A welding joint virtual fatigue data generation method based on physical laws and double-discriminative adversarial networks

Abstract

The application discloses a welding joint fatigue performance data generation method based on physical laws and a double-discriminative adversarial network. In view of the fact that existing data enhancement methods can only interpolate between existing samples and cannot generate data of untested working conditions, the application introduces the average stress of the stress concentration coefficient to correct the strain-life model and calculate the physical fatigue life, and uses the logarithm value of the physical fatigue life as the anchor point of the generator. The generator takes the working condition feature vector, the logarithm value of the physical fatigue life and a random noise vector as inputs, outputs the logarithm life correction value, and superimposes the anchor point to obtain a virtual sample. The data distribution discriminator constrains the generated sample to conform to the distribution of the real test data, and the physical consistency discriminator takes the deviation of the logarithm fatigue life of the sample and the logarithm fatigue life of the S-N curve under the same stress range as the discriminative feature, and constrains the generated sample to conform to the S-N curve law. The double constraints on the generator are formed through adversarial training. Finally, interpolation synthesis is performed on the sample along the classification boundary in the logarithm life interval with scarce samples, so that the data in each life interval is balanced. The application can generate virtual fatigue data covering untested working conditions and being physically reasonable under the premise of reducing destructive tests.

Description

Technical Field

[0001] This invention relates to the field of fatigue life assessment and health monitoring technology for welded structures, and more specifically, to a method for generating virtual fatigue data of welded joints based on physical laws and a dual-discriminant adversarial network. Background Technology

[0002] Fatigue life prediction of welded joints increasingly relies on data-driven deep learning models, the performance of which is highly dependent on large-scale, reliable, and evenly distributed fatigue test data. However, fatigue testing of welded joints is costly and time-consuming, resulting in a scarcity of real test samples, and the data exhibits a severe imbalance in distribution between the high-cycle and low-cycle fatigue regions, which limits the accuracy and generalization ability of the prediction models.

[0003] This problem is particularly prominent in high-strength steel welded structures of critical equipment such as ships, marine engineering, and bridges. On the one hand, the preparation of full-size fatigue specimens and long-cycle cyclic loading tests for high-strength steel welded joints are costly and time-consuming, and it is often difficult to obtain a sufficient number of effective data points under a single working condition. On the other hand, differences in welding processes, weld geometry, and initial defects between different batches of specimens introduce additional data discrepancies, further weakening the reliability of data under small sample conditions. In addition, for ultra-high cycle fatigue conditions with extremely high cycle counts, fatigue tests for the corresponding conditions are difficult to carry out in engineering due to limitations in testing machine time and loading frequency, resulting in a long-term lack of measured data for such conditions.

[0004] In existing technologies, data-driven methods have been applied to predict the fatigue life of welded joints. For example, patent application CN114595625A discloses a data-driven method for analyzing the fatigue performance of welded structures: a multi-scale fatigue performance database is established by organizing fatigue performance data from experiments and published literature; the database is cleaned and normalized to remove missing and duplicate values; the correlation between various influencing factors and fatigue life is analyzed using Pearson correlation coefficient and maximum information coefficient; the extreme gradient boosting algorithm is used to quantitatively analyze the weights of influencing factors; and deep convolutional neural networks are used to predict SN curve parameters, thereby extracting fatigue life under specific stress ranges. However, this type of method relies on existing fatigue performance databases, only cleaning and screening existing data, making it difficult to generate new sample data reflecting untested working conditions, and it does not impose physical constraints at the source of data generation. Another approach is to use generative adversarial networks to directly generate data from random noise, but the purely data-driven generation process lacks physical constraints and is prone to generating distorted data that violates the basic laws of fatigue in working conditions where real samples are sparse. Furthermore, existing fatigue life prediction and data augmentation methods often use strain-life models for smooth base material specimens when incorporating physical laws, without taking into account the stress concentration effect at geometric notches such as weld toes and weld roots of weld joints. This makes it difficult to accurately reflect the fatigue behavior of weld joints under notch stress concentration.

[0005] Therefore, there is an urgent need for a method to generate virtual fatigue data of welded joints that uses a physical model as a priori, applies distributional realism constraints and physical consistency constraints at the data generation source, and can generate new data covering untested working conditions. Summary of the Invention

[0006] The purpose of this invention is to address the technical problems of scarce and unevenly distributed real experimental data in existing fatigue life prediction methods, and the fact that existing data augmentation methods can only interpolate between existing samples and cannot cover untested working conditions. This invention provides a virtual fatigue data generation method for welded joints based on physical laws and a dual-discriminant adversarial network. The method uses the Morrow mean stress-corrected strain-life model with introduced stress concentration factors as the physical fatigue damage model. By simultaneously introducing stress concentration factors into both the stress amplitude and mean stress terms of this model, the calculation of physical fatigue life can reflect the stress concentration effect at geometric notches such as the weld toe and weld root of the welded joint, thus providing a physical baseline consistent with the fatigue behavior of the welded joint for the subsequent generation process. Unlike existing fatigue data augmentation methods that directly apply strain-life models for smooth base material samples to welded joints, this method's physical baseline incorporates the essential differences in notch stress concentration between welded joints and smooth samples from the outset.

