Heavy rail local strength online prediction method

Through the multi-pass rolling model and Gleeble thermal simulation experimental machine, a dynamic and static recrystallization simulation and machine learning algorithm was combined with the machine learning algorithm, a heavy rail local intensity online prediction method was established, solving the problems of slow detection cycle and low accuracy in the existing technology. , achieving high-precision online forecasting and improved production stability.

CN118917148APending Publication Date: 2024-11-08ANSTEEL BEIJING RES INST CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411053962.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing technology detects yield strength and tensile strength in heavy rail production, which has problems such as offline detection, slow cycle and consumes manpower and material resources. The microstructure evolution calculation simulation relies on rough physical metallurgy models, which is difficult to meet the requirements of accuracy and real-time forecasting. .

Method used

Multi-pass rolling model combined with Gleeble thermal simulation experimental machine for dynamic and static recrystallization process simulation, establish a cellular automaton model containing dynamic recrystallization and static softening modules, and build databases and mappings through machine learning or deep learning algorithms. Model to achieve online prediction of local intensity of heavy tracks.

Benefits of technology

It realizes high-precision online strength performance forecasting, reduces the number of manual detections, improves production stability, and leaves room for the future development of cellular automaton technology to improve the computing speed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917148A_ABST
    Figure CN118917148A_ABST
Patent Text Reader

Abstract

The invention relates to a heavy rail local strength online prediction method, which comprises the following steps: setting a cellular automaton calculation model and a finite element calculation model as offline calculation models, adopting an orthogonal batch calculation mode to obtain enough data points calculated by an offline cellular automaton, and utilizing a deep learning model as a medium to predict the local strength of a heavy rail. The model is converted into a model which can be completely used for online real-time forecasting; according to the method, the calculation precision of the model can be ensured through the cellular automaton, the total calculation time requirement required by online real-time forecasting can be met, and the problem of calculation time increase caused by model complication after the subsequent cellular automaton technology development is also considered; the problem that the calculation time of the whole model is affected after the development of the cellular automaton technology in the future is solved, the number of manual detection times can be reduced, and online strength performance forecasting is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of steel rolling, and in particular to an online prediction method for the local strength of a heavy rail. Background Art

[0002] After heavy rail production, specific parts need to be tested for yield strength and tensile strength. This test is an offline test with a slow test cycle and consumes a lot of manpower and material resources. If the yield strength and tensile strength of the parts to be tested can be accurately predicted online, it can save a lot of production costs for the company. At the same time, the yield strength and tensile strength can be associated with the production process parameters, and early warnings can be issued for abnormally produced heavy rails, thereby improving the production stability of the production line.

[0003] During the production process of heavy rails, there are macroscopic deformation processes and macroscopic temperature change processes. In this process, there are microscopic structural changes, usually dynamic recrystallization and static recrystallization processes. The final grain size and temperature change process determine the yield strength and tensile strength of the heavy rail. From a macroscopic perspective, the local temperature change process can be calculated by commercial software such as Fluent. From a microscopic perspective, the microscopic size of the grain can be calculated by microscopic methods. The most developed method is the cellular automaton method. In terms of digitalization, the parameters of hot-rolled strip and other products of concern are currently predicted mainly through methods such as big data analysis. However, due to the relatively rough physical metallurgical model, false positive and false negative prediction results often occur, which is difficult to meet the actual use on site.

[0004] At present, the computational simulation of the microstructure evolution of steel is still mainly based on offline simulation. In order to improve the prediction accuracy of the model, it is necessary to increase the complexity of the model and consider more dynamic recrystallization and static recrystallization processes, thereby forming more loop nesting, which leads to a longer final calculation time. At the same time, the macroscopic temperature change curve and stress-strain curve calculation on which the microstructure prediction depends are also calculated by finite element commercial software. The single calculation time is slow and cannot be integrated with the microscopic calculation model. For example, the invention patent with publication number CN101591729 B describes the establishment of a dynamic recrystallization cellular automaton model, but it does not consider processes such as static recrystallization, and the calculation accuracy is low. Another example is the invention patent with publication number CN114417664A, which proposes an online simulation and visualization method for the microstructure evolution of hot-rolled steel based on cellular automata. The calculation model of its cellular automaton is also too simple. Although it is written in Fortran language and uses an Open MP multi-processor compilation solution to achieve the level of online application, it still cannot completely solve the requirements of online application of improved calculation accuracy and calculation time. When the cellular automaton model becomes complicated, the method of patent CN114417664A will gradually lose the possibility of online application. Summary of the invention

