Search

CN-122020156-A - Stress-strain data enhancement method and system based on condition generation countermeasure network

CN122020156ACN 122020156 ACN122020156 ACN 122020156ACN-122020156-A

Abstract

The invention belongs to the technical field of intersection of material science and artificial intelligence, and particularly relates to a stress-strain data enhancement method and system based on a condition generation countermeasure network, wherein the stress-strain data enhancement method based on the condition generation countermeasure network comprises the following steps of constructing a training data set, constructing a condition generation countermeasure network initial model, designing a dynamic weighting fusion physical constraint loss function of the condition generation countermeasure network initial model, training the condition generation countermeasure network initial model by utilizing the training data set, and obtaining a converged optimal model; and inputting the target strain and random noise into the optimal model to obtain a target data set generated by the optimal model. The stress strain data enhancement method based on the condition generation countermeasure network at least solves the problems of insufficient experimental data of material mechanics and physical constraint loss of the traditional enhancement method.

Inventors

  • ZHAO RUIBIN
  • QIE XIWANG
  • ZHANG GUANG
  • SUN YAN
  • ZHANG MEIJUAN

Assignees

  • 北京航空材料研究院股份有限公司

Dates

Publication Date
20260512
Application Date
20251229

Claims (10)

  1. 1. A method for enhancing stress-strain data of a condition-based generation countermeasure network, comprising the steps of: (1) Acquiring original stress and strain data of metal, and constructing a training data set; (2) Constructing a condition generation countermeasure network initial model, wherein the condition generation countermeasure network initial model comprises a generator and a discriminator; (3) Designing the conditions to generate a dynamic weighted fusion physical constraint loss function of the initial model of the countermeasure network; (4) Training the condition by using the training data set to generate an initial model of the countermeasure network, so as to obtain a converged optimal model; (5) And inputting the target strain and random noise into the optimal model to obtain a target data set generated by the optimal model.
  2. 2. The condition-based generation countermeasure network stress-strain data enhancement method of claim 1, wherein construction of the training data set includes: de-weighting by np. Unique and ordering the raw stress, strain data in ascending order of strain; Generating high density data points using cubic spline interpolation, generating a training data set, and The training dataset was Z-score normalized.
  3. 3. The method for enhancing stress-strain data based on a condition generating countermeasure network of claim 1, wherein the interpolation formula for generating high-density data points using cubic spline interpolation is: ; In the formula, For the post-interpolation strain, In order to correspond to the stress forces, For the raw data point strain, Is an interpolation coefficient; Alternatively, the process may be carried out in a single-stage, 。
  4. 4. The method for enhancing stress-strain data of a condition-based generation countermeasure network according to claim 1, wherein in the construction process of the initial model of the condition-based generation countermeasure network, the generator inputs random noise and condition strain, outputs virtual stress, and the network structure is: Linear(101,128)→LeakyReLU(0.2)→Linear(128,256)→BatchNorm1d(256)→LeakyReLU(0.2)→Linear(256,1)→Tanh(); the discriminator inputs stress and conditional strain, outputs true and false probabilities, and has a network structure of: Linear(2,256)→LeakyReLU(0.2)→Linear(256,128)→LeakyReLU(0.2)→Linear(128,1)→Sigmoid(); optionally, the generator input random noise dimension is 100.
  5. 5. The condition-based generation countermeasure network stress-strain data enhancement method of claim 1, wherein an equation of the dynamic weighted fusion physical constraint loss function is expressed as: ; in which monotonicity dynamic weights Wherein, the method comprises the steps of, 、 Convex dynamic weight Wherein, the method comprises the steps of, Takes the maximum strain as 1.0; , Monotonicity coincidence rate Convexity compliance rate Each by PyTorch automatic differential calculations: , ; Constraint tolerance for monotonicity; Alternatively, the process may be carried out in a single-stage, The value of the elastic phase is more than or equal to 0.9, and the value of the dynamic recrystallization phase is less than or equal to 0.6; Alternatively, the process may be carried out in a single-stage, The value of the elastic phase is less than or equal to 0.3, and the value of the hardening phase is more than or equal to 0.8; Alternatively, the derivative is calculated by automatic differentiation The calculation map is reserved, the second derivative is led, and the gradient calculation error is less than 3%.
  6. 6. A method of enhancing stress-strain data for a condition-based generation countermeasure network as claimed in claim 1 or 5, wherein the generator total loss function is: ; wherein the antagonism loss function of the generator is: ; For the weight of the physical constraint, ; The discriminator fight loss function is: 。
  7. 7. The method for enhancing stress-strain data of a condition-based generation countermeasure network of claim 1, wherein the training process of the condition-based generation countermeasure network initial model includes: adopting an Adam optimizer, setting the batch size as 64, alternately training a discriminator and a generator, and minimizing the target of the discriminator The generator aims at minimizing Storing the model every 100 epoch to obtain the optimal model; optionally, model training iterations are greater than or equal to 5000 times, when the total loss of the generator Training is stopped when continuous 100 epoch fluctuation is less than or equal to 0.001 and the physical compliance rate is more than 98%.
  8. 8. The condition-based generation countermeasure network stress-strain data enhancement method of claim 1, further comprising: Evaluating the quality of the generated target data set; Optionally, the quality assessment index includes: physical rationality, namely, the point proportion of monotonicity coincidence rate is more than or equal to-0.05 is more than 98 percent, and the point proportion of convexity coincidence rate is more than or equal to 0 is more than 95 percent; Data similarity Wasserstein distance from original stress, strain data <0.15, pearson correlation coefficient >0.92.
  9. 9. A condition-generating countermeasure network-based stress-strain data enhancement system as claimed in any one of claims 1 to 8, wherein the condition-generating countermeasure network-based stress-strain data enhancement method is operated on the basis of the data enhancement system, the data enhancement system comprising: the data acquisition and processing module is used for acquiring original stress and strain data of the metal, constructing a training data set and acquiring target strain and random noise; CGAN a model module for constructing a condition generation countermeasure network initial model, wherein the condition generation countermeasure network initial model comprises a generator and a discriminator; The loss calculation module is used for designing the condition generation and resisting the dynamic weighting fusion physical constraint loss function of the network initial model and calculating; And the training generation module is used for training the condition by using the training data set to generate an optimal model converging against the initial model of the network, and generating a target data set according to the target strain and random noise.
  10. 10. The condition-based generation countermeasure network stress-strain data enhancement system of claim 8, further comprising: A quality assessment module for assessing the quality of the generated target data set; Optionally, the loss calculation module integrates PyTorch an automatic differentiation tool, calculates the first and second derivatives in real time, and dynamically adjusts And (3) with ; Optionally, the data acquisition and processing module includes: The data preprocessing module is used for carrying out interpolation and de-duplication processing on the original stress and strain data; and the normalization module is used for performing Z-score normalization and inverse normalization on the training data set.

