Steel fatigue performance migration prediction method and device, medium and computer equipment

Abstract

This application discloses a method, apparatus, medium, and computer equipment for predicting the transfer of fatigue performance in steel. The method includes: acquiring a tensile dataset of steel (including alloy composition, heat treatment parameters, and target tensile strength values) and a fatigue dataset of steel (including target fatigue life values ​​for the same samples). The tensile dataset is used as the source domain, and the fatigue dataset is used as the target domain. A tensile strength prediction model for the source domain is constructed and trained using the target tensile strength value as a supervisory signal; a fatigue life prediction model for the target domain is constructed and trained using the target fatigue life value as a supervisory signal. During prediction, the tensile strength value of the steel to be tested is predicted using the source domain model, and this value is used as a dynamic feature input into the fatigue life prediction model for the target domain. The fatigue life value is obtained through transfer prediction. By using the tensile strength value as an intermediary variable, a complete prediction link of material composition-mechanical properties-fatigue behavior is established, alleviating the problem of insufficient fatigue data and effectively predicting the fatigue performance of steel.

Description

Technical Field

[0001] This application relates to the field of steel fatigue performance calculation technology, and in particular to a method, device, medium, and computer equipment for predicting the migration of steel fatigue performance. Background Technology

[0002] Steel fatigue refers to the phenomenon where steel, under long-term alternating stress or strain, will eventually fracture due to cumulative damage, even if the stress is below its yield strength. This type of failure is sudden, insidious, and cumulative, and is one of the main forms of steel structure failure. Fatigue performance is determined by various factors, including the steel's chemical composition, manufacturing process, loading factors, and service environment. Fatigue strength refers to the maximum alternating stress amplitude that a material can withstand without fatigue failure within a specified number of cycles (life). It is the core indicator of steel fatigue performance, directly determining the material's resistance to fatigue failure under cyclic loading. Its influence is reflected in multiple dimensions, including lifespan, crack propagation rate, and design safety margin. Summary of the Invention

[0003] In view of this, this application provides a method, device, medium, and computer equipment for predicting the fatigue performance migration of steel. By using tensile strength value as an intermediary variable, a complete prediction link of material composition-mechanical properties-fatigue behavior is established, which alleviates the problem of insufficient fatigue data and can effectively predict the fatigue performance of steel.

[0004] According to one aspect of this application, a method for predicting the migration of fatigue properties in steel is provided, the method comprising:

[0005] A steel tensile dataset and a steel fatigue dataset are obtained separately. The steel tensile dataset includes the alloy composition, heat treatment parameters and tensile strength target values ​​of various steel samples. The steel fatigue dataset includes the fatigue life target values ​​of the same steel samples.

[0006] The steel tensile dataset is used as the source domain for evaluating the tensile strength of steel, and the steel fatigue dataset is used as the target domain for evaluating the fatigue life of steel.

[0007] A source domain tensile strength prediction model is constructed for the source domain, and the source domain tensile strength prediction model is trained based on the steel tensile dataset. During the training process, the tensile strength target value in the steel tensile dataset is used as the source domain supervision signal.

[0008] A target domain fatigue life prediction model is constructed for the target domain. The target domain fatigue life prediction model is trained based on the steel fatigue dataset. During the training process, the fatigue life target value in the steel fatigue dataset is used as the target domain supervision signal.

[0009] The trained source domain tensile strength prediction model is used to predict the tensile strength values ​​of the steel to be predicted based on its alloy composition and heat treatment process parameters. The predicted tensile strength values ​​are then used as dynamic features to input into the trained target domain fatigue life prediction model. This allows the trained target domain fatigue life prediction model to predict the fatigue life value of the steel to be predicted based on the input dynamic features. Specifically, the source domain tensile strength prediction model is used to capture the nonlinear relationship between the alloy composition, heat treatment parameters, and tensile strength of the steel to be predicted. The target domain fatigue life prediction model is used to predict the fatigue life value based on the tensile strength value predicted by the source domain tensile strength prediction model.

[0010] According to another aspect of this application, a device for predicting the migration of fatigue properties in steel is provided, the device comprising:

[0011] The data acquisition module is used to acquire steel tensile dataset and steel fatigue dataset respectively. The steel tensile dataset includes the alloy composition, heat treatment parameters and tensile strength target values ​​of various steel samples. The steel fatigue dataset includes the fatigue life target values ​​of the same steel samples.

[0012] The source domain and target domain determination module is used to use the steel tensile dataset as the source domain for evaluating the tensile strength of the steel, and the steel fatigue dataset as the target domain for evaluating the fatigue life of the steel.

[0013] The source domain model training module is used to construct a source domain tensile strength prediction model for the source domain and train the source domain tensile strength prediction model based on the steel tensile dataset. During the training process, the tensile strength target value in the steel tensile dataset is used as the source domain supervision signal.

[0014] The target domain model training module is used to construct a target domain fatigue life prediction model for the target domain. It trains the target domain fatigue life prediction model based on the steel fatigue dataset and uses the fatigue life target value in the steel fatigue dataset as the target domain supervision signal during the training process.

[0015] The strength transfer fatigue life prediction module is used to predict the tensile strength value of the steel to be predicted based on the alloy composition and heat treatment process parameters of the steel under test using a trained source domain tensile strength prediction model. The predicted tensile strength value is then used as a dynamic feature input to a trained target domain fatigue life prediction model, so that the trained target domain fatigue life prediction model can predict the fatigue life value of the steel under test based on the input dynamic features. The source domain tensile strength prediction model is used to capture the nonlinear relationship between the alloy composition, heat treatment parameters and tensile strength of the steel under test. The target domain fatigue life prediction model is used to transfer and predict the fatigue life value based on the tensile strength value predicted by the source domain tensile strength prediction model.

[0016] According to another aspect of this application, a medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting the migration of fatigue properties in steel.

[0017] According to another aspect of this application, a computer device is provided, including a medium, a processor, and a computer program stored on the medium and executable on the processor, wherein the processor executes the program to implement the above-described method for predicting the migration of fatigue properties in steel.

[0018] By employing the above technical solution, this application provides a method, apparatus, medium, and computer equipment for predicting the transfer of fatigue performance in steel. This method acquires a tensile dataset (including alloy composition, heat treatment parameters, and target tensile strength values) and a fatigue dataset (including target fatigue life values ​​for the same samples). The tensile dataset is used as the source domain, and the fatigue dataset as the target domain. A tensile strength prediction model for the source domain is constructed and trained using the target tensile strength value as a supervisory signal. A fatigue life prediction model for the target domain is also constructed and trained using the target fatigue life value as a supervisory signal. During prediction, the tensile strength value of the steel to be tested is predicted using the source domain model, and this value is used as a dynamic feature input into the target domain fatigue life prediction model. The fatigue life value is then obtained through transfer prediction. By using the tensile strength value as an intermediary variable, a complete prediction link between material composition, mechanical properties, and fatigue behavior is established, alleviating the problem of insufficient fatigue data and effectively predicting the fatigue performance of steel.

[0019] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0021] Figure 1 A flowchart illustrating a method for predicting the migration of fatigue properties in steel, as provided in an embodiment of this application, is shown.

