A method and system for predicting phase composition of low-alloy high-strength steel based on classification and regression coupling

Abstract

The application discloses a kind of based on classification and regression coupling's low alloy high-strength steel phase composition prediction method and system, and relates to metallurgical technology field.The method includes: obtaining pre-trained phase composition classification model and phase score regression model;The chemical composition features and process parameter features of the target to be predicted are input into the phase composition classification model to obtain classification prediction result;According to classification prediction result, the phase score prediction of the regression branch triggered to the target microstructure indicating existing is carried out, and the phase score of the target microstructure indicating non-existing is directly assigned as zero;The initial phase score prediction result is corrected, so that each phase score is non-negative and the total sum does not exceed 100%.The present application realizes end-to-end fast prediction and reverse design of multiphase organization, which can significantly shorten the research and development cycle and reduce experimental cost.

Description

Technical Field

[0001] This invention relates to the field of metallurgical technology, and in particular to a method and system for predicting the phase composition of low-alloy high-strength steel based on classification and regression coupling. Background Technology

[0002] Low-alloy high-strength steel (HSLA) is a key basic material in the national strategies of "materials first" and "manufacturing powerhouse," and is widely used in automobiles, energy, shipbuilding, bridges, and high-end equipment manufacturing. Its mechanical properties and safety in use are highly dependent on the type and distribution of its microstructure—especially the composition ratio and spatial distribution of ferrite, pearlite, and various types of bainite.

[0003] In actual production, the microstructure design of some HSLA steels is in the "transition zone" of phase transformation: due to the wide range of chemical composition fluctuations (especially alloying elements such as C, Mn, Cr, and Mo), coupled with the differences in hot rolling process parameters (final rolling temperature, cooling rate, red-hot temperature, etc.), steels of the same grade or similar composition may form completely different microstructure types. Taking Q550MPa grade bridge steel as an example, affected by the finished product thickness, chemical composition, hot rolling process, and air cooling process, the microstructure exhibits the following four typical situations: (1) Ferrite + Pearlite (F+P); (2) Ferrite + Pearlite + Feathered Bainite (Upper Bainite, UB) (F+P+UB); (3) Ferrite + Pearlite + Feathered Bainite + Granular Bainite (GB) (F+P+UB+GB); (4) Ferrite + Pearlite + Granular Bainite (GB) (F+P+GB).

[0004] Theoretically, when the cooling rate is fast enough or the carbon content is extremely low, a microstructure of ferrite + bainite (F+B) may occur. However, considering the cooling conditions of the medium plate and the Q550 data used, identifiable pearlite was present at 1 / 4 thickness in each batch; and when pearlite was extremely scarce (<2%), the metallographic calibration error was large, and separate classification would increase data noise. Furthermore, acicular ferrite (AF) refers to carbon-rich ferrite produced during continuous cooling, where carbon ultimately precipitates as cementite needles between ferrite grains, with a morphology similar to feathery bainite (UB). The classification of these two types is controversial in the metallographic community. Meanwhile, the M / A islands in granular bainite (GB) are extremely small (1~3μm), and the manual cost and repeatability of accurately calibrating and statistically analyzing their sparsity in metallographic images are high; ImageJ pixel statistics make it difficult to accurately separate M / A islands from the ferrite matrix; and the quantitative relationship between the sparsity of M / A islands and process parameters is difficult to model under small sample conditions.

[0005] In summary, from the perspective of engineering applications, the present invention makes the following treatments: (1) it does not consider the case where pearlite structure is completely absent; (2) it combines acicular ferrite (AF) and feathery bainite (UB) as the same structure type for statistical analysis, so as to simplify the classification system and improve the operability of the prediction model; (3) it no longer refines the distinction of the sparsity of M / A islands, so as to meet the engineering accuracy requirements while ensuring statistical efficiency.

[0006] Among the aforementioned microstructure types, the presence of feathery bainite (UB) often reduces the toughness of steel, while granular bainite (GB), in an appropriate proportion, can improve the strength-toughness balance. Therefore, accurate prediction of microstructure type and quantitative control of phase fraction have become the core challenges in the performance regulation of HSLA steel.

[0007] Traditional microstructure prediction mainly relies on: (1) continuous cooling transformation (CCT) diagrams, combined with thermal simulation experiments to determine phase transformation temperature and time; (2) calculation based on empirical formulas of physical metallurgical mechanisms; and (3) direct observation under a metallographic microscope by preparing metallographic structures. The above methods have the following shortcomings: (1) the establishment of CCT diagrams requires a large number of thermal simulation experiments, which are time-consuming and costly, and usually only cover a specific composition range; (2) empirical formulas are mostly single-phase microstructure models for specific steel grades, which are difficult to handle complex situations where multiple phases coexist and multiple bainite types appear at the same time; (3) metallographic observation cannot predict in advance and can only be carried out after production, which is a passive detection rather than an active design.

