A method for predicting the mechanical properties of steel

Abstract

A method for predicting the mechanical properties of steel is disclosed. The prediction is based on a pre-defined prediction model, and the variables input to the prediction model include: original variables based on historical data; and generated variables based on historical data and empirical formulas for the manufacturing process. This invention, from the perspective of data sources, focuses on optimizing the data structure input to the prediction model as its fundamental innovation. It optimizes both the vertical integration of empirical and intelligent models and the horizontal optimization of each model. Furthermore, to achieve online model updates and maintain good prediction accuracy, a sampling inspection mechanism is implemented. The model is updated based on the sampling results and frequency, ensuring the real-time effectiveness and accuracy of the model's predictions.

Description

Technical Field

[0001] This invention belongs to the field of metallurgical technology and industrial big data, and specifically relates to a method for predicting the mechanical properties of steel. Background Technology

[0002] Common steel products include four main categories: profiles, plates, pipes, and metal products. Specific products include rebar, wire rod, seamless pipes, cold-rolled sheets / coils, hot-rolled sheets / coils, and square steel. The production process is complex, involving multiple steps. Taking medium and heavy plate production as an example, the main stages include slab selection, slab heating, rolling, post-rolling accelerated cooling, hot straightening, cooling bed cooling, shearing, and cold straightening. If the mechanical properties of the product can be predicted before production, the production process can be modified and adjusted in a timely manner, thereby improving the yield.

[0003] The mechanical properties of steel, also known as its mechanical properties, mainly include tensile strength, yield strength, and impact toughness. Tensile strength refers to the strength corresponding to the highest point on the uniaxial tensile stress-strain curve, i.e., the maximum stress value that the steel material can withstand. Yield strength refers to the strength corresponding to the yielding phenomenon on the uniaxial tensile stress-strain curve of steel. Impact toughness refers to the material's ability to absorb plastic deformation work and fracture work under impact loads.

[0004] In steel production, two factors affect mechanical properties. The first is the composition of the steel material, specifically the proportion of trace elements added besides iron; the second is the process parameters at each stage of production, such as the final rolling temperature, slab temperature, roughing mill inlet and outlet temperatures, and finishing mill inlet and outlet temperatures in hot-rolled products. Methods for predicting mechanical properties can be broadly categorized into three types: traditional methods, expert system methods, and artificial intelligence methods. Traditional design primarily employs experimental verification using local samples, but this method is inefficient, lacks accuracy, and has poor versatility. Expert methods establish a rule base based on long-term accumulated rules relating material composition, processes, and properties, designing specific reasoning mechanisms to predict the mechanical properties of metallic materials. However, they suffer from drawbacks such as insufficient knowledge acquisition. Artificial intelligence represents a future direction for mechanical property prediction, with various intelligent algorithms such as machine learning and neural networks. However, due to the use of small datasets, it is difficult to fully represent the mapping relationships between parameters and to reasonably handle the coupling problems between input parameters.

[0005] The former two methods focus on process experience modeling, which involves constructing mathematical models based on laboratory physical metallurgy experiments to predict the mechanical properties of steel. The latter focuses on intelligent algorithm modeling, which uses relevant algorithms to model and predict from a data perspective. Both process experience modeling and intelligent algorithm modeling have their advantages and disadvantages in predicting mechanical properties.

[0006] While traditional process experience models can achieve a certain level of prediction accuracy, the variables considered in the models are fixed, and the process conditions are easily changed, so it is difficult to guarantee portability and robustness.

[0007] The invention application with application number CN201010195215.5 discloses "a method for establishing a prediction model for the tensile strength of cold-rolled steel strip". This method incorporates relevant parameters from the production process, such as chemical composition, final rolling temperature, coiling temperature, cold rolling reduction ratio, annealing time, and leveling elongation, into a regression neural network for analysis and model establishment. The steps are as follows: First, the process parameters for cold-rolled steel strip production are stored in a database; second, the tensile strength of the produced cold-rolled steel strip samples is measured using a stretching machine, and the measurement results are stored in the database; finally, a prediction model is established.

[0008] Invention application CN 202010966431.9 discloses "A Method for Predicting the Mechanical Properties of Automobile Beam Steel Based on LDA Theory," which provides a method for establishing a prediction model for the mechanical properties (yield strength, tensile strength, and elongation) of steel. The production process of hot-rolled high-strength steel generates a large amount of historical production data, and a model training data sample set is constructed based on this historical data. The training data sample set includes characteristic attributes (process and composition parameters) and their corresponding mechanical property parameters. A mechanical property prediction model is established using the training data sample set, and finally, the characteristic attribute parameters (process and composition parameters) of the prediction data sample set are input into the mechanical property prediction model to obtain the mechanical properties of the predicted data samples.

