint between damage and parameters refers to the positive correlation constraint between each sub-damage and the corresponding characteristic parameter, that is, the greater the temperature and load in physical cognition, the greater the creep damage, fatigue damage, and oxidation damage. For creep damage, it is related to the tensile peak temperature. T up , Maximum tensile stress σ up1 , tensile holding time t up and tensile loading time t tens Related; for fatigue damage, with stress range Δ σ , total strain range Δ ε Related; for oxidation damage, with compression valley temperature T down , plastic strain range Δ ε p , compression holding time t down and compressed loading time t comp Related, the inequality is as follows: （3）； In step S4, the physical information-neural network selectively inputs the twelve-dimensional input features into the creep damage network block, the fatigue damage network block and the oxidation damage network block in the network structure, and outputs potential creep damage factors, fatigue damage factors and oxidation damage factors respectively. The three damage factors constitute the damage layer of the physical information-neural network. The damage layer is used as the input of the life prediction network block to realize the prediction of fatigue life. In step S4, the physical information-neural network imposes corresponding physical constraints on the damage layer: for pure creep test data, the fatigue damage is set to 0, and for pure fatigue test data, the creep damage is set to 0; the physical information-neural network adds the three types of physical constraints as physical loss terms to the loss function of the physical information-neural network, and together with the data loss term, forms a composite loss function L ( θ ): （4） in, L d ( θ ) is the data loss function term, L p ( θ ) is the physical loss function term, λ is the physical weight coefficient to be optimized, which is used to adjust the relative weight of data loss and physical loss; Data loss function L d ( θ )Using the mean square error loss function: （5） in, n represents the number of data points, is the predicted value of the physical information-neural network model, is the actual value of the test; Physical loss function L p ( θ ) is obtained by transforming three types of physical constraints with the help of ReLU function.

2. The machine learning prediction method for complex fatigue life of materials based on simple test data according to claim 1 is characterized in that: In step S5, for inputting characteristic temperature related parameters T in , stress related parameters σ in , strain related parameters ε in , time-related parameters t in The following methods are used for normalization: （6） Among them, the symbols without a horizontal bar are the parameters before normalization, and the symbols with a horizontal bar are the parameters after normalization; is the melting point temperature of the material under standard atmospheric pressure, is the yield strength of the material at the test temperature, is the elastic modulus of the material at the test temperature, It is the time period of a single cycle; For output features, the number of fatigue life cycles of pure fatigue test N Converted into hours, compared with the creep life time of pure creep test t To form a unified output, the output characteristics of pure fatigue test and pure creep test are normalized by the following method: （7） in, and are the creep life time before and after normalization, Essentially the logarithm of the lifetime in seconds.

3. The machine learning prediction method for complex fatigue life of materials based on simple test data according to claim 1 is characterized in that: In the step S5, the simple load-life data set is randomly disrupted and then divided into subsets according to the types of pure creep test and pure fatigue test, and each subset is divided into a training set and a test set in proportion using a stratified sampling method.

4. The machine learning prediction method for complex fatigue life of materials based on simple test data according to claim 1 is characterized in that: In step S5, the data stream is forward propagated in a single training of the physical information-neural network, and then the composite loss function of formula (4) is L ( θ ) for back propagation, and the gradient descent algorithm is used to iteratively update the network parameters. The number of hidden layers and neurons in each layer of the creep damage network block, fatigue damage network block, oxidation damage network block and life prediction network block are determined by trial and error. The remaining hyperparameters include the initial learning rate α , weight decay coefficient W d and the physical weight coefficient λ ,The grid search algorithm is used to optimize the hyperparameters of the physical information-neural network on the training set, and the maximum number of iterations is fixed to 2000.

5. The machine learning prediction method for complex fatigue life of materials based on simple test data according to claim 2 is characterized in that: In step S6, the complex load life data set has the same twelve-dimensional input features as in step S1, and the output feature is the number of fatigue life cycles. N ; The input characteristics and output characteristics of creep-fatigue interaction test data of various loading types are normalized according to formula (6) and formula (7); the temperature-related parameters and time-related parameters of thermomechanical fatigue test data of various phases are normalized according to formula (6), the maximum tensile stress, loading end stress, and maximum compressive stress in the stress-related parameters are normalized according to formula (6) using the yield strength at the corresponding peak and valley temperatures, and the stress range and strain-related parameters in the stress-related parameters are normalized according to formula (6) using the corresponding yield strength and elastic modulus at the median temperature, and the output characteristics are normalized according to formula (7).

Images