Alloy fatigue life machine learning prediction method considering manufacturing defect characteristics

Abstract

The invention belongs to the technical field of material fatigue life prediction, and particularly relates to an alloy fatigue life machine learning prediction method considering manufacturing defect characteristics, which comprises the following steps: establishing a data set containing a plurality of alloys; normalizing the data set, and segmenting the data set into a training set, a test set and a verification set; fitting and checking the defect features, and sampling to obtain normalized defect feature samples conforming to distribution; physical constraints between the fatigue life-defect characteristics and the fatigue life-load are coded; designing a non-full-connection physical information neural network considering manufacturing defect features, and training the network; and performing life prediction on the data set by using a neural network. The method is simple, efficient and highly innovative, constructs the physical information neural network considering the manufacturing defect characteristics in a designable manner, forms a high-precision prediction method for the fatigue life, and forms a method for calculating the failure probability for the dispersity of material defects.

Description

Technical Field

[0001] This invention belongs to the field of material fatigue life prediction technology, specifically relating to a machine learning method for predicting alloy fatigue life that takes into account manufacturing defect characteristics. Background Technology

[0002] Modern aero-engines have partially adopted additively manufactured components such as compressors and turbine blades, offering advantages such as one-piece molding, simplified manufacturing, weight reduction, and shorter manufacturing cycles. However, the complex thermal history of additive manufacturing results in persistent microscopic defects that can easily become fatigue sources. The complex material deposition and solidification processes in additive manufacturing lead to significant randomness and dispersion of internal defects, constituting a key uncertainty factor affecting component performance. Therefore, incorporating the morphology and location of defects to accurately predict and assess fatigue life under alternating loads is fundamental to the use of additively manufactured and repaired components.

[0003] Considering the randomness of manufacturing defect characteristics, the failure probability of a group of additively manufactured test pieces in the same batch is an important characteristic. That is, under a given stress level and number of cycles, the probability of fatigue failure of the test piece can be measured. It can be quantitatively characterized by a probability mass function or a cumulative distribution function, providing a core basis for evaluating the reliability level of additively manufactured components and carrying out life prediction.

[0004] Current research on additive manufacturing defects focuses on defect detection, characterization, and quantitative description, but still lacks a quantitative correlation between defects and fatigue failure behavior. For the problem of fatigue failure caused by defects, the powerful designability and universality of deep neural network models provide a new approach to constructing fatigue life prediction methods that consider manufacturing defect characteristics. Developing machine learning models driven by both physical and data information, incorporating physical constraints, is an inevitable trend.

[0005] Chinese patent CN202411247910.X discloses a machine learning prediction method for the impact fatigue life of metals in additive manufacturing, combining traditional impact fatigue life prediction methods with productivity tools such as artificial intelligence and machine learning. However, the above method does not consider the impact of defects on the fatigue life of additively manufactured parts, and has significant limitations. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention starts with experimental data such as simple fatigue loads, defect characteristics, and material physical properties, encodes physical constraints, and constructs a physical information neural network that considers manufacturing defect characteristics to achieve high-precision prediction of fatigue life. Furthermore, considering the dispersion of material defects, a method for calculating failure probability is developed through statistical methods.