[0009] The invention application with application number CN202110568941.5 discloses "a method for predicting the mechanical properties of steel microstructure based on machine learning". It adopts a deep convolutional neural network model and regression algorithm in machine learning to predict the mechanical properties of steel by learning a metallographic image dataset: the metallographic microstructure image of steel obtained in the experiment is combined with the corresponding mechanical property data measured in the experiment to form a metallographic image-performance dataset. Then, feature extraction is performed on the metallographic image based on a pre-trained deep convolutional neural network. The extracted feature values ​​are further filtered. Finally, the filtered features are mapped to the mechanical property data, and regression mapping is achieved through a regression algorithm. Summary of the Invention

[0010] To address the above problems, this invention provides a method for predicting the mechanical properties of steel, the specific technical solution of which is as follows:

[0011] A method for predicting the mechanical properties of steel, characterized in that:

[0012] The prediction is based on the set prediction model.

[0013] The variables used as input to the prediction model include:

[0014] Original variables based on historical data;

[0015] Generative variables based on historical data and empirical formulas for processes.

[0016] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0017] The prediction model consists of an intelligent algorithm model.

[0018] or

[0019] It is composed of at least two intelligent algorithm models through model fusion.

[0020] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0021] The prediction also establishes a feedback adjustment and update mechanism for the prediction model through an online adjustment system.

[0022] The online adjustment mechanism is based on a set triggering mechanism, including: triggering based on the evaluation result of each prediction and triggering based on the cumulative number of checks set for the prediction model.

[0023] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0024] The prediction is performed according to the following steps:

[0025] S1: Based on the mechanical properties of the corresponding steel to be predicted, obtain the relevant process parameters and historical production data of the corresponding mechanical properties from the database;

[0026] S2: Based on historical production data and according to the empirical formulas of the process, the calculation of the mechanical properties to be predicted is completed, forming a result dataset, which is the generated variable; the data in the historical production data other than the data used in the calculation of the empirical formulas of the process are the original variables;

[0027] S3: Input the original variables and generated variables as input data into the prediction model to complete the training and prediction.

[0028] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0029] The model fusion is performed using the stacking method.

[0030] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0031] Both the evaluation result of each prediction and the cumulative number of checks set for the prediction model can run in parallel.

[0032] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0033] The aforementioned empirical formula is an empirical formula for a process.

[0034] or

[0035] It is a set of process empirical formulas consisting of at least two process empirical formulas.

[0036] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0037] The triggering based on the evaluation result of each prediction is specifically as follows:

[0038] The model's performance is evaluated based on the prediction results, using data collected within a specified time period according to the set inspection frequency.

[0039] When the evaluation results meet the set evaluation criteria, online adjustments will not be triggered.

[0040] When the evaluation results do not meet the set evaluation criteria, the prediction model is triggered to be retrained.

[0041] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0042] The triggering based on the cumulative number of checks set for the prediction model is specifically as follows:

[0043] The system uses data collected within a set time period as input to the prediction model based on the set inspection frequency, and establishes a system to check the performance of the prediction model. When the number of inspections of the same prediction model reaches a set threshold, the prediction model is triggered to be retrained.

[0044] According to the present invention, a method for predicting the mechanical properties of steel is characterized in that:

[0045] The historical data consists of process parameters and composition data from the entry of raw materials into the production line to the completion of rolling;

[0046] Data is collected according to the collection cycle set by each steel grade and model.

[0047] This invention proposes a method for predicting the mechanical properties of steel. This method effectively integrates expert experience models into intelligent algorithm models, improving the low training efficiency and accuracy of commonly used steel mechanical property prediction models. Starting from the data source perspective, its fundamental innovation lies in optimizing the data structure input to the prediction model. It establishes a structure that uses an experience model to generate variables. This innovative setting, by reducing the number of variables and achieving a pruning-like effect, improves the training efficiency of subsequent models while ensuring training accuracy. Furthermore, it optimizes both the vertical integration of the experience model and the intelligent model and the horizontal integration of the experience model and the intelligent model individually. To achieve online automatic updates of the prediction model and maintain good prediction accuracy, a sampling inspection mechanism is implemented. The model is updated based on the sampling results and frequency, ensuring the real-time effectiveness and accuracy of the model's predictions. Simultaneously, by constructing a hybrid model combining process experience models and intelligent algorithm models, process designers can accurately predict the mechanical properties of a certain type of steel, thereby improving product performance by modifying processes in advance, designing optimal process parameters, and increasing product yield. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the prediction steps of the present invention;

[0049] Figure 2 This is a schematic diagram of the hybrid model structure in the working principle of this invention;

[0050] Figure 3 This is a schematic diagram of the model fusion structure in the working principle of this invention. Detailed Implementation

[0051] The method for predicting the mechanical properties of steel according to the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] A method for predicting the mechanical properties of steel.