[0005] The present invention provides an online prediction method for the local strength of heavy rails, which can not only ensure the calculation accuracy of the model through cellular automata, but also meet the total calculation time requirement required for online real-time prediction. It also takes into account the problem of increased calculation time caused by the complexity of the model after the subsequent development of cellular automata technology, solves the problem of the impact of the future development of cellular automata technology on the calculation time of the entire model, and can reduce the number of manual inspections to achieve online strength performance prediction.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for online prediction of local strength of heavy rail comprises the following steps:

[0008] A method for online prediction of local strength of heavy rail comprises the following steps:

[0009] S1. Establishing a first model: establishing a multi-pass rolling model using the pass profiles of the heavy rails. The input conditions of the model include the heavy rail model, the material properties of the heavy rails and the incoming temperature of the heavy rails. The output information includes the final shape of the heavy rails and the stress and strain curves of each local position.

[0010] S2. Establishing the first experiment: using a Geleeble thermal simulation test machine to simulate the rolling process of the first model established in S1, conducting a constant temperature and constant strain rate hot compression experiment on the Gleeble thermal simulation test machine to study the dynamic recrystallization process, and conducting a constant temperature and constant strain rate hot compression + heat preservation experiment on the Gleeble thermal simulation test machine to study the static recrystallization process;

[0011] S3. Establish a second model: a cellular automaton model calculation model including dynamic recrystallization and static softening modules, simulate the first experimental process described in S2, use the austenite grain boundary obtained in S2 as an input condition, calculate the final grain size result, and use the result of the first experiment obtained in S2 to verify the result and correct the program, so as to obtain a cellular automaton model calculation model with a rolling process error of less than 5% covering the heavy rail steel grades to be studied;

[0012] S4, establishing a first database: using an orthogonal test method, arranging and combining the production parameter changes encountered in production, and establishing a database of permutations and combinations suitable for calculation; selecting one of the combinations, using the stress, strain, and temperature curves obtained by the first model established by S1, and using the S3 method to process the samples collected on site before rolling as input conditions, and finally using the cellular automaton model calculation model developed by S3 to calculate the grain distribution and grain size under the corresponding working conditions; repeating the above process to obtain the grain distribution and grain size calculation results of each grouping, and placing them in the first database;

[0013] S5. Establish the third model: transform the data in the database established in S4 into an input-output mapping space through machine learning or deep learning algorithms;

[0014] S6, embed the third model constructed in S5 into the on-site production model, collect parameters such as steel type, rolling force, temperature, etc. in production in real time, and generate prediction results of local yield strength and tensile strength of heavy rails through the third model;

[0015] S7. During the use of the model, when inaccurate predictions of certain extreme working conditions are encountered, the first database established in S4 is corrected.

[0016] Furthermore, the step S1 specifically includes the following steps:

[0017] S11, select finite element software to construct the model, wherein the roller adopts a rigid body and the rolled piece adopts an elastic body;

[0018] S12. Create a rolled piece model required for simulation based on actual production process, including temperature, material properties and heat transfer coefficient;

[0019] S13, setting a driving force for a certain period of time behind the rolled piece, so as to force the rolled piece to enter the roll gap formed by the rollers through friction, and use the friction to complete the subsequent rolling simulation;

[0020] S14, according to different parts of interest, set corresponding sets or paths during modeling, and use post-processing software to obtain calculation results of the corresponding sets or paths;

[0021] S15, post-processing the calculation results obtained in S14 to obtain input conditions for subsequent calculation models.