[0008] In recent years, machine learning methods have been gradually introduced into the field of steel materials. However, there is no existing technology that applies machine learning to steel grades with narrow composition and process windows in HSLA steel. There is also no systematic method for coupled classification prediction and phase fraction regression prediction of microstructure. Furthermore, the reverse design from target microstructure to composition-process parameters has not been realized. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method and system for predicting the phase composition of low alloy high-strength steel based on classification and regression coupling.

[0010] To solve the above technical problems, the technical solution of the present invention is as follows: In a first aspect, the present invention provides a method for predicting the phase composition of low-alloy high-strength steel based on classification and regression coupling, comprising: Obtain a pre-trained phase composition classification model and a phase fraction regression model; wherein, the phase composition classification model is used to predict the existence state of the target microstructure; and the phase fraction regression model is used to predict the phase fraction of the target microstructure. The chemical composition characteristics and process parameter characteristics of the target to be predicted are obtained. The chemical composition characteristics and process parameter characteristics are input into the phase composition classification model to obtain the classification prediction results. Based on the classification prediction results, the phase fraction of the target microstructure indicated by the classification prediction results is predicted by the phase fraction regression model, and the phase fraction of the target microstructure indicated by the classification prediction results is assigned to zero, so as to obtain the initial phase fraction prediction results. The initial phase fraction prediction results are corrected so that the corrected phase fractions are all non-negative and the sum of the phase fractions does not exceed 100%, thus obtaining the final phase fraction prediction results.

[0011] Secondly, this invention provides a training method for a machine learning classification and regression coupled prediction model for the phase composition of low-alloy high-strength steel, comprising the following steps: Obtain historical production datasets of low-alloy high-strength steel. The historical production datasets include chemical composition features, process parameter features, and corresponding microstructure classification labels and phase fraction labels for multiple samples. A phase composition classification model was trained using historical production datasets to obtain the trained phase composition classification model, which is used to predict the existence state of feathery bainite and granular bainite. A phase fraction regression model was trained using historical production datasets to obtain the trained phase fraction regression model. The phase fraction regression model is used to predict the phase fractions of pearlite, feathery bainite, and granular bainite. In particular, the regression branches for feathery bainite and granular bainite in the phase fraction regression model are only trained on samples where the corresponding microstructure classification label indicates their presence, and samples where the corresponding microstructure classification label indicates their absence are used as zero-value anchor points in the training.

[0012] Thirdly, the present invention provides a machine learning classification and regression coupled system for predicting the phase composition of low-alloy high-strength steel, comprising: The model acquisition module is used to acquire a pre-trained phase composition classification model and a phase fraction regression model. The phase composition classification model is used to predict the presence state of the target microscopic tissue. The phase fraction regression model is a model obtained by training only on training samples where the target microscopic tissue exists and using training samples where the target microscopic tissue does not exist as zero-value anchors in the training. The classification and prediction module is used to obtain the chemical composition characteristics and process parameter characteristics of the target to be predicted, input the chemical composition characteristics and process parameter characteristics into the phase to form a classification model, and obtain the classification and prediction results; The regression prediction module is used to predict the phase fraction of the target microstructure indicated by the classification prediction results through the phase fraction regression model, and to directly assign the phase fraction of the target microstructure indicated by the classification prediction results to zero, so as to obtain the initial phase fraction prediction result. The result correction module is used to correct the initial phase fraction prediction results so that the corrected phase fractions are all non-negative and the sum of the phase fractions does not exceed 100%, thus obtaining the final phase fraction prediction results.

[0013] Fourthly, this invention provides a reverse design method for the composition and processing of low-alloy high-strength steel based on a coupled model, comprising the following steps: A phase composition classification model and a phase fraction regression model are obtained. The phase composition classification model is used to predict the existence state of feathery bainite and granular bainite, and the phase fraction regression model is used to predict the phase fraction of pearlite, feathery bainite and granular bainite. The regression branch of the phase fraction regression model for feathery bainite and granular bainite is a model that is trained only on training samples that are indicated to exist by the corresponding classification label and uses training samples that are not indicated to not exist by the corresponding classification label as zero value anchor points for training. Obtain the target microstructure constraints and the preset composition and process search space, which is jointly limited by the value range of chemical composition parameters and the value range of process parameters. Candidate component-process combinations are generated within the component-process search space. These candidate component-process combinations are then input into the phase composition classification model and the phase score regression model, respectively, to obtain candidate classification prediction results and candidate phase score prediction results. When the candidate classification prediction results and candidate phase fraction prediction results meet the target microstructure constraints, the corresponding candidate component process combination will be output as the target component process scheme.

[0014] The beneficial effects of this invention are: (1) This invention constructs a two-layer architecture of phase composition classification model and phase fraction regression model. The phase composition classification model uses two independent binary branches to predict the existence state of feathery bainite and granular bainite respectively, and the phase fraction regression model uses three independent regression branches to predict the phase fractions of pearlite, feathery bainite and granular bainite respectively. Under the premise of ensuring the stability of the ferrite-pearlite basic structure, this two-layer architecture accurately captures the quantitative influence of alloy composition and process parameters on bainite type and phase fraction, realizing end-to-end rapid prediction from composition-process to microstructure, without relying on CCT images or thermal simulation experiments, and reducing the time for a single prediction from several hours of traditional metallographic observation to seconds.