Description

Stress-strain data enhancement method and system based on condition generation countermeasure network Technical Field The invention belongs to the technical field of intersection of material science and artificial intelligence, and particularly relates to a stress-strain data enhancement method and system based on a condition generation countermeasure network. Background The stress-strain curve is core data for representing the mechanical properties of materials, but the experimental acquisition faces the challenges of high cost, few samples, uneven distribution and the like, namely high-performance materials are tested by precision equipment such as a Gleeble thermal simulation tester, the cost of a single experiment is more than ten thousand yuan, and the data under extreme working conditions (such as high temperature of 1200 ℃ and high strain rate of 10 s -1) are scarce, so that the generalization capability of a machine learning model is poor. The traditional data enhancement method has obvious defects that an interpolation method (linear spline and cubic spline) can only generate a smooth curve between existing data points, the data distribution range cannot be expanded, the nonlinear characteristic of a material cannot be captured, a statistical model (Gaussian process) depends on a preset distribution assumption, the traditional GAN is difficult to adapt to complex physical processes such as material yield, work hardening and the like, and the traditional GAN lacks physical constraint, is easy to generate pseudo data such as stress decline along with strain, second derivative of hardening stage is negative and the like, and cannot be used for engineering analysis. Therefore, the technical pain points of insufficient data and low model training quality of the existing material mechanics experiment are needed to be solved. Disclosure of Invention The invention aims to provide a stress strain data enhancement method and a system based on a condition generation countermeasure network, wherein the stress strain data enhancement method based on the condition generation countermeasure network provided by the invention has the advantages that the monotonicity coincidence rate of data generated by the stress strain data enhancement method based on the condition generation countermeasure network is more than 98%, the convexity coincidence rate is more than 95%, the correlation coefficient with real data is more than 0.92, the material mechanics data set can be effectively expanded, the method is suitable for material design, and at least the problems of insufficient material mechanics experimental data and physical constraint loss of a traditional enhancement method are solved. The first aspect of the present invention provides a method for enhancing stress-strain data of a condition-based generation countermeasure network, comprising the steps of: (1) Acquiring original stress and strain data of metal, and constructing a training data set; (2) Constructing a condition generation countermeasure network initial model, wherein the condition generation countermeasure network initial model comprises a generator and a discriminator; (3) Designing the conditions to generate a dynamic weighted fusion physical constraint loss function of the initial model of the countermeasure network; (4) Training the condition by using the training data set to generate an initial model of the countermeasure network, so as to obtain a converged optimal model; (5) And inputting the target strain and random noise into the optimal model to obtain a target data set generated by the optimal model. In some embodiments, the construction of the training dataset includes de-duplicating and ordering the raw stress, strain data in ascending order of strain by np, generating a training dataset using cubic spline interpolation to generate high density data points, and Z-score normalization of the training dataset. In some embodiments, the interpolation formula for generating high density data points using cubic spline interpolation is: In the formula (I), in the formula (II), For the post-interpolation strain,In order to correspond to the stress forces,For the raw data point strain,Is an interpolation coefficient. In some embodiments of the present invention, in some embodiments,。 In some embodiments, in the construction process of the initial model of the condition generation countermeasure network, the generator inputs random noise and condition strain, outputs virtual stress, and the network structure is as follows: Linear(101,128)→LeakyReLU(0.2)→Linear(128,256)→BatchNorm1d(256)→LeakyReLU(0.2)→Linear(256,1)→Tanh(); the discriminator inputs stress and conditional strain, outputs true and false probabilities, and has a network structure of: Linear(2,256)→LeakyReLU(0.2)→Linear(256,128)→LeakyReLU(0.2)→Linear(128,1)→Sigmoid()。 In some embodiments, the generator input random noise dimension is 100. In some embodiments, the equation for the dynamical