Fatigue crack life prediction method and system based on physical information neural network

Abstract

The application provides a fatigue crack life prediction method and system based on a physical information neural network, comprising: a data acquisition step: through finite element numerical simulation, crack length and fatigue life expansion data under different working conditions are acquired; a data set construction step: a data set for training a neural network is constructed according to the expansion data; a neural network construction step: according to the data set, a physical information neural network is constructed and trained to obtain a physical information neural network model; a life prediction step: inputting working condition parameters to be measured into the physical information neural network model, fatigue crack prediction results are obtained and outputted. The application solves the problems of long time consumption of finite element calculation and data dependency of a data-driven model, and realizes accurate prediction of fatigue crack expansion residual life under a small sample.

Description

Technical Field

[0001] This invention belongs to the field of structural health detection technology, specifically relating to a fatigue crack life prediction method and system based on physical information neural networks. Background Technology

[0002] Predicting the remaining fatigue crack propagation life is crucial for the safety of large structural components such as aircraft, rockets, and bridges. Fatigue fracture is one of the main forms of failure in engineering structures, often characterized by its suddenness and insidious nature. Once it occurs, it can lead to catastrophic accidents and cause huge economic losses.

[0003] Therefore, accurately assessing the remaining lifespan of a structure during its service life is not only a core requirement for implementing the "damage tolerance" design concept, but also a key technical means to enable condition-based structural maintenance, establish reasonable inspection cycles, and ensure safe operation throughout the entire life cycle.

[0004] However, existing prediction methods still face many challenges in practical engineering applications. Traditional fracture mechanics-based physical models, while possessing clear physical meaning, often rely on costly finite element calculations, making it difficult to quickly predict crack propagation life. On the other hand, emerging pure data-driven deep learning methods, while possessing powerful feature extraction capabilities, heavily depend on massive amounts of experimental labeled data and suffer from model "black box" characteristics and poor generalization ability.

[0005] In fields such as aerospace, it is extremely difficult to obtain high-cost, long-cycle full-life fatigue test data. This "small sample" dilemma severely restricts the application of traditional deep learning methods.

[0006] This problem urgently needs to be solved. Summary of the Invention

[0007] In view of the deficiencies in the prior art, the purpose of this invention is to provide a fatigue crack life prediction method and system based on physical information neural networks.

[0008] A fatigue crack life prediction method based on a physical information neural network according to the present invention includes: Data acquisition steps: Obtain extended data on crack length and fatigue life under different working conditions through finite element numerical simulation; Dataset construction steps: Construct a dataset for training the neural network based on the expanded data; Neural network construction steps: Based on the dataset, construct and train the physical information neural network to obtain the physical information neural network model; Life prediction steps: Input the parameters of the working condition to be tested into the physical information neural network model, and obtain and output the fatigue crack prediction results.

[0009] Preferably, in the data acquisition step, ABAQUS-FRANC3D co-simulation is used to obtain extended data on crack length and fatigue life under different working conditions; The ABAQUS-FRANC3D co-simulation includes: Step A1: Call ABAQUS to obtain the current stress and displacement fields; Step A2: Call FRANC3D to read and obtain the stress intensity factor based on the stress and displacement field; Step A3: Based on the stress intensity factor, calculate the crack propagation data using the Paris formula, and repeat step A1 until the crack reaches the preset critical size.

[0010] Preferably, the dataset construction step includes: Step B1: Normalize the expanded data to obtain the normalized result; Step B2: According to a preset ratio, the preprocessed results are sequentially divided to obtain the training set and the test set; The preset ratio is as follows: select the first 20% of the normalized crack propagation data as the training set; select the remaining 80% of the data as the test set.

[0011] Preferably, the neural network construction step includes: Step C1: Based on the training set, construct a total loss function that includes physical information loss; Step C2: Based on the total loss function, use the fully connected neural network as the physical information neural network model; Step C3: Update the weights and bias parameters of the neural network through backpropagation until the physical information neural network model converges, thus obtaining the trained physical information neural network model. In step C1, the expression for the total loss function is:

[0012] in, Represents the total loss function. Indicates data loss items, and Both are balance coefficients; Indicates physical residual loss. This represents the loss due to monotonicity constraints. The expression for the data loss term is:

[0013] in, No. The true lifespan of a single neuron Indicates the predicted first The fatigue life of a single neuron; This represents the total number of training samples; The expression for the physical residual loss is:

[0014] in, Represents a logarithmic function; Indicates the crack length; Indicates the predicted fatigue life; and These are all empirical constants; Indicates the geometric correction factor; Represents the differential symbol. Indicates the first Each crack length; Indicates stress amplitude; The expression for the monotonicity constraint loss is:

[0015] Among them, subscript Represents the first in the dataset One training sample; This indicates taking the maximum value.