[0053] The prediction is based on the set prediction model.

[0054] The variables used as input to the prediction model include:

[0055] Original variables based on historical data;

[0056] Generative variables based on historical data and empirical formulas for processes.

[0057] in,

[0058] The historical data consists of process parameters and composition data from the entry of raw materials into the production line to the completion of rolling;

[0059] Data is collected according to the collection cycle set by each steel grade and model.

[0060] in,

[0061] The prediction model consists of an intelligent algorithm model.

[0062] or

[0063] It is composed of at least two intelligent algorithm models through model fusion.

[0064] in,

[0065] The prediction also establishes a feedback adjustment and update mechanism for the prediction model through an online adjustment system.

[0066] The online adjustment mechanism is based on a set triggering mechanism, including: triggering based on the evaluation result of each prediction and triggering based on the cumulative number of checks set for the prediction model.

[0067] Among them, such as Figure 1 As shown, the prediction is performed according to the following steps:

[0068] S1: Based on the mechanical properties of the corresponding steel to be predicted, obtain the relevant process parameters and historical production data of the corresponding mechanical properties from the database;

[0069] S2: Based on historical production data and according to the empirical formulas of the process, the calculation of the mechanical properties to be predicted is completed, forming a result dataset, which is the generated variable; the data in the historical production data other than the data used in the calculation of the empirical formulas of the process are the original variables;

[0070] S3: Input the original variables and generated variables as input data into the prediction model to complete the training and prediction.

[0071] in,

[0072] The model fusion is performed using the stacking method.

[0073] in,

[0074] Both the evaluation result of each prediction and the cumulative number of checks set for the prediction model can run in parallel.

[0075] in,

[0076] The aforementioned empirical formula is an empirical formula for a process.

[0077] or

[0078] It is a set of process empirical formulas consisting of at least two process empirical formulas.

[0079] in,

[0080] The triggering based on the evaluation result of each prediction is specifically as follows:

[0081] The model's performance is evaluated based on the prediction results, using data collected within a specified time period according to the set inspection frequency.

[0082] When the evaluation results meet the set evaluation criteria, online adjustments will not be triggered.

[0083] When the evaluation results do not meet the set evaluation criteria, the prediction model is triggered to be retrained.

[0084] in,

[0085] The triggering based on the cumulative number of checks set for the prediction model is specifically as follows:

[0086] The system uses data collected within a set time period as input to the prediction model based on the set inspection frequency, and establishes a system to check the performance of the prediction model. When the number of inspections of the same prediction model reaches a set threshold, the prediction model is triggered to be retrained.

[0087] Working principle, process and implementation examples

[0088] This technical solution proposes a method for predicting the mechanical properties of steel. It constructs a hybrid model based on a process experience model and an intelligent algorithm model. The prediction results obtained from the experience model are used as input to the intelligent algorithm model, becoming new variables. This model is then trained using historical production data, and the trained results are used as the final mechanical property prediction model. This approach fully utilizes the high accuracy of the experience model under certain conditions while leveraging the advantages of intelligent algorithms in data processing and mining, thus constructing a hybrid model capable of accurately predicting the mechanical properties of steel. The use of ensemble learning to fuse the hybrid model addresses the problem of single intelligent algorithms being overly sensitive to data. Furthermore, a dual-threshold online adjustment mechanism is designed to automatically optimize the prediction model. The specific implementation method is as follows:

[0089] The hybrid model for predicting the mechanical properties of steel consists of two parts: model construction and implementation method.

[0090] 1. Model Building

[0091] 1.1 Hybrid Model Design

[0092] S represents the set of variables used for predicting the mechanical properties of steel, where S = {x1, x2, ..., x...} p}, x i(i = 1, 2, ..., p) represents process parameters, composition parameters, and statistical quantities generated by multiple parameters (e.g., carbon equivalent), etc. i There are n samples, specifically represented as x i ={x 1i ,x 2i ,...,x ni} T U represents the set of variables for the mechanical properties of steel, where y i (i = 1, 2, ..., p1) represents performance indicators such as tensile strength, yield strength, and impact toughness. f and g represent the process experience model and the intelligent algorithm model, respectively, where f1, f2, ..., f... q This represents multiple empirical process models. Empirical model f i (i = 1, 2, ..., q) using k i The mechanical property of U obtained from the prediction of each variable is: Take the union of the variable sets used to calculate these q empirical process models, i.e., S1 = S 11 ∪S 12 ∪...∪S 1q ={x1,x2,...,x k S2 = S - S1 = {x} k+1 ,x k+2 ,...,x p This serves as part of the input for training the intelligent algorithm model. As input variables to the intelligent algorithm model g, a hybrid model is obtained. The variable set {x1, x2, ..., x} k} and {x k+1 ,x k+2 ,...,x p The components are independent of each other. A hybrid model h is trained using historical production data, and the trained result is used as the final hybrid model for predicting the mechanical properties of steel.