[0022] Figure 2 This illustration shows a schematic diagram of the prediction results of the adversarial strength (source domain) using the TrAdaBoost.R2 transfer learning algorithm provided in an embodiment of this application, with the gradient boosting decision tree as the base learner;

[0023] Figure 3 This illustration shows a schematic diagram of the prediction process of a target domain fatigue life prediction model provided in an embodiment of this application;

[0024] Figure 4 This illustration shows a prediction result of fatigue intensity (target domain) using the TrAdaBoost.R2 transfer learning algorithm and gradient boosting decision tree as the base learner, according to an embodiment of this application.

[0025] Figure 5 A schematic diagram of the structure of a steel fatigue performance migration prediction device provided in an embodiment of this application is shown. Detailed Implementation

[0026] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0027] This embodiment provides a method for predicting the migration of fatigue properties in steel, such as... Figure 1 As shown, the method includes:

[0028] Step 101: Obtain steel tensile dataset and steel fatigue dataset respectively. The steel tensile dataset includes the alloy composition, heat treatment parameters and tensile strength target values ​​of various steel samples. The steel fatigue dataset includes the fatigue life target values ​​of the same steel samples.

[0029] In the above embodiments of this application, fatigue testing is currently time-consuming and expensive, and a test may take several days or even weeks. Furthermore, fatigue strength data is relatively scarce, especially for new materials or complex loading conditions. Considering that tensile strength and fatigue strength have a good correlation, and that tensile testing has the advantages of being fast, relatively inexpensive, highly standardized, and having abundant and easily obtainable data, fatigue strength can be predicted by using tensile strength.

[0030] With the rapid development of materials science and computing technology, machine learning techniques have been introduced into fatigue performance prediction research. Machine learning methods, with their ability to handle multivariate data and diverse algorithms, have made significant progress in predicting and analyzing fatigue strength. In the embodiments described above, fatigue performance can be predicted based on a tensile performance dataset. The TrAdaBoost.R2 transfer learning algorithm is used, selecting a gradient boosting decision tree as the base learner, with tensile performance data as the source domain and fatigue performance data as the target domain. The model's predictive performance is then evaluated; that is, a transfer learning strategy is designed to predict fatigue strength.

[0031] First, relevant data on steel can be collected from literature and databases. Methods including missing value handling, outlier detection, and standardization are used to screen and clean the raw data to form a dataset of steel tensile and fatigue properties. This data can include tensile property indices (tensile strength), fatigue property indices (fatigue strength), chemical elemental composition (nine chemical elements such as C, Cr, Mn, Mo, and P), and heat treatment process parameters (normalizing temperature, quenching temperature, and tempering temperature). This data serves as the basis for subsequent analysis. Specifically:

[0032] For obtaining tensile data sets of steel, such as in experiments on high-performance building structural steel, the test steel (e.g., WGJ490DNHO steel) is smelted in a vacuum induction furnace, and its chemical composition (C, Si, Mn, Cr, Ni, etc.), heat treatment process (quenching temperature, tempering time), and tensile properties (tensile strength, yield strength, elongation after fracture) are recorded. Examples of the recorded data include:

[0033] Alloy composition: C = 0.133 wt%, Si = 0.27 wt%, Mn = 1.37 wt%, Cr = 0.008 wt%.

[0034] Heat treatment parameters: quenching temperature 850℃, tempering temperature 500℃, holding time 2 hours.

[0035] Tensile strength: The tensile strength was measured to be 570 MPa through a tensile test.

[0036] For example, when conducting tensile tests on reinforcing bars, the data recorded is as follows:

[0037] Alloy composition: C = 0.25 wt%, Mn = 1.50 wt%, Si = 0.80 wt%.

[0038] Heat treatment parameters: hot rolling process, no additional heat treatment.

[0039] Tensile strength: The measured tensile strength is 540 MPa and the yield strength is 400 MPa.

[0040] For acquiring fatigue datasets for steel, such as performing strain-controlled fatigue tests on the same batch of steel (e.g., WGJ490DNHO steel), the data is recorded as follows:

[0041] Fatigue life: When the number of cycles reaches 1×10⁶ under a stress amplitude of 200MPa without failure, it is defined as "infinite life".

[0042] Test conditions: frequency 10Hz, temperature 25℃, load mode tension-compression cycle.

[0043] Next, data cleaning and preprocessing are performed. This involves sequentially handling missing values, detecting outliers, and standardizing the initially collected data to improve data quality and consistency. Missing value handling includes removing all records containing missing or null values. Outlier handling involves using box plots and histograms to detect and process outliers. Standardization involves applying Z-scores to each feature of the dataset, while retaining the original values ​​for tensile strength and fatigue strength. Furthermore:

[0044] For missing value handling, in the tensile dataset, if the tensile strength of a sample is missing, KNN imputation (based on the mean of neighboring samples) can be used to fill the gap. In the fatigue dataset, if fatigue life is missing, it can be replaced by the average life of samples from the same batch.

[0045] For outlier detection, the IQR method can be used to detect outliers in tensile strength for tensile datasets. For example, a sample with a tensile strength of -50 MPa (clearly unreasonable) is identified as an outlier and removed. For fatigue datasets, the Z-score method can be used to identify outliers in fatigue life; for example, a sample with a fatigue life of 10 cycles (far below the mean) is marked as an outlier.

[0046] For standardization, the contents of elements such as C, Mn, and Si can be Z-score standardized for alloy composition to achieve a mean of 0 and a variance of 1. For heat treatment parameters, the quenching temperature can be normalized to the [0, 1] range, for example, 850℃ can be normalized to 0.85.

[0047] In particular, outlier detection can also utilize box plots and histogram analysis methods to identify and handle data outliers. First, the quartiles of each data point are calculated, and based on the interquartile range (IQR), an outlier limit of 1.5 times the IQR is defined, removing outlier data points that exceed the range. For tensile strength and fatigue strength, their normal ranges are determined and outlier data points are removed. Histograms and fitted curves are used to display the frequency distribution and clarify the outlier threshold.

[0048] Next, data integration and structuring are performed, specifically including:

[0049] Data matching requires ensuring a one-to-one correspondence between samples in the tensile dataset and the fatigue dataset. For example, the experimental steel WGJ490DNHO is numbered S-001 in the tensile dataset and is also labeled S-001 in the fatigue dataset. A steel tensile dataset, for example:

[0050] Table 1

[0051]

[0052] Steel fatigue datasets, for example:

[0053] Table 2

[0054] S-001 200 10Hz 25℃ 1e6

[0055] To this end, a steel tensile dataset and a steel fatigue dataset are constructed so that a tensile strength prediction model (such as a gradient boosting tree) can be trained using the cleaned tensile dataset. The input is alloy composition and heat treatment parameters, and the output is tensile strength. A fatigue life prediction model is also trained based on the fatigue dataset. The input includes the tensile strength predicted from the source domain and the original alloy composition, and the output is fatigue life.

[0056] Through the above steps, steel tensile and fatigue datasets can be systematically acquired and processed, providing high-quality data support for transfer learning models.