[0015] (2) In the training phase, the regression branches for feathery bainite and granular bainite in the phase fraction regression model are only trained on training samples indicated by the corresponding classification labels, avoiding the dilution of regression accuracy by a large number of zero-value samples; at the same time, training samples indicated by the classification labels as non-existent are used as zero-value anchor points for training, so that the regression model has a stable regression tendency near zero values. In the prediction phase, the corresponding regression branch is triggered to predict the phase fraction only for the target microstructure indicated by the classification prediction results, and the phase fraction of the target microstructure indicated by the classification prediction results as non-existent is directly assigned to zero, skipping the regression process of that phase. This coupling mechanism makes the classification model and the regression model no longer two independent parallel models, but form an organic whole of "classification-constrained regression", effectively solving the problem of insufficient prediction accuracy of the regression model for sparse phases in the case of multi-phase coexistence.

[0016] (3) In the post-processing stage of the present invention, the initial phase fraction prediction results are corrected so that the corrected phase fractions are all non-negative and the sum of the phase fractions does not exceed 100%. When the sum of the predicted values ​​of pearlite fraction, feathery bainite fraction and granular bainite fraction exceeds the threshold, the predicted values ​​of the three are compressed proportionally. This post-processing mechanism ensures that the prediction results always meet the physical meaning that the phase fractions are non-negative and the sum does not exceed the threshold, avoiding the physically unreasonable results that may be output by a purely data-driven model.

[0017] (4) The present invention constructs a reverse design functional module. When the target microstructure requirement is proposed, it can perform iterative search in the composition and process search space jointly defined by the value range of chemical composition parameters and the value range of process parameters in the form of objective function. Combined with particle swarm multi-objective optimization, Bayesian optimization or grid search strategy, it can recommend a combination of composition and process that meets the target structure constraint in reverse. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the phase composition prediction method for low-alloy high-strength steel based on classification and regression coupling provided in Example 1. Figure 2 This is a schematic diagram of the confusion matrix of the UB classification results provided in Example 3; Figure 3 This is a schematic diagram of the confusion matrix of GB classification results provided in Example 3; Figure 4A schematic diagram of the evaluation index for the classification model provided in Example 3; Figure 5 A schematic diagram of regression results for pearlite fraction (Fp), feathery bainite fraction (Fub), and granular bainite fraction (Fgb) is provided for Example 3; Figure 6 Metallographic images of hot-rolled specimen 1 provided for reverse design verification in this embodiment of the invention; Figure 7 Metallographic images of hot-rolled specimen 2 used for reverse design verification in this embodiment of the invention. Detailed Implementation

[0020] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0021] Example 1: This example uses Q550MPa grade bridge steel as the research object to explain in detail the specific implementation process of the phase composition prediction method for low-alloy high-strength steel based on classification and regression coupling. See [link to example]. Figure 1 The prediction method specifically includes the following steps: Step S101: Obtain the pre-trained phase composition classification model and phase fraction regression model; wherein, the phase composition classification model is used to predict the existence state of the target microstructure; the phase fraction regression model is used to predict the phase fraction of the target microstructure.

[0022] Specifically, the phase composition classification model includes a first classification branch and a second classification branch. The first classification branch is used to predict the presence state of feathery bainite (UB), and the output is the predicted probability (0~1) of the presence of UB; the second classification branch is used to predict the presence state of granular bainite (GB), and the output is the predicted probability (0~1) of the presence of GB. The two classification branches are independent of each other and are trained with binary log loss as the optimization objective.

[0023] The phase fraction regression model consists of three independent regression branches, which are used to predict the pearlite fraction (Fp), feathery bainite fraction (Fub), and granular bainite fraction (Fgb), respectively, and output the predicted values ​​(0~100%) of the corresponding phase fractions. The three regression branches are independent of each other and are trained with squared loss as the optimization objective.

[0024] It should be noted that the regression branches for feathery bainite and granular bainite in the phase fractional regression model are trained only on training samples whose corresponding classification labels indicate their presence, and with training samples whose corresponding classification labels indicate their absence used as zero-value anchors in the training. Specifically: (a) Sample selection: Taking the feathered bainite score (Fub) regression branch as an example, this regression branch is only trained on training samples where the classification label indicates the presence of UB (label is 1). This avoids the dilution of regression accuracy by a large number of samples where UB does not exist (label is 0, Fub=0), allowing the regression model to focus on learning the mapping relationship of "what is the UB score when UB exists".

[0025] (b) Zero-value anchors: Training samples whose classification labels indicate that UB does not exist (label is 0) are used as anchors with a target value of 0 during training. This gives the regression model a stable regression tendency near the zero value, ensuring that when the classification model determines that UB does not exist, even if the regression model is mistakenly triggered, its output tends to be 0 rather than a random value.

[0026] (c) The training method for the granular bainite fraction (Fgb) regression branch is the same as that for the Fub regression branch.