[0093] The analysis is based on the above process. The hybrid model can leverage the advantages of either a single process empirical model or an intelligent algorithm model. If the empirical model f performs well in prediction, its results can be incorporated as variables into the intelligent algorithm model. In the process, the optimal model parameters are obtained through training with intelligent algorithm models, and h is... The coefficients of x are very close to 1, and the coefficients of other variables x are also very close to 1. k+1 ,x k+2 ,...,x pThe coefficients are all close to 0, meaning the final hybrid model is very close to the empirical model that has achieved good prediction results. Similarly, if the prediction performance of a single intelligent algorithm model g is good, and the hybrid model h is trained after incorporating the empirical model, the coefficients in h... The coefficients of all variables are very close to 0, and the coefficients of other variables x are also very close to 0. k+1 ,x k+2 ,...,x p The coefficients are all close to those obtained from training a single intelligent algorithm, meaning that the final hybrid model is very close to the intelligent algorithm model that has achieved good prediction results.

[0094] Furthermore, hybrid models can compensate for the shortcomings of single models. (Single empirical model) Due to advancements in model size and mechanistic research, it is impossible to include all relevant variables in the model, i.e., k i While the values ​​may not be very large, the steel production process involves multiple steps, requiring the incorporation of more variables into the model to obtain more accurate predictions of mechanical properties. This is the biggest drawback of a single empirical process model. Intelligent algorithm models can solve this problem. Due to significant improvements in computing power and algorithm performance, intelligent algorithm models (such as machine learning) can perform well when handling high-dimensional data. However, a single intelligent algorithm model does not incorporate the physical meaning of the variables, resulting in predictions lacking sufficient interpretability. Furthermore, historical production data is not very accurate due to limitations in detection methods, data communication, and sample randomness, affecting prediction accuracy. The inclusion of empirical models can provide intelligent algorithms with initial values ​​close to the actual mechanical properties, thus avoiding potential pitfalls in the optimization iteration process and making it easier for the model to obtain optimal parameters.

[0095] In summary, the hybrid model leverages the strengths of each individual model while mitigating its weaknesses, ultimately achieving relatively high prediction accuracy. Its flowchart is shown below. Figure 2 As shown.

[0096] 1.2 Model Fusion Design

[0097] In many cases, a single intelligent algorithm cannot achieve good results in predicting the mechanical properties of steel, and its robustness to different data is not high, making it prone to overfitting. This invention uses the Stacking method from ensemble learning to improve this deficiency. The goal of ensemble learning is to learn a stable model that performs well in all aspects, and to obtain a better and more comprehensive strongly supervised model by combining the supervised models from multiple intelligent algorithms. Stacking is a typical hierarchical fusion model.

[0098] D represents the dataset used for predicting and modeling the mechanical properties of steel, where D = {x} .1 ,x .2,...,x .n} T x .i (i = 1, 2, ..., n) represents the i-th sample with p variables. The p variables include p1 process parameters, composition parameters, and statistics generated from multiple parameters, as well as p2 performance index parameters, i.e., predictor variables. The specific design of the Stacking method here is as follows:

[0099] (1) Divide D into D1 containing p1 process parameters and D2 containing p2 performance indicators. Divide D into training set D1 and training set D2 according to the ratio of test set / total samples = α. A and test set D B Training set D A The sample size is (1-α)n, and the test set D B The sample size is αn.

[0100] (2) Set the cross-validation fold number β.

[0101] (3) The set of models used for regression prediction: Model = {reg1, reg2, ..., reg t}

[0102] (4) D A Based on the logic of cross-validation, the training set is divided into (β-1) folds and the validation set into 1 folds, for a total of β groups.

[0103] (5) Use the (β-1) folded training set from each group separated in (4) to pair the regs selected from the Model in (3). i The training dataset (i = 1, ..., t) is used to obtain a trained model. This trained model is then validated on the validation set to obtain the desired results. αn data points are obtained by predicting the test set using β cross-validations. The prediction results for the validation set are β×αn, and the prediction results for the test set are β×αn.

