Metal material fatigue life prediction method driven by multi-scale and multi-depth neural network

Abstract

The present invention discloses a method for predicting the fatigue life of metal materials driven by a multi-scale, multi-depth neural network. The method comprises the following steps: constructing a metal material fatigue test dataset; considering the additional hardening effect of metal materials under non-proportional loads, adding the maximum normal stress of the critical plane to the dataset, and preprocessing the dataset; establishing a multi-scale, multi-depth neural network framework; considering that physical constraints can guide neural network training to increase prediction accuracy, and establishing a physical loss function with soft adjustment capabilities; building a multi-scale, multi-depth neural network-driven metal material fatigue life prediction model based on the multi-scale, multi-depth neural network framework and the physical loss function; continuously iteratively optimizing the model to obtain an optimal prediction model; and using the optimal prediction model to predict the fatigue life of the metal material. The present invention not only has the advantages of high prediction accuracy, but also enhances the stability and generalization of prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of material fatigue life prediction methods, and specifically provides a metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network. Background Art

[0002] Fatigue damage in engineering materials is a common cause of component failure. In practical engineering, many mechanical components are subject to cyclic loading, which is particularly common in vehicles and structures such as aircraft, high-speed trains, automobiles, bridges, and ships. Therefore, in-depth research on the fatigue of engineering materials is particularly important. To date, traditional methods for predicting the fatigue life of metal materials can be mainly divided into energy-based methods, equivalent strain-based methods, and critical plane-based methods. However, these prediction methods often add additional parameters as the complexity of the material loading path increases. This results in improved prediction accuracy but reduced robustness and generalization capabilities.

[0003] In recent years, neural networks have been widely used in fatigue life prediction for metal materials due to their superior learning and generalization capabilities. However, due to the lack of physical constraints, these models often exhibit significant fluctuations in prediction accuracy. Furthermore, previous research using neural networks to predict fatigue life for metal materials has mostly extracted and processed sample data at only a single scale or depth, failing to fully capture both deep and shallow features of the data. This results in poor prediction accuracy and stability. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a multi-scale and multi-depth neural network driven metal material fatigue life prediction method that can effectively solve the problems of large fluctuations in prediction accuracy, low robustness, and incomplete feature extraction and processing in existing metal material fatigue life prediction methods.

[0005] The technical solution adopted by the present invention is: a metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network, which specifically includes the following steps:

[0006] S1: Construction of metal material fatigue test dataset;

[0007] Collect fatigue test data sets of metal materials under different stress levels and multiple loading paths; including material performance parameters, dimensions, load parameters, loading paths, and fatigue life parameters;

[0008] S2: calculation of maximum normal stress on critical plane;

[0009] Based on the performance parameters and dimensions of the metal material, a finite element model is established; cyclic loads at different stress levels are applied to the specimens, and according to the critical plane theory, the maximum normal stress σ of the critical plane of the specimen under different loading paths is obtained. n,max ; This stress reflects the additional hardening effect of the material under non-proportional load;

[0010] S3: The maximum normal stress σ on the critical plane n,max , the load parameters obtained in step S1 are used as input data, fatigue life N e As output data; input data and output data together constitute the data set of the model;

[0011] S4: Dataset preprocessing for the model;

[0012] S41: Use the maximum and minimum normalization method to normalize the model data set to eliminate the impact of data set dimension differences on the model prediction results;

[0013] S42: Use K-fold cross-validation method to process the data set of the model to obtain training set and test set;

[0014] S5: Construct a fatigue life prediction model for metal materials driven by a multi-scale and multi-depth neural network;

[0015] S51: Construct a feature extraction module to extract features at multiple scales (channel scale, feature dependency scale);

[0016] S52: Construct a feature processing module to process features at multiple scales (global scale, local scale, feature enhancement scale) and multiple depths (deep features, shallow features);

[0017] S53: Incorporate physical constraints into the loss function to construct a physical loss function with soft adjustment function;

[0018] S6: Iterative optimization model, using the optimal multi-scale and multi-depth neural network model to predict the fatigue life of metal materials;

[0019] Preferably, in step S1, the fatigue test data set includes material performance parameters: elastic modulus (E), Poisson's ratio (ν); load parameters: strain amplitude (ε a ,γ a ), stress amplitude (σ a ,τ a ), mean stress σ m ; Fatigue life parameter: fatigue life N e .