[0057] Step 102: Use the steel tensile dataset as the source domain for evaluating the tensile strength of the steel, and use the steel fatigue dataset as the target domain for evaluating the fatigue life of the steel.

[0058] Furthermore, due to the high cost and long cycle time of fatigue life testing (e.g., high-frequency fatigue testing requires tens of thousands of cycles), the sample size of fatigue datasets is much smaller than that of tensile datasets. Utilizing the rich alloy composition-tensile strength relationships in tensile datasets to assist the learning of fatigue life prediction models can mitigate the risk of overfitting caused by insufficient data in the target domain. For example, if the fatigue dataset contains only 100 samples while the tensile dataset has 1000 samples, transfer learning can indirectly utilize the material properties in the tensile data to improve the generalization ability of the fatigue model.

[0059] Furthermore, since there is an implicit correlation between tensile strength and fatigue life (e.g., high-strength materials may become more brittle, reducing fatigue life), the alloy composition-tensile strength relationship learned by the source domain model can serve as an "intermediate feature" for the target domain model, helping it capture the complex nonlinear relationship between fatigue life and material composition. For example, if the source domain model finds that increasing C content improves tensile strength, the target domain model can further learn how high C content shortens fatigue life by affecting grain boundary structure. Directly training a fatigue life prediction model requires simultaneously learning the three-dimensional relationship between alloy composition, heat treatment parameters, and fatigue life, resulting in high computational complexity. The source domain model maps the high-dimensional input (alloy composition + heat treatment parameters) to low-dimensional tensile strength features, while the target domain model only needs to learn the two-dimensional relationship between tensile strength and fatigue life, simplifying the model structure.

[0060] Therefore, by using a source-target domain transfer learning framework, it is possible to "use low-cost stretching data to guide high-value fatigue prediction", thereby improving the efficiency of materials research and development and the reliability of products.

[0061] Step 103: Construct a source domain tensile strength prediction model for the source domain, and train the source domain tensile strength prediction model based on the steel tensile dataset. During the training process, the tensile strength target value in the steel tensile dataset is used as the source domain supervision signal.

[0062] Next, by using the tensile strength target value to provide the "correct answer" for the model, the mapping relationship between input features (alloy composition, heat treatment parameters) and output target (tensile strength) should be clearly defined. For example, if a sample alloy composition is C=0.2%, Mn=1.5%, and the heat treatment parameters are quenching at 850℃ and tempering at 500℃, its tensile strength target value is 600MPa. The model needs to adjust the parameters to make the predicted value close to 600MPa. Based on the tensile strength target value, the deviation between the predicted value and the true value (such as the mean square error, MSE) is calculated, driving the model to adjust the parameters through backpropagation to reduce the overall error. For example, if the model predicts that the tensile strength of a sample is 580MPa, while the target value is 600MPa, the MSE is (580-600)² = 400. The model can prioritize correcting the feature weights that cause this error (such as increasing the correlation weight of Mn content or quenching temperature).

[0063] By supervising the target value of tensile strength, the source domain model can efficiently and accurately capture the nonlinear relationship between alloy composition and tensile strength, providing high-quality dynamic feature input for the subsequent target domain fatigue life prediction model, thereby realizing "using low-cost tensile data to guide high-value fatigue prediction".

[0064] Optionally, the source domain tensile strength prediction model is constructed based on a gradient boosting tree model, which includes multiple decision trees. In step 103, the source domain tensile strength prediction model is trained based on the steel tensile dataset, specifically including:

[0065] Step 1031: Encode the alloy composition in the steel tensile dataset into an alloy composition feature vector, and encode the heat treatment process parameters in the steel tensile dataset into a heat treatment process parameter feature vector.

[0066] Step 1032: Use the alloy composition feature vector and the heat treatment process parameter feature vector as input features, and the tensile strength value as the target value to construct a training dataset.

[0067] Step 1033: Initialize the hyperparameters of the gradient boosting tree model and train the initialized gradient boosting tree model using the training dataset. The hyperparameters include at least one of the following: number of decision trees, learning rate, and maximum depth.

[0068] Step 1034: During the process of training the initialized gradient boosting tree model using the training dataset, for each iteration, the gradient boosting tree model trains any new decision tree to fit the residual based on the residual between the currently predicted tensile strength value and the target value, and then superimposes the tensile strength value predicted by the new decision tree with the existing gradient boosting tree model until the preset number of iterations is reached.

[0069] In the above embodiments of this application, for example, the alloy composition of a certain steel is: C (0.25%), Si (0.40%), Mn (1.20%), Cr (0.80%). The range of each element in the statistical historical data is, for example: C (0.10-0.35%), Si (0.10-0.60%), Mn (0.80-1.60%), Cr (0.20-1.20%). Then, the alloy composition feature vector of each alloy component is calculated, for example: [(0.25-0.10) / (0.35-0.10), (0.40-0.10) / (0.60-0.10), (1.20-0.80) / (1.60-0.80), (0.80-0.20) / (1.20-0.20)] = [0.6, 0.6, 0.5, 0.6].

[0070] Encoding of heat treatment parameters (heat treatment parameter feature vector), for example:

[0071] Original parameters: quenching temperature 850℃, holding time 45min, cooling rate 18℃ / s. After standardization, the quenching temperature is (850-800) / (900-800)=0.5, the holding time is (45-30) / (60-30)=0.5, the cooling rate is (18-10) / (20-10)=0.8, and the heat treatment parameter eigenvector is [0.5, 0.5, 0.8].

[0072] The training dataset is constructed by merging feature vectors. For example, the alloy composition feature vector [0.6, 0.6, 0.5, 0.6] + the heat treatment parameter feature vector [0.5, 0.5, 0.8] is combined to obtain a single input feature [0.6, 0.6, 0.5, 0.6, 0.5, 0.5, 0.8]. The target value is the measured tensile strength of 680 MPa.

[0073] Next, initialize the hyperparameters of the gradient boosting tree model, such as number of decision trees = 100, learning rate = 0.1, maximum depth = 4, and perform the first round of training, that is, obtain the initial prediction value, such as the mean of all samples is 650 MPa, and calculate the residual: 680 - 650 = 30 MPa. Then, train the first decision tree to fit the residual distribution and update the model: 650 + 0.1 * 30 = 653 MPa. The iterative process is carried out in sequence.

[0074] Specifically, the termination condition can be set to reach a preset number of iterations of 100 or the error of the validation set (obtained from the training dataset) does not decrease for 5 consecutive rounds.

[0075] Step 104: Construct a target domain fatigue life prediction model for the target domain. Train the target domain fatigue life prediction model based on the steel fatigue dataset. During the training process, use the fatigue life target value in the steel fatigue dataset as the target domain supervision signal.

[0076] Next, when constructing the fatigue life prediction model for the target domain, we can first encode the features of the steel fatigue dataset, such as material composition and process parameters, into input vectors, and use the fatigue life target value as a supervision signal. During training, the model predicts fatigue life using the feature vector as input. By calculating the error (e.g., mean squared error) between the predicted value and the true target value, we use gradient descent to adjust the model parameters, gradually bringing the predicted value closer to the true value. The target domain supervision signal provides the true label, guiding the model to learn the mapping relationship between features and fatigue life, ultimately resulting in a model that can accurately predict the fatigue life of steel in the target domain.