[0016] Preferably, in step C3, the weights and bias parameters of the neural network are updated through backpropagation, i.e., a stochastic gradient descent algorithm is used to determine and update the optimal neural network weights and bias parameters (which are physical parameters) by iteratively minimizing the loss function; wherein, for the physical information neural network model, the weights and bias parameters are updated by backpropagation. The first in the layer The _ neuron, the _ ... The expressions for calculating the weights and bias parameters in each connection, from top to bottom, are as follows:

[0017] in, Indicates a time step. Indicates the learning rate. Indicates the momentum coefficient; This represents the error term in backpropagation; and They represent the first Layer connection weights and bias terms; Indicates the increment symbol; The symbol represents the partial derivative.

[0018] A fatigue crack life prediction system based on a physical information neural network according to the present invention includes: Data acquisition module: acquires extended data on crack length and fatigue life under different working conditions through finite element numerical simulation; Dataset building module: Builds a dataset for training neural networks based on expanded data; Neural network construction module: Based on the dataset, construct and train the physical information neural network to obtain the physical information neural network model; Life prediction module: Input the parameters of the working condition to be tested into the physical information neural network model, obtain and output fatigue crack prediction results.

[0019] Preferably, in the data acquisition module, ABAQUS-FRANC3D co-simulation is used to acquire extended data on crack length and fatigue life under different working conditions; The ABAQUS-FRANC3D co-simulation includes: Module A1: Calls ABAQUS to obtain the current stress and displacement fields; Module A2: Calls FRANC3D to read and obtain the stress intensity factor based on the stress and displacement field; Module A3: Based on the stress intensity factor, calculate the crack propagation data using the Paris formula, and re-trigger module A1 until the crack reaches the preset critical size.

[0020] Preferably, the dataset construction module includes: Module B1: Normalizes the extended data to obtain the normalized result; Module B2: According to a preset ratio, the preprocessed results are sequentially divided to obtain a training set and a test set; The preset ratio is as follows: select the first 20% of the normalized crack propagation data as the training set; select the remaining 80% of the data as the test set.

[0021] Preferably, the neural network building module includes: Module C1: Based on the training set, construct a total loss function that includes physical information loss; Module C2: Based on the total loss function, the fully connected neural network is used as the physical information neural network model; Module C3: Updates the weights and bias parameters of the neural network through backpropagation until the physical information neural network model converges, thus obtaining the trained physical information neural network model; In module C1, the expression for the total loss function is:

[0022] in, Represents the total loss function. Indicates data loss items, and Both are balance coefficients; Indicates physical residual loss. This represents the loss due to monotonicity constraints. The expression for the data loss term is:

[0023] in, No. The true lifespan of a single neuron Indicates the predicted first The fatigue life of a single neuron; This represents the total number of training samples; The expression for the physical residual loss is:

[0024] in, Represents a logarithmic function; Indicates the crack length; Indicates the predicted fatigue life; and These are all empirical constants; Indicates the geometric correction factor; Represents the differential symbol. Indicates the first Each crack length; Indicates stress amplitude; The expression for the monotonicity constraint loss is:

[0025] Among them, subscript Represents the first in the dataset One training sample; This indicates taking the maximum value.

[0026] Preferably, in module C3, the weights and bias parameters of the neural network are updated through backpropagation. Specifically, a stochastic gradient descent algorithm is used to iteratively minimize the loss function to determine and update the optimal weights and bias parameters, which are physical parameters. For the physical information neural network model, the weights and bias parameters are... The first in the layer The _ neuron, the _ ... The expressions for calculating the weights and bias parameters in each connection, from top to bottom, are as follows:

[0027] in, Indicates a time step. Indicates the learning rate. Indicates the momentum coefficient; This represents the error term in backpropagation; and They represent the first Layer connection weights and bias terms; Indicates the increment symbol; The symbol represents the partial derivative.

[0028] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention only needs to utilize the early stage of the entire life cycle, i.e. the first 20% of the data, to achieve high-precision extrapolation of the remaining 80% of the life cycle, thereby improving data efficiency and solving the problem of difficulty in obtaining long-term trial data in engineering.