[0022] Furthermore, the step S2 specifically includes the following steps:

[0023] S21. Conduct a constant temperature and constant strain rate hot compression experiment on a Gleeble thermal simulation test machine to study the dynamic recrystallization process: select a cylindrical sample of φ8×12mm, and place graphite or tantalum sheets at both ends of the compression sample before hot compression;

[0024] S22. After hot compression, measure the diameter of the middle bulging part and the end diameter of the sample in the height direction;

[0025] S23, using thermocouples to track the actual temperature changes of the compressed samples in real time, and the selected hot compression temperature and strain rate cover the temperature and strain rate experienced by the heavy rail steel grades to be studied during continuous hot rolling;

[0026] S24. In the compression test, the sample is heated to 1100-1200°C at a rate of 5-10°C / s and kept at this temperature for 3-5 minutes; then cooled to a deformation temperature of 900-1150°C at a rate of 5-10°C / s and kept at this temperature for 1-3 minutes; finally, the sample is heated at a temperature of 900-1150°C and a temperature of 0.001-10s. -1 The sample was compressed and deformed to a true strain of 0.7 to 0.9 at a strain rate, and quenched immediately after deformation;

[0027] S25. A constant temperature and constant strain rate hot compression + heat preservation experiment was carried out on a Gleeble thermal simulation test machine to study the static recrystallization process. The sample size, lubrication conditions and temperature measurement requirements of the experiment were the same as those of the constant temperature and constant strain rate hot compression experiment.

[0028] S26, heat the sample to 1100-1200℃ at a rate of 5-10℃ / s, keep it warm for 3-5min; then cool it to the deformation temperature at a rate of 3-5℃ / s, keep it warm for 1-3min; 900-1150℃ and 0.001-10s -1 The sample is compressed and deformed at a strain rate until the true strain is 0.15 to 0.3; then the deformation is stopped, and the sample is kept warm for a corresponding time of 1 to 30 seconds, and quenched immediately after the holding is completed;

[0029] S27. After the experiment is completed, the sample is split longitudinally with a wire cutting machine, and the sample cross section is mechanically ground using 400#, 800#, 1500# and 2000# sandpaper in sequence, and then polished to a mirror state using a diamond polishing liquid with a particle size of 2.5μm. Finally, a saturated picric acid aqueous solution is used to corrode the original austenite grain boundary in a 50-70℃ water bath for 3-5 minutes until the sample surface color darkens, and then it is immediately cleaned with alcohol and blown dry with a hair dryer;

[0030] S28. Place the corroded sample under an optical microscope for observation and take photos under a 200x or 500x objective lens. Take at least 10 photos for each sample. Statistic the obtained microstructure photos to obtain the average grain size of the original austenite.

[0031] Furthermore, the step S3 specifically includes the following steps:

[0032] S31. Establish a dislocation density evolution model in the dynamic recrystallization model: adopt the Kocks-Mecking model, that is, the KM dislocation density model, and the specific formula is as follows:

[0033]

[0034] Among them, ρ is the dislocation density, ε is the strain, k1 is the hardening index, and k2 is the softening index;

[0035] S32, establish a dynamic recrystallization nucleation model in a dynamic recrystallization model, using the following model, the specific formula is as follows:

[0036]

[0037] in, is the nucleation rate, C DRX , m is the model constant, Q act is the dynamic recrystallization activation energy, is the strain rate, R is the ideal gas state constant, T is the deformation temperature, T m is the melting point used to calculate the shear modulus at a specific temperature;

[0038] S33. Establishing a dynamic recrystallization grain growth model in a dynamic recrystallization model:

[0039] v i =M i f i =M i [τ(ρ m -ρ i )-γ i κ i ] (3)

[0040] Among them, v i is the growth rate of dynamically recrystallized grains, M i is the grain boundary mobility, f i is the growth driving force, ρ m is the matrix dislocation density, ρ i is the dislocation density of the ith unit cell, γ i is the grain boundary energy of the i-th unit cell, κ i is the grain boundary curvature;

[0041] S34. Establishing a static recovery model in a static softening model:

[0042] The static reply SRV formula is:

[0043] ρ(t1)=ρ(t0)exp(-k2ε res ) (4)

[0044] Among them, ε res is the residual strain, ρ(t1) and ρ(t0) are the dislocation densities at the current and next moment, t0 = t1 + Δt, k2 is a constant;

[0045] S35. Establishing a static recrystallization model in a static softening model:

[0046] The nucleation rate of static recrystallization SRX is calculated as:

[0047]

[0048] in, is the nucleation probability, C SRX is the model constant, E is the residual deformation energy, and E min is the critical residual deformation energy, V Ω is the non-recrystallized volume fraction;