[0020] Further preferably, in step S2, the normal stress of any plane is as shown in formula (1):

[0021]

[0022] Where: σ θ is the normal stress of the plane; σ x is the normal stress; θ is the angle between any material plane and the specimen axis; τ xy is the shear stress;

[0023] For different loading paths, plane positions from 0° to 180° are defined with a step size of 0.1° to accurately capture the stress state on each plane; for each specific plane position, the corresponding stress component is calculated; based on the critical plane theory, the maximum normal stress on the critical plane is obtained.

[0024] Further preferably, in step S41, the data is scaled proportionally using a maximum-minimum normalization method to eliminate the influence of dimensional differences between the data on the model prediction results; the maximum-minimum normalization formula is shown in formula (2):

[0025]

[0026] Where: X N is the normalized input parameter; X is the initial input parameter; X min and X max are the minimum and maximum values ​​of the input parameters, respectively.

[0027] Further preferably, in step S42, the data set is processed using a K-fold cross-validation method to divide the data set into a training set and a test set, and the auxiliary model is used for parameter fitting and hyperparameter optimization, so as to obtain the optimal neural network-based metal material fatigue life prediction model; the specific steps are as follows:

[0028] a. First, the data is randomly divided into k equal-sized and non-overlapping subsets;

[0029] b. Then, k-1 copies of these subsets are selected as training data to train the model, and the remaining copy is used as validation data to evaluate the performance of the model;

[0030] c. Repeat this process k times, each time using a different subset as the validation set;

[0031] d. Calculate the model error on different validation sets and accumulate k error values;

[0032] e. Finally, calculate the average of these error values ​​to get the average error of cross validation;

[0033] f. Optimize model hyperparameters based on the average error.

[0034] Further preferably, in step S51, in the feature extraction module, an innovative multi-head attention mechanism is adopted, which integrates the new channel attention and evolutionary attention to efficiently extract multi-scale features, that is, extracting features from the channel scale of the feature map and the dependency scale between features; then, these features are integrated through a fully connected neural network, which consists of a Dropout function, a linear layer and a Relu activation function to enhance the expressiveness and generalization ability of the features.

[0035] In the channel attention module, the input data is first average-pooled through two pooling windows of different sizes to achieve feature fusion of the input data channel dimension; then, these pooled data are further processed by a 1×1 two-dimensional convolution operator to achieve multi-scale analysis of the feature channels; the processed feature maps are merged and weighted summed with the original data. This process aims to generate more accurate spatial features through feature processing at different spatial scales, so as to improve the efficiency and accuracy of the model in completing regression tasks; in order to overcome the degradation problem in deep network training, a residual network structure is introduced to achieve self-repair of the model, ensuring that the model performance increases instead of decreases as the number of network layers increases, thereby effectively avoiding network degradation.

[0036] The evolutionary attention mechanism consists of two sub-modules: sequence attention and IP attention, which aims to obtain the logical evolution relationship between the input sequence and fatigue life. First, the input data is preprocessed by upsampling (bilinear interpolation) to obtain a smoother information transition. In the sequence attention module, the data is first processed by the dilated convolution operator to extract the long-term characteristic trends of the data. These features are then processed nonlinearly through the normalization layer, linear layer and Relu activation function to enhance the model's feature recognition ability. Finally, the processed features are weighted summed by the Sigmoid function. This step is crucial for strengthening effective features because it allows the model to dynamically adjust its weights according to the importance of the features, thereby further improving the model's sensitivity to key features and prediction accuracy.