[0104] (6) Expand the prediction results of the validation set to obtain The validation set prediction matrix is ​​denoted as A1. The test set prediction results are weighted and averaged to obtain the αn×1 test set prediction matrix, denoted as B1.

[0105] (7) Select models from (3) in sequence and perform steps (4)-(6) to obtain the validation set prediction matrix: A = [A1,...,A t The test set prediction matrix is: B = [B1,...,B1] t ].

[0106] (8) Using A as the new training set and B as the new test set, select the regression model reg. new Training them completes the model fusion.

[0107] The above model fusion design is based on the hybrid model design. It integrates multiple hybrid models through ensemble learning, absorbing the advantages of various conventional algorithms to improve the prediction accuracy and robustness of the final model. The flowchart is as follows: Figure 3 As shown.

[0108] 1.3 Design of Online Adjustment Mechanism

[0109] As industrial production progresses, the production data collected for predicting the mechanical properties of steel may drift over time due to factors such as product characteristics, environmental conditions, operating conditions, equipment maintenance and replacement, and sensor drift. This can lead to inaccuracies in prediction models previously trained offline based on historical data. Timely updates to the prediction model are one way to address this issue. However, frequent updates consume and waste computing resources, and due to the low data acquisition frequency, they cannot improve predictions in a short period. Therefore, a robust online adjustment mechanism is needed to automatically optimize the prediction model.

[0110] This invention employs a two-layer threshold online adjustment mechanism, as detailed below:

[0111] (1) D represents the dataset used for predicting the mechanical properties of steel. The collection time interval is determined according to the collection frequency of various types of data to ensure appropriate model training data. E represents the dataset used to test the adaptability of the prediction model. count(D) represents the data size of D, and count(E) represents the data size of E. Generally, count(E) ∈ [20% count(D), 40% count(D)].

[0112] (2) Set the sampling frequency for model inspection to f. c During sampling inspection, historical data within the most recent time period T1 in production operation is selected as E, and the reg obtained from the model fusion design in section 1.2 is used. new To predict E, calculate the coefficient of determination R. 2 When R 2 When the value is below 0.75, a model retraining instruction is triggered.

[0113] (3) When the number of sampling inspections exceeds 10 and the model has not been adjusted, the model retraining instruction is triggered.

[0114] The above-mentioned online adjustment mechanism design allows for online adjustment of the model during normal steel production, meeting the requirement for good steel performance prediction capabilities.

[0115] 2. Hybrid Model Implementation Method

[0116] The method used is illustrated through the prediction of the mechanical properties of a certain type of steel. First, the set of variables related to the mechanical properties of the steel (pipeline steel, hot-rolled plate, or cold-rolled plate, etc.) is denoted as S, and the set of variables for the specific mechanical properties to be predicted (tensile strength, yield strength, or impact toughness, etc.) is denoted as U. Data collection for different production lines, steel grades, and models needs to be conducted separately according to their respective collection cycles to establish adaptation with model training. The current production process is analyzed, and the p-items of key process parameters that affect U are selected from S, including p1 (chemical composition, C, Mn, Si, etc.) and p2 (other factors, such as temperature and cooling rate). Next, modeling is performed through the following steps:

[0117] (1) Export the historical production data of the corresponding production line in the T1 time interval from the database, which includes the p process parameters and mechanical properties U, to form dataset D1;

[0118] (2) Find the empirical formula f for predicting mechanical property U. If multiple empirical models exist, denote them as f1, f2, ..., f q The k process parameters required for these models are all included in the p process parameters mentioned above.

[0119] (3) For dataset D1, calculate the mechanical properties U one by one using the q empirical formulas f in (2) to obtain the corresponding calculated values. Taking the union of the variable sets used to calculate these empirical process models, we get {x1, x2, ..., x...} k}. The remaining pk variables {x k+1 ,x k+2 ,...,x p}, Together with the mechanical properties U, they form a new dataset D2.

[0120] (4) Set the set of intelligent algorithms used: Model = {reg1, reg2, ..., reg t The ensemble learning method used is Stacking.

[0121] (5) D2 U is the independent variable, and D2 is segmented according to the Stacking method. The intelligent algorithms in Model (4) are used sequentially for prediction training, including variable selection, variable normalization, and parameter adjustment. The trained model is then... As a hybrid model.

[0122] (6) Based on the idea of ​​Stacking, the hybrid model is fused, and the final fused model is used to predict the mechanical properties of the new data, which can obtain accurate prediction values ​​and help subsequent decision-making.