[0049] S36. Establish a grain coarsening model in the static softening model: When the residual deformation energy in the material is completely consumed, the static recovery SRV, metadynamic recrystallization MDRX and static recrystallization SRX processes are completed, and the grains continue to grow under the single drive of the interface energy until the next deformation begins or the material is cooled. At this time, the calculation formula for the grain growth rate becomes:

[0050] v i =M i f i =M i γ i κ i (6)

[0051] Furthermore, the step S5 specifically includes the following steps:

[0052] S51. Use machine learning or deep learning methods to build a multi-layer training model, and divide the input variables in the data in the database established by S4, including the stress-strain related data obtained by the Gleeble thermal simulation test machine, and the grain size parameters and output variables calculated by the offline calculation model of the cellular automaton constructed by S3: the original microstructure photo results obtained in S2 into a training set and a validation set by a random method to ensure that data grouping does not have an adverse effect on the training of the model;

[0053] S52. Using the training set to train a machine learning or deep learning method;

[0054] S53. Use the validation set to validate the machine learning or deep learning method;

[0055] S54. Determine whether the verification result meets the production requirements. If not, adjust the machine learning or deep learning method.

[0056] S55. Through repeated iterations, the required mapping space of input and output is finally obtained.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1) The offline model has high calculation accuracy. Since the offline calculation part of this model uses the finite element model and the relevant data of the Gleeble thermal simulation experiment to provide data input for the cellular automaton, the obtained cellular automaton model is more consistent with the actual rolling phase transformation process of the material, especially the cellular automaton that includes dynamic recrystallization and static softening modules, which has higher calculation accuracy than ordinary cellular automata;

[0059] 2) The online model has a fast calculation speed. The online model calculation part uses the mapping model trained by the machine learning algorithm. While improving the online model calculation speed, it also leaves room for program upgrades to further improve the simulation accuracy of the cellular automation program in the future;

[0060] 3) The construction idea of ​​the model of the method described in the present invention can be applied to other spatial resolution simulation methods of microstructure evolution that may appear in the future. The model is applicable to a wider range of steel types and production processes. Since the cellular automaton program of this model is an offline calculation, the processing processes of different types of rails, hot-rolled strip steel, and even metals such as copper and magnesium can also be matched with corresponding cellular automata. Even if the offline calculation speed slows down due to the increase in phase change factors considered in the cellular automaton, it does not affect the calculation speed of the online model, so it has good scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flow chart of the method of the present invention.

[0062] Figure 2 It is a schematic diagram of the thermal compression and heat preservation experimental process of the present invention.

[0063] Figure 3 Schematic diagram of the sample before and after hot compression according to the present invention.

[0064] Figure 4 It is the calculation flow chart of the cellular automaton described in the present invention.

[0065] Figure 5 It is a schematic diagram of the original microstructure of the present invention.

[0066] Figure 6 It is the final static softened tissue of the present invention. DETAILED DESCRIPTION

[0067] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0068] See Figure 1 , is a flow chart of the method of the present invention. The present invention provides an online prediction method for the local strength of heavy rails, comprising the following steps:

[0069] S1. Establish the first model, and use the hole types of each pass of the heavy rail to establish a multi-pass rolling model. The input conditions of the model are the heavy rail model, the material properties of the heavy rail, the incoming temperature of the heavy rail and other information. The output information is the final shape of the heavy rail and the stress and strain curves of each local position. The model is a macro model. In order to perform large-scale calculations, it is set as a parameterized model, and the python language is used for batch pre-processing and post-processing to reduce model construction errors and improve information processing efficiency. The specific steps include the following:

[0070] S11. Select finite element software. Depending on the different genuine finite element software purchased by different enterprises and schools, you can use software such as ansys, ABAQUS, MARC, etc. to build the model, in which the roll uses a rigid body and the rolled piece uses an elastic body;

[0071] S12. Create a rolling model required for simulation according to the actual production process, including parameters such as temperature, material properties, and heat transfer coefficient;

[0072] S13, setting a driving force for a certain period of time behind the rolled piece, so as to force the rolled piece to enter the roll gap formed by the rollers through friction, and use the friction to complete the subsequent rolling simulation;

[0073] S14, according to different parts of interest, set the corresponding set or path during modeling, and use post-processing software to obtain the calculation result of the corresponding set or path;

[0074] S15, post-processing the calculation results obtained in S14 to obtain input conditions that can be used for subsequent calculation models.