[0037] In IP attention, for each input sample, the one-dimensional data sequence is first randomly rearranged to construct a two-dimensional feature map in order to capture the mutual dependence of data features in the spatial dimension; then, after upsampling, pooling operations are performed in the two orthogonal directions of the X-axis and the Y-axis, thereby simplifying the two-dimensional feature map into two one-dimensional sequences; after the dimensionality conversion, the two sequences will interact with each other in a single dimension, and their mapping capabilities are further enhanced through nonlinear transformations; the features after nonlinear transformation are then reorganized and restored to the form before the sequence dimension conversion; finally, these features are weightedly summed with the original input data through the Sigmoid function to achieve further feature enhancement; this process is designed to enable the neural network to capture the feature dependencies between different positions on the feature map, so as to deeply explore the logical deduction relationship between strain amplitude, stress amplitude, maximum stress of critical surface and fatigue life in the spatial dimension.

[0038] Further preferably, in step S52, in the feature processing module, after being processed by a two-dimensional convolution operator with a 1×1 convolution kernel, the features are respectively sent to three feature processing mechanisms of different scales; these mechanisms are responsible for extracting local features, global features, and enhanced features, respectively, to comprehensively capture the multi-scale and multi-depth features of the data, that is, to capture deep and shallow features from the global scale, local scale, and feature enhancement scale; subsequently, these features are fused in the channel dimension to integrate feature information of different scales and depths; the fused features are further processed by a feature attention mechanism composed of a linear layer and a Relu activation function, which purifies the features to enhance the model's ability to recognize key information; finally, through the global average pooling mechanism, the model outputs the fatigue life prediction value N of the metal material p , ensuring the stability and reliability of the prediction results.

[0039] Further preferably, in step S53, the physical loss function expression of the soft adjustment function is as shown in formula (3):

[0040] L=λ1L h +λ2L p (3)

[0041] Where: L is the physical loss function; λ1 and λ2 are the weights of the two loss functions; L h is the Huber loss function; L p is the physical constraint function; L h and L p The expressions of are shown in formula (4) and formula (5):

[0042]

[0043] Where: N eis the fatigue life verified in the experiment; N p is the fatigue life predicted by the model after training; δ and d are adjustment parameters, which can adjust the degree of model training penalty by changing their values.

[0044] The physical loss function controls the participation of physical constraints in neural network training by adjusting the values ​​of λ1 and λ2, thus realizing soft adjustment function.

[0045] Further preferably, in step S6, the fatigue test data obtained in step S1, including the strain amplitude (ε a ,γ a ), stress amplitude (σ a ,τ a ), mean stress σ m , and the maximum normal stress value σ of the critical surface calculated in step S2 n,max , these data are used as input to the model; fatigue life N p is used as the output of the model; then, the input data is preprocessed according to the data processing method described in step S4; then, the model constructed in step S5 is trained and iteratively optimized to obtain the optimal multi-scale and multi-depth neural network driven metal material fatigue life prediction model; finally, the optimal multi-scale and multi-depth neural network driven metal material fatigue life prediction model is used to predict the fatigue life of the metal material under cyclic load; the prediction results are visualized through a scatter plot for easy analysis and understanding.

[0046] The beneficial effects of the present invention are: by integrating a new neural network architecture and a physical loss function, a novel multi-scale and multi-depth neural network-driven fatigue life prediction model for metal materials is constructed; the model uses stress amplitude, strain amplitude, average stress and maximum normal stress on the critical plane as input parameters, and the fatigue life of the metal material as the output result, thereby achieving a fatigue life prediction of metal materials with higher prediction accuracy and stronger prediction stability; compared with traditional models, the present invention effectively solves the problems of insufficient prediction accuracy, poor prediction stability and limited robustness; therefore, the application of the method of the present invention to predict the fatigue life of metal materials under cyclic loads can establish a more accurate and stable mapping relationship between fatigue parameters and fatigue life. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of an embodiment of the present invention;

[0048] Figure 2 This is the overall framework diagram of the prediction model of the present invention;

[0049] Figure 3 This is the spatial attention framework diagram of the present invention;

[0050] Figure 4 This is the evolutionary attention framework diagram of the present invention;

[0051] Figure 5 This is a comparison chart of the predicted life and the tested life under cyclic load in an example of the present invention; DETAILED DESCRIPTION

[0052] In order to allow professionals in this field to more clearly understand the implementation details of the present invention, the specific contents of the present invention will be described in detail below with reference to relevant diagrams.