[0007] The method decomposes the generation of virtual sample fatigue life into a composite form of physical fatigue life and logarithmic life correction. The generator takes the feature vector of the working condition, the logarithmic value of the physical fatigue life, and the random noise vector as input, and outputs only the logarithmic life correction. The logarithmic life correction is defined as the logarithmic residual of the virtual sample's logarithmic fatigue life relative to the physical fatigue life. The logarithmic life correction is added to the logarithmic value of the physical fatigue life to obtain the logarithmic fatigue life of the virtual sample. This residual learning architecture allows physical laws to bear the main trend of fatigue life changing with working conditions, while the generator only describes its dispersion. Compared with existing pure data-driven generative adversarial networks that directly generate fatigue life from random noise, it has a stable physical extrapolation capability in working condition regions where real test samples are scarce. It can generate virtual samples that conform to the basic laws of fatigue on untested working conditions that do not overlap with the real sample working conditions.

[0008] The method constrains the logarithmic lifetime correction within a range centered at zero and with a half-width equal to the physical divergence band half-width. The physical divergence band half-width is adaptively determined based on the logarithmic residual standard deviation of the actual fatigue test sample relative to its physical fatigue life. The physical divergence band half-width is proportional to the residual standard deviation, and the scaling factor is a preset scaling factor. This adaptive physical divergence band allows the width of the physical divergence band to adaptively scale with the dispersion of the fatigue data of the welded joint itself. This allows the generator to output virtual samples that cover a reasonable dispersion range of the actual fatigue test data, while suppressing the generator to output outrageous extrapolations that violate the basic laws of fatigue.

[0009] The method simultaneously sets up a data distribution discriminator and a physical consistency discriminator, which together constitute a dual adversarial constraint on the generator. The data distribution discriminator takes the concatenated vector of the working condition feature vector and the logarithmic fatigue life as input, and constrains the generated samples to conform to the joint distribution of the real fatigue test sample set. The physical consistency discriminator takes the deviation between the logarithmic fatigue life of the sample and the logarithmic fatigue life of the SN curve under the same stress range as the discrimination feature. The SN curve is obtained by fitting the real fatigue test sample set according to the joint type. The physical consistency discriminator constrains the logarithmic deviation of the generated sample relative to the SN curve to fall within the allowable physical divergence range of the welded joint fatigue data. The loss function of the generator is taken as the weighted sum of the corresponding adversarial loss terms of the two discriminators. Compared with the existing single discriminator generator adversarial network, which only constrains the authenticity of the data distribution, the dual discriminator adversarial mechanism of this method enables the generated samples to simultaneously satisfy the distribution consistency and physical consistency, which significantly reduces the probability of generating distorted samples in the working condition region where real samples are scarce.

[0010] After the generator training is completed, the method performs boundary sample oversampling equalization processing on the virtual fatigue dataset containing untested working conditions according to the logarithmic fatigue life interval. The dataset is divided into several life intervals according to the logarithmic fatigue life. For sparse life intervals with a sample number lower than the equalization threshold, the samples with larger absolute values ​​of logarithmic life residuals in the sparse life interval are used as boundary samples for oversampling, so that the sample number of each logarithmic fatigue life interval tends to be balanced, overcoming the typical sample sparsity imbalance problem in the long life and short life regions of welded joint fatigue data.

[0011] In some implementations, the physical fatigue damage model is not limited to the Morrow mean stress-corrected strain-life model that incorporates stress concentration factors, but may also employ other mean stress-corrected strain-life models that incorporate stress concentration effects; the specific network structures of the generator, the data distribution discriminator, and the physical consistency discriminator include one or more of multilayer perceptrons, convolutional neural networks, or Transformer networks; the working condition feature vectors of each sample in the real fatigue test sample set may also include welding method, plate thickness, and joint type, wherein the welding method and joint type are incorporated into the working condition feature vector after being thermally encoded.