[0075] S2. Establish a first experiment, use a Geleeble thermal simulation test machine to simulate the simulated rolling process of the first model established in S1. The experiment is divided into two parts, simulating dynamic recrystallization and static recrystallization processes respectively: a constant temperature and constant strain rate hot compression experiment is performed on the Gleeble thermal simulation test machine to study the dynamic recrystallization process, and a constant temperature and constant strain rate hot compression + insulation experiment is performed on the Gleeble thermal simulation test machine to study the static recrystallization process. Cut and corrode the test samples after the experiment to obtain the average grain size and distribution of the original austenite, as well as the average grain size and distribution after simulated rolling. Specifically, the following steps are included:

[0076] S21. Conduct a constant temperature and constant strain rate hot compression experiment on a Gleeble thermal simulation test machine to study the dynamic recrystallization process: select a cylindrical sample of φ8×12mm. Before hot compression, place graphite or tantalum sheets at both ends of the compressed sample to reduce the uneven deformation caused by the friction between the indenter and the sample.

[0077] S22. After hot compression, measure the diameter of the middle bulging part and the end diameter of the sample in the height direction to evaluate the bulging degree and make friction correction for flow stress;

[0078] S23. Use thermocouples to track the actual temperature changes of the compressed samples in real time, so as to evaluate the influence of temperature rise caused by deformation and make temperature correction for flow stress. The selected hot compression temperature and strain rate cover the temperature and strain rate experienced by the heavy rail steel grades to be studied during continuous hot rolling.

[0079] S24. In the compression test, the sample was heated to 1200°C at a rate of 5°C / s and kept at that temperature for 3 minutes; then cooled to the deformation temperature of 900°C at a rate of 5°C / s and kept at that temperature for 1 minute; finally, the sample was heated to 1200°C at a rate of 5°C / s and kept at that temperature for 3 minutes. -1 ~10s -1 The sample was compressed and deformed to a true strain of 0.7 at a strain rate and quenched immediately after deformation;

[0080] S25, see Figure 2 , a constant temperature and constant strain rate hot compression + heat preservation experiment was carried out on a Gleeble thermal simulation test machine to study the static recrystallization process. The sample size, lubrication conditions and temperature measurement requirements of the experiment were the same as those of the constant temperature and constant strain rate hot compression experiment;

[0081] S26, see Figure 3 First, the sample was heated to 1200°C at a rate of 5°C / s and kept at that temperature for 3 min; then it was cooled to the deformation temperature at a rate of 5°C / s and kept at that temperature for 1 min; then the sample was heated to 900°C and kept at that temperature for 0.001 to 10 s. -1The sample is compressed and deformed at a strain rate until the true strain is 0.15; then the deformation is stopped, and the sample is kept warm for a corresponding time of 1 to 30 seconds, and quenched immediately after the holding is completed;

[0082] S27. After the experiment is completed, the sample is split longitudinally with a wire cutting machine, and the sample cross section is mechanically ground using 400#, 800#, 1500# and 2000# sandpaper in sequence, and then polished to a mirror state using a diamond polishing liquid with a particle size of 2.5μm. Finally, a saturated picric acid aqueous solution is used to corrode the original austenite grain boundary in a 50-70℃ water bath for 3-5 minutes until the sample surface color darkens, and then it is immediately cleaned with alcohol and blown dry with a hair dryer;

[0083] S28. Place the corroded sample under an optical microscope for observation and take photos under a 200x or 500x objective lens. Take at least 10 photos for each sample. Statistic the obtained microstructure photos to obtain the average grain size of the original austenite. The relevant photos can be further used to guide the development of physical metallurgical models.

