Stress-strain data enhancement method and system based on conditional generative adversarial network

Abstract

The invention belongs to the cross technical field of material science and artificial intelligence, and particularly relates to a stress-strain data enhancement method and system based on a conditional generative adversarial network, and the method comprises the following steps: constructing a training data set, constructing a conditional generative adversarial network initial model, designing a dynamic weighted fusion physical constraint loss function of the conditional generative adversarial network initial model, and training the conditional generative adversarial network initial model by using the training data set to obtain a convergent optimal model; and inputting the target strain and the 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 conditional generative adversarial network at least solves the problems that experimental data of material mechanics is insufficient and physical constraints of a traditional enhancement method are missing.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of materials science and artificial intelligence, and specifically relates to a stress-strain data enhancement method and system based on conditional generative adversarial networks. Background Technology

[0002] Stress-strain curves are core data characterizing the mechanical properties of materials, but their experimental acquisition faces challenges such as high cost, small sample size, and uneven distribution: testing high-performance materials requires precision equipment, such as the Gleeble thermal simulation test chamber, with a single experiment costing over ten thousand yuan; extreme conditions (such as 1200℃ high temperature, 10 s) also present challenges. -1 The scarcity of data at high strain rates leads to poor generalization ability of machine learning models.

[0003] Traditional data augmentation methods have significant drawbacks: interpolation methods (linear, cubic spline) can only generate smooth curves between existing data points, failing to expand the data distribution range and capture the nonlinear characteristics of materials; statistical models (Gaussian processes) rely on preset distribution assumptions, making it difficult to adapt to complex physical processes such as material yielding and work hardening; traditional GANs lack physical constraints and are prone to generating pseudo-data such as "stress decreases with strain" and "the second derivative is negative in the hardening stage," making them unsuitable for engineering analysis.

[0004] Therefore, the technical pain points of insufficient experimental data in materials mechanics and low model training quality urgently need to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide a stress-strain data augmentation method and system based on conditional generative adversarial networks. The stress-strain data augmentation method based on conditional generative adversarial networks provided by this invention generates data with a monotonicity compliance rate of >98%, a convexity compliance rate of >95%, and a correlation coefficient with real data of >0.92. It can effectively expand the materials mechanics dataset, is applicable to materials design, and at least solves the problems of insufficient experimental data in materials mechanics and lack of physical constraints in traditional augmentation methods.

[0006] The first aspect of this invention provides a method for enhancing stress-strain data based on conditional generative adversarial networks, comprising the following steps: (1) Obtain the original stress and strain data of the metal and construct a training dataset; (2) Construct an initial model of a conditional generative adversarial network, wherein the initial model of the conditional generative adversarial network includes a generator and a discriminator; (3) Design the dynamic weighted fusion physical constraint loss function for the initial model of the conditional generative adversarial network; (4) Train the initial model of the conditional generative adversarial network using the training dataset to obtain the converged optimal model; (5) Input the target strain and random noise into the optimal model to obtain the target dataset generated by the optimal model.

[0007] In some embodiments, the construction of the training dataset includes: deduplicating the original stress and strain data using np.unique and sorting them in ascending order of strain; generating high-density data points using cubic spline interpolation to generate the training dataset; and performing Z-score normalization on the training dataset.

[0008] In some embodiments, the interpolation formula for generating high-density data points using cubic spline interpolation is as follows: In the formula, For interpolated strain, For the corresponding stress, For the original data points, strain, These are the interpolation coefficients.

[0009] In some embodiments, .

[0010] In some embodiments, during the construction of the initial model of the conditional generative adversarial network, the generator takes random noise and conditional strain as input and 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 takes stress and conditional strain as input and outputs true / false probabilities. The network structure is as follows: Linear(2,256)→LeakyReLU(0.2)→Linear(256,128)→LeakyReLU(0.2)→Linear(128,1)→Sigmoid().

[0011] In some embodiments, the generator input random noise dimension is 100.

[0012] In some embodiments, the equation for the dynamically weighted fusion physical constraint loss function is expressed as: In the formula: monotonic dynamic weights ;in, , Convex dynamic weights ;in, For maximum strain, a value of 1.0 is used; , Monotonicity compliance rate Convexity compliance rate Each is automatically differentiated using PyTorch: , ; This is a monotonicity constraint tolerance.

