Proxy model establishment method for fatigue strength of steel
By constructing random forest and artificial neural network models, and combining chemical composition and processing parameters, a surrogate model was established, which solved the problem of insufficient accuracy in predicting the fatigue strength of alloy materials, and achieved a more efficient and accurate fatigue performance assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG JINDUN PRESSURE VESSEL
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-28
AI Technical Summary
Existing traditional methods are insufficient to accurately reflect the differences in fatigue performance of alloy materials under different compositions and processing conditions, resulting in insufficient accuracy in fatigue strength prediction and failing to provide a reliable basis for engineering design and material selection.
By constructing a random forest model and an artificial neural network model, and combining chemical composition, processing parameters, and Young's modulus, a surrogate model is established to achieve feature enhancement and accurately reflect differences in fatigue performance.
It improves the accuracy of fatigue strength prediction for alloy materials, reduces testing costs and time, is applicable to more similar materials and application scenarios, and provides a reliable basis for engineering design.
Smart Images
Figure CN121938490A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of alloy materials technology, specifically to a method for establishing a proxy model for the fatigue strength of steel. Background Technology
[0002] The fatigue strength of alloy materials is a key indicator for evaluating their service performance under cyclic loading, and it plays an extremely important role in mechanical engineering, aerospace, and automotive manufacturing. Accurately understanding the fatigue properties of alloy materials is of profound significance for ensuring the safe and reliable operation of various mechanical structures and components, life assessment, and design optimization. Generally, the fatigue strength of a specific alloy material can be obtained through material fatigue testing, which is the maximum stress under a certain number of alternating loads without fatigue fracture. However, the fatigue strength of alloy materials is affected by many factors such as their chemical composition, microstructure, processing technology, and service environment, making accurate prediction of fatigue strength extremely difficult.
[0003] Existing traditional methods for predicting the fatigue strength of alloy materials cannot fully explore the correlation between chemical composition, processing parameters and material properties. They are unable to accurately reflect the differences in fatigue performance under different compositions and processing conditions, resulting in insufficient prediction accuracy and failing to provide a reliable basis for engineering design and material selection. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a method for establishing a surrogate model for the fatigue strength of steel, solving the problem that traditional prediction methods are unable to accurately reflect the differences in fatigue performance under different compositions and process conditions, resulting in insufficient prediction accuracy.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for establishing a proxy model for the fatigue strength of steel, comprising:
[0006] S1. Perform uniaxial tensile tests on alloy material samples and calculate the Young's modulus of the material; perform fatigue tests on alloy material samples and obtain the fatigue strength of the material; combine the Young's modulus, fatigue strength, chemical composition of the alloy material, and processing parameters to form a dataset.
[0007] S2. Perform data preprocessing operations such as cleaning and standardizing the dataset;
[0008] S3. Construct a random forest model, using the chemical composition and processing parameters of the alloy material as features and Young's modulus as labels, and train the random forest model to obtain a predictive surrogate model for the Young's modulus of the alloy material.
[0009] S4. Construct an artificial neural network model, using the chemical composition, processing parameters, and Young's modulus of the alloy material as features. For the missing Young's modulus of the alloy material in the dataset, use a predictive surrogate model to predict and fill it in. Use fatigue strength as a label and train the artificial neural network model to obtain a surrogate model for the fatigue strength of the alloy material.
[0010] By adopting the above technical solution, feature enhancement is achieved by introducing Young's modulus, which can fully explore the correlation information between chemical composition, processing parameters and material properties, thereby accurately reflecting the differences in fatigue performance under different composition and processing conditions, effectively improving the prediction accuracy of fatigue strength of alloy materials, and solving the problem of insufficient prediction accuracy caused by the difficulty in capturing the above correlation in traditional prediction methods.
[0011] Preferably, in S1: the uniaxial tensile test and fatigue test use the same batch of samples, and the average value is obtained through multiple tests; the chemical composition covers major and trace elements with different mass percentages in the alloy material; the processing parameters include static and dynamic parameters of heat treatment and welding.