[0027] (d) The pearlite score (Fp) regression branch is not constrained by the classification label and is trained on the full training samples because pearlite is an intrinsic tissue (widely present) and does not require classification judgment.

[0028] Both the phase-based classification model and the phase-based regression model were constructed using the XGBoost algorithm. During training, a Bayesian optimization algorithm was used to iteratively optimize the model's hyperparameters, including the maximum tree depth (max_depth), learning rate (learning_rate), L2 regularization parameter (lambda), subsample ratio (subsample), and column sample ratio (colsample_bytree). The evaluation metric for the classification model was AUC (area under the ROC curve), and the evaluation metric for the regression model was RMSE (root mean square error). Both models had an early_stopping_rounds limit of 50 rounds; training was stopped if the validation set metrics showed no improvement within 50 rounds.

[0029] The model's input features are chemical composition features and process parameter features. The chemical composition features include the mass percentages of 14 elements: C, Si, Mn, P, S, N, Cu, Al, Cr, Nb, Mo, Ni, V, and Ti. The process parameters include rolling temperature, reduction rate, cooling control parameters, and red-heating temperature for each pass, totaling 53 feature vectors.

[0030] Step S102: Obtain the chemical composition characteristics and process parameter characteristics of the target to be predicted, input the chemical composition characteristics and process parameter characteristics into the phase composition classification model, and obtain the classification prediction results; Specifically, the chemical composition characteristics and process parameter characteristics of the target to be predicted are obtained. The chemical composition characteristics and process parameter characteristics are input into the first classification branch and the second classification branch, respectively, to obtain the first prediction probability (the probability of UB existing) and the second prediction probability (the probability of GB existing). The first prediction probability and the second prediction probability are combined as the classification prediction result.

[0031] Specifically, when the first predicted probability is greater than a preset threshold of 0.5, UB is determined to exist; when the second predicted probability is greater than a preset threshold, GB is determined to exist. The classification prediction results indicate the presence status of feathery bainite and granular bainite in the target to be predicted.

[0032] Step S103: Based on the classification prediction results, the phase fraction of the target microstructure indicated by the classification prediction results is predicted using the phase fraction regression model. The phase fraction of the target microstructure indicated by the classification prediction results is directly assigned to zero, thus obtaining the initial phase fraction prediction results.

[0033] Specifically, based on the classification prediction results obtained in step S102, conditionally triggered phase score prediction is performed: (a) Pearlite score prediction: Since pearlite is an intrinsic tissue (widely present) and is not constrained by the classification results, the pearlite score is always predicted through the pearlite score (Fp) regression branch in the phase score regression model to obtain the pearlite score prediction value.

[0034] (b) Feathered bainite fraction prediction: If the classification prediction result indicates that UB exists (the first prediction probability is greater than the threshold), the feathered bainite fraction (Fub) regression branch in the phase fraction regression model is triggered to predict the feathered bainite fraction; if the classification prediction result indicates that UB does not exist (the first prediction probability is less than or equal to the threshold), the feathered bainite fraction is directly assigned to zero, and the prediction process of the Fub regression branch is skipped.

[0035] (c) Granular bainite fraction prediction: If the classification prediction result indicates that GB exists (the second prediction probability is greater than the threshold), the granular bainite fraction (Fgb) regression branch in the phase fraction regression model is triggered to predict the granular bainite fraction; if the classification prediction result indicates that GB does not exist (the second prediction probability is less than or equal to the threshold), the granular bainite fraction is directly assigned to zero, and the prediction process of the Fgb regression branch is skipped.

[0036] The results of the above predictions and assignments are used as the initial phase fraction prediction results.

[0037] Step S104: Correct the initial phase fraction prediction results so that the corrected phase fractions are all non-negative and the sum of the phase fractions does not exceed 100%, and obtain the final phase fraction prediction results.

[0038] Specifically, the initial phase fraction prediction results obtained in step S103 are corrected for physical constraints: (a) Non-negative correction: If the predicted value of any phase fraction is less than 0, the corresponding value is corrected to 0.

[0039] (b) Upper limit correction: When the sum of the predicted values ​​of pearlite fraction, feathery bainite fraction and granular bainite fraction is greater than 100%, the predicted values ​​of the three are compressed proportionally so that the sum of the compressed phase fractions does not exceed 100%. Specifically, the compression factor is 100% divided by the sum of the predicted values ​​of the three, and each predicted phase fraction is multiplied by the compression factor.

[0040] The corrected result is the final phase fraction prediction result, including pearlite fraction (Fp), feathery bainite fraction (Fub) and granular bainite fraction (Fgb). Ferrite fraction (F) is obtained by subtracting the sum of Fp, Fub and Fgb from 100%.

[0041] It should be noted that after obtaining the final phase fraction prediction results, the SHAP interpretability analysis method can be used to quantify the contribution of chemical composition characteristics and process parameter characteristics to the prediction results of the phase composition classification model and the phase fraction regression model, respectively, to reveal the influence of alloying elements and process parameters on bainite formation.