[0013] In some embodiments, The value is ≥0.9 in the elastic stage and ≤0.6 in the dynamic recrystallization stage.

[0014] In some embodiments, The value is ≤0.3 in the elastic stage and ≥0.8 in the hardening stage.

[0015] In some embodiments, when automatically differentiating and calculating the derivative, by... Preserve the computation graph, ensure the second derivative is differentiable, and keep the gradient calculation error <3%.

[0016] In some embodiments, the generator's total loss function is: In the formula, the adversarial loss function of the generator is: ; For physical constraint weights, ; The adversarial loss function of the discriminator is: .

[0017] In some embodiments, the training process of the initial model of the conditional generative adversarial network includes: using the Adam optimizer, setting the batch size to 64, and training the discriminator and generator alternately, with the discriminator aiming to minimize... The generator's objective is to minimize The model is saved every 100 epochs to obtain the optimal model.

[0018] In some embodiments, the model is trained ≥5000 times, when the total loss of the generator... Training is stopped when the fluctuation is ≤0.001 for 100 consecutive epochs and the physical conformity rate is >98%.

[0019] In some embodiments, the data augmentation method further includes: evaluating the quality of the generated target dataset.

[0020] In some embodiments, the quality assessment indicators include: physical rationality: the percentage of points with a monotonicity compliance rate ≥ -0.05 > 98%, and the percentage of points with a convexity compliance rate ≥ 0 > 95%; data similarity: Wasserstein distance from the original stress and strain data < 0.15, and Pearson correlation coefficient > 0.92.

[0021] A second aspect of this invention provides a stress-strain data augmentation system based on conditional generative adversarial networks (GANs). The stress-strain data augmentation method based on GANs described in the first aspect operates on the data augmentation system, which includes: The data acquisition and processing module is used to acquire the original stress and strain data of the metal and construct the training dataset; as well as to acquire the target strain and random noise. The CGAN model module is used to construct an initial model of a conditional generative adversarial network, which includes a generator and a discriminator. The loss calculation module is used to design and calculate the dynamic weighted fusion physical constraint loss function of the initial model of the conditional generative adversarial network. The training and generation module is used to train the initial model of the conditional generative adversarial network using the training dataset to obtain a converged optimal model, and to generate a target dataset based on the target strain and random noise.

[0022] In some embodiments, the data augmentation system further includes a quality assessment module for assessing the quality of the generated target dataset.

[0023] In some embodiments, the loss calculation module integrates PyTorch's automatic differentiation tool to calculate the first and second derivatives in real time and dynamically adjust... and .

[0024] In some embodiments, the data acquisition and processing module includes: a data preprocessing module for interpolating and deduplicating the original stress and strain data; and a standardization module for performing Z-score standardization and destandardization on the training dataset.

[0025] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor implements the method described above when executing the computer program.

[0026] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0027] A fifth aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0028] Advantages of this invention: 1) In this invention, a dynamic weighted fusion constraint is designed. By using strain adaptive weights, monotonicity and convexity constraints are coupled into a unified loss, which adapts to the characteristics of the material in the full deformation stage and solves the problem of poor adaptability of traditional fixed weights.

[0029] 2) In this invention, a dual physical-data driven approach is designed. CGAN ensures data diversity, and dynamic constraints ensure physical rationality, thereby achieving the dual goals of "high-quality generation + interpretability".

[0030] 3) The data augmentation method in this invention achieves engineering adaptation, the tolerance design and parameter optimization conform to the actual material experiment, and the generated data can be directly used for finite element simulation and material design, reducing experimental costs by 80%.

[0031] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0032] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating the stress-strain data augmentation method based on conditional generative adversarial networks in an embodiment of the present invention. Detailed Implementation

[0034] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0035] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0036] In the description of the embodiments of this invention, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this invention, "multiple" means two or more, unless otherwise explicitly defined.

[0037] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0038] In the description of the embodiments of this invention, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0039] In the description of the embodiments of the present invention, the term "multiple" refers to two or more (including two), similarly, "multiple groups" refers to two or more (including two groups), and "multiple pieces" refers to two or more (including two pieces).