[0077] Optionally, the target domain fatigue life prediction model includes an input layer and a hidden layer. The input layer includes a main branch and a newly added strength prediction branch. The main branch receives alloy composition feature vectors, and the strength prediction branch receives tensile strength values. After adjustment, these values ​​are concatenated with the alloy composition feature vectors received by the main branch to form dynamic feature input. The hidden layer is a gradient boosting tree model. In step 104, the target domain fatigue life prediction model is trained based on the steel fatigue dataset, including:

[0078] Step 1041: When training the target domain fatigue life prediction model based on the steel fatigue dataset, a two-stage feature fusion strategy is adopted, and the training process is implemented through iteration. The two-stage feature fusion strategy includes receiving the alloy composition feature vector directly as input through the main branch, receiving the tensile strength value generated by the source domain tensile strength prediction model through the strength prediction branch, and dynamically concatenating the alloy composition feature vector with the alloy composition feature vector received by the main branch to form a dynamic feature input. The dynamic feature input is then predicted and output through the gradient boosting tree model in the hidden layer.

[0079] In the above embodiments of this application, the main branch directly receives the alloy composition feature vector as a static input describing the basic properties of the material; the strength prediction branch receives the tensile strength value generated by the source domain tensile strength prediction model, forming a dynamic input related to the material strength. At this time, the features of the two branches have not yet interacted directly, only completing the independent transmission of data streams. The tensile strength value output by the strength prediction branch is concatenated with the alloy composition feature vector of the main branch according to the feature dimension to form a dynamic feature input containing material composition and strength information. The gradient boosting tree model in the hidden layer uses this dynamic input as a reference and predicts fatigue life through a weighted combination of multiple decision trees. Each tree focuses on the interaction of different feature subsets, and the final output is the cumulative value of the prediction results of all trees.

[0080] Specifically, during iterative training and feature optimization, the target domain fatigue life prediction model is initialized using a steel fatigue dataset. The main branch inputs alloy composition features, and the strength prediction branch inputs the tensile strength values ​​predicted by the source domain tensile strength prediction model. After concatenation, the fatigue life prediction value is output through a gradient boosting tree. During iteration, the residual between the predicted value and the measured fatigue life (target fatigue life value) is calculated, and the gradient boosting tree fits the residual by adding a new decision tree. As iteration progresses, the main branch adjusts the weights of the alloy composition features, and the strength prediction branch simultaneously optimizes the source domain model parameters, making the concatenated dynamic features more accurately reflect the key factors of fatigue life. Training stops when the validation set residuals converge (e.g., the residual decrease is <1% for 5 consecutive rounds) or the maximum number of iterations is reached.

[0081] Through a two-stage fusion, the target domain fatigue life prediction model utilizes both the physical characteristics of the alloy composition and introduces cross-domain transferred strength information as a dynamic supplement, thus solving the problem of insufficient information from a single feature domain. The iterative residual learning mechanism of the gradient boosting tree further strengthens the nonlinear interaction between features, improving the robustness of fatigue life prediction.

[0082] Step 105: The trained source domain tensile strength prediction model is used to predict the tensile strength values ​​of the alloy composition and heat treatment process parameters of the steel to be predicted. The predicted tensile strength values ​​are then used as dynamic features to input into the trained target domain fatigue life prediction model. This allows the trained target domain fatigue life prediction model to predict the fatigue life value of the steel to be predicted based on the input dynamic features. The source domain tensile strength prediction model is used to capture the nonlinear relationship between the alloy composition, heat treatment parameters, and tensile strength of the steel to be predicted. The target domain fatigue life prediction model is used to predict the fatigue life value based on the tensile strength value predicted by the source domain tensile strength prediction model.

[0083] Next, in the cross-domain prediction, a pre-trained source domain tensile strength prediction model is used to predict the tensile strength value of the steel to be predicted based on its alloy composition and heat treatment process parameters. This model captures the complex relationship between material composition, process, and strength through nonlinear mapping, outputting domain-adaptive strength features. Subsequently, the predicted tensile strength value is used as a dynamic feature, concatenated with the original alloy composition features, and input into a pre-trained target domain fatigue life prediction model. The target domain model, based on a gradient boosting tree structure, uses the strength information in the dynamic features as a key intermediate variable to transfer learning the nonlinear relationship between fatigue life and material properties, ultimately outputting the fatigue life value of the steel to be predicted, achieving efficient transfer from source domain knowledge to target domain tasks.

[0084] Optionally, in step 105, the prediction of the tensile strength value of the steel to be predicted using the trained source domain tensile strength prediction model includes:

[0085] Step 1051: Obtain the alloy composition feature vector corresponding to the alloy composition of the steel to be predicted, and the heat treatment process parameter feature vector corresponding to the heat treatment process parameters.

[0086] Step 1052: Normalize the alloy composition feature vector and the heat treatment process parameter feature vector respectively, and then concatenate the normalized alloy composition feature vector and the heat treatment process parameter feature vector into a single input feature.

[0087] Step 1053: Predict the tensile strength value corresponding to the single input feature using the trained source domain tensile strength prediction model.

[0088] In the above embodiments of this application, such as Figure 2 As shown, Figure 2 To use the TrAdaBoost.R2 transfer learning algorithm, gradient boosting decision trees are used as base learners to predict the adversarial strength (source domain). Figure 2 In R, Sourcedomain (TS) refers to the source domain. In the context of transfer learning or cross-domain modeling, the source domain refers to a domain with a large amount of labeled data or a known model that performs well. Here, it represents the steel tensile dataset used to train the tensile strength prediction model. Bestfitline is the best-fit line, a straight line drawn from the data points that best represents the data trend. It can be determined using methods such as least squares and is used to show the linear relationship between predicted and experimental values. Predictedvalue is the predicted value, i.e., the result calculated by the model. Experimentalvalue is the experimental value, the data obtained through actual experiments. 2=0.9851 is the coefficient of determination, ranging from 0 to 1. The closer to 1, the better the model fits the data; here, 0.9851 indicates a high degree of model fit. MAE = 12.2961 MPa is the mean absolute error, the average absolute value of the error between the predicted and experimental values, measured in megapascals (MPa), reflecting the average deviation between the predicted and experimental values. RMSE = 16.9943 MPa is the root mean square error, the square root of the average of the squares of the errors between the predicted and experimental values, also measured in megapascals (MPa). It gives greater weight to larger errors and is more sensitive to larger prediction errors.

[0089] Specifically, by decoupling and normalizing domain features, the dimensional differences between alloy composition and heat treatment parameters are eliminated, improving the stability of model training; multi-source feature splicing and fusion of material genes and process parameters enhances the ability to model composition-process coupling relationships; combined with source domain transfer learning, cross-steel tensile strength prediction is achieved, which can systematically reduce experimental costs and accelerate process optimization.

