A multi-modal data driven method for metal multiaxial fatigue life prediction

Abstract

The present invention discloses a multi-modal data-driven metal multi-axial fatigue life prediction method, which takes multi-modal data as the data-driven basis and uses a fusion machine learning model to provide a new fatigue life prediction framework. The method can serve as an alternative model for fatigue life prediction. First, the present invention defines multi-precision data: experimental data obtained through actual experiments and simulation data obtained through physical simulation. The present invention obtains the mechanical properties of the material through preliminary experiments, and establishes a reliable finite element analysis (FEA) model based on these properties. Subsequently, a high-reliability data set is generated using the finite element analysis model. In order to process data with multi-modal characteristics, the present invention includes constructing a hybrid prediction model DL‑LGBM‑DRS based on dynamic regression selection. The model can simultaneously learn the characteristics of discrete and time series samples with high accuracy, so that the hybrid prediction model DL‑LGBM‑DRS can efficiently and accurately predict cyclic fatigue life under various multi-axial loading conditions.

Description

Technical Field

[0001] The present invention belongs to big data analysis and application, and relates to the field of fatigue life prediction based on multimodal physical information, especially low-cycle fatigue life prediction technology under complex multiaxial loads, and more particularly to a multimodal data-driven metal multiaxial fatigue life prediction method. Background Art

[0002] Existing fatigue life prediction methods under multiaxial loading face two key challenges that need to be addressed: the complex physical relationships underlying fatigue failure under multiaxial loading make prediction difficult; and predictions based on statistical or experimental methods are expensive. In multiaxial fatigue situations where only horizontal loading (such as proportional or constant-amplitude cyclic loading) is considered, the yield criterion, known as the equivalent stress-strain criterion, can be used to replace the equivalent stress. Some researchers have defined multiple equivalent parameters for external and non-proportional conditions based on the maximum prism and minimum Frobenius criteria. However, for more complex loading conditions, the life expectancy given by this method is unreliable. Later, it was discovered that stress along the shear direction significantly increases the fatigue crack growth rate, leading to the proposal of the critical plane method. The critical plane method evaluates multiaxial fatigue based on damage parameters calculated from stress and / or strain components and has become a popular method due to its effectiveness in multiaxial fatigue problems. The critical plane method can also determine the location of fatigue crack initiation under various cyclic fatigue states.

[0003] With technological advancements, existing technologies include the use of finite element methods with built-in physics principles for fatigue analysis under multiaxial loading. Finite element analysis (FEA) is a widely used method based on physical laws in fatigue analysis. FEA can simulate fatigue conditions under a complex loading spectrum, such as the stress-strain distribution under full multiaxial non-proportional loading. The fatigue state of a surface can also be well described based on roughness and geometric properties. In addition to mechanical analysis, many studies have focused on crack growth. The extended finite element method (XFEA), based on fracture mechanics, can reveal crack growth trends in metal components. Furthermore, fatigue analysis software such as Fe-Safe can be used to directly perform life assessment using stress and strain information output from FEA. To improve efficiency and reliability, researchers are seeking to develop alternative methods that avoid the extensive nonlinear operations of FEA. Considering that uncertainties (such as material properties, geometry, and loading) can lead to scatter in fatigue life, statistical methods have been proposed to obtain more reasonable fatigue life distributions. In the case of probabilistic fatigue life prediction, Monte Carlo simulation (MCS) is often used as a general prediction framework. For multivariate uncertainty assessment, it is challenging to handle a large number of input influencing parameters and output results. MCS can solve this problem by combining the weighting law in structural reliability.