[0123] Set the parameters required for the online adjustment mechanism, and update the model promptly when the threshold conditions are met.

[0124] Example

[0125] In the following embodiments, to verify the feasibility of this technical solution, the dataset was divided into a training set and a test set. The feasibility and effectiveness of the technical solution were verified.

[0126] 1. Example 1

[0127] This embodiment uses hot-rolled steel for method verification. The tensile strength of the hot-rolled steel is predicted. Analysis of the production process yields 22 process parameters, including 8 chemical composition parameters (C, Mn, Si...) and 14 other factors (final rolling temperature, cooling time, etc.). Modeling is performed through the following steps:

[0128] (1) 3452 historical production data containing these process parameters and tensile strength were obtained from the database. The training set and the test set were divided in a 9:1 ratio. The training set was denoted as dataset D1.

[0129] (2) The empirical formulas for hot-rolled steel are:

[0130] f1=279.3+6.86C+0.0588Mn+7.35P+1.96Si+K 硬化

[0131] Includes 5 process parameters, K 硬化 Depending on the rolling conditions and the quantity of finished products, if there is no work hardening and the size is small, it can be ignored.

[0132] f2 = 3.3(X) M HV M +X B HV B +(X F +X P HV F+P )

[0133] HV M =127+949ω C +27ω Si +11ω Mn +8ω Ni +16ω Cr +21lgv M

[0134] HVB =-323+185ω C +330ω Si +153ω Mn +65ω Ni +144ω Cr +191ω Mo +(89+53ω C -55ω Si -22ω Mn -10ω Ni -20ω Cr -33ω Mn logv B

[0135] HV F+P =42+223ω C +53ω Si +30ω Mn +12.6ω Ni +7ω Cr +19ω Mo +(10-19ω Si +4ω Ni +8ω Cr +130ω v logv F+P

[0136] It includes 7 process parameters: HV is the Vickers hardness (HV) at a certain location; XM, XB, and XF+P are the percentage contents (%) of each phase; and HVM, HVB, and HVF+P are the Vickers hardness (HV) of martensite, bainite, and ferrite-pearlite.

[0137] (3) For dataset D1, the tensile strength is calculated for each data point using the empirical formula in (2), and the corresponding calculated values ​​are obtained. Taking the union of the variable sets used to calculate these empirical process models requires 8 process parameters. The remaining 14 process parameters... The actual tensile strength in D1 forms a new dataset D2.

[0138] (4) Set the set of intelligent algorithms used as Model = {regression tree, random forest, ..., neural network} and the ensemble learning method used as Stacking.

[0139] (5) Combine the process parameters in D2 and As the independent variable, tensile strength is used as the dependent variable. Data samples of D2 are segmented according to the Stacking method. The intelligent algorithms in Model (4) are used sequentially for prediction training. Taking the neural network algorithm as an example, the training process includes processing of classification variables, setting of hidden layers, setting of activation functions, etc. The trained model is then... As a hybrid model.

[0140] (6) Based on the Stacking concept, the multiple hybrid models obtained in (5) are fused, and the final fused hybrid model is used to predict the mechanical performance of new data. The performance of the empirical model, neural network model, and hybrid model is verified using the test set divided in (1), with the coefficient of determination r. 2 As evaluation metrics, the empirical model scored 0.64, the neural network model scored 0.70, and the hybrid model scored 0.79. This verifies the effectiveness of the hybrid model method designed in this patent.

[0141] (7) After verifying the effectiveness of the model, online adjustment mechanism parameters need to be set, and the model should be updated in a timely manner when the threshold conditions are met.

[0142] 2. Example 2

[0143] This embodiment uses pipeline steel for method verification. The yield strength of the pipeline steel is predicted. Analysis of the production process yields 20 process parameters, including 11 chemical components (C, Mn, Si, ...) and 9 other factors (furnace tapping temperature, initial rolling temperature, initial cooling temperature, etc.). Modeling is performed through the following steps:

[0144] (1) 21,853 historical production data containing these process parameters and yield strength were obtained from the database. The training set and the test set were divided in a 9:1 ratio. The training set was denoted as dataset D1.

[0145] (2) The empirical formula for the yield strength of pipeline steel is: Where [C%], [Nb%], and [Mn%] represent the percentage content of chemical components, d f V represents the grain size of ferrite (μm). p T represents the volume fraction of pearlite. c h represents the curling temperature, and h represents the thickness (mm).

[0146] (3) For dataset D1, calculate the yield strength for each data point using the empirical formula in (2) to obtain the corresponding calculated value. Taking the union of the variable sets for calculating these empirical process models requires 7 process parameters. The remaining 13 process parameters... The actual yield strength in D1 forms a new dataset D2.