[0090] Optionally, such as Figure 3 As shown, the target domain fatigue life prediction model has an input layer, and the alloy composition of the steel to be predicted corresponds to an alloy composition feature vector. In step 105, the predicted tensile strength value is used as a dynamic feature input to the trained target domain fatigue life prediction model, including:

[0091] Step 1054: Add an intensity prediction branch to the input layer of the target domain fatigue life prediction model.

[0092] Step 1055: The tensile strength value predicted by the source domain tensile strength prediction model is linearly transformed to the same numerical range as the alloy composition feature vector through the newly added strength prediction branch. After the dimension of the transformed tensile strength value is consistent with the alloy composition feature vector by copying or padding, it is concatenated with the alloy composition feature vector to obtain the target domain fatigue performance prediction model after dynamic feature input training.

[0093] In the above embodiments of this application, such as Figure 4 As shown, Figure 4 To use the TrAdaBoost.R2 transfer learning algorithm, gradient boosting decision trees are used as base learners to predict fatigue intensity (target domain). Figure 4In R, `test` represents the test data, `train` represents the training data, and `Bestfitline` is the best-fit line, a straight line fitted to the data points in a scatter plot to show the linear relationship between two variables. `Targetdomain` (FS) represents the target domain. `Predictedvalue` is the predicted value, and `Experimentalvalue` is the experimental value, obtained through actual experiments. 2 =0.9970, MAE=2.9986MPa, RMSE=3.4480MPa are respectively: R 2 The coefficient of determination (COD) measures how well a model fits the data, ranging from 0 to 1; a value closer to 1 indicates a better fit. The mean absolute error (MAE) is the average of the absolute errors between predicted and experimental values, used to assess the accuracy of the prediction; a smaller value indicates a more accurate prediction. The root mean square error (RMSE) is the square root of the average of the sum of squared errors between predicted and experimental values, also used to measure the accuracy of the prediction; a smaller value is better.

[0094] Specifically, by constructing feature vectors using alloy composition, the material composition characteristics can be accurately characterized, providing basic feature information for the model. Adding a strength prediction branch and linearly transforming the tensile strength value enables cross-domain data alignment, allowing strength information to be adapted to composition features. By copying or filling a unified dimension and then splicing them together, multi-source features are integrated, enhancing feature expression capabilities and forming dynamic feature inputs. This helps the fatigue life prediction model in the target domain to fully utilize composition and strength information, improve the prediction accuracy of fatigue performance of different steels, and reduce experimental costs.

[0095] Optionally, in step 105, before predicting the tensile strength values ​​of the alloy composition and heat treatment process parameters of the steel to be predicted using the trained source domain tensile strength prediction model, the following steps are also included:

[0096] Step 106: The SHAP value method is used to perform global feature importance analysis on the trained source domain tensile strength prediction model. The global feature importance analysis quantifies the marginal contribution of alloy composition and heat treatment process parameters to the predicted tensile strength value. At the same time, the dynamic feature interpretation of the target domain fatigue life prediction model is performed. Based on the dynamic feature interpretation, the synergistic effect of the dynamic feature after splicing the alloy composition feature vector and the predicted tensile strength value on fatigue life is analyzed, and the SHAP value interpretation results of the source domain tensile strength prediction model and the target domain fatigue life prediction model are obtained.

[0097] Step 107: When the SHAP interpretation results of both the source domain tensile strength prediction model and the target domain fatigue life prediction model are verified and show stable feature contribution distribution, it is determined that the source domain tensile strength prediction model and the target domain fatigue life prediction model meet the interpretability requirements.

[0098] In the embodiments described above, the SHAP value analysis method can be used to assess the importance of feature variables. Model interpretability is achieved by quantifying the marginal impact of features on the prediction results. Specifically, for the source domain tensile strength prediction model, its input features include the alloy composition of the steel (e.g., the content of various elements) and heat treatment process parameters (e.g., quenching temperature, tempering time, etc.). The SHAP value is based on the Shapley value concept in game theory, calculated by considering the contribution of each feature to the model's prediction results across all possible combinations of features. For each sample and each feature, the marginal contribution of that feature to the predicted tensile strength value under all possible subset combinations of features is calculated. Then, these marginal contributions are weighted and averaged to obtain the SHAP value of that feature for that sample. By calculating the absolute average of the SHAP values ​​of all samples or other statistical measures, the marginal contribution of each alloy composition feature and heat treatment process parameter feature to the predicted tensile strength value can be quantified. For example, a large average absolute value of the SAP values ​​for carbon content characteristics indicates a significant impact of carbon content on the predicted tensile strength; conversely, a small SAP value for a heat treatment parameter suggests a relatively small impact on the prediction result. Global feature importance analysis helps understand the importance ranking of features in the source domain tensile strength prediction model, thus guiding the focus on key factors during materials development. For instance, if certain alloying elements are found to significantly improve tensile strength, their content can be prioritized in subsequent materials design.

[0099] For example, the dynamic feature input of the target domain fatigue life prediction model is the result of concatenating the alloy composition feature vector with the predicted tensile strength value. When calculating the Shapley value, it is necessary to consider the contribution of each original feature (alloy composition feature and tensile strength feature) and their interactions to the predicted fatigue life value in this concatenated dynamic feature. Similarly, based on the concept of Shapley value, the marginal contribution of each feature to the predicted fatigue life value under all possible feature combinations is calculated to obtain the corresponding Shapley value. By analyzing the Shapley value, the synergistic effect of the dynamic feature resulting from concatenating the alloy composition feature vector and the predicted tensile strength value on fatigue life can be analyzed. For example, some alloy composition features and tensile strength features may have a positive synergistic effect, meaning that when these features coexist and take specific values, they will significantly increase or decrease the predicted fatigue life value; while some features may have a negative synergistic effect.

[0100] The role of dynamic feature interpretation is to gain a deeper understanding of the decision-making basis of fatigue life prediction models in the target domain, revealing the complex relationship between alloy composition, tensile strength, and fatigue life. This helps to optimize material design and heat treatment processes to improve the fatigue life of steel. For example, if it is found that the combination of tensile strength and certain alloy compositions has a positive impact on fatigue life, these parameters can be adjusted during the production process.

[0101] Through the above embodiments of this application, the interpretation results of the SHAP values ​​of the source domain tensile strength prediction model and the target domain fatigue life prediction model can be obtained. These results not only help to understand the internal mechanism of the model and improve its interpretability, but also provide theoretical basis and guidance for the material design and process optimization of steel, thereby improving the performance and quality of steel.

[0102] Optionally, the target domain fatigue life prediction model achieves cross-domain knowledge transfer through the TrAdaBoost.R2 algorithm.

[0103] In the embodiments described above, the target domain fatigue life prediction model employs the TrAdaBoost.R2 algorithm to achieve cross-domain knowledge transfer. This algorithm effectively utilizes the abundant tensile strength-related data in the source domain (steel tensile dataset) and transfers its knowledge to the target domain (steel fatigue dataset), solving the problem of insufficient data in the target domain. By adaptively adjusting sample weights, it strengthens the learning of target domain data while rationally utilizing useful information from the source domain, thereby improving the model's prediction accuracy for fatigue life. It also enhances the model's generalization ability, enabling it to better adapt to fatigue life prediction of different steel samples and reducing experimental costs.