[0004] While statistical knowledge-based methods can predict fatigue life with high accuracy and efficiency, they still have several limitations in terms of implementation and prediction results: (i) difficulty in obtaining sufficient statistical data; (ii) lower interpretability and fidelity compared to physics-based methods; and (iii) computational cost is expensive in situations with considerable uncertainty. Therefore, the growing demand for fatigue prediction requires novel hybrid approaches. Physically informed machine learning (PIML) is an emerging paradigm that incorporates physical principles as "prior" knowledge to significantly improve the performance of machine learning models. One approach to PIML is to embed physical or empirical laws (such as ordinary differential equations) into machine learning networks. Therefore, in practical engineering applications, insufficient data dimensionality and quantity will seriously affect the training performance of PIML. As mentioned above, finite element analysis can serve as an appropriate alternative to data acquisition based on physical laws, which is common in many engineering projects. PIML based on finite element analysis has proven to be a very reliable and efficient method in the absence of data.

[0005] Multiaxial fatigue life prediction is complex and involves multiple factors, particularly those with multimodal inputs. The fundamental physical properties of a material determine its fatigue behavior and evolution under complex external loads. These loads are time-series and iteratively influence structural responses through geometric effects. Numerous studies have demonstrated a strong correlation between physical responses and fatigue life, yet methods for simultaneously leveraging both discrete and time-series characteristics for prediction are lacking.

[0006] In summary, the existing technology mainly has the following deficiencies:

[0007] 1. Data-driven approaches lack physical guidance;

[0008] 2. Data-driven methods lack data sources in complex load fields;

[0009] 3. Fatigue life prediction is costly, slow and has low accuracy. Summary of the Invention

[0010] Purpose of the invention: In view of the above-mentioned deficiencies in the prior art, the present invention provides a multi-modal data driven metal multi-axial fatigue life prediction method.

[0011] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions.

[0012] A multi-modal data driven metal multiaxial fatigue life prediction method comprises the following steps:

[0013] S1. Based on the preliminary experiments, a database of properties of aerospace materials to be analyzed is established. The physical properties of the materials are obtained by fitting engineering equations with the experimental data. These properties are then added to the data set as discrete characteristic properties.

[0014] S2. Build a finite element model based on the physical properties of the material. By combining finite element analysis with actual test results, fatigue life is determined as a label to establish a highly reliable training dataset based on simulation. This training dataset is used by machine learning algorithms to learn physical relationships.

[0015] S3, the data set includes multi-axial load spectrum with time series characteristics, time series physical response and material and external environment information with discrete characteristics;

[0016] This step involves constructing a hybrid prediction algorithm DL-LGBM-DRS based on dynamic regression selection, thereby realizing the prediction of multiaxial fatigue life of metal materials under multimodal input characteristics. The algorithm establishment steps include:

[0017] S31. By introducing a gradient-based unilateral sampling method and a mutually exclusive feature merging method, we construct a LightGBM model. Then, we use the GBDT algorithm to combine multiple decision trees to process dense discrete data, achieving high interpretability and generalization.

[0018] S32. Constructing a deep learning model based on a long short-term memory network based on a recurrent neural network, wherein the long short-term memory network is constructed by a forget gate, an input gate, and an output gate to realize a long-term memory function;

[0019] S33. Build a dynamic regression selection algorithm to fuse the LightGBM model and the deep learning model to form a fatigue life prediction algorithm that processes multimodal inputs.

[0020] The dynamic regression network is as follows:

[0021] Input set F, training set T, test set S, weight vector W, define the nearest neighbors as K, and the independent model number as n;

[0022] The output first defines MAE as mean absolute error:

[0023]

[0024] Where y is the actual value of the training sample, is the predicted value of the sample;

[0025] Initialize the mean absolute error, define error Error = 0 and empty set A, then for each sample x in the test set S i, find the competition interval ψ of the sample corresponding to the training set T, and the value predicted by each independent model in each set group F Calculate the mean absolute error about the competition interval ψ. The calculation process includes:

[0026] Definition a n =w×MAE n , then A=A∪a n , until the prediction results and error means of all models are:

[0027]

[0028] and Represents the update of the set, for each predicted value in the set group If a n >(a max -a min ) / 2, then perform model selection and perform the following updates:

[0029]

[0030] Define N as the size of the cluster mean, for each an Perform model combination, namely:

[0031]

[0032] Now define:

[0033]

[0034] The final error is Error fin =Error / |X|;

[0035] The dynamic regression selection algorithm finally evaluates the weight of each model, which is defined by the final error Errorfin above, f(x j ) is the observed value of the corresponding sample in the test set, is the normalized result;

[0036] For each sample, the value with the smaller error is judged to be the optimal one, and the local optimal prediction result is the final prediction result of the dynamic regression network algorithm.