[0042] Example 2: This example provides a training method for a machine learning classification and regression coupled prediction model for the phase composition of low-alloy high-strength steel. The method specifically includes the following steps: Step S201: Obtain the historical production dataset of low-alloy high-strength steel. The historical production dataset includes the chemical composition characteristics, process parameter characteristics, and corresponding microstructure classification labels and phase fraction labels of multiple samples.

[0043] Specifically, 307 hot-rolled specimens of Q550D grade with thicknesses ranging from 20 mm to 60 mm, containing complete production data and micrographs, were collected and compiled. Each data point includes: (1) Chemical composition: including the mass percentage of 14 elements, namely C, Si, Mn, P, S, N, Cu, Al, Cr, Nb, Mo, Ni, V and Ti.

[0044] (2) Process parameters: including key hot rolling process parameters such as rolling temperature, reduction rate, cooling control parameters and reddening temperature for each pass, totaling 53 feature vectors.

[0045] (3) Microstructure information: including structure type (classified into four cases: F+P, F+P+UB, F+P+UB+GB, F+P+GB) and volume fraction of each phase (phase fraction, unit: %).

[0046] The microstructure information was obtained as follows: Five micrographs were taken at a constant longitudinal position at 1 / 4 of the sample thickness under a metallographic microscope, and at transverse positions at 1 / 10, 3 / 10, 5 / 10, 7 / 10, and 9 / 10 of the sample thickness. LabelImg was used to perform metallographic calibration on the images, delineating pearlite, feathery bainite, and granular bainite regions. ImageJ's statistical function was used to calculate the pixel percentage of each phase (the ferrite phase was calculated by subtracting the other three from 100%). The phase composition of the same sample was the average of the statistical results from the five images.

[0047] The following preprocessing operations were performed on the original data: (1) Missing value processing: For samples with a small number of missing values, the mean of the feature column was used to fill them; (2) Outlier removal: Outliers of chemical composition or phase fraction were identified and removed by box plot method (IQR 1.5 times criterion); (3) Normalization: Min-Max normalization was performed on all input features to eliminate the influence of dimensional differences on model training; (4) Feature encoding: The tissue type was labeled according to the four classification standard to obtain classification labels.

[0048] The dataset was divided into a training set (246 records) and a test set (61 records) in an 8:2 ratio for subsequent model training and performance evaluation.

[0049] Step S202: Train the phase composition classification model using historical production datasets to obtain the trained phase composition classification model. The phase composition classification model is used to predict the existence state of feathery bainite and granular bainite. Specifically, ferrite and pearlite are intrinsic structures in Q550 steel (commonly present). Therefore, this application uses two independent binary classification models to predict: ① whether feathery bainite (UB) exists (label: 0 - not present, 1 - present); ② whether granular bainite (GB) exists (label: 0 - not present, 1 - present).

[0050] The model input features are chemical composition (content of 14 elements) and process parameters (rolling temperature, reduction rate, cooling control, and red-heating temperature for each pass), totaling a 53-dimensional feature vector. The main hyperparameters of the XGBoost classifier are optimized iteratively and automatically by expanding the selected hyperparameter range using a Bayesian optimization algorithm. The main hyperparameters are as follows: objective is binary:logistic (binary logistic loss); max_depth (maximum tree depth) is self-optimized; learning_rate (learning rate) is self-optimized; lambda (L2 regularization parameter) is self-optimized with a lower limit of 1.0 to prevent overfitting; subsample (subsampling ratio) is self-optimized with an upper limit of 0.9 to prevent overfitting; colsample_bytree (column sampling ratio) is self-optimized with an upper limit of 0.9 to prevent overfitting; eval_metric is AUC (area under the ROC curve); early_stopping_rounds is 50 rounds; training stops if the validation set AUC does not improve within 50 rounds.

[0051] Step S203: Train the phase fraction regression model using the historical production dataset to obtain the trained phase fraction regression model. The phase fraction regression model is used to predict the phase fractions of pearlite, feathery bainite and granular bainite. In the phase fraction regression model, the regression branch for feathery bainite and granular bainite is only trained on samples where the corresponding microstructure classification label indicates their presence, and samples where the corresponding microstructure classification label indicates their absence are used as zero anchor points for training.

[0052] Specifically, three independent XGBoost regression models were constructed for pearlite fraction (Fp), feathery bainite fraction (Fub), and granular bainite fraction (Fgb). The outputs are the predicted values ​​(0~100%) for the corresponding phase fractions.

[0053] The main hyperparameters of the XGBoost regressor are optimized iteratively and automatically by expanding the range of the selected hyperparameters using a Bayesian optimization algorithm. The configuration is as follows: objective is reg:squarederror (squared loss); max_depth (maximum tree depth) is self-optimized; learning_rate (learning rate) is self-optimized; lambda (L2 regularization parameter) is self-optimized with a lower limit of 1.5 to prevent overfitting; subsample (subsampling ratio) is self-optimized with an upper limit of 0.9 to prevent overfitting; colsample_bytree (column sampling ratio) is self-optimized with an upper limit of 0.9 to prevent overfitting; eval_metric is RMSE (root mean square error); early_stopping_rounds is 50 rounds.