[0012] The core innovations of this invention are as follows: First, at the source of data generation, a physical fatigue damage model is embedded into the generation path as a priori anchor point. The generation of virtual sample fatigue life is decomposed into a composite form of physical fatigue life and logarithmic life correction, allowing physical laws to bear the main trend of working condition-life, while the generator only learns random dispersion. This enables stable extrapolation to untested working conditions even when real samples are scarce. Second, to address the stress concentration effect at geometric notches such as weld toes and weld roots in welded joints, a stress concentration coefficient is introduced into the physical fatigue damage model, enabling the calculation of physical fatigue life to reflect the essential differences in fatigue behavior between welded joints and smooth base material samples. Third, a data distribution discriminator and a physical consistency discriminator are constructed. The system employs a dual-resistance mechanism: the former constrains the authenticity of the generated sample distribution, while the latter uses the deviation between the logarithmic fatigue life of the sample and the logarithmic fatigue life of the SN curve under the same stress range as a discriminant feature to constrain the physical consistency of the samples, ensuring that the generated samples simultaneously satisfy both types of constraints. Fourth, the logarithmic lifetime correction is constrained within a physical divergence band centered at zero, and the half-width of the physical divergence band is adaptively determined using the standard deviation of the logarithmic residuals of the actual test samples relative to the physical fatigue life, allowing necessary random dispersion while suppressing out-of-spectrum extrapolation. Fifth, the virtual fatigue dataset covering untested conditions undergoes boundary sample oversampling equalization processing according to the logarithmic fatigue life interval, alleviating the sample sparsity imbalance between the long-life and short-life regions. For the specific mathematical expression, parameter value range, and training implementation details of the above technical solutions, please refer to the claims and specific embodiments. Beneficial effects

[0013] Compared with the prior art, the present invention has at least the following beneficial effects:

[0014] This invention uses the physical fatigue life calculated by a physical fatigue damage model as the generator anchor point, allowing the generator to learn only the logarithmic residual of the actual fatigue life relative to the physical fatigue life. The physical fatigue life then bears the main responsibility for the lifespan variation under different operating conditions. This reduces reliance on large-scale destructive testing when real samples are scarce, and generates virtual fatigue data covering untested conditions under the guidance of physical laws. A data distribution discriminator and a physical consistency discriminator form a collaborative constraint on the generator, ensuring that the generated samples maintain physical rationality while conforming to the real data distribution, preventing the generator from deviating from physical laws in order to fit the distribution. Furthermore, the logarithmic life correction output by the generator is constrained within a half-width adaptively determined physical divergence band, matching the dispersion of the generated lifespan to the inherent dispersion of fatigue life. Finally, life interval equalization improves the balance of sample distribution across life intervals, thereby providing high-quality training data for the fatigue life prediction model while reducing destructive testing, thus improving the model's prediction accuracy and generalization ability. Attached Figure Description Figure 1 This is a flowchart illustrating the overall implementation of the method of the present invention; Figure 2 This is a schematic diagram of the adversarial structure between the physical law generator and the dual discriminator in this invention; Figure 3 This is a schematic diagram of the Borderline-SMOTE lifetime range equalization process in this invention; Figure 4 This is a comparison diagram of the distribution of real and virtual samples on the stress range-log lifetime plane in an embodiment of the present invention; Figure 5 This is a comparison chart of the accuracy of the downstream fatigue life prediction model before and after data augmentation in an embodiment of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and a specific embodiment. The described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0016] This embodiment uses the generation of virtual fatigue data for EH36 high-strength steel welded joints as an example to illustrate the implementation process of the present invention in detail. like Figure 1 As shown, the specific steps of the present invention are as follows:

[0017] Step 1: Construct a real sample set and calculate the physical fatigue life.

[0018] The actual fatigue test data in this embodiment were compiled from laboratory welded joint fatigue test data and fatigue performance data from publicly available literature. A total of 189 sets of actual fatigue test samples of EH36 high-strength steel welded joints under different working conditions were collected. The characteristic vector of the working conditions for each set of samples includes: welding method, plate thickness, joint type, stress ratio R, stress range ΔS, and stress concentration factor K. t The welding method and joint type are individually thermally encoded (each with 3 categories), and then concatenated with the normalized plate thickness, stress ratio, stress range, and stress concentration factor to obtain a 10-dimensional feature vector x for the working condition; the corresponding label is the logarithmic fatigue life lgN. true The values ​​for each working condition were determined based on the statistical distribution of the established dataset. Typical ranges are as follows: stress range ΔS ∈ (0, 600) MPa (mean 257.4, standard deviation 91.4), stress ratio R ∈ [-1, 1) (mean -0.2), plate thickness ∈ (0, 90) mm (mean 26.7), and stress concentration factor K. t ∈(0.5, 5) (mean 2.1). The stress concentration factor K for each sample in this embodiment. t The data is taken from the established dataset. In the absence of measured values, the stress concentration factor can also be estimated based on the joint type and plate thickness using the formula of the nominal stress method. The calculation formulas corresponding to the geometry of butt joints, T-joints and cross joints are used respectively.