[0037] Furthermore, the steps of establishing the hybrid prediction model DL-LGBM-DRS for dynamic regression selection in step S3 include:

[0038] S31. Use the GBDT algorithm to combine multiple decision trees to process dense discrete data and achieve high interpretability and generalization. The calculation is as follows:

[0039] The standard decision tree is as follows:

[0040]

[0041] In the formula, X1, X2 are the training subsets of each branch, yi is the actual value of the sample, is the branch prediction value of the sample;

[0042] The GBDT algorithm requires the following sub-steps:

[0043] 1) Initialize the weak learner and define the loss function L(y,f(x)) and the prediction function g0(x):

[0044]

[0045] Where C m Expressed as the mean of sample y in m trees;

[0046] 2) For M iterations, the residual is estimated as:

[0047]

[0048] The regression tree then starts from {(x i ,r im ), i=1...n} are generated, and the nodes of m trees and J leaves are R mj (j=1,2,...,J), then cmj Find the optimal value and minimize the loss function at the same time, that is:

[0049]

[0050] At this time c mj R mj The minimum square loss is:

[0051]

[0052] I is the indicator function, if x∈R mj Then I = 1; otherwise I = 0;

[0053] 3) The strong learner will be updated as:

[0054]

[0055] 4) The iteration stops when the decision function converges, and the final model with higher accuracy is:

[0056]

[0057] After that, the LightGBM model was constructed by introducing the gradient-based single-side sampling method (GOSS) and the mutually exclusive feature merging (EFB) method.

[0058] Furthermore, step S1 includes performing a preliminary estimation of the physical properties of the required aviation materials using quasi-static uniaxial tensile tests, uniaxial tension-compression tests, pure torsion tests, and multiaxial fatigue tests, fitting the test data using the Manson-Coffin equation and the Ramberg-Osgood equation, and then extracting the equation characteristics to obtain the uniaxial and cyclic fatigue properties of the materials.

[0059] Furthermore, the combined finite element analysis method described in step S2 specifically includes the following operations:

[0060] During the physical response calculation phase, the finite element solver ABAQUS / Standard was used with an embedded mixed cycle model to achieve the physical calculation effect of the Chaboche mixed cycle model. At the same time, the Ramberg-Osgood coefficient was used for plastic correction to ensure the accuracy of the physical response calculation under multi-axial loading.

[0061] During the fatigue life calculation phase, the finite element solver Simulia / Fe-Safe was used, and the Brown-Miller multiaxial load fatigue life estimation method was adopted. The load spectrum and physical response were input to calculate the fatigue life. The final result included three physical response quantities: displacement, stress, and strain. These data with time series characteristics will be used to establish a data set.

[0062] Furthermore, the discrete data in step S1 includes:

[0063] Basic properties: including material density and elastic modulus;

[0064] Hardening model parameters: including yield strength, tensile strength, Poisson's ratio, elongation and area reduction;

[0065] Cyclic fatigue parameters: including uniaxial tensile and compressive strain amplitude, uniaxial tensile and compressive stress amplitude, cyclic hardening coefficient and cyclic hardening exponent.