[0053] The present invention proposes a method for predicting fatigue life of metal materials driven by a multi-scale and multi-depth neural network. Figure 1 , the specific steps are outlined as follows.

[0054] S1: Construction of metal material fatigue test dataset;

[0055] The fatigue data of AZ61A magnesium alloy under four different loading paths (complete reverse tension and compression, complete reverse cyclic torsion, proportional loading, and non-proportional loading) provided in the literature published abroad [1] Yu Q, Zhang J, Jiang Y, Li Q.'Multiaxial fatigue of extruded AZ61A magnesium alloy, International Journal of Fatigue, 2011; 33: 437-47. were collected, including the dimensions, elastic modulus (E), Poisson's ratio (ν), strain amplitude (ε a ,γ a ), stress amplitude (σ a ,τ a ), mean stress σ m , and fatigue life N e .

[0056] S2: calculation of maximum normal stress on critical plane;

[0057] An accurate finite element model was constructed based on the dimensions, elastic modulus (E), and Poisson's ratio (ν) of the metal material. By simulating experimental conditions, cyclic loads including tension, torsion, and tension-torsion combined directions were applied to the specimen to obtain the maximum normal stress σ on the critical plane of the specimen under different loading paths. n,max This stress parameter reflects the additional hardening effect of the material under non-proportional loads, providing a more comprehensive description of the stress state for fatigue analysis of the material. The normal stress of any plane is shown in formula (1):

[0058]

[0059] Where: σ θ is the normal stress of the plane; σ x is the normal stress; θ is the angle between any material plane and the specimen axis; τ xy is the shear stress.

[0060] For different loading paths, plane positions from 0° to 180° are defined with a step size of 0.1° to accurately capture the stress state on each plane; for each specific plane position, the corresponding stress component is calculated; based on the critical plane theory, the maximum normal stress on the critical plane is obtained.

[0061] This process ensures that the fatigue behavior of materials under various complex loading conditions can be accurately evaluated, providing key stress parameters for fatigue life prediction.

[0062] S3: The maximum normal stress σ on the critical plane n,max , strain amplitude (ε a ,γ a ), stress amplitude (σ a ,τ a ), mean stress σ m As input data, fatigue life N e As output data; input data and output data together constitute the data set of the model;

[0063] S4: Dataset preprocessing for the model;

[0064] S41: Use the maximum and minimum normalization method to normalize the model data set. This method effectively eliminates the dimensional differences between data by scaling them proportionally, providing a unified data scale for the model, thereby improving the generalization ability and prediction accuracy of the model. The normalization formula is shown in formula (2):

[0065]

[0066] Where: X N is the normalized input parameter; X is the initial input parameter; X min and X max are the minimum and maximum values ​​of the input parameters, respectively.

[0067] S42: Use K-fold cross-validation method to process the data set of the model to obtain training set and test set;

[0068] The K-fold cross-validation method is used to divide the data set into a training set and a validation set to assist the model in parameter fitting and hyperparameter optimization, thereby obtaining the optimal neural network-based metal material fatigue life prediction model. The specific steps are as follows:

[0069] a. First, the data is randomly divided into k equal-sized and non-overlapping subsets;

[0070] b. Then, k-1 copies of these subsets are selected as training data to train the model, and the remaining copy is used as validation data to evaluate the performance of the model;

[0071] c. Repeat this process k times, each time using a different subset as the validation set;

[0072] d. Calculate the model error on different validation sets and accumulate k error values;

[0073] e. Finally, calculate the average of these error values ​​to get the average error of cross validation;

[0074] f. Optimize model hyperparameters based on the average error.