[0019] In this embodiment, the relevant material parameters are those corresponding to EH36 high-strength steel: elastic modulus E = 206000 MPa; fatigue strength coefficient σ′ f =900MPa, fatigue ductility coefficient ε′ f =0.40, fatigue strength index b = -0.09, fatigue ductility index c = -0.58 (the above four fatigue parameters are calibrated by standard fatigue tests). From this, the cyclic hardening index is obtained. Cyclic hardening coefficient

[0020] For each set of working conditions, the physical fatigue life N is calculated based on its stress range, stress ratio, and stress concentration factor. p Taking one set of working conditions as an example: the stress range ΔS = 300MPa, then the nominal stress amplitude S a =150MPa; stress ratio R = -0.2, then the average stress Stress concentration factor K t =2.1. Substituting into the physical fatigue damage model equation:

[0021] The left-hand side of the equation is approximately 1.99 × 10⁻⁶. -3 The bisection method is used for numerical solution, yielding 2N.p ≈7.81×10 4 That is, physical fatigue life N p ≈3.90×10 4 The second cycle, its logarithm lgN p ≈4.59. The above calculation was performed on all 189 working conditions to obtain the physical fatigue life of each sample.

[0022] Step 2: Build the generator.

[0023] like Figure 2 As shown, the generator employs a multilayer perceptron, with its inputs being a 10-dimensional feature vector x representing the working conditions and a 1-dimensional logarithmic value lgN representing the physical fatigue life. p The 27-dimensional vector is obtained by concatenating a 16-dimensional random noise vector z; the hidden layer consists of two layers, each with 128 neurons, using the ReLU activation function; the output layer has one neuron, which is activated by hyperbolic tangent and then multiplied by the divergence band half-width τ = 0.6 to obtain the logarithmic lifetime correction δ ∈ [-0.6, 0.6]. The logarithmic fatigue lifetime of the virtual sample is lgN. p +δ. Using N from step 1 p ≈3.90×10 4 Taking the operating condition as an example, if the generator outputs δ = 0.05, then the virtual sample logarithmic lifetime is 4.59 + 0.05 = 4.64, corresponding to a virtual lifetime of approximately 4.37 × 10⁻⁶. 4 The next cycle falls within a reasonable divergence band centered on the physical fatigue life. The half-width τ of the divergence band is adaptively determined based on the residual of the logarithmic fatigue life relative to the physical fatigue life of the actual sample. The logarithmic fatigue life residual of the actual sample is denoted as the difference between its logarithmic fatigue life and the logarithmic value of the physical fatigue life. In this embodiment, it approximately follows a zero-mean normal distribution on the welded joint dataset, with a residual standard deviation s. r Approximately 0.30. To ensure that the logarithmic lifetime correction output by the generator can cover the (1-α) proportion of the physical dispersion of the actual fatigue data, the half-width of the divergence band is taken as the normal quantile corresponding to this confidence level, i.e., τ = z. 1-α / 2 ·s r In the formula z 1-α / 2 Let (1-α / 2) be the quantile of the standard normal distribution, and (1-α) be the preset physical dispersion coverage. Therefore, the half-width of the divergence band is not empirically determined, but rather determined by the inherent dispersion of fatigue life. r This is determined in conjunction with the required physical dispersion coverage. In this embodiment, the coverage (1-α) is set at 95.4%, corresponding to z 1-α / 2Since τ equals 2.00, τ equals 2.00 × 0.30, or 0.60. The physical meaning of a coverage rate of 95.4% is that within a divergence band centered on the physical fatigue life and with a half-width of τ, approximately 95.4% of the true fatigue life dispersion can be accommodated. This ensures that the generated samples reflect the inherent dispersion of fatigue life while avoiding an excessively wide divergence band that deviates from the fatigue law, thanks to the constraint of physical quantiles. If the coverage rate is increased or decreased, the half-width of the divergence band changes with z. 1-α / 2 Increase or decrease accordingly: τ corresponds to s when the coverage rate is 86.6% and 98.8%, respectively. r The 1.5 times and 2.5 times of the divergence band form the adjustable range of the half-width of the divergence band.

[0024] Step 3: Construct a dual discriminator.

[0025] Data distribution discriminator D d A multilayer perceptron is used. The input is an 11-dimensional vector (a concatenation of 10-dimensional working condition features and 1-dimensional log-lifetime). The hidden layers consist of two layers with 128 and 64 neurons, both activated by LeakyReLU (negative slope 0.2). The output layer consists of a single neuron activated by Sigmoid, and the output sample is the probability that it follows the true distribution.