[0066] Beneficial Effects: Compared with existing technologies, the multimodal data-driven metal multiaxial fatigue life prediction method provided by this invention can establish a highly reliable dataset through physics-based knowledge and finite element methods when a training dataset is insufficient. Furthermore, by leveraging multimodal data input, fatigue life prediction under multiaxial loading can be performed with greater speed and accuracy, benefiting practical engineering applications, particularly those requiring more real-time prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1It is a schematic diagram of the framework flow of the method of the present invention;

[0068] Figure 2 This is a network structure diagram of the dynamic regression classification hybrid prediction method proposed by the present invention;

[0069] Figure 3 It is a fitting schematic diagram of the preliminary experimental embodiment of the present invention;

[0070] Figure 4 It is the algorithm effect diagram in the embodiment;

[0071] Figure 5 This is a diagram showing the effects of the present invention in predicting fatigue life of other aviation materials. DETAILED DESCRIPTION

[0072] In order to illustrate the technical solution disclosed in the present invention in detail, a specific introduction is given below in conjunction with the accompanying drawings.

[0073] This invention relates to big data methods and applications, specifically to the field of fatigue life prediction based on physics information, particularly low-cycle fatigue life prediction technology under complex multiaxial loading, and more specifically, a multimodal data-driven method for multiaxial fatigue life prediction of metals. The fatigue life prediction framework constructed in this invention utilizes physics-guided machine learning methods as an alternative model for fatigue life prediction, enabling high accuracy and improved computational efficiency in predicting fatigue life under multiaxial loading.

[0074] Combine Figure 1 The hybrid prediction framework structure shown in the figure, the method described in the present invention uses a high-reliability data set based on physical information and finite element analysis for training. With the help of test experiments, the required physical properties of aerospace materials within a specific range can be obtained. On this basis, the combined finite element analysis method is used to obtain the physical response of the parts associated with the geometric properties and the low-cycle fatigue life under the corresponding load conditions. However, this method requires a lot of computing cost and time cost. In this regard, the present invention uses ABAQUS / CAE and Simulia Fe-Safe solver to perform finite element solution to obtain a high-reliability data set associated with fatigue life for subsequent training of the agent model based on machine learning.

[0075] Considering that LightGBM, as an ensemble learning model, has a more efficient computing speed than other base models and is very flexible in dealing with discrete data, its modeling ability for time series data is limited. The deep learning method of long short-term memory network excels at time series and dependencies, and can automatically perform feature engineering, but it has the disadvantage of high requirements for computing resources. Therefore, the present invention designs a fusion prediction strategy that can combine the advantages of both, fully utilize the input features while improving computing efficiency and prediction accuracy.

[0076] For fusion, this paper proposes a dynamic regression selection algorithm based on the K-nearest neighbor algorithm, which can select a regression model suitable for dynamic system modeling. It combines regression analysis with time series prediction and dynamically adjusts the model to adapt to data changes.

[0077] Finally, the LightGBM model, long short-term memory network and dynamic regression selection algorithm are combined to train the two models on the previous data set, and the mean absolute error is used as an indicator to select the optimal model for each prediction sample to obtain the final prediction result.

[0078] Specifically, the present invention provides a multi-modal data driven metal multi-axial fatigue life prediction method comprising the following steps:

[0079] (1) Metal materials, especially multiphase metals used in complex environments such as aerospace and deep-sea exploration, have complex physical properties. Furthermore, determining the distribution of their physical responses based on actual experiments or conducting fatigue life tests on them is extremely costly. To obtain highly reliable results in the future, the present invention first establishes a database of properties of the aerospace materials to be analyzed based on preliminary experiments. The physical properties of the materials are then obtained by fitting engineering equations with the experimental data.

[0080] This step includes using quasi-static uniaxial tensile tests, uniaxial tension-compression tests, pure torsion tests, and multiaxial fatigue tests to make a preliminary estimate of the physical properties of the required aviation materials, fitting the test data using the Manson-Coffin equation and the Ramberg-Osgood equation, and then extracting the equation characteristics to obtain the uniaxial and cyclic fatigue properties of the material.