[0012] Preferably, in step S1: the reference number of fatigue strength cycles is adjusted according to the application scenario, and the corresponding number of cycles parameter is marked in the dataset; the fatigue test determines the stress amplitude of the corresponding number of cycles by plotting the stress-life curve.
[0013] Preferably, in step S2: the dataset is first grouped by alloy material type using a clustering algorithm, and then outlier identification and missing value filling are performed separately for each group of data; after normalizing the chemical composition data, standardization processing is then performed.
[0014] Preferably, in step S3, when constructing the random forest model, a perturbation factor with small random fluctuations in feature values is introduced into the training set; when optimizing the hyperparameters of the random forest model, the feature importance distribution entropy is used as an auxiliary objective, so that the model prioritizes features that have a significant impact on Young's modulus. The hyperparameters include the number of decision trees, the maximum depth, the minimum number of samples required to split internal nodes, and the minimum number of samples required at leaf nodes.
[0015] Preferably, in step S3, the predictive proxy model for Young's modulus calculates and outputs the range of predicted value fluctuations at different confidence levels through sampling, providing an uncertainty reference for subsequent fatigue strength prediction.
[0016] Preferably, in step S4, the artificial neural network model adopts a structure that combines embedding layers, skip connections, and attention mechanisms: the input layer sets independent embedding layers for different types of features, skip connections are added between hidden layers to perform feature fusion, and dynamic weights are assigned to different features through an attention mechanism.
[0017] Preferably, in step S4, a phased training strategy is adopted. In the first phase, the basic model is trained with complete data. In the second phase, the missing data filled by the predictive proxy model is introduced for fine-tuning, and the error accumulation is reduced by adjusting the weight of the loss function.
[0018] Preferably, when new alloy material test data are obtained, only the decision tree branches in the random forest model related to the new chemical composition and new processing technology are retrained, while the node parameters of the corresponding features in the artificial neural network are adjusted.
[0019] Preferably, a proxy model establishment system for steel fatigue strength includes the following modules:
[0020] The data integration module is used to conduct uniaxial tensile tests on alloy material samples to obtain Young's modulus and fatigue tests to obtain fatigue strength, and to integrate these data with the chemical composition and processing parameters of the alloy material to form a dataset.
[0021] The data processing module is used to perform preprocessing tasks such as cleaning and standardization on the integrated dataset;
[0022] The predictive surrogate model module is used to build a random forest model. It is trained with chemical composition and processing parameters as input features and Young's modulus as output label to generate a predictive surrogate model of Young's modulus.
[0023] The fatigue strength proxy model module is used to build an artificial neural network model. It takes chemical composition, processing parameters and Young's modulus as input features and fatigue strength as output label for training. For Young's modulus missing in the dataset, it uses the predictive proxy model to predict and fill in the missing values, thus generating the fatigue strength proxy model.
[0024] The model update module is used to retrain only the decision tree branches in the random forest model that are related to the new chemical composition and new processing technology when new alloy material test data are acquired, while adjusting the node parameters of the corresponding features in the artificial neural network.
[0025] This invention provides a method for establishing a surrogate model for the fatigue strength of steel. It has the following beneficial effects:
[0026] 1. In this invention, by constructing a random forest model and an artificial neural network model, a staged prediction method is adopted. First, the Young's modulus is predicted based on the chemical composition and processing parameters. Then, the Young's modulus and the aforementioned parameters are used together as features to predict fatigue strength, thereby achieving feature enhancement to mine more related information and improving the prediction accuracy of fatigue strength of alloy materials. Compared with traditional methods, it can more accurately reflect the fatigue performance under different conditions and provide a reliable basis for engineering design and material selection.
[0027] 2. In this invention, the prediction of fatigue strength of traditional alloy materials relies on a large amount of experimental data, which is time-consuming and costly. By using limited experimental data to establish a random forest model and an artificial neural network model, the dependence on a large amount of experimental data is reduced. With the help of efficient training and prediction of machine learning models, fatigue strength assessment can be completed in a short time, reducing experimental costs and time and improving engineering efficiency.