[0040] To address the issues of insufficient experimental data in existing materials mechanics and the lack of physical constraints in traditional enhancement methods, this invention provides an enhancement method that integrates material physical properties with deep learning. Through an innovative constraint mechanism, it ensures that the generated data not only conforms to the laws of mechanics but also possesses diversity, providing high-quality training data for material performance prediction models and finite element simulations. This method is applicable to materials mechanics research in fields such as aerospace, automotive, and medical.

[0041] The first aspect of this invention provides a stress-strain data augmentation method based on conditional generative adversarial networks, the data augmentation method comprising the following steps: (1) Obtain the original stress and strain data of the metal and construct a training dataset.

[0042] (2) Construct an initial model of a conditional generative adversarial network, which includes a generator and a discriminator.

[0043] (3) Design the dynamic weighted fusion physical constraint loss function for the initial model of the conditional generative adversarial network.

[0044] (4) Use the training dataset to train the initial model of the conditional generative adversarial network to obtain the converged optimal model.

[0045] (5) Input the target strain and random noise into the optimal model to obtain the target dataset generated by the optimal model.

[0046] In this embodiment of the invention, the stress-strain data enhancement method based on conditional generative adversarial networks generates data with a monotonicity compliance rate of >98%, a convexity compliance rate of >95%, and a correlation coefficient with real data of >0.92. This method can effectively expand the materials mechanics dataset and is suitable for materials design.

[0047] This invention designs a dynamic weighted fusion physical constraint loss function, which couples monotonicity and convexity constraints into a unified loss through "strain adaptive weights", thereby achieving precise constraints during the full deformation stage of the material.

[0048] It should be noted that the metal can be either a titanium alloy or an aluminum alloy, and the prediction method in this invention is applicable to all types of metal materials.

[0049] Combination Figure 1 The stress-strain data augmentation method based on conditional generative adversarial networks in this invention is described in detail below.

[0050] Training dataset construction (I) Data Acquisition and Preprocessing In this embodiment of the invention, the original stress of the metal is obtained ( ),strain( The original stress and strain data are loaded, deduplicated using np.unique(strain, return_index=True), and sorted in ascending order of strain. Then, high-density data points (interpolation points) are generated using cubic spline interpolation.

[0051] Specifically, stress and strain data in CSV format are loaded, and duplicate strain points (generally redundant data caused by the precision of experimental equipment) are removed using `np.unique`. The data is then sorted in ascending order of strain to ensure the basic monotonicity of the data. Cubic spline interpolation is performed using `scipy.interpolate.CubicSpline` to generate high-density data points. The interpolation curve satisfies the continuity of the original data points and the boundary conditions of the second derivative (natural splines), avoiding non-physical fluctuations. Points near the yield point (data fluctuation region) and high-strain regions (data sparse region) are supplemented to enhance the data detail.

[0052] In some embodiments, the generated high-density data points are 3 to 5 times the original data.

[0053] In this embodiment of the invention, the interpolation formula for generating high-density data points using cubic spline interpolation is as follows: .

[0054] In the formula, For interpolated strain, For the corresponding stress, For the original data points, strain, These are the interpolation coefficients.

[0055] In some embodiments, the interpolation system satisfies Natural spline conditions.

[0056] (II) Construction and Standardization of Training Dataset In this embodiment of the invention, the original data points and interpolation points are merged, and duplicates are removed again to generate the training dataset. .

[0057] In some embodiments, data quality is ensured through data integrity checks (such as no missing values ​​and strain range covering 0 to 1.0), providing high-quality input for the CGAN model.

[0058] In this embodiment of the invention, the generated training dataset Perform Z-score standardization: ;in, The mean strain, The standard deviation of the strain.

[0059] Specifically, strain and stress data are converted into PyTorch tensors (dimensions). , (Total data), Z-score standardization is used to eliminate dimensional differences: , ;in, The mean strain, The standard deviation of strain, The mean stress is... Let be the standard deviation of stress. After standardization, the data magnitude is unified to . This avoids the model from becoming overly reliant on high-level features (such as stress at the MPa level).

[0060] In this embodiment of the invention, the standardized parameters are retained for destandardization.

[0061] Specifically, retain , , , Used for restoring the physical dimensions of generated data: , Ensure that the output data can be directly used for engineering calculations.