[0054] In the phase fractional regression model, the regression branches for feathery bainite and granular bainite are trained only on samples where the corresponding microstructure classification label indicates their presence, and samples where the corresponding microstructure classification label indicates their absence are used as zero-value anchors in the training. Specifically: (a) Fub regression branch: Regression training is performed only on samples with a UB classification label of 1 (UB exists), and samples with a UB classification label of 0 (UB does not exist) are used as zero anchor points with a target value of 0 for training.

[0055] (b) Fgb regression branch: Regression training is performed only on samples with GB classification label of 1 (GB exists), and samples with GB classification label of 0 (GB does not exist) are used as zero anchor points with a target value of 0 for training.

[0056] (c) Fp regression branch: trained on the full training samples, without being constrained by the classification labels.

[0057] Example 3: This example provides a machine learning classification and regression coupled system for predicting the phase composition of low-alloy high-strength steel. The system specifically includes a model acquisition module, a classification prediction module, a regression prediction module, and a result correction module.

[0058] Specifically, the model acquisition module acquires pre-trained phase composition classification and phase fraction regression models. The module loads pre-trained model files (e.g., .H5 format files) from a model repository (such as a local file system, a remote model server, or a model management platform). The phase composition classification model includes a first classification branch (for predicting the presence of UB) and a second classification branch (for predicting the presence of GB), both constructed using the XGBoost binary classification algorithm. The phase fraction regression model includes three independent regression branches (for predicting Fp, Fub, and Fgb, respectively), all constructed using the XGBoost regression algorithm.

[0059] In addition, the model acquisition module is also responsible for verifying the integrity and version consistency of the model, ensuring that the loaded model file matches the input feature dimensions and output format required by the current prediction task.

[0060] In the phase fractional regression model, the regression branches for feathery bainite and granular bainite are trained only on training samples whose corresponding classification labels indicate their presence, and with training samples whose corresponding classification labels indicate their absence serving as zero-value anchors. This training method enables the regression model to have high accuracy in the positive sample region and a stable regression tendency near the zero value.

[0061] The classification and prediction module is used to obtain the chemical composition characteristics and process parameter characteristics of the target to be predicted. The chemical composition characteristics and process parameter characteristics are input into the phase to form a classification model and obtain the classification and prediction results.

[0062] Specifically, the classification prediction module receives chemical composition data (mass percentage of 14 elements) and process parameter data (rolling temperature, reduction rate, cooling control parameters, and reddening temperature of each pass) of the target to be predicted, either input by the user or automatically obtained from the production system (such as MES / L2 system). After performing normalization preprocessing on the input data in the same manner as in the training phase, the 53-dimensional feature vector is input into the first classification branch and the second classification branch respectively.

[0063] The first classification branch outputs the predicted probability (0~1) of the presence of UB, and the second classification branch outputs the predicted probability (0~1) of the presence of GB. The classification prediction module combines the two predicted probabilities to form the classification prediction result. When the predicted probability is greater than the preset threshold of 0.5, the corresponding microstructure is determined to exist; otherwise, it is determined not to exist.

[0064] The regression prediction module is used to predict the phase fraction of target microstructures indicated by the classification prediction results through a phase fraction regression model, and to directly assign a value of zero to the phase fraction of target microstructures indicated by the classification prediction results, thus obtaining the initial phase fraction prediction results.

[0065] Specifically, the regression prediction module executes the following logic: (a) Pearlite score prediction: Regardless of the classification prediction result, the 53-dimensional feature vector is always input into the Fp regression branch to obtain the pearlite score prediction value.

[0066] (b) Feathered bainite score prediction: If the classification prediction result indicates that UB exists, the 53-dimensional feature vector is input into the Fub regression branch to obtain the predicted value of the feathered bainite score; if the classification prediction result indicates that UB does not exist, the feathered bainite score is directly assigned to zero and the Fub regression branch is not called.

[0067] (c) Granular bainite fraction prediction: If the classification prediction result indicates that GB exists, the 53-dimensional feature vector is input into the Fgb regression branch to obtain the granular bainite fraction prediction value; if the classification prediction result indicates that GB does not exist, the granular bainite fraction is directly assigned to zero, and the Fgb regression branch is not called.

[0068] The results of the above predictions and assignments are combined to form the initial phase fraction prediction results (Fp, Fub, Fgb).

[0069] The result correction module is used to correct the initial phase fraction prediction results so that the corrected phase fractions are all non-negative and the sum of the phase fractions does not exceed 100%, thus obtaining the final phase fraction prediction results.

[0070] Specifically, the result correction module executes the following correction logic: (a) Non-negative correction: Iterate through each phase fraction in the initial phase fraction prediction results. If the predicted value of any phase fraction is less than 0, then correct the corresponding value to 0.

[0071] (b) Upper limit correction: Calculate the sum of the predicted values ​​of pearlite fraction, feathery bainite fraction and granular bainite fraction. If the sum is greater than 100%, calculate the compression factor (100% divided by the sum of the three predicted values), and multiply the predicted value of each phase fraction by the compression factor so that the sum of the compressed phase fractions does not exceed 100%.