[0028] 3. In this invention, data preprocessing, cross-validation and hyperparameter optimization techniques are used in model construction to ensure model stability and generalization ability. By comprehensively considering the chemical composition, processing parameters and Young's modulus of alloy materials, the model can better adapt to different types of alloy materials and processing conditions, thus making it applicable to more similar materials and application scenarios, and has broad application value. Attached Figure Description
[0029] Figure 1 A flowchart illustrating a method for establishing a proxy model for the fatigue strength of steel according to the present invention;
[0030] Figure 2 The flowchart shows the data acquisition and integration process for establishing a proxy model of steel fatigue strength according to the present invention.
[0031] Figure 3 The data preprocessing flowchart is shown in the surrogate model establishment method for the fatigue strength of steel according to the present invention.
[0032] Figure 4 This is a flowchart of the random forest model construction process for a surrogate model establishment method for steel fatigue strength according to the present invention.
[0033] Figure 5 This is a flowchart illustrating the artificial neural network model construction process for a surrogate model establishment method for steel fatigue strength according to the present invention.
[0034] Figure 6 The system architecture diagram for establishing a proxy model for the fatigue strength of steel according to the present invention is shown. Detailed Implementation
[0035] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, 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 scope of protection of the present invention.
[0036] Example 1:
[0037] Please see the appendix Figure 1 -Appendix Figure 5This invention provides a method for establishing a surrogate model for the fatigue strength of steel, comprising:
[0038] S1. Perform uniaxial tensile tests on alloy material samples and calculate the Young's modulus of the material; perform fatigue tests on alloy material samples and obtain the fatigue strength of the material; combine the Young's modulus, fatigue strength, chemical composition of the alloy material, and processing parameters to form a dataset.
[0039] In S1: the uniaxial tensile test and fatigue test use the same batch of samples, and the average value is obtained through multiple tests; the chemical composition covers the major and trace elements with different mass percentages in the alloy material; the processing parameters include the static and dynamic parameters of heat treatment and welding.
[0040] In S1: The reference number of cycles for fatigue strength is adjusted according to the application scenario, and the corresponding number of cycles parameter is marked in the dataset; the stress amplitude for the corresponding number of cycles is determined by plotting the stress-life curve in the fatigue test.
[0041] Specifically, firstly, a uniaxial tensile test is performed on the alloy material sample to calculate the Young's modulus of the material. At the same time, a fatigue test is performed on the alloy material sample to obtain the fatigue strength of the material. The obtained Young's modulus, fatigue strength, chemical composition and processing parameters of the alloy material are combined to form a dataset.
[0042] During the S1 operation, the same batch of alloy material samples are used for uniaxial tensile tests and fatigue tests to reduce the interference of material batch differences on the test results, ensure the consistency and reliability of test data, and multiple uniaxial tensile tests and fatigue tests are performed on samples of the same specification. Finally, the average value of multiple test results is taken as the final Young's modulus and fatigue strength value to reduce the influence of random errors during the test.
[0043] The chemical composition involved in S1 covers major and trace elements with different mass percentages in alloy materials, specifically including the mass percentages of metallic or non-metallic elements such as C, Si, P, S, Fe, Ni, and Cu. Since the performance of alloy materials is determined by the various elements they contain, these elements will affect the microstructure and macroscopic properties of alloy materials to varying degrees, thereby affecting their fatigue strength. Incorporating the information of these elements into the dataset helps to more comprehensively explore the correlation between material composition and fatigue strength.
[0044] Processing parameters include static and dynamic parameters of processes such as heat treatment, welding, surface treatment, and additive manufacturing. Static parameters of heat treatment processes include normalizing temperature and quenching temperature, while dynamic parameters include quenching time and tempering time. Static parameters of welding processes include welding current and welding voltage, while dynamic parameters include welding speed. Since processing technology is one of the key factors affecting the properties of alloy materials, different process parameters will lead to differences in the internal microstructure of the material, thereby changing its fatigue properties. Incorporating these parameters into the dataset will reflect the influence of processing technology on the fatigue strength of alloy materials.