[0007] This invention is implemented by providing a machine learning method for predicting the fatigue life of alloys that takes into account manufacturing defect characteristics, comprising the following steps: S1: Establish a dataset containing multiple alloys. The dataset includes load levels, alloy manufacturing defect characteristics, material property parameters, and fatigue life parameters as model outputs. S2: Normalize the load levels, alloy manufacturing defect characteristics, material property parameters and fatigue life parameters in the dataset. Fit and test the statistical distribution of the normalized manufacturing defect characteristics. Obtain manufacturing defect characteristic samples that conform to the distribution by sampling using the Monte Carlo sampling method. S3: Encode the physical constraints between fatigue life-manufacturing defect characteristics and fatigue life-load level, including the negative correlation constraint between normalized stress and fatigue life and the constraint of the monotonic relationship between normalized manufacturing defect characteristics and fatigue life. S4: Design a non-fully connected physical information neural network that takes into account the characteristics of manufacturing defects. The neural network consists of a load level network block, a manufacturing defect feature network block, a material property network block, and a fatigue life prediction network block. S5: Split the dataset into training, testing, and validation sets, perform data augmentation on the training set, and train the neural network; S6: Apply the trained neural network to predict the lifespan of the dataset. Based on the statistical distribution of manufacturing defect features obtained from step S2, calculate the fatigue failure probability of a certain alloy under a specified load level, and complete the fatigue life prediction.

[0008] Preferably, in step S1, collecting datasets of various alloys from previous test and materials handbooks specifically includes: Load level related parameters: maximum stress S, initial stress intensity factor Strain ratio R, fracture strain ; Manufacturing defect feature related parameters: Manufacturing defect feature size D d = The manufacturing defect feature is located at a depth d from the surface, the ratio of the minor axis to the major axis of the fitted ellipse of the manufacturing defect feature is a / c, the ratio of the contour perimeter L1 to the ellipse perimeter L2 is L1 / L2, and the relative position of the manufacturing defect feature is h. Material property parameters: Yield strength YS, tensile strength , namely UTS, Young's modulus E; Fatigue life related parameters: Fatigue life cycle count N f .

[0009] Further preferably, in step S2, the load level related parameter S, The normalization method is as follows: (1) Wherein, the symbol without the top bar represents the parameter before normalization, and the symbol with the top bar represents the parameter after normalization, and the same applies below; Y is the stress factor, Y=0.65 when the manufacturing defect is located on the sample surface, otherwise Y=0.5, D g This represents the average grain size. Manufacturing defect characteristic related parameter D d The normalization methods for , d, and h are as follows: (2) Where r is the radius of the sample cross-section at the location of the manufacturing defect; Fatigue life related parameter N f The normalization method is as follows: (3).

[0010] Further preferably, in step S2, the normalized manufacturing defect characteristic related parameters are fitted with a generalized extreme value distribution using the maximum likelihood estimation method. The probability density function and cumulative distribution function of the generalized extreme value distribution are as follows: (4) (5) Where, x i (i=1, 2, 3, 4, 5) respectively take the relevant parameters of manufacturing defect characteristics , x i The location, size, and shape parameters of the generalized extremum distribution; The Kolmogorov-Smirnov test was performed on the generalized extreme value distribution fitted to the manufacturing defect characteristics to verify whether the data satisfies the assumed distribution. The Monte Carlo sampling method is used to perform inverse transform sampling on the generalized extreme value distribution of the manufacturing defect characteristic correlation parameters obtained by fitting, resulting in a large-capacity sample that conforms to the current distribution. The inverse function of the cumulative distribution function of the generalized extreme value distribution, i.e., the quantile function of the generalized extreme value distribution, is as follows: (6).

[0011] Further preferably, in step S3, the physical constraint inequality between fatigue life-manufacturing defect characteristics and fatigue life-load level is as follows: (7); Physical constraints correspond to the following: the greater the stress in physical cognition, the shorter the fatigue life; the larger the size of the manufacturing defect feature, the shorter the fatigue life; the farther the manufacturing defect feature is from the edge, the longer the fatigue life; and the greater the roundness of the manufacturing defect feature, the longer the fatigue life.

[0012] In a further preferred embodiment, in step S4, the non-fully connected physical information neural network selectively inputs the 12-dimensional model input parameters into the front-end load level network block, manufacturing defect feature network block, and material property network block, and outputs potential load level factors, manufacturing defect feature factors, and material property factors, respectively, forming a factor layer. The factor layer serves as the input to the back-end fatigue life prediction network block, thereby realizing alloy fatigue life prediction that takes into account manufacturing defect features.