[0062] Construct an initial model of a conditional generative adversarial network (CGAN). In this embodiment of the invention, the CGAN model is defined as follows: 1) Generator

[0063] The core function of the generator is to learn the physical mapping relationship between strain and stress, and guide the direction of data generation through conditional input to avoid the randomness of unconstrained GANs.

[0064] In this embodiment of the invention, the generator Input random noise (e.g., dimension 100) and conditional strain Output virtual stress The network structure is as follows: Linear(101,128)→LeakyReLU(0.2)→Linear(128,256)→BatchNorm1d(256)→LeakyReLU(0.2)→Linear(256,1)→Tanh().

[0065] In some embodiments, a multilayer perceptron (MLP) architecture is employed, with the input being "100-dimensional random noise". +1-dimensional conditional strain Virtual stress is generated through nonlinear transformation. .

[0066] Network layer design adapts to the nonlinear characteristics of the material: the hidden layers use the LeakyReLU(0.2) activation function (to alleviate gradient vanishing), BatchNorm1d(256) to accelerate training, and the output layer is normalized to 0.2 using Tanh(). It matches the standardized stress range.

[0067] 2) Discriminator

[0068] In this embodiment of the invention, the discriminator Input stress and conditional strain, output true / false probabilities, network structure is as follows: Linear(2,256)→LeakyReLU(0.2)→Linear(256,128)→LeakyReLU(0.2)→Linear(128,1)→Sigmoid().

[0069] In some embodiments, the MLP architecture is also used, with the input being "1-dimensional stress". +1-dimensional conditional strain The output data represents the probability (0~1) that the data is the real data.

[0070] The hidden layer uses the LeakyReLU(0.2) activation function, and the output layer uses Sigmoid() to output probabilities. By distinguishing between real data and generated data, the generator performance is optimized in reverse to ensure the statistical authenticity of the generated data.

[0071] Construction of Dynamic Weighted Fusion Physical Constraint Loss Function In this embodiment of the invention, monotonicity and convexity constraints are coupled into a unified loss through "strain adaptive dynamic weights", which solves the problem that traditional fixed weights cannot adapt to the multi-stage characteristics of materials.

[0072] In this embodiment of the invention, the expression for the dynamic weighted fusion physical constraint loss function is: .

[0073] In the formula, the monotonicity constraint term is: ;in, The first derivative of virtual stress with respect to strain, calculated automatically by PyTorch, characterizes the rate of change of stress with strain. In materials mechanics, it must be non-negative (allowing a 5% tolerance to accommodate small fluctuations during the yield stage); 0.05 is the monotonicity constraint tolerance (accommodating small fluctuations during the material's yield stage); monotonicity dynamic weights. ;in, , .

[0074] In some embodiments, for the elastic phase ( ≤0.01), strain The smaller, The larger, The closer to 1, ≥0.9, strengthen the monotonicity constraint.

[0075] In some embodiments, for the dynamic recrystallization stage ( >0.2), strain The larger, It dropped to 0.6. ≤0.6, allowing for slight softening, and adapting to the actual deformation law of the material.

[0076] In the formula, the convexity constraint term is: ;in, The second derivative of virtual stress with respect to strain characterizes the work hardening rate; it must be non-negative during the material hardening stage (the curve exhibits an "upward convex" shape); convexity dynamic weights. ;in, , , For the maximum strain, the value is 1.0.

[0077] In some embodiments, for the elastic phase ( ≤0.01), strain The smaller, It dropped to 0.3. ≤0.3, weakens convexity constraint (second derivative close to 0 in elastic stage).

[0078] In some embodiments, for the hardening stage (0.01 < <0.2), strain The larger, Rising to 0.8 ≥0.8, strengthens convexity constraint, and ensures that work hardening characteristics conform to physical laws.

[0079] In this embodiment of the invention, , Automatic differentiation calculation via PyTorch: ; .

[0080] When automatically calculating derivatives, by... Preserve the computation graph to ensure the second derivative is differentiable, gradient calculation error <3%, and avoid numerical oscillations.

[0081] In the formula, the mathematical expectation The average value of the batch data is taken to make the loss statistically significant, avoid the interference of noise from a single sample, and ensure the stability of the constraint.

[0082] In this embodiment of the invention, dynamic weight adaptation is performed across multiple stages, through... and The strain dependence enables precise constraints during the entire deformation stage of the material: Elastic phase ( ): , Strong monotonicity and weak convexity ensure a linear stress-strain relationship.