[0045] The reference number of cycles for fatigue intensity should be adjusted according to the actual application scenario; for example, 10 cycles can be selected. 7 The cycle number is used as a reference point, and the corresponding cycle number parameter is clearly marked in the dataset to accurately associate fatigue strength data under different cycle numbers in subsequent model training and application. The fatigue test determines the stress amplitude of the corresponding cycle number by plotting the stress-life curve, which is the fatigue strength. Specifically, fatigue specimens of alloy materials are prepared according to test standards such as GB / T4337, and room temperature fatigue tests are carried out using a fatigue testing machine. The fracture cycle number of each specimen and the stress-strain data during the test are recorded, the stress amplitude of the specimen is calculated, and the stress-life curve is plotted. The corresponding stress amplitude is determined on the curve according to the selected reference cycle number, which is the fatigue strength of the alloy material.
[0046] S2. Perform data preprocessing operations such as cleaning and standardizing the dataset;
[0047] In S2: First, the dataset is grouped by alloy material type using a clustering algorithm. Then, outlier identification and missing value imputation are performed separately for each group of data. After normalizing the chemical composition data, standardization processing is then performed.
[0048] Specifically, in the data preprocessing process, the dataset is first grouped by alloy material type using a clustering algorithm. Since different types of alloy materials have significant differences in chemical composition, microstructure and performance, targeted processing after grouping by type can reduce interference between different types of material data, making subsequent outlier identification and missing value imputation more targeted, thereby improving the accuracy of data processing. For each group of data, outlier identification and missing value imputation operations need to be performed separately.
[0049] For outlier identification, commonly used statistical methods in the industry are adopted. By analyzing the distribution characteristics of the data, data that deviates from the normal range are identified as outliers and deleted to avoid the adverse effects of outlier data on model training.
[0050] For missing values, the statistical median is used to fill them in, which can preserve the overall distribution characteristics of the data and reduce data bias caused by improper handling of missing values. For the chemical composition data in the dataset, normalization is performed first, and then standardization is performed. Normalization of chemical composition data is performed because the mass percentage of different elements in alloy materials varies greatly. Normalization can convert the content of each element into a relative proportion, making the proportion information of different elements more comparable and making it easier for the model to better capture the correlation between chemical components.
[0051] After normalization, all characteristic data, including chemical composition, processing parameters, and Young's modulus, are standardized. Because these characteristic data vary significantly in magnitude and dimension—for example, the mass percentage of chemical composition is usually expressed as a percentage, while process parameters such as temperature and time have their own units and magnitudes—standardization converts these data into a distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: ;in and These are the feature data before and after standardization, respectively. and The mean and standard deviation of the data are used to eliminate the influence of differences in units and orders of magnitude, which helps the model algorithm run stably and accelerates the convergence speed.
[0052] S3. Construct a random forest model, using the chemical composition and processing parameters of the alloy material as features and Young's modulus as labels, and train the random forest model to obtain a predictive surrogate model for the Young's modulus of the alloy material.
[0053] In S3, when constructing the random forest model, a perturbation factor with small random fluctuations in feature values is introduced into the training set. When optimizing the hyperparameters of the random forest model, the feature importance distribution entropy is used as an auxiliary objective to make the model prioritize features that have a significant impact on Young's modulus. The hyperparameters include the number of decision trees, the maximum depth, the minimum number of samples required to split internal nodes, and the minimum number of samples required at leaf nodes.
[0054] In S3, the predictive proxy model for Young's modulus calculates and outputs the range of predicted values at different confidence levels through sampling, providing an uncertainty reference for subsequent fatigue strength prediction.
[0055] Specifically, when building a random forest model, it is necessary to introduce a perturbation factor with small random fluctuations in feature values into the training set. This simulates the subtle changes that feature parameters may occur in actual working conditions, enhances the model's adaptability to data perturbations, and thus improves the model's generalization performance, avoiding overfitting of the model to specific values in the training data.