[0081] As attached Figure 2 The following is an example of material property fitting for TC4 alloy. The relationship between the Manson-Coffin equation and the Ramberg-Osgood equation has a good correlation with the test results. Figure 2 (a) is the fitting result of the Ramberg-Osgood equation, through which the stress-strain relationship of the experimental TC4 alloy in the uniaxial state is obtained. Figure 2 (b) is the Manson-Coffin equation, which provides the relationship between material strain and low-cycle fatigue life. This allows us to determine the material fatigue properties required for simulation.

[0082] (2) After obtaining the basic physical properties, a finite element model that can provide a data set is needed to meet subsequent data-driven requirements. The present invention establishes a finite element model based on the physical properties of the material. By combining finite element analysis with actual test comparison, fatigue life is determined as a label to establish a training data set. The data set is used for hidden relationship training of the machine learning algorithm.

[0083] This step establishes a high-accuracy finite element model based on physical properties, uses combined finite element analysis, compares it with actual experiments, and uses deterministic fatigue life as a label to establish a highly reliable training data set.

[0084] The combined finite element analysis method includes the following operations:

[0085] During the physical response calculation phase, the finite element solver ABAQUS / Standard was used with an embedded mixed cycle model to achieve the physical calculation effect of the Chaboche mixed cycle model. The Ramberg-Osgood coefficient was also used for plastic correction. This method ensures the accuracy of the physical response calculation under multiaxial loading.

[0086] In the fatigue life calculation stage, the finite element calculation solver Simulia / Fe-Safe is used, and the Brown-Miller multi-axial load fatigue life estimation method is adopted to input the load spectrum and physical response to calculate the fatigue life.

[0087] (3) Fatigue prediction under multi-axial load conditions is a very complex problem. One of the points is that the input information has multimodal characteristics. It can be simply classified as load spectrum input information with time series characteristics, and it also belongs to external environment information in fatigue life prediction. In addition, it includes discrete physical response information generated after the material is subjected to external loading. In order to better train the data set containing the above input features obtained by combining the finite element method in the previous step, the present invention proposes a hybrid prediction model DL-LGBM-DRS based on deep learning and LightGBM algorithm to construct dynamic regression selection, thereby realizing the prediction of low-cycle fatigue life under multimodal input environment.

[0088] In summary, this step is based on deep learning and LightGBM model, and proposes a hybrid prediction model DL-LGBM-DRS with dynamic regression selection. The network structure is shown in the attached figure. Figure 3 As shown in Figure 3, the prediction of low-cycle fatigue life under multi-modal input environment is achieved.

[0089] The steps for establishing the hybrid prediction model DL-LGBM-DRS with dynamic regression selection include:

[0090] (3.1) The GBDT algorithm is used to combine multiple decision trees to process dense discrete data and achieve high interpretability and generalization. The calculation is as follows:

[0091] The standard decision tree is as follows:

[0092]

[0093] In the formula, X1, X2 are the training subsets of each branch respectively yi , is the actual value of the sample, is the branch prediction value of the sample.

[0094] The GBDT algorithm requires the following sub-steps:

[0095] (3.1.1) Initialize the weak learner and define the loss function L(y,f(x)) and the prediction function g0(x):

[0096]

[0097] Where C m Expressed as the mean of the sample y in m trees

[0098] (3.1.2) For M iterations, the residual is estimated as:

[0099]

[0100] The regression tree then starts from {(x i ,r im ), i=1...n} are generated, and the nodes of m trees and J leaves are R mj (j=1,2,...,J), then cmj Find the optimal value and minimize the loss function at the same time, that is:

[0101]

[0102] At this time c m j is R m The minimum square loss of j is:

[0103]

[0104] I is the indicator function, if x∈R mj Then I = 1; otherwise I = 0.

[0105] (3.1.3) The strong learner will be updated as:

[0106]

[0107] (3.1.4) The iteration stops when the decision function converges, and the final model with higher accuracy is:

[0108]

[0109] After that, the LightGBM model was constructed by introducing the gradient-based single-side sampling method (GOSS) and the mutually exclusive feature merging (EFB) method.