[0083] Plastic stage ( ): , It has a medium monotonicity and strong convexity, making it suitable for work-hardened convex curves.

[0084] Recrystallization stage ( ): , It exhibits weak monotonicity and moderate convexity, allowing for slight softening and conforming to the stress reduction trend.

[0085] It is worth mentioning that the tolerance design in this embodiment of the invention conforms to engineering practice. The monotonicity constraint introduces a tolerance of 0.05 to accommodate the small stress fluctuations caused by dislocation movement during the yielding stage of the material (common in experimental data), and avoids excessive constraints that cause model training oscillations. The convexity constraint has no additional tolerance because the non-negativity of the second derivative during the work hardening stage is a core property of the material and must be strictly adhered to.

[0086] Model training and target dataset generation In this embodiment of the invention, loss function optimization during model training includes: Discriminator against loss The binary cross-entropy loss is used to distinguish between real and generated data. The formula is as follows: The goal is to minimize the misjudgment rate of real data and the misjudgment rate of generated data.

[0087] Total generator loss The formula for combining adversarial loss and physical constraint loss is as follows: Among them, combating losses (The goal is to deceive the discriminator); The physical constraint weights are determined through orthogonal experiments to balance adversarial learning with physical laws.

[0088] In this embodiment of the invention, the model training process and parameter settings include: Both the generator and discriminator use the Adam optimizer with a learning rate of [missing information]. Momentum parameters (Classic parameters adapted for GAN training).

[0089] Batch size: 64 (to balance memory usage and training stability).

[0090] Number of iterations: ≥5000 epochs, alternating training of the discriminator (minimizing) ) and generator (minimizer) (This saves the model checkpoint every 100 epochs to avoid training interruption.)

[0091] Convergence criterion: When When the fluctuation is ≤0.001 for 100 consecutive epochs and the physical conformity rate is >98%, the model is considered to have converged, training is stopped, and the optimal model is obtained.

[0092] In this embodiment of the invention, the process of generating the target dataset includes: inputting the target strain, such as an equally spaced strain sequence (0~1.0, interval 0.001) and random noise into the obtained optimal model, generating virtual stress data, restoring the physical dimensions through denormalization, saving it as an Excel file, and obtaining the target dataset.

[0093] In some embodiments, the generated data strain range is extended to 1.3 times that of the original data to supplement extreme condition data.

[0094] Quality assessment In this embodiment of the invention, the quality of the generated data is evaluated using a two-dimensional approach: "physical rationality + data similarity".

[0095] Physical rationality: Monotonicity compliance rate ( The percentage of points > 98%, convexity compliance rate ( (The percentage of points) > 95%.

[0096] Data similarity: Wasserstein distance from real data < 0.15, Pearson correlation coefficient > 0.92, ensuring that the generated data can be used for engineering applications.

[0097] In this embodiment of the invention, the quality of the original data is optimized by cubic spline interpolation, and a dynamic weighted fusion physical constraint loss function is designed to construct a CGAN model with strain as the condition. The monotonicity and convexity of the material are coupled into a unified constraint loss through dynamic weights, thereby achieving a balance between physical laws and data diversity, and at least solving the defect of traditional GANs in generating pseudo data.

[0098] The second aspect of this invention provides a stress-strain data augmentation system based on conditional generative adversarial networks (GANs). The key difference lies in the fact that the stress-strain data augmentation method based on GANs described in the first aspect operates within this data augmentation system, which includes: The data acquisition and processing module is used to acquire the original stress and strain data of the metal and construct the training dataset; as well as to acquire the target strain and random noise.

[0099] The CGAN model module is used to build an initial model for a conditional generative adversarial network, which includes a generator and a discriminator.

[0100] The loss calculation module is used to design and calculate the dynamically weighted fusion physical constraint loss function for the initial model of the conditional generative adversarial network.

[0101] The training and generation module is used to generate an initial adversarial network model using the training dataset to obtain a converged optimal model, and to generate a target dataset based on target strain and random noise.

[0102] In some embodiments, the data acquisition and processing module includes: The data preprocessing module is used to interpolate and deduplicate the original stress and strain data.

[0103] The standardization module is used to perform Z-score standardization and destandardization on the training dataset.