[0056] When optimizing the hyperparameters of the random forest model, the feature importance distribution entropy is used as an auxiliary objective. The feature importance distribution entropy can reflect the balance of the model's attention to each feature. Using this as an auxiliary objective can guide the model to focus on features that have a significant impact on Young's modulus during the parameter optimization process, thereby more accurately capturing the correlation between key factors and Young's modulus.
[0057] Among them, the hyperparameters to be optimized include the number of decision trees, the maximum depth, the minimum number of samples required to split internal nodes, and the minimum number of samples required at leaf nodes. These parameters are used to determine the structural complexity and learning ability of the model. By setting them reasonably, the model can achieve a balance between fitting effect and generalization ability. The Young's modulus prediction surrogate model needs to be calculated by sampling to output the fluctuation range of the predicted value under different confidence levels. This can quantify the uncertainty of the prediction results and provide a reference for subsequent fatigue strength prediction, making it easier for engineers to make decisions based on the reliability of the prediction results in practical applications.
[0058] S4. Construct an artificial neural network model, using the chemical composition, processing parameters, and Young's modulus of the alloy material as features. For the missing Young's modulus of the alloy material in the dataset, use a predictive surrogate model to predict and fill it in. Use fatigue strength as a label and train the artificial neural network model to obtain a surrogate model for the fatigue strength of the alloy material.
[0059] In S4, the artificial neural network model adopts a structure that combines embedding layers, skip connections, and attention mechanisms: the input layer sets up independent embedding layers for different types of features, skip connections are added between hidden layers to perform feature fusion, and dynamic weights are assigned to different features through the attention mechanism.
[0060] In S4, a phased training strategy is adopted. In the first phase, the basic model is trained with complete data. In the second phase, the missing data of the predictive agent model is introduced for fine-tuning. The error accumulation is reduced by adjusting the weight of the loss function.
[0061] Specifically, when constructing the artificial neural network model, the chemical composition, processing parameters, and Young's modulus of the alloy material are used as features, and fatigue strength is used as a label. The dataset is divided into training and testing sets in an 8:2 ratio for model training and performance evaluation, respectively. For the missing Young's modulus values of the alloy material in the dataset, a predictive surrogate model is used to predict and fill in the missing values, ensuring the completeness of the feature data input to the model. The Young's modulus of the alloy material reflects the microstructure and performance characteristics of the material, and also affects the deformation and stress of the material under fatigue load. Adding the Young's modulus to the features enhances the features and helps the neural network model to explore the correlation between the chemical composition and macroscopic mechanical properties of the alloy material from multiple dimensions, thereby enhancing the interpretability of the model.
[0062] The artificial neural network model adopts a structure that combines embedding layers, skip connections and attention mechanisms. The input layer sets up independent embedding layers for different types of features, converting various features into vector forms that are more suitable for neural network processing. Skip connections are added between hidden layers to fuse features, promote the interaction and integration of features at different levels, and dynamically assign weights to different features through the attention mechanism, so that the model focuses on features that have a significant impact on fatigue intensity.
[0063] An artificial neural network model is built using the open-source PyTorch framework in the Python platform. The model structure includes an input layer, hidden layers, and an output layer. The number of neurons in the input layer is the same as the number of features. To explore non-linear relationships, 2 to 3 hidden layers can be selected. The ReLU activation function is used. It is used to perform nonlinear transformation on the output of the hidden layer of a neural network. In artificial neural networks, the output of each neuron is a linear combination of the input and weights. However, the relationship between the chemical composition, processing parameters and fatigue strength of alloy materials is nonlinear. By using the ReLU function to map input values less than 0 to 0, and input values greater than or equal to 0 remain unchanged, nonlinear characteristics can be introduced into the neural network, enabling the model to fit nonlinear relationships and thus better capture the complex mapping law between the multidimensional characteristics of alloy materials and fatigue strength.