[0075] S5: Construct a fatigue life prediction model for metal materials driven by a multi-scale and multi-depth neural network;

[0076] The feature extraction module and feature processing module are used to capture and process deep features and shallow features from the channel scale, feature dependency scale, global scale, and local scale to construct a metal material fatigue life prediction model driven by a multi-scale and multi-depth neural network.

[0077] S51: Construct feature extraction module, multi-scale (channel scale, feature dependency scale)

[0078] Extract features;

[0079] In the feature extraction module, an innovative multi-head attention mechanism is adopted, which combines the novel channel attention and evolutionary attention to efficiently extract multi-scale features, that is, extracting features from the channel scale of the feature map and the dependency scale between features; then, these features are integrated through a fully connected neural network consisting of a Dropout function, a linear layer, and a Relu activation function to enhance the expressiveness and generalization ability of the features.

[0080] In the channel attention module, the input data is first average-pooled through two pooling windows of different sizes to achieve feature fusion of the input data channel dimension; then, these pooled data are further processed by a 1×1 two-dimensional convolution operator to achieve multi-scale analysis of the feature channels; the processed feature maps are merged and weighted summed with the original data. This process aims to generate more accurate spatial features through feature processing at different spatial scales, so as to improve the efficiency and accuracy of the model in completing regression tasks; in order to overcome the degradation problem in deep network training, a residual network structure is introduced to achieve self-repair of the model, ensuring that the model performance increases instead of decreases as the number of network layers increases, thereby effectively avoiding network degradation.

[0081] The evolutionary attention mechanism consists of two sub-modules: sequence attention and IP attention, which aims to obtain the logical evolution relationship between the input sequence and fatigue life. First, the input data is preprocessed by upsampling (bilinear interpolation) to obtain a smoother information transition. In the sequence attention module, the data is first processed by the dilated convolution operator to extract the long-term characteristic trends of the data. These features are then processed nonlinearly through the normalization layer, linear layer and Relu activation function to enhance the model's feature recognition ability. Finally, the processed features are weighted summed by the Sigmoid function. This step is crucial for strengthening effective features because it allows the model to dynamically adjust its weights according to the importance of the features, thereby further improving the model's sensitivity to key features and prediction accuracy.

[0082] In IP attention, for each input sample, the one-dimensional data sequence is first randomly rearranged to construct a two-dimensional feature map in order to capture the mutual dependence of data features in the spatial dimension; then, after upsampling, pooling operations are performed in the two orthogonal directions of the X-axis and the Y-axis, thereby simplifying the two-dimensional feature map into two one-dimensional sequences; after the dimensionality conversion, the two sequences will interact with each other in a single dimension, and their mapping capabilities are further enhanced through nonlinear transformations; the features after nonlinear transformation are then reorganized and restored to the form before the sequence dimension conversion; finally, these features are weightedly summed with the original input data through the Sigmoid function to achieve further feature enhancement; this process is designed to enable the neural network to capture the feature dependencies between different positions on the feature map, so as to deeply explore the logical deduction relationship between strain amplitude, stress amplitude, maximum stress of critical surface and fatigue life in the spatial dimension.

[0083] S52: Construct a feature processing module to process features at multiple scales (global scale, local scale, feature enhancement scale) and multiple depths (deep features, shallow features);

[0084] In the feature processing module, after being processed by the two-dimensional convolution operator of the 1×1 convolution kernel, the features are sent to three feature processing mechanisms of different scales respectively; these mechanisms are responsible for extracting local features, global features and enhanced features respectively to fully capture the multi-scale and multi-depth features of the data, that is, to capture deep and shallow features from the global scale, local scale and feature enhancement scale; then, these features are fused in the channel dimension to integrate feature information of different scales and depths; the fused features are further processed by the feature attention mechanism composed of linear layer and Relu activation function, which purifies the features to enhance the model's ability to recognize key information; finally, through the global average pooling mechanism, the model outputs the fatigue life prediction value N of the metal material p , ensuring the stability and reliability of the prediction results.