[0104] In this embodiment of the invention, the data augmentation system further includes a quality assessment module for assessing the quality of the generated target dataset.

[0105] In some implementations, the loss calculation module integrates PyTorch's automatic differentiation tool to calculate the first and second derivatives in real time and dynamically adjust... and This ensures that the constraint strength is adapted to different deformation stages of the material.

[0106] In this embodiment of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, wherein when the program is executed by a processor, the program implements the steps of the stress-strain data augmentation method based on conditional generative adversarial networks as described in the first aspect.

[0107] Unless otherwise defined, the technical terms used in the following embodiments have the same meaning as commonly understood by those skilled in the art. Unless otherwise specified, the experimental reagents used in the following embodiments are all conventional biochemical reagents; the raw materials, instruments, and equipment used in the following embodiments can all be obtained commercially or through existing methods; unless otherwise specified, the amounts of experimental reagents used are the amounts used in conventional experimental operations; unless otherwise specified, the experimental methods are conventional methods. It should be further noted that the following description is merely exemplary and not a specific limitation of the present invention.

[0108] Example 1 Using TA15 titanium alloy (800℃~1200℃, 0.001 s) -1 ~1 s -1 Taking stress and strain data enhancement as an example, the stress and strain data enhancement method based on conditional generative adversarial networks specifically includes the following steps: 1) Raw data processing: TA15 titanium alloy was loaded at 1000℃ for 0.1 s.-1 The original data under the working condition (100 points, strain 0~1.0, stress 75 MPa~120 MPa) was deduplicated and 95 points were retained. Cubic spline interpolation was used to generate 300 points. After merging and deduplicating, 378 training samples were obtained. The data fusion confidence (deviation between original and interpolated points) was >95%.

[0109] 2) Model initialization: The generator has an input dimension of 101 (100+1) and an output dimension of 1; the discriminator has an input dimension of 2 (1+1) and an output dimension of 1; the optimizer Adam ( Batch size 64.

[0110] 3) Loss function settings: , , , , , .

[0111] 4) Model training: Set the iteration to 5000 epochs, with the first 1000 epochs as follows: The physical compliance rate increased from 85% to 98% when the value decreased from 2.3 to 0.5; 1000~5000 epochs: It remains stable at 0.4~0.5, with a physical compliance rate of 98%~99%.

[0112] 5) Input the equally spaced strain sequences (0~1.0, interval 0.001) and random noise (dimension 100) into the optimal model obtained after training to obtain the target dataset generated by the optimal model.

[0113] Generate data quality assessment: Physical rationality: Monotonicity compliance rate is 98.7%, convexity compliance rate is 95.3%, and there is no spurious data showing that stress decreases with strain.

[0114] Data similarity: Wasserstein distance to the real data is 0.12, Pearson correlation coefficient is 0.94, and curve fit is extremely high.

[0115] Data diversity: The strain range of the generated data is extended to 0~1.3 (original 0~1.0), and the stress range is extended to 60MPa~130 MPa (original 75 MPa~120 MPa), supplementing the data of extreme working conditions with low stress and high strain.

[0116] The data augmentation method in Embodiment 1 of the present invention is compared with traditional GAN ​​and cubic spline interpolation. The results are shown in Table 1 below.

[0117] Table 1

[0118] As can be seen from Table 1, the data augmentation method in this invention is superior to traditional methods in terms of physical rationality and data diversity. Although interpolation has high similarity, it cannot expand the data range. Traditional GANs have insufficient diversity and high physical error rate.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A stress-strain data augmentation method based on conditional generative adversarial networks, characterized in that, Includes the following steps: (1) Obtain the original stress and strain data of the metal and construct a training dataset; (2) Construct an initial model of a conditional generative adversarial network, wherein the initial model of the conditional generative adversarial network includes a generator and a discriminator; (3) Design the dynamic weighted fusion physical constraint loss function for the initial model of the conditional generative adversarial network; (4) Train the initial model of the conditional generative adversarial network using the training dataset to obtain the converged optimal model; (5) Input the target strain and random noise into the optimal model to obtain the target dataset generated by the optimal model.