[0013] In a further preferred embodiment, in step S5, the physical information neural network adds the physical constraint inequality from step S3 as a physical loss function term to the loss function of the neural network, forming a composite loss function together with the data loss function term. L ( i To ensure the physical consistency of the neural network: (8); in, L d ( i ) represents the data loss function term. L p ( i ) represents the physical loss function term. l These are the weighting coefficients to be optimized, used to adjust the relative weights of data loss and physical loss. L p ( i The specific form is as follows: (9); Where K represents the number of data points, and ReLU is the linear rectifier unit function. ; Data loss function term L d ( i The mean squared error loss function is used: (10); in, The model provides commonly used logarithmic fatigue life predictions. This is the actual value.

[0014] In a further preferred embodiment, in step S5, the model input parameter dataset and the model output parameter dataset are split, the validation set ratio is 0.2, five 5-fold cross-validation is used, the training set and test set are not strictly divided, the dataset excluding the test set is randomly divided into five parts, and one different part is selected each time as the training set, and the rest are used as the validation set. Dynamic data augmentation is performed on each randomly selected training set. Data augmentation employs two methods: adding Gaussian noise and random feature scaling. The Gaussian noise addition method is as follows: (11); Where x and y are the original input and output vectors of the model undergoing data augmentation, respectively. These are the input and output vectors after adding Gaussian noise, respectively. They are vectors of small variables, These are its components; Φ is a normal distribution. , where is the standard deviation of the normal distribution of Gaussian noise in the current round. p It is the ratio of the current round to the maximum number of rounds. Using a value of 0.005, perform the following random feature scaling every 3 rounds: (12); in, The input vector is scaled by features. U It is a uniform distribution, and s is the scaling factor.

[0015] Further preferred, in step S5, the data stream is forward-propagated during a single training iteration of the physical information neural network, and then the composite loss function of equation (8) is back-propagated. Adaptive momentum estimation is selected to iteratively update the network parameters. The number of hidden layers in each neural network block is determined by trial and error. Initially, Bayesian optimization algorithm is used on the training set to adjust the number of hidden layer nodes and the initial learning rate of hyperparameters. α Weight decay coefficient W d Adam optimizer first and second order moment decay coefficients β 1. β 2. Hidden layer drop rate keepdrop and physical weight coefficient l After 50 iterations, the optimal number of hidden layer nodes is fixed; then, the hyperparameters of the neural network are optimized on the training set using the Bayesian optimization algorithm, with the maximum number of iterations fixed at 200.

[0016] In a further preferred embodiment, in step S6, the sample obtained in step S2 is used as the input of the manufacturing defect feature network block. By fixing the required load level and material properties, the fatigue life corresponding to the sample can be calculated by the trained neural network model. Furthermore, in step S6, the Gaussian kernel density estimation method is used to obtain the probability mass function of the fatigue life of the sampled sample under the target load level and material properties. The cumulative distribution function (PMF) of fatigue life is obtained by numerically integrating the composite Simpson formula. The value corresponding to the PMF at a certain number of cycles is the fatigue failure probability after a specified number of cycles under the target load level and material properties.

[0017] The implementation of steps S2-S6 relies on, but is not limited to, the Python language, the Java language, or the commercial computing software Matlab.

[0018] Compared with the prior art, the advantages of the present invention are as follows: This invention starts with experimental data on simple fatigue loads, manufacturing defect characteristics, and material properties, encodes physical constraints, and constructs a physical information neural network that considers manufacturing defect characteristics, forming a high-precision prediction method for fatigue life. Furthermore, considering the dispersion of material defects, a statistical method is used to develop a method for calculating the failure probability. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the method provided by this invention applied to a nickel-based superalloy; Figure 2 This is a diagram of the non-fully connected physical information neural network architecture designed in step S4 of the present invention; Figure 3 The neural network trained according to this invention provides fatigue life prediction results considering manufacturing defect characteristics under simple fatigue loads, including training and testing sets. Figure 4 The cumulative failure curve of a certain nickel-based superalloy under a specified load is calculated by the neural network trained by this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer and easier to understand, the following detailed description, in conjunction with the accompanying drawings, uses the fatigue life and failure probability assessment of a nickel-based superalloy considering manufacturing defects as an example to further illustrate the specific embodiments of this invention. It should be noted that the specific embodiments described herein are for illustrative purposes only and not for limiting the scope of this invention.