[0026] Physical consistency discriminator D p Structure and D d The input is the same, but it additionally incorporates 2D physical features: physical residual r = lgN - lgN p And the deviation Δ between the sample logarithmic lifetime and the logarithmic lifetime of the SN curve under the same stress range. SN The system has 13 dimensions. The training positive samples consist of 189 sets of real test samples; the training negative samples are physically distorted samples, constructed by replacing the logarithmic fatigue life of the real samples with values ​​whose absolute residuals relative to the physical fatigue life fall within the interval [τ+0.2, τ+1.0] (i.e., [0.8, 1.6]) and whose signs are random, totaling 189 sets. The lower bound of the absolute residual is set to τ+0.2 rather than directly adjacent to τ to leave a safety margin and prevent critical disturbance samples from falling into the physical divergence band due to numerical calculation errors, thus avoiding label noise in both positive and negative samples. In this embodiment, the SN curve is based on the nominal stress method and fitted according to joint type: for butt joints, T-joints, and cross joints, linear regression is performed on the logarithmic fatigue life of the stress range of the test samples for each joint type, yielding an SN curve in the form lgN = lgC-m·lgΔS. Each of the three joint types corresponds to one set of (C, m) parameters. Furthermore, to quantitatively characterize the consistency between the generated samples and the physical laws represented by the SN curves, a normalized physical consistency score is defined: d is denoted as the deviation between the logarithmic fatigue life of the sample and the logarithmic fatigue life of the SN curve for the corresponding joint type under the same stress range, and s is denoted as the standard deviation of the residuals from the regression of the SN curve for the corresponding joint type. SNIn this embodiment, its value is approximately 0.30, which is comparable to the inherent dispersion of logarithmic fatigue life. Therefore, the physical consistency score is defined as follows: Its value range is (0, 1): when the sample falls exactly on the SN curve, d equals 0 and the score is 1; the further the deviation, the closer the score is to 0. The physical consistency discriminator uses standardized deviation. As a normalization discrimination feature, the aforementioned cleaning thresholds of 5 times and 10 times the average deviation of the entire dataset are used. Based on a ratio of approximately 0.8 between the average absolute deviation and the standard deviation, this is roughly equivalent to removing samples whose standardized deviation exceeds 4 times and 8 times the residual standard deviation, respectively. The real samples have already been cleaned according to objective criteria before being added to the database. These objective criteria are: removing samples whose deviation between the fatigue life value to be determined and the SN curve life value under the same stress range exceeds 5 times the average deviation of the entire dataset, and removing samples whose deviation exceeds 10 times. After this cleaning, the absolute value of the residuals of the real samples relative to the physical fatigue life does not exceed the half-width τ of the divergence band, while the absolute values ​​of the residuals of the aforementioned negative samples all exceed τ, thus ensuring that positive and negative samples do not overlap in the residual dimension and can be stably distinguished.

[0027] Step 4: Competitive training.

[0028] An alternating optimization strategy is adopted: in each iteration, the generator is fixed first, and the two discriminators are updated once each; then the two discriminators are fixed again, and the generator is updated once. The generator loss function is... L G =-E[lnD d (G)]-E[lnD p (G)]+0.1×(max(0,|δ|-0.6)) 2

[0029] In this embodiment, the weight μ of the physical anchoring regularization term is set to 0.1, and the half-width τ of the divergence band is set to 0.6. μ is used to balance the adversarial loss of the generator and the physical anchoring regularization term: if μ is too high, the physical anchoring term will dominate the optimization, and the generator will tend to directly reproduce the physical fatigue lifetime, weakening the fit to the real data distribution; if μ is too low, the physical anchoring effect will be insufficient, and the generated lifetime will easily deviate from the physical fatigue lifetime. In this embodiment, μ = 0.1 is used, so that the physical anchoring term imposes a moderate constraint on the generated lifetime without suppressing adversarial learning. The Adam optimizer is used, and the learning rate η is set to 2 × 10. -4 The first-order moment decay coefficient is set to 0.5, the batch size to 64, and the number of training iterations to 5000; the maximum information coefficient deviation threshold is set to 0.05. This embodiment uses loss function stability as the training convergence criterion: when the changes in the loss functions of the generator, data distribution discriminator, and physical consistency discriminator are all less than the loss stability threshold of 1×10⁻⁶ within 100 consecutive iterations... -3When training converges, the generator parameters are saved. Based on this, the maximum information coefficient deviation between the virtual dataset and the real sample set for each working condition dimension is less than 0.05, which is used as a verification condition for the statistical consistency of the generated virtual fatigue dataset. When training converges, the output probabilities of the data distribution discriminator and the physical consistency discriminator for the generated samples approach 0.5, serving as an auxiliary reference.

[0030] Step 5: Batch generation and merging.