[0110] (3.2) According to the recurrent neural network (RNN) model, the present invention constructs a deep learning model based on the long short-term memory network, wherein the long short-term memory network is constructed through a forget gate, an input gate, and an output gate to realize a long-term memory function.

[0111] (3.3) Using the dynamic regression selection method, the LightGBM model and the deep learning model are integrated to obtain a low-cycle fatigue life prediction algorithm with multimodal input. The specific steps of constructing the dynamic regression network algorithm include:

[0112] S33. Design a dynamic regression selection algorithm that can effectively integrate the LightGBM model and the deep learning model to form a fatigue life prediction algorithm that processes multimodal inputs;

[0113] The dynamic regression network is as follows:

[0114] Input set F, training set T, test set S, weight vector W, define the nearest neighbors as K, and the independent model number as n;

[0115] The output first defines MAE as mean absolute error:

[0116]

[0117] Where, y is the actual value of the training sample, is the predicted value of the sample;

[0118] Initialize the mean absolute error, define error Error = 0 and empty set A, then for each sample x in the test set S i , find the competition interval ψ of the sample corresponding to the training set T. For each independent model prediction value in each set group F Compute the mean absolute error with respect to the competition interval ψ:

[0119] Definition a n =w×MAE n , then A=A∪a n ;

[0120] Until the prediction results and error mean of all models are:

[0121]

[0122] For each predicted value in the cluster If a n >(a max -a min ) / 2 Then the model selection is carried out, at this time:

[0123]

[0124] Define N as the size of the cluster mean, for each a n Perform model combination, namely:

[0125]

[0126] Now define:

[0127]

[0128] The final error is

[0129] Error fin =Error / |X|

[0130] The dynamic regression selection algorithm finally evaluates the weight of each model, which is defined by the final error Errorfin above. f(xj) is the observed value of the corresponding sample in the test set, For the two models used in the present invention, for each sample, the value with the smaller error is determined to be the optimal value, and the local optimal prediction result is the final prediction result of the dynamic regression network algorithm.

[0131] Example

[0132] First, the proposed method was tested on a fatigue test dataset of a shaft sample to verify the effectiveness of the proposed model. The models selected for comparison were LSTM network without inputting loading spectrum (DL w / o loadingspecturms), independent deep learning network, independent LightGBM, XGBoost and random forest. The evaluation indicators were root mean square error (RMSE), mean absolute error (MAE), regression coefficient (R 2 ) and calculation speed. The experimental results are respectively in the appendix of the manual. Figure 4 In display.

[0133] The fusion model proposed in this invention has better performance in predicting simulated fatigue life. The prediction performance of LGBM is good, showing its high learning efficiency on discretized feature data, and it is a potential fatigue behavior prediction model. The results of deep learning are worthy of attention. The accuracy of the deep learning model using the fusion of loading spectrum and physical response as input is slightly lower than that of the proposed DRS model. However, after removing the loading spectrum, the accuracy of deep learning dropped significantly, while the error increased significantly. This difference highlights the importance of load spectrum information as input. In addition, we also compared other GBDT-based algorithms, including XGBoost and random forest, and their gradual decline in prediction performance is consistent with the pattern we found in previous studies. In general, the errors of the machine learning models are within an acceptable range, and the prediction accuracy of the fusion model is higher than that of the single model. In terms of time cost, due to the addition of the decision layer that plays a role in late fusion, the fusion model is slower than the independent model. However, compared with the finite element analysis method, the machine learning model used for environmental applications in this study has an incomparable speed advantage. Appendix Figure 4 Figure c shows a comparison between the fatigue life predicted by the machine learning model and the fatigue life calculated using the finite element analysis method. Clearly, the proposed hybrid predictive DL-LGBM-DRS model shows very favorable results compared to other models when predicting fatigue life based on the generated dataset. This demonstrates the feasibility of using the DL-LGBM-DRS model as an alternative model for fatigue life prediction, maintaining high accuracy while also improving efficiency.