[0072] (c) Ferrite fraction calculation: Ferrite fraction (F) = 100% - Fp - Fub - Fgb.

[0073] The corrected result is the final phase fraction prediction result, which is output to the user or written to the production system database.

[0074] Performance evaluation metrics for classification models: Accuracy, Precision, Recall, F1 score, and AUC. Performance evaluation metrics for regression models: Coefficient of determination (R²), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE).

[0075] In this embodiment: Result of the classification model implementation: Feathered bainite (UB) presence classification: test set accuracy 90.2%, precision 82.8%, recall 96.0%, F1 score 88.9%, AUC 0.931; Granular bainite (GB) presence classification: test set accuracy 91.8%, precision 98.2%, recall 93.1%, F1 score 95.6%, AUC 0.948.

[0076] The above results demonstrate that the XGBoost classification model has a very high ability to identify the presence of the two types of bainite and can effectively distinguish samples of different microstructure types.

[0077] Results of regression model implementation: Pearlite fraction (Fp) regression: R²=0.921, RMSE=2.28%, MAE=1.80%; Regression of feathered bainite fraction (Fub): R²=0.872, RMSE=2.63%, MAE=2.15%; Regression of granular bainite fraction (Fgb): R²=0.895, RMSE=1.91%, MAE=1.73%.

[0078] The R² values ​​of all three regression models are greater than 0.85. Considering the errors in microstructure identification and industrial data noise, the models have high reliability in predicting the fractions of each phase and can meet the accuracy requirements for quantitative prediction of microstructure in industrial production.

[0079] Example 4: This example provides a reverse design method for the composition and process of low-alloy high-strength steel based on a coupled model. The method specifically includes the following steps: Step S301: Obtain the phase composition classification model and the phase fraction regression model. The phase composition classification model is used to predict the existence state of feathery bainite and granular bainite, and the phase fraction regression model is used to predict the phase fraction of pearlite, feathery bainite and granular bainite. The regression branch of the phase fraction regression model for feathery bainite and granular bainite is a model that is trained only on training samples that are indicated to exist by the corresponding classification label and uses training samples that are not indicated to not exist by the corresponding classification label as zero anchor points for training. Step S302: Obtain the target microstructure constraints and the preset composition and process search space. The composition and process search space is jointly limited by the value range of chemical composition parameters and the value range of process parameters. Step S303: Generate candidate component process combinations within the component process search space, and input the candidate component process combinations into the phase composition classification model and the phase score regression model respectively to obtain candidate classification prediction results and candidate phase score prediction results; Step S304: When the candidate classification prediction results and candidate phase fraction prediction results meet the target microstructure constraints, the corresponding candidate component process combination is output as the target component process scheme.

[0080] With the objective constraint of a granular bainite fraction between 60% and 80% and the prohibition of feathery bainite, and a carbon equivalent (Ceq) less than 20, particle swarm optimization was used to design 40mm thick Q550D grade bridge steel. The reverse design module, based on a phase composition prediction model coupled with classification and regression, retrieved the recommended composition and process within the following window by calling the trained model: C 0.10~0.15wt.%, Mn 1.00~1.50wt.%, Cr 0.20~0.40wt.%, Si 0.20~0.40wt.%, final rolling temperature 820~850℃, and red-heating temperature 500~550℃.

[0081] The verification experiment selected two sets of parameters within the recommended window for smelting and hot rolling. No feathery bainite appeared in the metallographic structure, which verified that the composition-process scheme output by the reverse design module can directly guide the actual trial rolling, and the designed metallographic structure was obtained, effectively reducing the number of process trial errors.

[0082] The above embodiments fully demonstrate that the XGBoost coupled classification-regression model of the present invention has excellent prediction accuracy, good generalization ability and clear industrial practical value in HSLA steel microstructure prediction and reverse design. It can effectively replace some physical metallographic experiments and significantly reduce R&D costs.

[0083] In addition to the above embodiments, the present invention may have other implementation methods; all technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A method for predicting the phase composition of low-alloy high-strength steel based on classification and regression coupling, characterized in that, include: Obtain a pre-trained phase composition classification model and a phase fraction regression model; wherein, the phase composition classification model is used to predict the existence state of the target microstructure; and the phase fraction regression model is used to predict the phase fraction of the target microstructure. Obtain the chemical composition characteristics and process parameter characteristics of the target to be predicted, input the chemical composition characteristics and process parameter characteristics into the phase composition classification model, and obtain the classification prediction results; Based on the classification prediction results, the phase fraction of the target microstructure indicated by the classification prediction results is predicted by the phase fraction regression model, and the phase fraction of the target microstructure indicated by the classification prediction results is assigned to zero to obtain the initial phase fraction prediction results. The initial phase fraction prediction results are corrected so that each phase fraction is non-negative and the sum of each phase fraction does not exceed 100%, thus obtaining the final phase fraction prediction results.