[0085] S53: Incorporate physical constraints into the loss function and construct a physical loss function with soft adjustment function;

[0086] In order to increase the optimization process of the model, physical constraints are integrated into the loss function, and a physical loss function with soft adjustment function is constructed; the physical loss function expression of the soft adjustment function is shown in formula (3):

[0087] L=λ1L h +λ2L p (3)

[0088] Where: L is the physical loss function; λ1 and λ2 are the weights of the two loss functions; L h is the Huber loss function; L p is the physical constraint function; L h and L p The expressions of are shown in formula (4) and formula (5):

[0089]

[0090] Where: N e is the fatigue life verified in the experiment; N p is the fatigue life predicted by the model after training; δ and d are adjustment parameters, which can adjust the degree of model training penalty by changing their values.

[0091] The physical loss function controls the participation of physical constraints in neural network training by adjusting the values ​​of λ1 and λ2, thus realizing soft adjustment function.

[0092] S6: Iterative optimization model, using the optimal multi-scale and multi-depth neural network model to predict the fatigue life of metal materials;

[0093] According to the fatigue test data obtained in step S1, including the strain amplitude (εa ,γ a ), stress amplitude (σ a ,τ a ), mean stress σ m , and the maximum normal stress value σ of the critical surface calculated in step S2 n,max , these data are used as input to the model; fatigue life N p is used as the output of the model; then, the input data is preprocessed according to the data processing method described in step S4; then, the model constructed in step S5 is trained and iteratively optimized to obtain the optimal multi-scale and multi-depth neural network driven metal material fatigue life prediction model, whose hyperparameters are shown in Tables 1, 2, and 3;

[0094] Table 1

[0095]

[0096] Table 2

[0097]

[0098] Table 3

[0099]

[0100] Next, the optimal neural network fatigue life prediction model is used to predict the fatigue life of metal materials. In order to intuitively display the prediction results, a scatter plot is used to visualize the prediction results. Figure 5 .

[0101] according to Figure 5 The results show that the model proposed in the present invention exhibits high prediction accuracy in four different loading paths; the probability of its prediction results being within 1.5 times the scatter band is 87%, and the probability of being within 1.5 times the scatter band is 93%; the prediction accuracy of the test set is lower than that of the training set, which is due to the generalization error inherent in the structural characteristics of the neural network; in summary, in the test set under the four loading paths, the prediction results of the model proposed in the present invention are basically within 1.5 times the scatter band, and the prediction accuracy is relatively stable.

[0102] Professionals in this field should understand that the above embodiments are only used to illustrate the basic principles of the present invention and do not mean that the scope of protection of the present invention is limited to this specific implementation mode; for those skilled in the art, the implementation of the present invention can be subject to various changes and adjustments; any modifications, adjustments and equivalent substitutions within the framework of the basic principles of the present invention should be deemed to be included in the scope of protection of the claims of the present invention.