[0134] Subsequently, the method described in the present invention was tested on other commonly used metal materials in aerospace to verify the effectiveness of the model in predicting fatigue life under multi-axial loads. The metals selected for comparison were 304 stainless steel, pure titanium, GH4169 and 7075-T651 aluminum alloys. The prediction results are shown in the attached manual. Figure 5 shown.

[0135] Figure 5 (a) shows the prediction results of 304 stainless steel, Figure 5 (b) is the prediction result of 7075-T651, Figure 5 (c) shows the prediction results of pure titanium, Figure 5(d) Prediction results for the GH4169 superalloy. The predicted sample is consistent with the TC4 alloy. The dashed lines represent the 1.5-fold and double scatter bands, verifying the accuracy of the prediction results. The results show that the fit results for 304 stainless steel, pure titanium, and the GH4169 superalloy are quite good; the fitting points lie within the double scatter bands. In the GH4169 case, it can be observed that most of the fitting points lie below the fitting line, indicating that the model's fatigue life prediction is relatively pessimistic. This phenomenon is consistent with the fatigue life trend calculated using the Brown-Miller method. The predicted results for the 7075-T651 aluminum alloy case deviate from the experimental results. Due to cost constraints, we conducted fewer experiments for the 7075-T651 alloy. This may result in a probabilistic distribution of the experimental fatigue life. In addition, due to its material properties, the 7075-T651 alloy exhibits a higher fatigue life under comparable loading conditions. In this case, the fusion method guided by the physical model is insufficient to predict high-cycle fatigue life. Because the mechanical properties of the materials were incorporated differently to save costs in this experiment, subtle differences in the material properties can lead to discrepancies in the final prediction results. Overall, the proposed method is satisfactory for predicting low-cycle fatigue life under multiaxial loading. Furthermore, the results obtained using the proposed framework are in good agreement with experimental fatigue life for different alloys. This demonstrates the potential of the proposed method for fatigue life prediction in other metallic materials.

Claims

1. A multi-modal data driven metal multiaxial fatigue life prediction method, characterized in that: The steps include: S1. Based on the preliminary experiments, a database of properties of aerospace materials to be analyzed is established. The physical properties of the materials are obtained by fitting engineering equations with the experimental data. These properties are then added to the data set as discrete characteristic properties. S2. Build a finite element model based on the physical properties of the material. By combining finite element analysis with actual test results, fatigue life is determined as a label to establish a highly reliable training dataset based on simulation. This training dataset is used by machine learning algorithms to learn physical relationships. S3, the data set includes multi-axial load spectrum with time series characteristics, time series physical response and material and external environment information with discrete characteristics; This step involves constructing a hybrid prediction algorithm DL-LGBM-DRS based on dynamic regression selection, thereby realizing the prediction of multiaxial fatigue life of metal materials under multimodal input characteristics. The algorithm establishment steps include: S31. By introducing a gradient-based unilateral sampling method and a mutually exclusive feature merging method, we construct a LightGBM model. Then, we use the GBDT algorithm to combine multiple decision trees to process dense discrete data, achieving high interpretability and generalization. S32. Constructing a deep learning model based on a long short-term memory network based on a recurrent neural network, wherein the long short-term memory network is constructed by a forget gate, an input gate, and an output gate to realize a long-term memory function; S33. Build a dynamic regression selection algorithm to fuse the LightGBM model and the deep learning model to form a fatigue life prediction algorithm that processes multimodal inputs. The dynamic regression network is as follows: Input set F, training set T, test set S, weight vector w, and independent model number n; The output first defines MAE as mean absolute error: Where, y is the actual value of the training sample, is the predicted value of the sample; Initialize the mean absolute error, define error Error = 0 and empty set A, then for each sample x in the test set S i , corresponding to the training set T to find the competition interval of the sample ψ , for each independent model prediction value in each cluster F Calculate the contention interval ψ The mean absolute error is calculated as follows: Definition a n =w×MAE n , then A=A∪a n , until the prediction results and error means of all models are: and Represents the update of the set, for each predicted value in the set group If a n >(a max -a min ) / 2, then perform model selection and perform the following updates: Define N as the size of the cluster mean, for each an Perform model combination, namely: Now define: The final error is Error fin =Error / |X|; The dynamic regression selection algorithm finally evaluates the weight of each model, which is defined by the final error Errorfin above, f(x j ) is the observed value of the corresponding sample in the test set, is the normalized result; For each sample, the value with the smaller error is judged to be the optimal one, and the local optimal prediction result is the final prediction result of the dynamic regression network algorithm.