2. The method for predicting the phase composition of low-alloy high-strength steel based on classification and regression coupling as described in claim 1, characterized in that, The phase composition classification model includes a first classification branch and a second classification branch; The step of inputting the chemical composition features and the process parameter features into the phase composition classification model to obtain classification prediction results includes: The chemical composition features and the process parameter features are respectively input into the first classification branch and the second classification branch to obtain the first prediction probability and the second prediction probability. The first prediction probability and the second prediction probability are combined as the classification prediction result. The first classification branch is used to predict the presence state of feathery bainite, and the second classification branch is used to predict the presence state of granular bainite.

3. The method for predicting the phase composition of low-alloy high-strength steel based on classification and regression coupling as described in claim 1, characterized in that, The phase fraction regression model includes three regression branches, which are used to predict the pearlite fraction, feathery bainite fraction, and granular bainite fraction, respectively.

4. The method for predicting the phase composition of low-alloy high-strength steel based on classification and regression coupling according to claim 1, characterized in that, Both the phase composition classification model and the phase score regression model are constructed using the XGBoost algorithm. During the training of the phase composition classification model and the phase score regression model, the Bayesian optimization algorithm is used to iteratively optimize the hyperparameters of the model. The hyperparameters include the maximum tree depth, learning rate, L2 regularization parameter, subsampling ratio, and column sampling ratio.

5. The method for predicting the phase composition of low-alloy high-strength steel based on classification and regression coupling according to claim 1, characterized in that, The chemical composition characteristics include the mass percentages of 14 elements: C, Si, Mn, P, S, N, Cu, Al, Cr, Nb, Mo, Ni, V, and Ti. The process parameters include the rolling temperature, reduction rate, cooling control parameters, and reddening temperature for each pass.

6. The method for predicting the phase composition of low-alloy high-strength steel based on classification and regression coupling according to claim 1, characterized in that, After obtaining the final phase fraction prediction result, the following is also included: The contribution of the chemical composition characteristics and process parameter characteristics to the prediction results of the phase composition classification model and the phase fraction regression model is quantified based on the SHAP interpretability analysis method.

7. A training method for a machine learning classification and regression coupled prediction model for the phase composition of low-alloy high-strength steel, characterized in that, Includes the following steps: Obtain historical production datasets of low-alloy high-strength steel, which include chemical composition characteristics, process parameter characteristics, and corresponding microstructure classification labels and phase fraction labels of multiple samples; The historical production dataset is used to train a phase composition classification model to obtain a trained phase composition classification model, which is used to predict the existence state of feathery bainite and granular bainite. The phase fraction regression model is trained using the historical production dataset to obtain the trained phase fraction regression model. The phase fraction regression model is used to predict the phase fractions of pearlite, feathery bainite, and granular bainite. In the phase fraction regression model, the regression branches for feathery bainite and granular bainite are only trained on samples where the corresponding microstructure classification label indicates their presence, and samples where the corresponding microstructure classification label indicates their absence are used as zero-value anchor points in the training.

8. A machine learning classification and regression coupled system for predicting the phase composition of low-alloy high-strength steel, characterized in that, include: The model acquisition module acquires a pre-trained phase composition classification model and a phase fraction regression model; wherein, the phase composition classification model is used to predict the presence state of the target microscopic tissue; the phase fraction regression model is a model obtained by training only on training samples where the target microscopic tissue exists, and using training samples where the target microscopic tissue does not exist as zero-value anchor points in the training. The classification and prediction module is used to acquire the chemical composition features and process parameter features of the target to be predicted, and input the chemical composition features and process parameter features into the phase composition classification model to obtain the classification and prediction results; The regression prediction module is used to predict the phase fraction of the target microstructure indicated by the classification prediction result based on the classification prediction result and through the phase fraction regression model. The phase fraction of the target microstructure indicated by the classification prediction result as not existing is directly assigned to zero to obtain the initial phase fraction prediction result. The result correction module is used to correct the initial phase fraction prediction result so that each phase fraction after correction is a non-negative value and the sum of each phase fraction does not exceed 100%, thus obtaining the final phase fraction prediction result.

9. A reverse design method for the composition and processing of low-alloy high-strength steel based on a coupled model, characterized in that, Includes the following steps: A phase composition classification model and a phase fraction regression model are obtained. The phase composition classification model is used to predict the existence state of feathery bainite and granular bainite. The phase fraction regression model is used to predict the phase fraction of pearlite, feathery bainite and granular bainite. The regression branch of the phase fraction regression model for feathery bainite and granular bainite is a model that is trained only on training samples that are indicated to exist by the corresponding classification label and uses training samples that are not indicated to not exist by the corresponding classification label as zero value anchor points for training. Obtain the target microstructure constraints and the preset composition and process search space, which is jointly limited by the value range of chemical composition parameters and the value range of process parameters; Candidate component process combinations are generated within the component process search space. The candidate component process combinations are then input into the phase composition classification model and the phase score regression model, respectively, to obtain candidate classification prediction results and candidate phase score prediction results. When the candidate classification prediction result and the candidate phase fraction prediction result satisfy the target microstructure constraint, the corresponding candidate component process combination will be output as the component process target scheme.

Images