2. The stress-strain data augmentation method based on conditional generative adversarial networks as described in claim 1, characterized in that, The construction of the training dataset includes: The original stress and strain data were deduplicated using np.unique and sorted in ascending order of strain. High-density data points are generated using cubic spline interpolation to create a training dataset; and The training dataset is Z-score normalized.

3. The stress-strain data augmentation method based on conditional generative adversarial networks as described in claim 1, characterized in that, The interpolation formula for generating high-density data points using cubic spline interpolation is as follows: ； In the formula, For interpolated strain, For the corresponding stress, For the original data points, strain, These are the interpolation coefficients; Optionally, .

4. The stress-strain data augmentation method based on conditional generative adversarial networks as described in claim 1, characterized in that, In the construction of the initial model of the conditional generative adversarial network, the generator takes random noise and conditional strain as input and outputs virtual stress. 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 takes stress and conditional strain as input and outputs true / false probabilities. The network structure is as follows: 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. The stress-strain data augmentation method based on conditional generative adversarial networks as described in claim 1, characterized in that, The equation for the dynamically weighted fusion physical constraint loss function is expressed as follows: ； Where: Monotonic dynamic weight ;in, , Convex dynamic weights ;in, For maximum strain, a value of 1.0 is used; , Monotonicity compliance rate Convexity compliance rate Each is automatically differentiated using PyTorch: , ; Tolerance for monotonicity constraints; Optionally, The value is ≥0.9 in the elastic stage and ≤0.6 in the dynamic recrystallization stage. Optionally, The value is ≤0.3 in the elastic stage and ≥0.8 in the hardening stage; Optionally, when automatically differentiating and calculating the derivative, by... Preserve the computation graph, ensure the second derivative is differentiable, and keep the gradient calculation error <3%.

6. The stress-strain data augmentation method based on conditional generative adversarial networks as described in claim 1 or 5, characterized in that, The generator's total loss function is: ; In the formula, the adversarial loss function of the generator is: ; For physical constraint weights, ; The adversarial loss function of the discriminator is: .

7. The stress-strain data augmentation method based on conditional generative adversarial networks as described in claim 1, characterized in that, The training process of the initial model of the conditional generative adversarial network includes: The Adam optimizer is used, with a batch size of 64. The discriminator and generator are trained alternately, and the objective of the discriminator is to minimize... The generator's objective is to minimize Save the model every 100 epochs to obtain the optimal model; Optionally, the model is trained for ≥5000 iterations, when the total loss of the generator... Training is stopped when the fluctuation is ≤0.001 for 100 consecutive epochs and the physical conformity rate is >98%.

8. The stress-strain data augmentation method based on conditional generative adversarial networks as described in claim 1, characterized in that, The data augmentation method further includes: The quality of the generated target dataset is evaluated; Optionally, quality assessment indicators include: Physical rationality: The percentage of points with a monotonicity compliance rate ≥ -0.05 is > 98%, and the percentage of points with a convexity compliance rate ≥ 0 is > 95%; Data similarity: Wasserstein distance from the original stress and strain data < 0.15, Pearson correlation coefficient > 0.

92.

9. A stress-strain data augmentation system based on conditional generative adversarial networks, characterized in that, The stress-strain data augmentation method based on conditional generative adversarial networks according to any one of claims 1 to 8 operates on the data augmentation system, the data augmentation system comprising: The data acquisition and processing module is used to acquire the original stress and strain data of the metal and construct the training dataset; as well as to acquire the target strain and random noise. The CGAN model module is used to construct an initial model of a conditional generative adversarial network, which includes a generator and a discriminator. The loss calculation module is used to design and calculate the dynamic weighted fusion physical constraint loss function of the initial model of the conditional generative adversarial network. The training and generation module is used to train the initial model of the conditional generative adversarial network using the training dataset to obtain a converged optimal model, and to generate a target dataset based on the target strain and random noise.

10. The stress-strain data augmentation system based on conditional generative adversarial networks as described in claim 8, characterized in that, The data augmentation system also includes: A quality assessment module is used to assess the quality of the generated target dataset; Optionally, the loss calculation module integrates PyTorch's automatic differentiation tool to calculate the first and second derivatives in real time and dynamically adjust... and ; Optionally, the data acquisition and processing module includes: The data preprocessing module is used to interpolate and deduplicate the original stress and strain data; The standardization module is used to perform Z-score standardization and destandardization on the training dataset.

Images