[0021] This invention provides a machine learning method for predicting the fatigue life of alloys that takes into account manufacturing defect characteristics, such as... Figure 1 This is a simplified flowchart of a method for evaluating the fatigue life and failure probability of a nickel-based superalloy considering manufacturing defects, as described in an embodiment of the present invention. The alloy used in this embodiment is a nickel-based superalloy, and the specific steps are as follows: S1: Establish a dataset containing multiple alloys. The dataset includes load levels, alloy manufacturing defect characteristics, material property parameters, and fatigue life parameters as model outputs. S2: Normalize the load levels, alloy manufacturing defect characteristics, material property parameters and fatigue life parameters in the dataset. Fit and test the statistical distribution of the normalized manufacturing defect characteristics. Obtain manufacturing defect characteristic samples that conform to the distribution by sampling using the Monte Carlo sampling method. S3: Encode the physical constraints between fatigue life-manufacturing defect characteristics and fatigue life-load level, including the negative correlation constraint between normalized stress and fatigue life and the constraint of the monotonic relationship between normalized manufacturing defect characteristics and fatigue life. S4: Design a non-fully connected physical information neural network that takes into account the characteristics of manufacturing defects. The neural network consists of a load level network block, a manufacturing defect feature network block, a material property network block, and a fatigue life prediction network block. S5: Split the dataset into training, testing, and validation sets, perform data augmentation on the training set, and train the neural network; S6: Apply the trained neural network to predict the lifespan of the dataset. Based on the statistical distribution of manufacturing defect features obtained from step S2, calculate the fatigue failure probability of a certain alloy under a specified load level, and complete the fatigue life prediction.

[0022] In one specific embodiment, the alloy mentioned in step S1 is selected as a nickel-based high-temperature alloy.

[0023] In a specific implementation case, the fatigue life data of the material mentioned in step S1 comes from experimental data, publicly published literature at home and abroad, and a material handbook.

[0024] In a specific implementation case, when the dataset is split in step S5, the proportion of the test set is 0.2, and the training set and validation set are automatically split in 5 five-fold cross-validation.

[0025] In a specific implementation case, steps S2-S6 are implemented using the Python language.

[0026] To demonstrate the application of the technology of this invention more clearly, the invention applies material data in practice. Taking a selected nickel-based superalloy as an example, the above method is used to predict its lifespan and failure probability under simple fatigue loads, considering manufacturing defects.

[0027] S1. A dataset of 142 sets was collected, consisting of load conditions, defect characteristics, material properties, and fatigue life data from experiments, publicly published domestic and international literature, and a materials handbook. This dataset includes the maximum stress S and the initial stress intensity factor. Strain ratio R, fracture strain Defect feature size D d = Defect depth from surface d, ratio of minor to major axis of the defect fitted ellipse a / c, profile perimeter L1, ellipse perimeter L2, relative position of defect h, yield strength. (YS), tensile strength (UTS), Young's modulus E; and fatigue life cycle count. N f .

[0028] S2. Normalize a portion of the data in the dataset to obtain a 12-dimensional input and a 1-dimensional output. The symbols without a top bar represent the parameters before normalization, while those with a top bar represent the parameters after normalization. The normalization methods are shown in equations (1), (2), and (3). The defect-related parameters are fitted with a generalized extreme value distribution using the maximum likelihood estimation method. The generalized extreme value distribution is shown in equations (4) and (5). After passing the KS test, the Monte Carlo sampling results are obtained using the inverse transformation method. The inverse transformation of the generalized extreme value distribution is shown in equation (6).