[0031] Using the trained generator, 3000 sets of feature vectors for working conditions are sampled within the working condition space. Sampling is performed uniformly within the range of real sample working conditions and extrapolated to untested working condition regions; during extrapolation, the constraint stress range does not exceed 0.9 times the elastic limit corresponding to the yield strength of the parent material, and the stress concentration factor K... t The value should not exceed 5 (i.e., the upper limit of the effective value for the physical fatigue damage model) to ensure the validity of the physical fatigue life calculation. In this embodiment, the sampling range for each working condition parameter is as follows: stress range ΔS is (0, 600) MPa, stress ratio R is [-1, 1), plate thickness is (0, 90] mm, and stress concentration factor K... t The range is (0.5, 5). The welding method is taken from the categories already included in the dataset, and the joint type is selected as butt joint, T-joint, or cross joint. During sampling, samples are uniformly collected within the range of real sample working conditions, and the values ​​of extrapolated working conditions do not exceed the above interval boundaries. The physical fatigue life of each sample working condition is calculated and the virtual logarithmic life is output by the generator to obtain 3000 sets of virtual samples. These are then merged with 189 sets of real samples to form a virtual fatigue dataset containing 3189 sets of samples.

[0032] Step 6: Lifetime range equalization.

[0033] like Figure 3 As shown, the virtual fatigue dataset is evenly divided into 5 intervals based on logarithmic fatigue life. The sample size for each interval is counted; in this embodiment, the sample sizes for the 5 intervals are approximately 185, 760, 1290, 760, and 194 (total 3189). The low-cycle fatigue region (lgN < 4, 185 groups) and the ultra-high-cycle fatigue region (lgN > 6, 194 groups) have significantly fewer samples than the middle intervals and are therefore identified as minority class intervals. The boundary samples v at the classification boundaries within the minority class intervals are further analyzed. b Within the same minority class interval, select its 5 nearest neighbor samples and synthesize a new sample according to the following formula: v new =v b +ζ·(v nn -v b ) Where ζ is a random number within [0, 1]; approximately 210 sets were synthesized in the low-frequency region and approximately 247 sets were synthesized in the ultra-high-frequency region, for a total of approximately 457 sets. After equalization, the dataset contains approximately 3646 sets of samples, with the difference in the number of samples in each interval not exceeding 10%.

[0034] Step 7: Verify the effect.

[0035] To illustrate the effectiveness of this invention, a deep convolutional neural network driven by both physical and data principles was used as the downstream fatigue life prediction model. While keeping the hyperparameters of the downstream model constant, the model was trained on three different training sets and evaluated on the same independent test set. The independent test set consisted of real fatigue test samples that were not used in the training, with a portion of the samples falling within the sparsely covered working condition region of the training set, used to examine the prediction performance of each training set on this type of working condition. The evaluation results were as follows: when training with only 189 real samples, the mean relative error (MRE) was 0.42 and the standard deviation of relative error (RMSE) was 0.46; when training with the existing Borderline-SMOTE to enhance the real samples, the MRE was approximately 0.38 and the RMSE was approximately 0.41; when training with the equalized virtual fatigue dataset of this invention, simulations showed an MRE of approximately 0.30 and an RMSE of approximately 0.33. Furthermore, the physical rationality of the generated virtual samples was evaluated: the physical violation rate was defined as the proportion of generated samples whose deviation from the SN curve life of the joint type under the same stress range exceeded the aforementioned objective law criterion threshold. The logarithmic lifetime correction δ output by the generator is constrained within the physical divergence band [-τ, τ], limiting the offset of the generated samples relative to the physical fatigue lifetime. Due to this constraint, the physical violation rate of the generated samples is lower than that of purely data-driven generation methods. Simultaneously, in this embodiment, the maximum information coefficient of the virtual dataset's "stress range - logarithmic lifetime" dimension remains within the range of [0.63, 0.73], with a deviation from 0.68 of the real sample set not exceeding 0.05, indicating that the generated data maintains a good correlation between operating conditions and lifetime. Figure 4 As shown, the distribution of virtual samples on the stress range-log lifetime plane is similar to that of real samples, and it covers some operating conditions where there are fewer real samples; for example... Figure 5 As shown, compared to training with only real samples, the MRE and RMSE on the test set are reduced when training with the dataset generated by this invention.