2. The multi-modal data driven metal multi-axial fatigue life prediction method according to claim 1, characterized in that: The calculation steps of the GBDT algorithm for processing the discrete part of the input in step S31 include: The GBDT algorithm is used to combine multiple decision trees to process dense discrete data and achieve high interpretability and generalization. The calculation is as follows: The standard decision tree is as follows: Where X1, X2 are the training subsets y of each branch respectively i , is the actual value of the sample, is the branch prediction value of the sample; The GBDT algorithm requires the following sub-steps: 1) Initialize the weak learner and define the loss function L(y,f(x)) and the prediction function g0(x): Where C m Expressed as the mean of sample y in m trees; 2) For M iterations, the residual is estimated as: The regression tree then starts from {(x i ,r im ), i=1...n} are generated, and the nodes of m trees and J leaves are R mj (j=1,2,...,J), then cmj Find the optimal value and minimize the loss function at the same time, that is: At this time c mj R mj The minimum square loss is: I is the indicator function, if x∈R mj Then I = 1; otherwise I = 0; 3) The strong learner will be updated as: 4) The iteration stops when the decision function converges, and the final model with higher accuracy is: After that, the construction of the LightGBM model was realized by introducing the gradient-based unilateral sampling method and the mutually exclusive feature merging method.

3. The multi-modal data driven metal multi-axial fatigue life prediction method according to claim 1, characterized in that: Step S1 includes using quasi-static uniaxial tensile tests, uniaxial tension-compression tests, pure torsion tests, and multiaxial fatigue tests to make a preliminary estimate of the physical properties of the required aviation materials. The test data are fitted using the Manson-Coffin equation and the Ramberg-Osgood equation, and then the equation characteristics are extracted to obtain the uniaxial and cyclic fatigue properties of the material.

4. The multi-modal data driven metal multi-axial fatigue life prediction method according to claim 1, characterized in that: The combined finite element analysis method described in step S2 specifically includes the following operations: During the physical response calculation phase, the finite element solver ABAQUS / Standard was used with an embedded mixed cycle model to achieve the physical calculation effect of the Chaboche mixed cycle model. At the same time, the Ramberg-Osgood coefficient was used for plastic correction to ensure the accuracy of the physical response calculation under multi-axial loading. During the fatigue life calculation phase, the finite element solver Simulia / Fe-Safe was used, and the Brown-Miller multiaxial load fatigue life estimation method was adopted. The load spectrum and physical response were input to calculate the fatigue life. The final result included three physical response quantities: displacement, stress, and strain. These data with time series characteristics will be used to establish a data set.

5. The multi-modal data driven metal multiaxial fatigue life prediction method according to claim 1, characterized in that: The discrete data described in step S1 include: Basic properties: including material density and elastic modulus; Hardening model parameters: including yield strength, tensile strength, Poisson's ratio, elongation and area reduction; Cyclic fatigue parameters: including uniaxial tensile and compressive strain amplitude, uniaxial tensile and compressive stress amplitude, cyclic hardening coefficient and cyclic hardening exponent.

Images