[0029] 2. This invention aims to provide a novel physical information neural network that solves the problems of long calculation time in finite element method and data dependence of data-driven models. It achieves accurate prediction of fatigue crack propagation remaining life under small sample conditions and improves the solution efficiency. Compared with the traditional finite element method, the proposed physical information neural network can improve the solution speed and greatly shorten the calculation time.

[0030] 3. This invention does not require prior knowledge of the accurate geometric correction coefficients. The model can automatically solve for these parameters during training with an error of less than 10%, demonstrating physical interpretability and realizing the reverse solution of geometric parameters. Attached Figure Description

[0031] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A schematic block diagram illustrating the method for obtaining the lifetime of complex structures using the finite element method provided by this invention. Figure 2 A schematic block diagram of the physical information neural network structure provided by the present invention; Figure 3 A schematic diagram of the physical information neural network prediction results provided by this invention; Figure 4 A schematic diagram illustrating the prediction accuracy of the physical information neural network provided by this invention. Detailed Implementation

[0032] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0033] This invention mainly solves the problems of excessive data sample requirements and excessive time consumption in the existing pure data-driven lifetime prediction process.

[0034] Specifically, this invention proposes a method for embedding physical information into a neural network, which integrates the classical mathematical model in fracture mechanics into a deep learning network, thereby achieving the interpretability of the neural network and the ability to accurately predict fatigue life under small sample conditions. That is, it can efficiently predict the remaining life in the early stage of crack propagation, i.e., the first 20%.

[0035] This invention obtains the fatigue crack propagation life under different working conditions through finite element analysis. After preprocessing the data, a training dataset is constructed. A nonlinear mapping relationship between crack length a and fatigue life N is established through the proposed novel physical information neural network, thereby achieving accurate prediction of fatigue crack propagation life and identification of geometric parameters.

[0036] This invention discloses a fatigue crack life prediction method based on a physical information neural network, mainly used for accurately predicting the crack propagation life of complex structures in the early stages of crack development. The purpose of this invention is to provide an efficient, fast, and reliable solution, thereby addressing the problems of time consumption and low accuracy in existing methods.

[0037] To achieve the above objectives, the technical solution of the present invention is: A fatigue crack life prediction method based on physical information neural network is proposed. This method introduces physical loss into the loss function of machine learning and inputs geometric parameters as trainable variables into the machine learning framework. This enables fast and accurate fatigue life prediction even when data is scarce and true values ​​of geometric parameters are lacking.

[0038] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0039] S1, Crack length and life data acquisition.

[0040] First, taking a rocket propellant tank as an example, a crack propagation dataset is constructed using the ABAQUS-FRANC3D numerical calculation method.

[0041] Specifically, adopting a method based on The M-integral method, which calculates stress intensity factors in combined Type I, Type II, and Type III crack modes, can precisely decouple crack components from the global energy release rate. .

[0042] The The expression for the integral is shown in Equation 1:

[0043] in, Let M represent the M integral of the real state and the auxiliary state, respectively; This represents a closed integral path along the perimeter of the crack tip; and These represent the stress tensor components of the actual state and the auxiliary state, respectively; Represents the partial derivative operator; and These represent the coordinate components in the Cartesian coordinate system; and The displacement vector components representing the actual state and the auxiliary state; Represents the interaction strain energy density; Represents the Kronecker function; Represents the weighting function; ds The infinitesimal element representing the integration path.

[0044] also, The integral can also be written in terms of the stress intensity factor. The expression:

[0045] in, and These represent Young's modulus and Poisson's ratio of the material, respectively. , , and , , The results represent the calculated stress intensity factors under three crack conditions, based on the finite element method and the asymptotic auxiliary field used to separate the hybrid modes.

[0046] This invention uses the maximum tensile stress rule to calculate the direction of crack propagation, and the calculation criteria are shown in equation (3-5):

[0047] in, Indicates circumferential stress; Represents the polar radius in polar coordinates, and represents the radial distance from the crack tip; Indicates the polar angle; and These represent the type I and type II stress intensity factors, respectively. Represents the shear stress in polar coordinates; Figure 1 This invention demonstrates the computational flow of co-simulation using ABAQUS and FRANC3D. The ABAQUS solver is used to calculate the current stress and displacement fields, which are then read and processed by FRANC3D. The stress intensity factor is solved using an integral method, and then the crack propagation is calculated according to the Paris formula, and the mesh is re-generated. This iterative process continues until the crack reaches the critical size.