[0104] By applying the technical solution of this embodiment, the limitations of traditional prediction methods can be overcome. A novel and efficient fatigue strength transfer learning prediction method is constructed using data processing and machine learning techniques. Specifically, it utilizes the TrAdaBoost.R2 transfer learning algorithm with gradient boosting decision trees as base learners to conduct in-depth research on the chemical composition and heat treatment process parameters of steel. A systematic evaluation of the transfer learning model is performed, constructing a high-precision fatigue performance prediction model. Furthermore, the SHAP value analysis method is used to further explore the complex relationship between the fatigue performance of steel and its chemical composition and heat treatment process parameters. This helps identify key features affecting fatigue performance. Through feature selection and model evaluation techniques, the correlation between variables and fatigue performance is effectively demonstrated, improving the prediction accuracy of fatigue performance.

[0105] In one specific embodiment, alloy composition feature variables, heat treatment process parameters, and corresponding target variables are extracted from the tensile and fatigue performance datasets of steel. The TrAdaBoost.R2 transfer learning algorithm, with a gradient boosting decision tree as the base learner, is used. The tensile performance dataset is used as the source domain, and the fatigue performance dataset as the target domain, which are then input into the model. One-third of the data in the target domain is used as the training set, and the remainder as the test set. The target domain is partitioned using a random seed. The prediction results for tensile and fatigue performance are evaluated using the coefficient of determination (R²), mean absolute coefficient (MAE), and root mean square error (RMSE). The SHAP (SHapley Additive Ex Planations) method is used to calculate the marginal contribution of each feature to the model results, revealing the influence of different elemental features and heat treatment processes on tensile and fatigue strength. The importance of a feature is assessed by the absolute magnitude of its SHAP value; a larger SHAP value indicates a higher importance of the feature in the model.

[0106] Therefore, a fatigue performance prediction method for steel based on transfer learning was developed. This method uses the chemical composition and heat treatment parameters of steel as a basis, and combines transfer learning algorithms to predict fatigue performance. By using data processing and analysis techniques, including outlier handling, feature selection, and the TrAdaBoost.R2 machine learning transfer learning algorithm, the fatigue performance of steel can be accurately predicted based on limited experimental data, reducing the time and material costs associated with traditional experimental methods. Furthermore, it reduces reliance on experimental conditions and theoretical assumptions, improving the accuracy and efficiency of prediction. The prediction method employs SHAP value analysis to further explore the influence of chemical composition and heat treatment parameters on fatigue performance, effectively demonstrating the correlation between these variables and fatigue performance. The prediction method has the advantages of high accuracy, low cost, and rapid implementation, and has wide applicability, making it suitable for optimizing the fatigue performance evaluation of steel and other similar materials.

[0107] In another specific embodiment, this can be applied to use room temperature fatigue performance prediction as the source domain and transfer the prediction to high-temperature fatigue performance (target domain). Specifically, a room temperature fatigue dataset covering various steel samples is collected, including alloy composition, heat treatment parameters, and target values ​​for room temperature fatigue life; simultaneously, a high-temperature fatigue dataset of the same steel samples under high-temperature conditions is collected, including target values ​​for high-temperature fatigue life. The room temperature fatigue dataset is used as the source domain to evaluate the room temperature fatigue performance of the steel; the high-temperature fatigue dataset is used as the target domain to evaluate the high-temperature fatigue performance of the steel. A source domain fatigue life prediction model is constructed for the source domain (room temperature fatigue performance). The input of this model is features such as the alloy composition and heat treatment parameters of the steel, and the output is the predicted value of room temperature fatigue life. This model is trained based on the room temperature fatigue dataset, using the target value of room temperature fatigue life as the source domain supervision signal, enabling the model to learn the relationship between various features and fatigue life at room temperature. A target domain fatigue life prediction model is constructed for the target domain (high-temperature fatigue performance), considering the incorporation of knowledge transferred from the source domain. This model is trained based on the high-temperature fatigue dataset, using the target value of high-temperature fatigue life as the target domain supervision signal. Using a trained source domain fatigue life prediction model, the room temperature fatigue life of the steel to be predicted is predicted based on its relevant characteristics (alloy composition, heat treatment parameters, etc.). The predicted room temperature fatigue life-related feature information (which can be appropriately processed, such as linear transformation, to adapt it to high-temperature-related features) is used as part of the dynamic features and concatenated with other features of the steel, then input into the trained target domain fatigue life prediction model. This model predicts the high-temperature fatigue life value of the steel based on transferred knowledge and its own learned high-temperature features. Based on the difference between the predicted results and the actual high-temperature fatigue life values, the source domain and target domain models are optimized and adjusted, such as adjusting model parameters and feature selection, to improve the accuracy of the transfer prediction. Appropriate evaluation metrics (such as root mean square error, coefficient of determination, etc.) are used to evaluate the predictive effect of the target domain model on high-temperature fatigue life, ensuring the effectiveness of knowledge transfer.

[0108] In another specific embodiment, this method can also be applied to predict high-cycle fatigue performance by transferring low-cycle fatigue performance. Similarly, alloy composition, heat treatment parameters, and low-cycle fatigue life target values ​​(such as strain amplitude and cycle count) are collected for various steel samples. High-cycle fatigue life target values ​​(such as stress amplitude and cycle count) are also collected for the same steel samples. It is ensured that the samples in both datasets are completely identical in terms of basic characteristics such as alloy composition and heat treatment process, differing only in fatigue life target values ​​and test conditions. The low-cycle fatigue dataset is used as the source domain to learn the relationship between basic material characteristics and low-cycle fatigue performance. The high-cycle fatigue dataset is used as the target domain to predict high-cycle fatigue life. Specifically, a regression model (such as a neural network or random forest) can be constructed, with the steel's alloy composition and heat treatment parameters as input, and the low-cycle fatigue life prediction value as output. The model is trained based on the low-cycle fatigue dataset, using the low-cycle fatigue life target value as a supervision signal to learn the nonlinear relationship between features and low-cycle fatigue life. Then, another regression model is constructed, with a structure similar to the source domain model, but adjusted for high-cycle fatigue characteristics (such as increasing sensitivity to features like high-frequency vibration and stress concentration). The model is trained based on a high-cycle fatigue dataset, with the high-cycle fatigue life target value as the supervision signal.

[0109] Using the trained source domain model (low-cycle fatigue prediction model), the alloy composition and heat treatment parameters of the steel to be predicted are used to predict low-cycle fatigue life, obtaining predicted values. These predicted values ​​are then used as dynamic features and concatenated with the original features (alloy composition, heat treatment parameters) to form a new input feature vector. The concatenated dynamic features are input into the target domain model (high-cycle fatigue prediction model). Through transfer learning, the relationship between the basic features captured by the source domain model and low-cycle fatigue performance is utilized to assist in predicting high-cycle fatigue life. Features in the low-cycle and high-cycle fatigue datasets are standardized or normalized to ensure consistent numerical ranges. If distribution differences exist (e.g., the amount of low-cycle fatigue data is much larger than that of high-cycle fatigue), domain adaptation techniques (e.g., adversarial training, minimizing the maximum mean difference) can be used to align feature distributions. Some parameters of the source domain model (e.g., shallow network weights) are transferred to the target domain model as initialization parameters to accelerate the convergence of the target domain model. The TrAdaBoost.R2 algorithm is used to dynamically adjust sample weights, strengthening the learning of the target domain data while suppressing features in the source domain that are irrelevant to the target domain.