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating virtual fatigue data of welded joints based on physical laws and a dual-discriminatory adversarial network, characterized in that, Includes the following steps: Step 1: Obtain a real fatigue test sample set of welded joints. Each sample includes a condition feature vector characterizing the load condition and the corresponding logarithmic fatigue life. The condition feature vector includes at least the stress range, stress ratio and stress concentration factor. The physical fatigue damage model is a Morrow mean stress modified strain-life model that introduces a stress concentration factor. It introduces a stress concentration factor into both the stress amplitude term and the mean stress term to reflect the geometric notch effect at the weld toe and weld root of the weld joint. Based on the working condition feature vector, the physical fatigue life corresponding to each sample is calculated through the physical fatigue damage model. The physical fatigue life serves as the physical prior anchor point for the subsequent generation process. Step 2: Construct a generator. The generator takes the feature vector of the working conditions, the logarithmic value of the physical fatigue life, and the random noise vector as input, and outputs only the logarithmic fatigue life correction. The logarithmic fatigue life correction is defined as the logarithmic residual of the virtual sample's logarithmic fatigue life relative to the physical fatigue life. The logarithmic fatigue life correction is added to the logarithmic value of the physical fatigue life to obtain the logarithmic fatigue life of the virtual sample. Thus, the generation of the virtual sample fatigue life is decomposed into a composite form with physical fatigue life as the deterministic principal term and the logarithmic fatigue life correction as the random residual term. The logarithmic fatigue life correction is constrained to an interval centered at zero with a half-width of the physical divergence band half-width. The physical divergence band half-width is adaptively determined based on the standard deviation of the logarithmic residual of the real fatigue test sample relative to its physical fatigue life, so that the width of the physical divergence band adaptively scales with the dispersion of the fatigue data of the welded joint itself. Step 3: Construct a data distribution discriminator and a physical consistency discriminator; the data distribution discriminator takes the concatenated vector of the working condition feature vector and the logarithmic fatigue life as input, and is used to determine whether the input sample conforms to the joint distribution of the real fatigue test sample set; the physical consistency discriminator takes the deviation between the logarithmic fatigue life of the input sample and the logarithmic fatigue life of the SN curve under the same stress range as the discrimination feature, the SN curve is obtained by fitting the real fatigue test sample set according to the joint type, and the physical consistency discriminator is used to determine whether the logarithmic deviation of the input sample relative to the SN curve falls within the allowable physical divergence range of the welded joint fatigue data; Step 4: Perform alternating adversarial training on the generator, the data distribution discriminator, and the physical consistency discriminator; the loss function of the generator is taken as the weighted sum of the adversarial loss terms corresponding to the data distribution discriminator and the physical consistency discriminator, where both weights of the weighted sum are non-negative and at least one is non-zero, so that the generator is simultaneously driven by both distribution realism constraints and physical consistency constraints; the data distribution discriminator and the physical consistency discriminator are trained with the objectives of classifying real samples as true and classifying virtual samples output by the generator as false, respectively. Until training converges; Step 5: Using the trained generator, generate virtual samples in batches within the condition space containing untested working conditions. The untested working conditions refer to working conditions whose condition feature vectors do not overlap with the condition feature vectors of any sample in the real fatigue test sample set. Merge the virtual samples with the real fatigue test sample set to form a virtual fatigue dataset. The parameters of each working condition in the condition space are within the effective value range of the physical fatigue damage model, so that the physical fatigue life corresponding to each working condition is solvable. Step 6: Divide the virtual fatigue dataset into several lifetime intervals according to the logarithmic fatigue lifetime. For sparse lifetime intervals with a sample number lower than the equilibrium threshold, oversample the samples with larger absolute values ​​of logarithmic lifetime residuals within the sparse lifetime intervals as boundary samples. The logarithmic lifetime residual is the difference between the logarithmic fatigue lifetime and the physical fatigue lifetime of the sample. Output the final virtual fatigue dataset where the sample number in each logarithmic fatigue lifetime interval tends to be balanced.