[0048] S2, Dataset Construction: To ensure the diversity and broad representativeness of the experimental data, multiple sets of working conditions with different crack shape ratios and stress ratios were constructed, and the corresponding crack propagation data were systematically analyzed. Considering that the weld is usually a weak area in the mechanical properties of welded structures, this paper fixes the initial crack location of all computational models at the weld mid-surface to study the fatigue crack propagation behavior in this region.

[0049] The input and output data were normalized to eliminate dimensionality differences and reduce the impact of extreme and outlier values, thereby maintaining model stability. This paper employs a max-min normalization method to map the input and output data to the dimensionless interval [0,1]. The normalization formula is defined as follows:

[0050] in, Represents the normalized data. , These represent the maximum and minimum values ​​of the data, respectively. Represents data.

[0051] To further evaluate the model's extrapolation and prediction capabilities under small sample conditions, this study designed a specific segmentation strategy based on the crack propagation stage. Unlike conventional random sampling, data from the entire crack propagation process are segmented according to crack depth. Temporal segmentation was performed. Specifically, the first 20% of the data in the early stage of crack propagation was selected as the training set to capture the physical laws of crack growth; while the remaining 80% of the later crack propagation data was used as the test set.

[0052] S3, Construction of the Physical Information Neural Network: First, a loss function containing physical information is constructed. Stress intensity factor. While it can effectively characterize the stress field intensity at the crack tip, under fatigue loading, the core parameter controlling crack propagation behavior is not a single-moment parameter. It is not the value, but the range of the stress intensity factor. During the cyclic loading process, Defined as maximum stress intensity factor max With minimum stress intensity factor min The difference. Within the framework of linear elastic fracture mechanics, the crack propagation rate / and Follow the Paris formula.

[0053]

[0054] In the formula, / This represents the crack propagation increment in a single cycle. and These are empirical constants related to material properties and the environment. Existing researchers have provided analytical solutions for the stress intensity factor of some common geometries, but for complex structures such as storage tanks, the finite element method is still needed for calculation. .

[0055] Generally, the range of stress intensity factor is shown in the following formula:

[0056] in, It is the geometry correction factor, a function that constructs the geometry and crack size. Therefore, the Paris formula can be written as:

[0057] The physical information loss function can be obtained from equation (8). :

[0058] Based on the principle of irreversible fatigue damage evolution, cumulative lifespan Must vary with crack length Monotonically increasing. To eliminate non-physical oscillation predictions, a monotonicity penalty term based on prior physical knowledge is introduced:

[0059] A data loss term based on mean squared error is established to ensure that the neural network can accurately reproduce the observed experimental data. This term quantifies the difference between the predicted fatigue life and the actual fatigue life on the training set.

[0060] in, That's the actual lifespan. It is the predicted fatigue life.

[0061] The loss function used in this invention is a composite loss function, as shown in equation (11).

[0062]

[0063] The total loss function is composed of three weighted parts: data-driven loss. Physical residual loss and monotonicity constraint loss . =0.1 and =0.01 is the balancing coefficient, used to adjust the contribution weight of different constraint terms in the gradient descent process.

[0064] After determining the loss function, a fully connected neural network was chosen as the primary neural network model. Nonlinear regression was performed through forward propagation, and gradient descent was achieved using backpropagation of the error, thereby reducing fatigue life prediction errors. Forward propagation is the fundamental process for information processing and nonlinear mapping in artificial neural networks. Mathematically, assuming the network's... The output vector of the layer is Then the first The output of the layer is:

[0065] in, Indicates the first The output of the layer, For the first The weight matrix of the layer, For bias vectors, For activation function; symbol This represents matrix multiplication.

[0066] In this model, GELU is used as the activation function in the hidden layers, while a linear activation function is used in the output layer. . This represents the slope of the linear activation function.

[0067] The mathematical expression for the GELU function is:

[0068] In the formula, The cumulative distribution function represents the standard normal distribution; Represents the integral symbol.

[0069] This invention employs a stochastic gradient descent algorithm to determine the optimal network weights and physical parameters by iteratively minimizing the loss function. For the... The first in the layer The _ neuron, the _ ... The adaptive calculation of weights and biases in each connection can be expressed as:

[0070] in, It is a time step. For learning rate, It is the momentum coefficient. This represents the error term in backpropagation.

[0071] S4, Training of the physical information neural network: Input the data from steps S1 and S2 into the network, update the weights and bias parameters of the neural network through backpropagation until the model converges, and obtain the trained physical information neural network model.