[0064] In the process of constructing an artificial neural network model to obtain a surrogate model for the fatigue strength of alloy materials, there is a risk of overlearning noise or details in the training data, leading to overfitting. That is, the model performs well on the training set but poorly on the unseen test set. To avoid model overfitting, L1 and L2 regularization are introduced in the optimizer, and the corresponding loss function formula is as follows: ;
[0065] In the above formula, Mean Squared Error (MSE) is used to measure the model's predicted values. Compared with the true value The overall deviation between them, of which The sample size directly reflects the model's fit to the training data. For L1 regularization, For L2 regularization terms, where The number of model parameters, For each weight parameter in the model, and These are the L1 and L2 regularization parameters, respectively.
[0066] L1 regularization penalizes the sum of the absolute values of the weight parameters, causing some weight parameters to approach 0, thus achieving the effect of feature selection and simplifying the model structure; L2 regularization penalizes the sum of the squares of the weight parameters, which can cause the weight parameters to tend to a smaller value overall, avoiding the model from over-relying on certain features and enhancing the model's generalization ability.
[0067] and The value of is used to control the contribution ratio of the regularization term to the total loss, and is optimized and adjusted through methods such as grid search. and If the value is too large, it will excessively suppress the model's fitting ability, leading to underfitting; if the value is too small, it will not effectively prevent overfitting. Therefore, it is necessary to set the value reasonably according to the actual data and model performance.
[0068] By adopting the above loss function, while ensuring that the model has a certain fitting ability to the training data, the complexity of the model can be effectively constrained, and overfitting can be avoided. This makes the trained artificial neural network model have better stability and generalization ability when predicting the fatigue strength of alloy materials, and can be applied more reliably to actual engineering scenarios.
[0069] A phased training strategy is adopted. In the first phase, the basic model is trained with complete data to enable the model to initially learn the correlation between features and fatigue strength. In the second phase, missing data is introduced for fine-tuning. Error accumulation is reduced by adjusting the weights of the loss function. The optimal model structure and parameter combination are obtained through grid search and cross-validation. The final model is trained and evaluated on the test set. The final fatigue strength model of alloy materials consists of a Young's modulus prediction surrogate model and a fatigue strength surrogate model.
[0070] Given the chemical composition and processing parameters of the alloy material, the Young's modulus is first predicted using a Young's modulus prediction surrogate model. Then, it is input together with the chemical composition and processing parameters into a fatigue strength surrogate model to obtain the fatigue strength prediction result.
[0071] When acquiring new alloy material test data, only the decision tree branches in the random forest model related to the new chemical composition and new processing technology are retrained, while the node parameters of the corresponding features in the artificial neural network are adjusted.
[0072] Specifically, when acquiring new alloy material test data, the existing random forest model and artificial neural network model are adjusted accordingly, rather than retraining the entire model. For the random forest model, the decision tree branches related to new chemical composition and new processing technology are retrained. Since the random forest model is composed of multiple decision trees, each decision tree branch corresponds to different feature division rules. New chemical composition or processing technology parameters will only affect the decision tree branches related to them. Other branches unrelated to the new features can still maintain good applicability. By retraining only the relevant branches, the model's adaptability to new data is ensured, while the amount of computation is greatly reduced and the model update efficiency is improved.
[0073] For artificial neural network models, the parameters of nodes associated with corresponding features are adjusted. In artificial neural network models, different input features correspond to different input nodes and subsequent hidden layer nodes connected to them. New chemical composition or processing parameters are used as new features, affecting the weights and biases of the nodes associated with them in the calculation process. By adjusting the parameters of the corresponding nodes, the neural network model can learn the relationship between new features and fatigue intensity, ensuring that the model can maintain good predictive performance after incorporating new data, and without having to re-optimize all the parameters of the entire network, further improving the efficiency of model updates.