[0147] (4) Set the set of intelligent algorithms used as Model = {regression tree, random forest, ..., neural network} and the ensemble learning method used as Stacking.

[0148] (5) Combine the process parameters in D2 and As the independent variable, yield strength is used as the dependent variable. Data samples of D2 are segmented according to the Stacking method. The intelligent algorithms in Model (4) are used sequentially for prediction training. Taking the random forest algorithm as an example, the training process includes the processing of categorical variables, the number of trees, and the number of tree layers, etc. The trained model is then... As a hybrid model.

[0149] (6) Based on the Stacking concept, the multiple hybrid models obtained in (5) are fused, and the final fused hybrid model is used to predict the mechanical performance of the new data. The performance of the empirical model, the random forest model, and the hybrid model is verified using the test set divided in (1), with the coefficient of determination r. 2 As performance metrics, the empirical model scored 0.77, the random forest model scored 0.76, and the hybrid model scored 0.82. This verifies the effectiveness of the hybrid model method designed in this patent.

[0150] (7) After verifying the effectiveness of the model, online adjustment mechanism parameters need to be set, and the model should be updated in a timely manner when the threshold conditions are met.

[0151] 3. Example 3

[0152] This embodiment uses non-quenched and tempered pipeline steel for method verification and predicts its yield strength. Analysis of the production process yielded 24 process parameters, including 7 chemical composition parameters (C, Mn, Si, ...) and 17 other factors (soaking temperature, upper cooling bed temperature, final rolling temperature, etc.). Modeling was performed through the following steps:

[0153] (1) 5643 historical production data containing these process parameters and yield strength were obtained from the database and divided into training set and test set in a 9:1 ratio to form dataset D1;

[0154] (2) The empirical formulas for pipeline steel are:

[0155]

[0156] It includes 8 process parameters, where %Si, %N, and %Mn are the percentage contents of the chemical composition, and f p d represents the volume fraction of the second-phase particles. p For the size of the second-phase particles, d a f is the ferrite grain size. aS is the ferrite volume fraction, and S0 is the pearlite lamellar spacing.

[0157] (3) For dataset D1, calculate the yield strength for each data point using the empirical formula in (2) to obtain the corresponding calculated value. Eight process parameters are required. The remaining 16 process parameters... The actual yield strength in D1 forms a new dataset D2.

[0158] (4) Set the set of intelligent algorithms used as Model = {regression tree, GBDT, ..., neural network} and the ensemble learning method used as Stacking.

[0159] (5) Combine the process parameters in D2 and As the independent variable, yield strength is used as the dependent variable. Data samples of D2 are segmented according to the Stacking method. The intelligent algorithms in Model (4) are used for prediction training in sequence. Taking the GBDT algorithm as an example, the training process includes processing of categorical variables, minimum number of partitions, number of leaf nodes, etc. The trained model is then... As a hybrid model.

[0160] (6) Based on the Stacking concept, the multiple hybrid models obtained in (5) are fused, and the final fused hybrid model is used to predict the mechanical properties of the new data. The performance of the empirical model, the GBDT model, and the hybrid model is verified using the test set divided in (1), with the coefficient of determination r. 2 As a metric, the empirical model scored 0.78, the GBDT model scored 0.84, and the hybrid model scored 0.85. This verifies the effectiveness of the hybrid model method designed in this patent.

[0161] (7) After verifying the effectiveness of the model, online adjustment mechanism parameters need to be set, and the model should be updated in a timely manner when the threshold conditions are met.

[0162] The above embodiments, through combined design experiments on multiple steel types, multiple mechanical performance indicators, and multiple intelligent algorithm models, verified the effectiveness and robustness of the hybrid model method.

[0163] This invention provides a method for predicting the mechanical properties of steel, which effectively combines a process experience model and an intelligent algorithm model. The results of the process experience model in the first stage serve as the input for the intelligent algorithm model in the second stage, thus fully utilizing the advantages of both to achieve high-precision prediction of the mechanical properties of steel. It cleverly integrates the two models to construct a hybrid model, thereby retaining the advantages of individual intelligent algorithm models while maximizing the use of expert experience and statistical regularities from historical production data to compensate for the shortcomings of individual models. Furthermore, to address the issues of robustness and low accuracy of individual intelligent algorithms, an ensemble learning method is employed to design the prediction structure of the fusion model. An online update logic for the prediction model is also designed, facilitating sustainable use and reducing unnecessary expenditure of manpower and resources.