[0029] S3. Based on the physical understanding of the stress-life and defect-life relationships, the physical constraints are encoded, and the physical constraints involved are as shown in equation (7).

[0030] S4. The constructed non-fully connected physical information neural network selectively inputs 12-dimensional input features into the front-end load, defect, and physical attribute network blocks, outputting potential load, defect, and physical attribute factors, respectively. The factor layer serves as the input to the back-end fatigue life prediction network block, achieving fatigue life prediction that considers manufacturing defect characteristics. The network architecture is as follows... Figure 2 The physical constraint inequality from step S3 is added as a physical loss penalty term to the loss function of the neural network, forming a composite loss function together with the data loss term. L ( i The physical loss function is shown in equation (8), and the total loss function is shown in equation (9).

[0031] S5. In step S5, the processed input-output dataset is split, and the validation set ratio is 0.2. Five 5-fold cross-validation is used, so the training set and test set are not strictly divided. Instead, the dataset excluding the test set is randomly divided into five parts. Each time, one different part is selected as the training set, and the rest are used as the validation set. Dynamic data augmentation is performed on the randomly selected training set each time to increase the size of the training set. Data augmentation uses two methods: adding Gaussian noise and random feature scaling. The method of adding Gaussian noise is shown in Equation (11), and the method of random feature scaling is shown in Equation (12). In the single training of the physical information neural network, the data flow is propagated forward, and then the composite loss function term of Equation (8) is propagated backward. The adaptive momentum estimation (Adam optimizer) is selected to iteratively update the network parameters. The number of hidden layers of each neural network block is determined by trial and error. Initially, the Bayesian optimization algorithm is used on the training set to adjust the number of hidden layer nodes and other hyperparameters, including the initial learning rate. α Weight decay coefficient W d Adam optimizer first and second order moment decay coefficients β 1. β 2. Hidden layer drop rate keepdrop and physical weight coefficient l After 50 iterations, the optimal number of hidden layer nodes was fixed. Then, the hyperparameters of the neural network were optimized on the training set using a Bayesian optimization algorithm, with the maximum number of iterations fixed at 200. The final results were: load network block with 3 hidden layers and 12, 18, and 10 nodes; defect feature network block with 3 hidden layers and 14, 11, and 5 nodes; physical property network block with 3 hidden layers and 14, 16, and 4 nodes; and fatigue life network block with 3 hidden layers and 4, 13, and 6 nodes. α =0.057997, W d =0.014498, β 1 = 0.966964 β 2 = 0.992311, keepdrop =2.658583e-5, λ= 0.012531.

[0032] S6. Perform lifetime prediction based on the trained physical information neural network. The cross-training and test set results of the network are as follows: Figure 3 As can be seen, all data from training and cross-validation fall within the 3-strip width, while almost all data from testing fall within the 2-strip width. The Monte Carlo sampling samples obtained in step S2 are used as input to the defect feature network block. In the current implementation, the material properties are YS=1078MPa, UTS=1200MPa, and E=210GPa. In the current implementation, S=970.2MPa (at this time...). All conditions are input into a physical information neural network. For the output, the density mass function of logarithmic fatigue life is obtained using the Gaussian kernel density estimation method. The cumulative distribution function of logarithmic fatigue life, i.e., the cumulative failure curve with respect to the number of cycles under the current conditions, is obtained by numerical integration using the composite Simpson formula. Figure 4 .

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of application of the present invention. Various adjustments can be made to the above embodiments of the present invention. Therefore, all simple and equivalent changes and modifications made based on the claims and description of this application are within the scope of protection of the claims of this patent.