[0110] Therefore, when the high-cycle fatigue dataset is small, transferring low-cycle fatigue knowledge can significantly improve the prediction accuracy of the target domain model. Alloy composition and heat treatment processes affect both low-cycle and high-cycle fatigue performance. The source domain model can help the target domain model identify these common characteristics. Through the above embodiments, the reliance on high-cycle fatigue experiments can be reduced, and computational simulation can replace some experiments, accelerating the materials research and development process.

[0111] Furthermore, as Figure 1 In terms of specific implementation, this application provides a device for predicting the migration of fatigue properties in steel, such as... Figure 5 As shown, the device includes:

[0112] The data acquisition module 201 is used to acquire steel tensile dataset and steel fatigue dataset respectively. The steel tensile dataset includes the alloy composition, heat treatment parameters and tensile strength target values ​​of various steel samples. The steel fatigue dataset includes the fatigue life target values ​​of the same steel samples.

[0113] The source domain and target domain determination module 202 is used to use the steel tensile dataset as the source domain for evaluating the tensile strength of the steel, and the steel fatigue dataset as the target domain for evaluating the fatigue life of the steel.

[0114] The source domain model training module 203 is used to construct a source domain tensile strength prediction model for the source domain and train the source domain tensile strength prediction model based on the steel tensile dataset. During the training process, the tensile strength target value in the steel tensile dataset is used as the source domain supervision signal.

[0115] The target domain model training module 204 is used to construct a target domain fatigue life prediction model for the target domain. The target domain fatigue life prediction model is trained based on the steel fatigue dataset, and during the training process, the fatigue life target value in the steel fatigue dataset is used as the target domain supervision signal.

[0116] The strength transfer fatigue life prediction module 205 is used to predict the tensile strength value of the steel to be predicted based on the alloy composition and heat treatment process parameters of the steel under test using the trained source domain tensile strength prediction model, and input the predicted tensile strength value as a dynamic feature into the trained target domain fatigue life prediction model, so that the trained target domain fatigue life prediction model can predict the fatigue life value of the steel under test based on the input dynamic features. The source domain tensile strength prediction model is used to capture the nonlinear relationship between the alloy composition, heat treatment parameters and tensile strength of the steel under test, and the target domain fatigue life prediction model is used to transfer and predict the fatigue life value based on the tensile strength value predicted by the source domain tensile strength prediction model.

[0117] It should be noted that other corresponding descriptions of the functional units involved in the steel fatigue performance migration prediction device provided in this application embodiment can be found in the following references. Figures 1 to 2 The corresponding descriptions in the method will not be repeated here.

[0118] Based on the above, Figures 1 to 2 Accordingly, this application also provides a medium on which a computer program is stored, which, when executed by a processor, implements the above-described method. Figures 1 to 2The method for predicting the migration of fatigue properties in steel is shown.

[0119] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile medium (such as a CD-ROM, USB flash drive, or portable hard drive) and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0120] Based on the above, Figures 1 to 2 The method shown, and Figure 5 To achieve the above objectives, the present application also provides a computer device, specifically a personal computer, server, network device, etc., as shown in the virtual device embodiment. This computer device includes a medium and a processor; the medium stores a computer program; the processor executes the computer program to achieve the above-described objectives. Figures 1 to 2 The method for predicting the migration of fatigue properties in steel is shown.

[0121] Optionally, the computer device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Bluetooth interfaces, Wi-Fi interfaces), etc.

[0122] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0123] The medium may also include an operating system and a network communication module. The operating system is a program that manages and stores the hardware and software resources of a computer device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the medium, as well as communication with other hardware and software within the physical device.

[0124] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented using hardware to obtain steel tensile datasets (including alloy composition, heat treatment parameters, and target tensile strength values) and steel fatigue datasets (including target fatigue life values ​​for the same samples). The tensile dataset is used as the source domain, and the fatigue dataset as the target domain. A tensile strength prediction model for the source domain is constructed and trained using the target tensile strength value as a supervisory signal; a fatigue life prediction model for the target domain is constructed and trained using the target fatigue life value as a supervisory signal. During prediction, the tensile strength of the steel to be tested is predicted using the source domain model, and this is input as a dynamic feature into the target domain model. The fatigue life value is obtained through transfer prediction. By using tensile strength as an intermediary variable, a complete prediction link of material composition-mechanical properties-fatigue behavior is established, alleviating the problem of insufficient fatigue data and effectively predicting the fatigue performance of steel.

[0125] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0126] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any modifications that can be made by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for predicting the migration of fatigue properties in steel, characterized in that, The method includes: A steel tensile dataset and a steel fatigue dataset are obtained separately. The steel tensile dataset includes the alloy composition, heat treatment parameters and tensile strength target values ​​of various steel samples. The steel fatigue dataset includes the fatigue life target values ​​of the same steel samples. The steel tensile dataset is used as the source domain for evaluating the tensile strength of steel, and the steel fatigue dataset is used as the target domain for evaluating the fatigue life of steel. A source domain tensile strength prediction model is constructed, which is based on a gradient boosting tree model, which consists of multiple decision trees. The alloy composition in the steel tensile dataset is encoded into an alloy composition feature vector, and the heat treatment process parameters in the steel tensile dataset are encoded into a heat treatment process parameter feature vector. The training dataset is constructed by using the alloy composition feature vector and the heat treatment process parameter feature vector as input features and the tensile strength value as the target value. Initialize the hyperparameters of the gradient boosting tree model and train the initialized gradient boosting tree model using the training dataset. The hyperparameters include at least one of the following: number of decision trees, learning rate, and maximum depth. During the training of the initialized gradient boosting tree model using the training dataset, for each iteration, the gradient boosting tree model trains any new decision tree to fit the residual based on the residual between the currently predicted tensile strength value and the target value. The tensile strength value predicted by the new decision tree is then superimposed on the existing gradient boosting tree model until the preset number of iterations is reached. During the training process, the target value of tensile strength in the steel tensile dataset is used as the source domain supervision signal. A target domain fatigue life prediction model is constructed, which includes an input layer and a hidden layer. The input layer includes a main branch and a newly added strength prediction branch. The main branch is used to receive alloy composition feature vectors, and the strength prediction branch is used to receive tensile strength values. After adjustment, the values ​​are concatenated with the alloy composition feature vectors received by the main branch to form dynamic feature input. The hidden layer is a gradient boosting tree model. When training a target domain fatigue life prediction model based on a steel fatigue dataset, a two-stage feature fusion strategy is adopted, and the training process is implemented iteratively. The two-stage feature fusion strategy includes receiving alloy composition feature vectors directly as input through the main branch, receiving tensile strength values ​​generated by the source domain tensile strength prediction model through the strength prediction branch, dynamically concatenating the alloy composition feature vectors with the alloy composition feature vectors received by the main branch to form dynamic feature inputs, predicting the output of the dynamic feature inputs through a gradient boosting tree model in the hidden layer, and using the fatigue life target value in the steel fatigue dataset as the target domain supervision signal during the training process. The trained source domain tensile strength prediction model is used to predict the tensile strength values ​​of the steel to be predicted based on its alloy composition and heat treatment process parameters. The predicted tensile strength values ​​are then used as dynamic features to input into the trained target domain fatigue life prediction model. This allows the trained target domain fatigue life prediction model to predict the fatigue life value of the steel to be predicted based on the input dynamic features. Specifically, the source domain tensile strength prediction model is used to capture the nonlinear relationship between the alloy composition, heat treatment parameters, and tensile strength of the steel to be predicted. The target domain fatigue life prediction model is used to predict the fatigue life value based on the tensile strength value predicted by the source domain tensile strength prediction model.