[0072] S5, Remaining fatigue crack propagation life prediction and assessment: Input the load parameters of the test condition and the current number of cycles into the trained model, predict the crack length change curve with the number of cycles, and calculate the number of cycles when the crack reaches the critical value by setting the critical crack length a, and then obtain the remaining fatigue life.

[0073] To quantitatively evaluate the fitting accuracy and generalization performance of the proposed PINN model in fatigue life prediction, especially its extrapolation ability for the last 80% of the data, this study uses the root mean square error (RMSE). Mean Absolute Percentage Error (MAE) , as the core evaluation indicator.

[0074] It is used to measure the degree of global deviation between predicted and actual values, and is particularly sensitive to outliers with large amplitude values. Its calculation formula is as follows:

[0075] Considering that fatigue life data typically spans multiple orders of magnitude, simple absolute error is insufficient to fully reflect the model's balanced performance across low and high lifespan ranges. Therefore, an additional error is introduced... To quantify the relative deviation of the prediction results. It has scale invariance, which can more intuitively reflect the percentage accuracy of the model's predictions:

[0076] Appendix Figure 3 and 4 This paper demonstrates a comparison between the proposed physical information neural network and traditional data-driven methods, showing that the proposed model achieves better prediction accuracy and lower latency. and It can accurately predict the remaining fatigue life.

[0077] The present invention also provides a fatigue crack life prediction system based on physical information neural network. The fatigue crack life prediction system based on physical information neural network can be implemented by executing the process steps of the fatigue crack life prediction method based on physical information neural network. That is, those skilled in the art can understand the fatigue crack life prediction method based on physical information neural network as a preferred embodiment of the fatigue crack life prediction system based on physical information neural network.

[0078] A fatigue crack life prediction system based on a physical information neural network according to the present invention includes: Data acquisition module: acquires extended data on crack length and fatigue life under different working conditions through finite element numerical simulation; Dataset building module: Builds a dataset for training neural networks based on expanded data; Neural network construction module: Based on the dataset, construct and train the physical information neural network to obtain the physical information neural network model; Life prediction module: Input the parameters of the working condition to be tested into the physical information neural network model, obtain and output fatigue crack prediction results.

[0079] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0080] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A fatigue crack life prediction method based on a physical information neural network, characterized in that, include: Data acquisition steps: Obtain extended data on crack length and fatigue life under different working conditions through finite element numerical simulation; Dataset construction steps: Construct a dataset for training the neural network based on the expanded data; Neural network construction steps: Based on the dataset, construct and train the physical information neural network to obtain the physical information neural network model; Life prediction steps: Input the parameters of the working condition to be tested into the physical information neural network model, and obtain and output the fatigue crack prediction results.

2. The fatigue crack life prediction method based on physical information neural network according to claim 1, characterized in that, In the data acquisition step, ABAQUS-FRANC3D co-simulation is used to obtain extended data on crack length and fatigue life under different working conditions; The ABAQUS-FRANC3D co-simulation includes: Step A1: Call ABAQUS to obtain the current stress and displacement fields; Step A2: Call FRANC3D to read and obtain the stress intensity factor based on the stress and displacement field; Step A3: Based on the stress intensity factor, calculate the crack propagation data using the Paris formula, and repeat step A1 until the crack reaches the preset critical size.

3. The fatigue crack life prediction method based on physical information neural network according to claim 1, characterized in that, The dataset construction step includes: Step B1: Normalize the expanded data to obtain the normalized result; Step B2: According to a preset ratio, the preprocessed results are sequentially divided to obtain the training set and the test set; The preset ratio is as follows: select the first 20% of the normalized crack propagation data as the training set; select the remaining 80% of the data as the test set.

4. The fatigue crack life prediction method based on physical information neural network according to claim 3, characterized in that, The neural network construction step includes: Step C1: Based on the training set, construct a total loss function that includes physical information loss; Step C2: Based on the total loss function, use the fully connected neural network as the physical information neural network model; Step C3: Update the weights and bias parameters of the neural network through backpropagation until the physical information neural network model converges, thus obtaining the trained physical information neural network model. In step C1, the expression for the total loss function is: in, Represents the total loss function. Indicates data loss items, and Both are balance coefficients; Indicates physical residual loss. This represents the loss due to monotonicity constraints. The expression for the data loss term is: in, No. The true lifespan of a single neuron Indicates the predicted first The fatigue life of a single neuron; This represents the total number of training samples; The expression for the physical residual loss is: in, Represents a logarithmic function; Indicates the crack length; Indicates the predicted fatigue life; and These are all empirical constants; Indicates the geometric correction factor; Represents the differential symbol. Indicates the first Each crack length; Indicates stress amplitude; The expression for the monotonicity constraint loss is: Among them, subscript Represents the first in the dataset One training sample; This indicates taking the maximum value.