2. The method according to claim 1, characterized in that, The physical fatigue life N mentioned in step 1 p The following equations are obtained by solving them numerically: Among them, K t S is the stress concentration factor. a S is the nominal stress amplitude and equal to half of the stress range. m The mean stress is determined by the stress ratio, E is the elastic modulus, and σ′ is the mean stress. f ε′ is the fatigue strength coefficient. f Here, b is the fatigue ductility coefficient, c is the fatigue strength index, and H is the fatigue ductility index. c The cyclic hardening coefficient is and satisfies h c It is the cyclic hardening index and satisfies The average stress S m From stress ratio R and nominal stress amplitude S a Determine using the following formula: The stress concentration factor K t The stress concentration factor is determined based on the joint type and plate thickness of the welded joint using a preset formula.

3. The method according to claim 1, characterized in that, The generator mentioned in step 2 is a neural network model, which is one or more of a multilayer perceptron, convolutional neural network, or Transformer network; the logarithmic lifetime correction is scaled to a preset physical divergence band [-τ, τ] after being mapped by the hyperbolic tangent activation function, where r is the half-width of the divergence band characterizing the inherent dispersion of fatigue lifetime.

4. The method according to claim 1, characterized in that, In step 3, the positive training samples for the physical consistency discriminator are the actual fatigue test sample set, and the negative training samples are physical distortion samples. The physical distortion samples are constructed by applying a random perturbation to the logarithmic fatigue life of the actual fatigue test samples. The random perturbation causes the absolute value of the logarithmic residual of the sample relative to its physical fatigue life after the perturbation to exceed the half-width of the physical divergence band, and the perturbation sign is random. The SN curve is based on the nominal stress method and fitted according to the joint type of the welded joint. For each joint type, a linear regression is performed between the logarithmic value of the stress range of the actual fatigue test samples of that joint type and the logarithmic fatigue life to obtain an SN curve in the form of lgN = lgC - m·lgΔS, where ΔS is the stress range, N is the fatigue life, and C and m are the SN curve parameters (C is the intercept parameter and m is the slope parameter). The input of the physical consistency discriminator includes the deviation between the logarithmic fatigue life of the sample and the logarithmic fatigue life of the SN curve of the joint type under the same stress range. When the changes in the loss functions of the generator, the data distribution discriminator, and the physical consistency discriminator within a preset number of consecutive rounds are all less than a preset loss stability threshold, the training is deemed to have converged. The maximum information coefficient is used to measure the correlation strength between each dimension of the working condition feature vector and the logarithmic fatigue life. The deviation between the maximum information coefficient of each dimension in the virtual fatigue dataset and the logarithmic fatigue life and the maximum information coefficient of the corresponding dimension in the real fatigue test sample set is less than a preset threshold, which is used as a verification condition for the statistical consistency of the generated virtual fatigue dataset.

5. The method according to claim 1, characterized in that, The loss function of the generator in step 4 is expressed as: L G -E[lnD d (G)]-E[lnD p (G)]+µ·E[R(δ)] Among them, D d (G) is the discriminant output of the data distribution discriminator for the generated samples, D p (G) is the discrimination output of the physical consistency discriminator on the generated samples, δ is the logarithmic lifetime correction, R(δ) is the physical anchoring regularization term applied to the logarithmic lifetime correction, μ is the balance hyperparameter, and E[·] represents the expectation of the batch of samples. The physical anchoring regularization term R(δ) is defined as follows: R(δ)=(max(0,|δ|-τ)) 2 Wherein, τ is the half-width of the physical divergence band; its value is zero when the absolute value of the logarithmic lifetime correction does not exceed the physical divergence band, and a second penalty is applied when it exceeds it.

6. The method according to claim 1, characterized in that, Step 6, the boundary sample oversampling equalization process, employs a boundary synthesis minority class oversampling technique, including: dividing the virtual fatigue dataset into several intervals based on logarithmic fatigue life; identifying intervals with a sample count below a preset threshold as minority class intervals; and synthesizing new samples from the boundary samples at the classification boundary within the minority class intervals using the following formula: v new =v b +ζ·(v nn -v b ) Among them, v b v is the feature vector of the boundary sample. nn For v b The feature vectors of nearest neighbors within the same minority class interval, where ζ is a random number in the interval [0, 1], v new This is the feature vector of the synthesized new sample.

7. The method according to claim 1, characterized in that, The working condition feature vector mentioned in step 1 also includes welding method, plate thickness and joint type; wherein the welding method and joint type are incorporated into the working condition feature vector after being thermally encoded. The welded joint is a butt joint, T-joint, or cross joint made of metal materials, and the virtual fatigue dataset is used to train a fatigue life prediction model for the welded joint.

8. The method according to claim 3, characterized in that, The half-width r of the physical divergence band is determined based on the standard deviation of the residuals between the logarithmic fatigue life and the logarithmic physical fatigue life of each sample in the real fatigue test sample set, and is taken as 1.5 to 2.5 times the standard deviation.

9. The method according to claim 1, characterized in that, The working condition sampling in step 5 satisfies the following constraints: the stress range does not exceed the preset proportion of the elastic limit corresponding to the yield strength of the base material of the welded joint; the stress concentration factor falls within the effective value range of the physical fatigue damage model; the stress ratio falls within the stress ratio value range of the real fatigue test sample set; the plate thickness falls within the plate thickness value range of the real fatigue test sample set; and the welding method and joint type are taken from the categories already included in the real fatigue test sample set, so as to ensure that the physical fatigue life of the generated virtual sample is effective.

10. A virtual fatigue data generation system for welded joints based on physical laws and a dual-discriminatory adversarial network, characterized in that, include: The physical law calculation module is used to obtain a real fatigue test sample set of welded joints and calculate the physical fatigue life based on the feature vector of working conditions through a preset physical fatigue damage model. The generation module is used to take the feature vector of the working condition, the logarithmic value of the physical fatigue life and the random noise vector as input, output the logarithmic life correction amount, and add the logarithmic life correction amount to the logarithmic value of the physical fatigue life to obtain the logarithmic fatigue life of the virtual sample. The dual-discrimination module includes a data distribution discriminator and a physical consistency discriminator, which are used to determine whether the input sample conforms to the distribution of real experimental data and whether it is consistent with the physical law represented by the SN curve, respectively. The adversarial training module is used to perform alternating adversarial training on the generation module and the dual-discrimination module with the goal of having the virtual samples output by the generation module simultaneously pass the tests of the data distribution discriminator and the physical consistency discriminator. The equalization module is used to perform boundary sample oversampling equalization on the generated virtual fatigue dataset according to the logarithmic fatigue life interval, and output the final virtual fatigue dataset.

Images