Claims

1. A multi-scale and multi-depth neural network driven metal material fatigue life prediction method includes the following steps: S1: Construction of metal material fatigue test dataset; Collect fatigue test data sets of metal materials under different stress levels and multiple loading paths; including material performance parameters, dimensions, load parameters, loading paths, and fatigue life parameters; S2: calculation of maximum normal stress on critical plane; Based on the performance parameters and dimensions of the metal material, a finite element model is established; cyclic loads at different stress levels are applied to the specimens, and according to the critical plane theory, the maximum normal stress σ of the critical plane of the specimen under different loading paths is obtained. n,max ; This stress reflects the additional hardening effect of the material under non-proportional load; S3: The maximum normal stress σ on the critical plane n,max , the load parameters obtained in step S1 are used as input data, fatigue life N e As output data; input data and output data together constitute the data set of the model; S4: Dataset preprocessing for the model; S41: Use the maximum and minimum normalization method to normalize the model data set to eliminate the impact of data set dimension differences on the model prediction results; S42: Use K-fold cross-validation method to process the data set of the model to obtain training set and test set; S5: Construct a fatigue life prediction model for metal materials driven by a multi-scale and multi-depth neural network; S51: Construct a feature extraction module to extract features at multiple scales (channel scale, feature dependency scale); S52: Construct a feature processing module to process features at multiple scales (global scale, local scale, feature enhancement scale) and multiple depths (deep features, shallow features); S53: Incorporate physical constraints into the loss function to construct a physical loss function with soft adjustment function; S6: Iteratively optimize the model and use the optimal multi-scale and multi-depth neural network model to predict the fatigue life of metal materials.

2. The metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network according to claim 1 is characterized in that: In step S1, the fatigue test data set includes material performance parameters: elastic modulus (E), Poisson's ratio (ν); load parameters: strain amplitude (ε a ,γ a ), stress amplitude (σ a ,τ a ), mean stress σ m ; Fatigue life parameter: fatigue life N e .

3. The metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network according to claim 1 is characterized in that: In step S2, the normal stress of any plane is expressed as formula (1): Where: σ θ is the normal stress of the plane; σ x is the normal stress; θ is the angle between any material plane and the specimen axis; τ xy is the shear stress; For different loading paths, plane positions from 0° to 180° are defined with a step size of 0.1° to accurately capture the stress state on each plane; for each specific plane position, the corresponding stress component is calculated; Based on the critical plane theory, the maximum normal stress on the critical plane is obtained.

4. The metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network according to claim 1 is characterized in that: In step S41, the data is scaled proportionally using the maximum and minimum normalization method to eliminate the influence of the dimension difference between the data on the model prediction results; the maximum and minimum normalization formula is shown in formula (2): Where: X N is the normalized input parameter; X is the initial input parameter; X min and X max are the minimum and maximum values ​​of the input parameters, respectively.

5. The metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network according to claim 1 is characterized in that: In step S42, the K-fold cross-validation method is used to process the data set, divide it into a training set and a test set, and assist the model in parameter fitting and hyperparameter optimization, thereby obtaining the optimal neural network-based metal material fatigue life prediction model. The specific steps are as follows: a. First, the data is randomly divided into k equal-sized and non-overlapping subsets; b. Then, k-1 copies of these subsets are selected as training data to train the model, and the remaining copy is used as validation data to evaluate the performance of the model; c. Repeat this process k times, each time using a different subset as the validation set; d. Calculate the model error on different validation sets and accumulate k error values; e. Finally, calculate the average of these error values ​​to get the average error of cross validation; f. Optimize model hyperparameters based on the average error.