[0164] In summary, this invention provides a method for predicting the mechanical properties of steel. It effectively integrates expert experience models into intelligent algorithm models, improving the low training efficiency and accuracy of commonly used steel mechanical property prediction models. Starting from the data source perspective, its fundamental innovation lies in optimizing the data structure input to the prediction model. It establishes a structure that utilizes an experience model to generate variables. This innovative setting, by reducing the number of variables and achieving a pruning-like effect, improves the training efficiency of subsequent models while ensuring training accuracy. Furthermore, it optimizes both the vertical integration of the experience model and the intelligent model and the horizontal integration of the experience model and the intelligent model individually. To achieve online automatic updates of the prediction model and maintain good prediction accuracy, a sampling inspection mechanism is implemented. The model is updated based on the sampling results and frequency, ensuring the real-time effectiveness and accuracy of the model's predictions. Simultaneously, by constructing a hybrid model combining process experience models and intelligent algorithm models, process designers can accurately predict the mechanical properties of a certain type of steel, thereby improving product performance by modifying processes in advance, designing optimal process parameters, and increasing product yield.

Claims

1. A method for predicting mechanical properties of steel materials, characterized in that: the prediction is based on a set of prediction models, the variables for inputting the prediction models include: original variables: process parameters and composition data based on historical data; generated variables: mechanical property prediction values calculated based on the historical data and process empirical formulas, wherein the process empirical formulas include one or more empirical models for outputting intermediate prediction results as inputs of the prediction models, the prediction models are hybrid models, including: one or more process empirical formulas for calculating generated variables from original variables; intelligent algorithm models, taking the generated variables and the remaining original variables as inputs, outputting final mechanical property prediction values, the prediction further establishes feedback adjustment and update of the prediction models through a set of online adjustment mechanisms, the online adjustment mechanisms are based on a set of trigger mechanisms, including: triggering according to evaluation results of each prediction, triggering according to a set number of cumulative checks of the prediction models.

2. The method for predicting mechanical properties of steel materials according to claim 1, characterized in that: the prediction models are composed of one intelligent algorithm model, or composed of at least two intelligent algorithm models through model fusion.

3. The method for predicting mechanical properties of steel materials according to claim 1, characterized in that: the prediction further establishes feedback adjustment and update of the prediction models through a set of online adjustment mechanisms, the online adjustment mechanisms are based on a set of trigger mechanisms, including: triggering according to evaluation results of each prediction: triggering retraining when a prediction determination coefficient (R2) is lower than 0.75; triggering according to a set number of cumulative checks of the prediction models: triggering retraining when the number of spot checks reaches 10 times and no adjustment is made.

4. The method for predicting mechanical properties of steel materials according to claim 1, characterized in that: the prediction is performed according to the following steps: S1: based on the mechanical properties of the corresponding steel materials to be predicted, obtaining relevant process parameters and historical production data of corresponding mechanical properties in a database; S2: based on the historical production data, calculating the mechanical properties to be predicted according to process empirical formulas to form a result data set, which is the generated variables; the data in the historical production data other than the data participating in the calculation of the process empirical formulas are the original variables; S3: inputting the original variables and the generated variables as input data into the prediction models to complete training and prediction.

5. The method for predicting mechanical properties of steel materials according to claim 2, characterized in that: the model fusion is performed according to the stacking method.

6. The method for predicting mechanical properties of steel materials according to claim 3, characterized in that: the triggering according to evaluation results of each prediction and the triggering according to a set number of cumulative checks of the prediction models are operated in a parallel manner.

7. The method for predicting mechanical properties of steel materials according to claim 4, characterized in that: the process empirical formulas are one process empirical formula, or a set of process empirical formulas composed of at least two process empirical formulas. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ 8. The method for predicting the mechanical properties of steel according to claim 3 or 6, characterized in that: the triggering according to the evaluation results of each prediction is specifically: collecting data in the set time period as the input of the prediction model according to the set checking frequency, and evaluating the performance of the model according to the prediction results; when the evaluation results meet the set evaluation standard, no online adjustment is triggered; when the evaluation results do not meet the set evaluation standard, the prediction model is triggered for retraining.

9. The method for predicting the mechanical properties of steel according to claim 3 or 6, characterized in that: the triggering according to the set cumulative number of checks of the prediction model is specifically: collecting data in the set time period as the input of the prediction model according to the set checking frequency, and establishing the check of the performance of the prediction model accordingly; when the number of checks of the same prediction model reaches the set threshold, the prediction model is triggered for retraining.

10. The method for predicting the mechanical properties of steel according to claim 1, characterized in that: the historical data is composed of process parameters and composition data from the entry of raw materials into the production line to the completion of rolling; it is formed by collecting according to the set collection period of each production steel grade and model.

Images