5. The fatigue crack life prediction method based on physical information neural network according to claim 4, characterized in that, In step C3, the weights and bias parameters of the neural network are updated through backpropagation. Specifically, a stochastic gradient descent algorithm is used to iteratively minimize the loss function to determine and update the optimal weights and bias parameters, which are physical parameters. For the physical information neural network model, the weights and bias parameters are... The first in the layer The _ neuron, the _ ... The expressions for calculating the weights and bias parameters in each connection, from top to bottom, are as follows: in, Indicates a time step. Indicates the learning rate. Indicates the momentum coefficient; This represents the error term in backpropagation; and They represent the first Layer connection weights and bias terms; Indicates the increment symbol; The symbol represents the partial derivative.

6. A fatigue crack life prediction system based on a physical information neural network, characterized in that, include: Data acquisition module: acquires extended data on crack length and fatigue life under different working conditions through finite element numerical simulation; Dataset building module: Builds a dataset for training neural networks based on expanded data; Neural network construction module: Based on the dataset, construct and train the physical information neural network to obtain the physical information neural network model; Life prediction module: Input the parameters of the working condition to be tested into the physical information neural network model, obtain and output fatigue crack prediction results.

7. The fatigue crack life prediction system based on physical information neural network according to claim 6, characterized in that, In the data acquisition module, ABAQUS-FRANC3D co-simulation is used to obtain extended data on crack length and fatigue life under different working conditions; The ABAQUS-FRANC3D co-simulation includes: Module A1: Calls ABAQUS to obtain the current stress and displacement fields; Module A2: Calls FRANC3D to read and obtain the stress intensity factor based on the stress and displacement field; Module A3: Based on the stress intensity factor, calculate the crack propagation data using the Paris formula, and re-trigger module A1 until the crack reaches the preset critical size.

8. The fatigue crack life prediction system based on physical information neural network according to claim 6, characterized in that, The dataset construction module includes: Module B1: Normalizes the extended data to obtain the normalized result; Module B2: According to a preset ratio, the preprocessed results are sequentially divided to obtain a training set and a test set; The preset ratio is as follows: select the first 20% of the normalized crack propagation data as the training set; select the remaining 80% of the data as the test set.

9. The fatigue crack life prediction system based on physical information neural network according to claim 8, characterized in that, The neural network building module includes: Module C1: Based on the training set, construct a total loss function that includes physical information loss; Module C2: Based on the total loss function, the fully connected neural network is used as the physical information neural network model; Module C3: Updates the weights and bias parameters of the neural network through backpropagation until the physical information neural network model converges, thus obtaining the trained physical information neural network model; In module C1, the expression for the total loss function is: in, Represents the total loss function. Indicates data loss items, and Both are balance coefficients; Indicates physical residual loss. This represents the loss due to monotonicity constraints. The expression for the data loss term is: in, No. The true lifespan of a single neuron Indicates the predicted first The fatigue life of a single neuron; This represents the total number of training samples; The expression for the physical residual loss is: in, Represents a logarithmic function; Indicates the crack length; Indicates the predicted fatigue life; and These are all empirical constants; Indicates the geometric correction factor; Represents the differential symbol. Indicates the first Each crack length; Indicates stress amplitude; The expression for the monotonicity constraint loss is: Among them, subscript Represents the first in the dataset One training sample; This indicates taking the maximum value.

10. The fatigue crack life prediction system based on a physical information neural network according to claim 9, characterized in that, In module C3, the weights and bias parameters of the neural network are updated through backpropagation. Specifically, a stochastic gradient descent algorithm is used to iteratively minimize the loss function to determine and update the optimal weights and bias parameters, which are physical parameters. For the physical information neural network model, the weights and bias parameters are... The first in the layer The _ neuron, the _ ... The expressions for calculating the weights and bias parameters in each connection, from top to bottom, are as follows: in, Indicates a time step. Indicates the learning rate. Indicates the momentum coefficient; This represents the error term in backpropagation; and They represent the first Layer connection weights and bias terms; Indicates the increment symbol; The symbol represents the partial derivative.

Images