6. The metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network according to claim 1 is characterized in that: In step S51, in the feature extraction module, an innovative multi-head attention mechanism is adopted, which combines the new channel attention and evolutionary attention to efficiently extract multi-scale features, that is, extract features from the channel scale of the feature map and the dependency scale between features; Then, These features are integrated through a fully connected neural network, which consists of a Dropout function, a linear layer, and a ReLU activation function to enhance the expressiveness and generalization ability of the features; In the channel attention module, the input data is first average-pooled across two pooling windows of different sizes to achieve feature fusion across the input data channel dimension. The pooled data is then further processed using a 1×1 two-dimensional convolution operator to achieve multi-scale analysis of the feature channels. The processed feature maps are merged and weighted summed with the original data. This process aims to generate more precise spatial features by processing features at different spatial scales, thereby improving the efficiency and accuracy of the model in completing regression tasks. To overcome the degradation problem in deep network training, a residual network structure is introduced to achieve self-healing of the model, ensuring that model performance improves rather than decreases as the number of network layers increases, thereby effectively avoiding network degradation. The evolutionary attention mechanism consists of two submodules: sequence attention and IP attention. It aims to obtain the logical evolution relationship between the input sequence and fatigue life. First, the input data is preprocessed through upsampling (bilinear interpolation) to obtain a smoother information transition. In the sequence attention module, the data is first processed by the dilated convolution operator to extract the long-term characteristic trends of the data. These features are then processed nonlinearly through normalization layers, linear layers, and ReLU activation functions to enhance the model's ability to recognize features. Finally, the processed features are weighted and summed using the Sigmoid function. This step is crucial for strengthening effective features because it allows the model to dynamically adjust the weights of features based on their importance, thereby further improving the model's sensitivity to key features and prediction accuracy. In IP attention, for each input sample, the one-dimensional data sequence is first randomly rearranged to construct a two-dimensional feature map in order to capture the interdependence of data features in the spatial dimension; Subsequently, after upsampling, pooling operations are performed in two orthogonal directions, the X-axis and the Y-axis, thereby simplifying the two-dimensional feature map into two one-dimensional sequences. After the two sequences are transformed, their features will interact on a single dimension and further enhance their mapping capabilities through nonlinear transformation. The features after nonlinear transformation are then reorganized and restored to their original form before the sequence dimension transformation. Ultimately, these features are weighted and summed with the original input data through a Sigmoid function to further enhance the features. This process aims to enable the neural network to capture the feature dependencies between different positions on the feature map, thereby deeply exploring the logical deduction relationship between strain amplitude, stress amplitude, maximum stress on the critical surface and fatigue life in the spatial dimension.

7. The metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network according to claim 1 is characterized in that: In step S52, in the feature processing module, after being processed by the two-dimensional convolution operator of the 1×1 convolution kernel, the features are respectively sent to three feature processing mechanisms of different scales; these mechanisms are responsible for extracting local features, global features and enhanced features, respectively, to fully capture the multi-scale and multi-depth features of the data, that is, to capture deep and shallow features from the global scale, local scale and feature enhancement scale; then, these features are fused in the channel dimension to integrate feature information of different scales and depths; the fused features are further processed by the feature attention mechanism composed of a linear layer and a Relu activation function, which purifies the features to enhance the model's ability to recognize key information; finally, through the global average pooling mechanism, the model outputs the fatigue life prediction value N of the metal material p , ensuring the stability and reliability of the prediction results.

8. The metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network according to claim 1 is characterized in that: In step S53, the physical loss function expression of the soft adjustment function is shown in formula (3): L=λ1L h +λ2L p (3) Where: L is the physical loss function; λ1 and λ2 are the weights of the two loss functions; L h is the Huber loss function; L p is the physical constraint function; L h and L p The expressions of are shown in formula (4) and formula (5): Where: N e is the fatigue life verified in the experiment; N p is the fatigue life predicted by the model after training; δ and d are adjustment parameters, which can adjust the degree of model training penalty by changing their values; The physical loss function controls the participation of physical constraints in neural network training by adjusting the values ​​of λ1 and λ2, thus realizing soft adjustment function.

9. The metal material fatigue life prediction method driven by a multi-scale and multi-depth neural network according to claim 1 is characterized in that: In step S6, the fatigue test data obtained in step S1, including the strain amplitude (ε a ,γ a ), stress amplitude (σ a ,τ a ), mean stress σ m , and the maximum normal stress value σ of the critical surface calculated in step S2 n,max , these data are used as input to the model; fatigue life N p is used as the output of the model; then, the input data is preprocessed according to the data processing method described in step S4; then, the model constructed in step S5 is trained and iteratively optimized to obtain the optimal multi-scale and multi-depth neural network driven metal material fatigue life prediction model; finally, the optimal multi-scale and multi-depth neural network driven metal material fatigue life prediction model is used to predict the fatigue life of the metal material under cyclic load; the prediction results are visualized through a scatter plot for easy analysis and understanding.

Images