Claims

1. A machine learning method for predicting the fatigue life of alloys considering manufacturing defect characteristics, characterized in that, Includes the following steps: S1: Establish a dataset containing multiple alloys. The dataset includes load levels, alloy manufacturing defect characteristics, material property parameters, and fatigue life parameters as model outputs. S2: Normalize the load levels, alloy manufacturing defect characteristics, material property parameters and fatigue life parameters in the dataset. Fit and test the statistical distribution of the normalized manufacturing defect characteristics. Obtain manufacturing defect characteristic samples that conform to the distribution by sampling using the Monte Carlo sampling method. S3: Encode the physical constraints between fatigue life-manufacturing defect characteristics and fatigue life-load level, including the negative correlation constraint between normalized stress and fatigue life and the constraint of the monotonic relationship between normalized manufacturing defect characteristics and fatigue life. S4: Design a non-fully connected physical information neural network that takes into account the characteristics of manufacturing defects. The neural network consists of a load level network block, a manufacturing defect feature network block, a material property network block, and a fatigue life prediction network block. S5: Split the dataset into training, testing, and validation sets, perform data augmentation on the training set, and train the neural network; S6: Apply the trained neural network to predict the lifespan of the dataset. Based on the statistical distribution of manufacturing defect features obtained from step S2, calculate the fatigue failure probability of a certain alloy under a specified load level, and complete the fatigue life prediction.

2. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 1, characterized in that, In step S1, data sets of various alloys are collected from previous experiments and materials handbooks, specifically including: Load level related parameters: maximum stress S, initial stress intensity factor Strain ratio R, fracture strain ; Manufacturing defect feature related parameters: Manufacturing defect feature size D d = The manufacturing defect feature is located at a depth d from the surface, the ratio of the minor axis to the major axis of the fitted ellipse of the manufacturing defect feature is a / c, the ratio of the contour perimeter L1 to the ellipse perimeter L2 is L1 / L2, and the relative position of the manufacturing defect feature is h. Material property parameters: Yield strength YS, tensile strength , namely UTS, Young's modulus E; Fatigue life related parameters: Fatigue life cycle count N f .

3. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 2, characterized in that, In step S2, the normalization method for the load level related parameters S is as follows: （1） Wherein, the symbol without the top bar represents the parameter before normalization, and the symbol with the top bar represents the parameter after normalization, and the same applies below; Y is the stress factor, Y=0.65 when the manufacturing defect is located on the sample surface, otherwise Y=0.5, D g This represents the average grain size. Manufacturing defect characteristic related parameter D d The normalization methods for , d, and h are as follows: （2） Where r is the radius of the sample cross-section at the location of the manufacturing defect; Fatigue life related parameter N f The normalization method is as follows: （3）。 4. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 3, characterized in that, In step S2, the normalized manufacturing defect characteristic related parameters are fitted with a generalized extreme value distribution using the maximum likelihood estimation method. The probability density function and cumulative distribution function of the generalized extreme value distribution are as follows: （4） （5） Where, x i (i=1, 2, 3, 4, 5) respectively take the relevant parameters of manufacturing defect characteristics , , , , , x i The location, size, and shape parameters of the generalized extremum distribution; The Kolmogorov-Smirnov test was performed on the generalized extreme value distribution fitted to the manufacturing defect characteristics to verify whether the data satisfies the assumed distribution. The Monte Carlo sampling method is used to perform inverse transform sampling on the generalized extreme value distribution of the manufacturing defect characteristic correlation parameters obtained by fitting, resulting in a large-capacity sample that conforms to the current distribution. The inverse function of the cumulative distribution function of the generalized extreme value distribution, i.e., the quantile function of the generalized extreme value distribution, is as follows: （6）。 5. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 3, characterized in that, In step S3, the physical constraint inequality between fatigue life-manufacturing defect characteristics and fatigue life-load level is as follows: （7）； Physical constraints correspond to stresses in physical cognition, which in turn lead to shorter fatigue life. The larger the size of the manufacturing defect feature, the shorter the fatigue life; the farther the manufacturing defect feature is from the edge, the longer the fatigue life. The greater the roundness of the manufacturing defect feature, the greater the fatigue life.

6. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 3, characterized in that, In step S4, the non-fully connected physical information neural network selectively inputs the 12-dimensional model input parameters into the front-end load level network block, manufacturing defect feature network block, and material property network block, and outputs potential load level factors, manufacturing defect feature factors, and material property factors, respectively, forming a factor layer. The factor layer serves as the input to the back-end fatigue life prediction network block, thereby realizing alloy fatigue life prediction that takes into account manufacturing defect features.

7. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 5, characterized in that, In step S5, the physical information neural network adds the physical constraint inequality from step S3 as a physical loss function term to the neural network's loss function, forming a composite loss function together with the data loss function term. L ( θ To ensure the physical consistency of the neural network: （8）； in, L d ( θ ) represents the data loss function term. L p ( θ ) represents the physical loss function term. λ These are the weighting coefficients to be optimized, used to adjust the relative weights of data loss and physical loss. L p ( θ The specific form is as follows: （9）； Where K represents the number of data points, and ReLU is the linear rectifier unit function. ; Data loss function term L d ( θ The mean squared error loss function is used: （10）； in, The model provides commonly used logarithmic fatigue life predictions. This is the actual value.

8. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 1, characterized in that, In step S5, the model input parameter dataset and the model output parameter dataset are split, the validation set ratio is 0.2, and 5-fold cross-validation is used. The training set and test set are not strictly divided. The dataset excluding the test set is randomly divided into 5 parts. Each time, a different part is selected as the training set and the rest are used as the validation set. Dynamic data augmentation is performed on each randomly selected training set. Data augmentation employs two methods: adding Gaussian noise and random feature scaling. The Gaussian noise addition method is as follows: （11）； Where x and y are the original input and output vectors of the model undergoing data augmentation, respectively. These are the input and output vectors after adding Gaussian noise, respectively. They are vectors of small variables, These are its components; Φ is a normal distribution. , where is the standard deviation of the normal distribution of Gaussian noise in the current round. p It is the ratio of the current round to the maximum number of rounds. Using a value of 0.005, perform the following random feature scaling every 3 rounds: （12）； in, The input vector is scaled by features. U It is a uniform distribution, and s is the scaling factor.

9. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 7, characterized in that, In step S5, the data stream is propagated forward during a single training iteration of the physical information neural network, and then the composite loss function of equation (8) is propagated backward. Adaptive momentum estimation is selected to iteratively update the network parameters. The number of hidden layers in each neural network block is determined by trial and error. Initially, Bayesian optimization algorithm is used on the training set to adjust the number of hidden layer nodes and the initial learning rate of hyperparameters. α Weight decay coefficient W d Adam optimizer first and second order moment decay coefficients β 1. β 2. Hidden layer drop rate keepdrop and physical weight coefficient λ After 50 iterations, the optimal number of hidden layer nodes is fixed; then, the hyperparameters of the neural network are optimized on the training set using the Bayesian optimization algorithm, with the maximum number of iterations fixed at 200.

10. The machine learning prediction method for alloy fatigue life considering manufacturing defect characteristics according to claim 1, characterized in that, In step S6, the sample obtained in step S2 is used as the input of the manufacturing defect feature network block. By fixing the required load level and material properties, the fatigue life corresponding to the sample can be calculated by the trained neural network model. Furthermore, in step S6, the Gaussian kernel density estimation method is used to obtain the probability mass function of the fatigue life of the sampled sample under the target load level and material properties. The cumulative distribution function (PMF) of fatigue life is obtained by numerically integrating the composite Simpson formula. The value corresponding to the PMF at a certain number of cycles is the fatigue failure probability after a specified number of cycles under the target load level and material properties.

Images