[0074] Example 2:
[0075] In the production of alloy structural components, it is necessary to quickly assess the fatigue strength of steel to ensure the safe service of components, taking into account fluctuations in chemical composition and adjustments in heat treatment parameters across different batches. However, traditional methods rely on empirical formulas or physical modeling, which struggle to accurately capture the impact of composition and process parameter changes on fatigue performance. Furthermore, when the composition or process parameters of a new batch of material exceed historical data ranges, numerous fatigue tests must be conducted again, resulting in low assessment efficiency and high costs, failing to meet the production line's need for rapid verification of material properties. To address these issues, this invention provides a proxy model establishment system for steel fatigue strength, the structure of which is as follows: Figure 6 As shown. The specific implementation process of this system is as follows:
[0076] The data integration module is used to conduct uniaxial tensile tests on alloy material samples to obtain Young's modulus and fatigue tests to obtain fatigue strength, and to integrate these data with the chemical composition and processing parameters of the alloy material to form a dataset.
[0077] The data processing module is used to perform preprocessing tasks such as cleaning and standardization on the integrated dataset;
[0078] The predictive surrogate model module is used to build a random forest model. It is trained with chemical composition and processing parameters as input features and Young's modulus as output label to generate a predictive surrogate model of Young's modulus.
[0079] The fatigue strength proxy model module is used to build an artificial neural network model. It takes chemical composition, processing parameters and Young's modulus as input features and fatigue strength as output label for training. For Young's modulus missing in the dataset, it uses the predictive proxy model to predict and fill in the missing values, thus generating the fatigue strength proxy model.
[0080] The model update module is used to retrain only the decision tree branches in the random forest model that are related to the new chemical composition and new processing technology when new alloy material test data are acquired, while adjusting the node parameters of the corresponding features in the artificial neural network.
[0081] Specifically, the data integration module is used to conduct uniaxial tensile tests on alloy material samples to obtain Young's modulus and fatigue tests to obtain fatigue strength. It integrates these data with the chemical composition and processing parameters of the alloy material to form a dataset. The uniaxial tensile test and fatigue test use the same batch of samples and take the average value of multiple tests. The fatigue strength is determined by plotting the stress-life curve to determine the stress amplitude of the corresponding number of cycles.
[0082] The data processing module is used to preprocess the integrated dataset. First, it groups the data according to the alloy material type using a clustering algorithm. Then, it performs outlier identification and missing value filling separately for each group of data. The chemical composition data is first normalized and then standardized together with other feature data.
[0083] The predictive surrogate model module is used to build a random forest model. It is trained with chemical composition and processing parameters as input features and Young's modulus as output label. When building the model, a perturbation factor with small random fluctuations in feature values is introduced. When optimizing hyperparameters, the feature importance distribution entropy is used as an auxiliary objective. The generated model outputs the range of predicted value fluctuations at different confidence levels.
[0084] The fatigue strength proxy model module is used to build an artificial neural network model. It uses chemical composition, processing parameters and Young's modulus as input features and fatigue strength as a label for training. The model adopts a structure that combines embedding layers, skip connections and attention mechanisms. It is trained in stages and built on the Python platform by calling the PyTorch framework. It uses a loss function with L1 and L2 regularization to avoid overfitting.
[0085] When acquiring new data, the model update module only retrains the decision tree branches in the random forest model that are related to new chemical components and new processing techniques, while adjusting the node parameters of the corresponding features in the artificial neural network.
[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for establishing a surrogate model for the fatigue strength of steel, characterized in that, include: S1. Perform a uniaxial tensile test on the alloy material sample and calculate the Young's modulus of the material. Fatigue tests were conducted on alloy material samples to obtain the fatigue strength of the materials. The Young's modulus, fatigue strength, chemical composition of alloy materials, and processing parameters are combined to form a dataset. S2. Perform data preprocessing operations such as cleaning and standardizing the dataset; S3. Construct a random forest model, using the chemical composition and processing parameters of the alloy material as features and Young's modulus as labels, and train the random forest model to obtain a predictive surrogate model for the Young's modulus of the alloy material. S4. Construct an artificial neural network model, using the chemical composition, processing parameters, and Young's modulus of the alloy material as features. For the missing Young's modulus of the alloy material in the dataset, use a predictive surrogate model to predict and fill it in. Use fatigue strength as a label and train the artificial neural network model to obtain a surrogate model for the fatigue strength of the alloy material.