2. The method according to claim 1, characterized in that, The target domain fatigue life prediction model has an input layer, and the alloy composition of the steel to be predicted corresponds to an alloy composition feature vector. The step of using the predicted tensile strength value as a dynamic feature input to train the target domain fatigue life prediction model includes: A strength prediction branch is added to the input layer of the target domain fatigue life prediction model. The tensile strength value predicted by the source domain tensile strength prediction model is linearly transformed to the same numerical range as the alloy composition feature vector through the new strength prediction branch. After the dimension of the transformed tensile strength value is consistent with the alloy composition feature vector by copying or padding, it is concatenated with the alloy composition feature vector to obtain the target domain fatigue performance prediction model after dynamic feature input training.

3. The method according to claim 1, characterized in that, The process of predicting the tensile strength of the steel by using the trained source domain tensile strength prediction model includes: Obtain the alloy composition feature vector corresponding to the alloy composition of the steel to be predicted, and the heat treatment process parameter feature vector corresponding to the heat treatment process parameters. The feature vectors of alloy composition and heat treatment process parameters are normalized respectively, and the normalized feature vectors of alloy composition and heat treatment process parameters are concatenated into a single input feature. The tensile strength value corresponding to the single input feature is predicted by the trained source domain tensile strength prediction model.

4. The method according to claim 1, characterized in that, Before using the trained source domain tensile strength prediction model to predict the tensile strength value of the steel to be predicted based on its alloy composition and heat treatment process parameters, the method further includes: The SHAP value method is used to perform global feature importance analysis on the trained source domain tensile strength prediction model. The global feature importance analysis quantifies the marginal contribution of alloy composition and heat treatment process parameters to the predicted tensile strength value. At the same time, dynamic feature interpretation is performed on the target domain fatigue life prediction model. Based on the dynamic feature interpretation, the synergistic effect of the dynamic feature after splicing the alloy composition feature vector and the predicted tensile strength value on fatigue life is analyzed, and the SHAP value interpretation results of the source domain tensile strength prediction model and the target domain fatigue life prediction model are obtained. When the SHAP interpretation results of both the source domain tensile strength prediction model and the target domain fatigue life prediction model are verified and show stable characteristic contribution distribution, the source domain tensile strength prediction model and the target domain fatigue life prediction model are deemed to have met the interpretability requirements.

5. The method according to any one of claims 1 to 4, characterized in that, The target domain fatigue life prediction model achieves cross-domain knowledge transfer through the TrAdaBoost.R2 algorithm.

6. A device for predicting the migration of fatigue properties in steel, characterized in that, The device includes: The data acquisition module is used to acquire steel tensile dataset and steel fatigue dataset respectively. The steel tensile dataset includes the alloy composition, heat treatment parameters and tensile strength target values ​​of various steel samples. The steel fatigue dataset includes the fatigue life target values ​​of the same steel samples. The source domain and target domain determination module is used to use the steel tensile dataset as the source domain for evaluating the tensile strength of the steel, and the steel fatigue dataset as the target domain for evaluating the fatigue life of the steel. The source domain model training module is used to construct a source domain tensile strength prediction model. This model is based on a gradient boosting tree model, which includes multiple decision trees. The alloy composition in the steel tensile dataset is encoded as an alloy composition feature vector, and the heat treatment process parameters in the steel tensile dataset are encoded as heat treatment process parameter feature vectors. The alloy composition feature vector and the heat treatment process parameter feature vector are used as input features, and the tensile strength value is used as the target value to construct a training dataset. The hyperparameters of the gradient boosting tree model are initialized, and the initialized gradient boosting tree model is trained using the training dataset. The hyperparameters include at least one of the following: the number of decision trees, the learning rate, and the maximum depth. During the training of the initialized gradient boosting tree model using the training dataset, for each iteration, the gradient boosting tree model trains any new decision tree to fit the residual between the currently predicted tensile strength value and the target value. The tensile strength value predicted by the new decision tree is then superimposed on the existing gradient boosting tree model until a preset number of iterations is reached. During training, the target tensile strength value in the steel tensile dataset is used as a source domain supervision signal. The target domain model training module is used to construct a target domain fatigue life prediction model. This model includes an input layer and a hidden layer. The input layer consists of a main branch and a newly added strength prediction branch. The main branch receives alloy composition feature vectors, and the strength prediction branch receives tensile strength values. After adjustment, these values ​​are concatenated with the alloy composition feature vectors received by the main branch to form dynamic feature input. The hidden layer is a gradient boosting tree model. When training the target domain fatigue life prediction model based on a steel fatigue dataset, a two-stage feature fusion strategy is adopted, and the training process is iterative. This two-stage feature fusion strategy involves directly receiving alloy composition feature vectors as input through the main branch, receiving tensile strength values ​​generated by the source domain tensile strength prediction model through the strength prediction branch, dynamically concatenating the alloy composition feature vectors with those received by the main branch to form dynamic feature input, and using a gradient boosting tree model in the hidden layer to predict the output of the dynamic feature input. During training, the fatigue life target value from the steel fatigue dataset is used as a target domain supervision signal. The strength transfer fatigue life prediction module is used to predict the tensile strength value of the steel to be predicted based on the alloy composition and heat treatment process parameters of the steel under test using a trained source domain tensile strength prediction model. The predicted tensile strength value is then used as a dynamic feature input to a trained target domain fatigue life prediction model, so that the trained target domain fatigue life prediction model can predict the fatigue life value of the steel under test based on the input dynamic features. The source domain tensile strength prediction model is used to capture the nonlinear relationship between the alloy composition, heat treatment parameters and tensile strength of the steel under test. The target domain fatigue life prediction model is used to transfer and predict the fatigue life value based on the tensile strength value predicted by the source domain tensile strength prediction model.

7. A medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method for predicting the fatigue performance migration of steel as described in any one of claims 1 to 5.

8. A computer device, comprising a medium, a processor, and a computer program stored on the medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for predicting the migration of fatigue properties of steel as described in any one of claims 1 to 5.

Images