2. The method for establishing a surrogate model for the fatigue strength of steel according to claim 1, characterized in that, In S1: the uniaxial tensile test and fatigue test use the same batch of samples, and the average value is obtained through multiple tests; the chemical composition covers the major and trace elements with different mass percentages in the alloy material; the processing parameters include static and dynamic parameters of heat treatment and welding.
3. The method for establishing a surrogate model for the fatigue strength of steel according to claim 1, characterized in that, In S1: the reference number of fatigue strength cycles is adjusted according to the application scenario, and the corresponding number of cycles parameter is marked in the dataset; the fatigue test determines the stress amplitude of the corresponding number of cycles by plotting the stress-life curve.
4. The method for establishing a surrogate model for the fatigue strength of steel according to claim 1, characterized in that, In S2: First, the dataset is grouped according to alloy material type using a clustering algorithm, and then outlier identification and missing value filling are performed separately for each group of data; after normalizing the chemical composition data, standardization processing is then performed.
5. The method for establishing a surrogate model for the fatigue strength of steel according to claim 1, characterized in that, In S3, when constructing the random forest model, a perturbation factor with small random fluctuations in feature values is introduced into the training set; when optimizing the hyperparameters of the random forest model, the feature importance distribution entropy is used as an auxiliary objective to make the model prioritize features that have a significant impact on Young's modulus. The hyperparameters include the number of decision trees, the maximum depth, the minimum number of samples required to split internal nodes, and the minimum number of samples required at leaf nodes.
6. The method for establishing a surrogate model for the fatigue strength of steel according to claim 1, characterized in that, In step S3, the Young's modulus prediction proxy model calculates and outputs the fluctuation range of the predicted value under different confidence levels through sampling, providing an uncertainty reference for subsequent fatigue strength prediction.
7. The method for establishing a surrogate model for the fatigue strength of steel according to claim 1, characterized in that, In S4, the artificial neural network model adopts a structure that combines embedding layers, skip connections, and attention mechanisms: the input layer sets independent embedding layers for different types of features, skip connections are added between hidden layers to perform feature fusion, and dynamic weights are assigned to different features through an attention mechanism.
8. The method for establishing a surrogate model for the fatigue strength of steel according to claim 1, characterized in that, In S4, a phased training strategy is adopted. In the first phase, the basic model is trained with complete data. In the second phase, the missing data of the predictive agent model is introduced for fine-tuning. The error accumulation is reduced by adjusting the weight of the loss function.
9. The method for establishing a surrogate model for the fatigue strength of steel according to claim 1, characterized in that, When acquiring new alloy material test data, only the decision tree branches in the random forest model related to the new chemical composition and new processing technology are retrained, while the node parameters of the corresponding features in the artificial neural network are adjusted.
10. A surrogate model establishment system for the fatigue strength of steel, characterized in that, The system for establishing a surrogate model for the fatigue strength of steel according to any one of claims 1-9 includes the following modules: The data integration module is used to conduct uniaxial tensile tests on alloy material samples to obtain Young's modulus and fatigue tests to obtain fatigue strength, and to integrate these data with the chemical composition and processing parameters of the alloy material to form a dataset. The data processing module is used to perform preprocessing tasks such as cleaning and standardization on the integrated dataset; The predictive surrogate model module is used to build a random forest model. It is trained with chemical composition and processing parameters as input features and Young's modulus as output label to generate a predictive surrogate model of Young's modulus. The fatigue strength proxy model module is used to build an artificial neural network model. It takes chemical composition, processing parameters and Young's modulus as input features and fatigue strength as output label for training. For Young's modulus missing in the dataset, it uses the predictive proxy model to predict and fill in the missing values, thus generating the fatigue strength proxy model. The model update module is used to retrain only the decision tree branches in the random forest model that are related to the new chemical composition and new processing technology when new alloy material test data are acquired, while adjusting the node parameters of the corresponding features in the artificial neural network.