[0084] S3. Build the second model, see Figure 4 The second model is a cellular automaton model calculation model including dynamic recrystallization and static softening modules. The model is established based on the physical metallurgical model. Its purpose is to simulate the first experimental process described in S2, take the austenite grain boundary obtained by S2 as the input condition, calculate the final grain size result, and use the results of the first experiment obtained by S2 to verify the results and correct the program, so as to obtain a cellular automaton model calculation model with an error of less than 5% in the rolling process of the heavy rail steel to be studied. These models are the underlying logic of the cellular automaton program calculation. With the deepening of the research, new models can be added to the cellular automaton calculation program at any time. However, with the increase of added models, the cycle of the cellular automaton program will also increase, and the calculation time will also increase accordingly. Specifically, the following steps are included:

[0085] S31. Establishment of dislocation density evolution model in dynamic recrystallization model

[0086] The present invention adopts the Kocks-Mecking model, that is, the KM dislocation density model, and the specific formula is as follows:

[0087]

[0088] Among them, ρ is the dislocation density, ε is the strain, k1 is the hardening index, and k2 is the softening index;

[0089] S32. Establishment of dynamic recrystallization nucleation model in dynamic recrystallization model

[0090] This method adopts the following model, and the specific formula is as follows:

[0091]

[0092] in, is the nucleation rate, C DRX , m is the model constant, Q act is the dynamic recrystallization activation energy, is the strain rate, R is the ideal gas state constant, T is the deformation temperature, T m is the melting point used to calculate the shear modulus at a specific temperature;

[0093] S33. Establishing a dynamic recrystallization grain growth model in a dynamic recrystallization model

[0094] The model used in this study is as follows:

[0095] v i =M i f i =M i [τ(ρ m -ρ i )-γ i κ i ] (9)

[0096] Among them, v i is the growth rate of dynamically recrystallized grains, M i is the grain boundary mobility, f i is the growth driving force, ρ m is the matrix dislocation density, ρ i is the dislocation density of the ith unit cell, γ i is the grain boundary energy of the i-th unit cell, κ i is the grain boundary curvature;

[0097] S34. Establishing a static recovery model in a static softening model

[0098] Static recovery SRV will cause a decrease in the material dislocation density, which can be described as:

[0099]

[0100] Among them, ε res is the residual strain, ρ(t1) and ρ(t0) are the dislocation densities at the current and next moment, t0 = t1 + Δt, k2 is a constant;

[0101] S35. Establishing a static recrystallization model in a static softening model

[0102] The nucleation rate of static recrystallization SRX is calculated as:

[0103]

[0104] in, is the nucleation probability, C SRX is the model constant, E is the residual deformation energy, and E min is the critical residual deformation energy, V Ω is the non-recrystallized volume fraction;

[0105] S36. Establishing the grain coarsening model in the static softening model

[0106] When the residual deformation energy in the material is completely consumed, the SRV, MDRX and SRX processes end. Subsequently, the grains continue to grow under the sole drive of the interface energy until the next deformation begins or the material is cooled to a lower temperature. This process is called GC. Since there is no deformation energy driving the GC process, the calculation formula for the grain growth rate at this time becomes:

[0107] v i =M i f i =M i γ i κ i (12)

[0108] S4. Establish a first database, use orthogonal test method to arrange and combine the production parameter changes that may be encountered in production, and establish a database of arrangements and combinations suitable for calculation; select one of the combinations, use the stress, strain, and temperature curves obtained by the first model established by S1, and use the S3 method to process the grain distribution of the samples collected on site before rolling as input conditions, and finally use the cellular automaton model calculation model developed by S3 to calculate the grain distribution and grain size under the corresponding working conditions, repeat the above process to obtain the grain distribution and grain size calculation results of each grouping, and place them in the first database.

[0109] S5. Establish a third model. The model can be based on a machine learning algorithm or a deep learning algorithm. The function of the model is to transform the data in the database established in S4 into an input-output mapping space through a machine learning or deep learning algorithm so as to be used for subsequent online computing. The model specifically includes the following steps:

[0110] S51. Use machine learning or deep learning methods to build a multi-layer training model, the purpose of which is to convert the input variables in the data in the database established by S4, including the stress-strain related data obtained by the Gleeble thermal simulation test machine, and the grain size parameters and output variables calculated by the offline calculation model of the cellular automaton constructed by S3: the original microstructure photos obtained in S2, see Figure 5 , Figure 6 ,The results are randomly divided into training set and validation set to ensure that data grouping does not have adverse effects on model training;

[0111] S52. Using the training set to train a machine learning or deep learning method;

[0112] S53. Use the validation set to validate the machine learning or deep learning method;

[0113] S54. Determine whether the verification result meets the production requirements. If not, fine-tune the machine learning or deep learning method, or try other model methods to improve the training accuracy.

[0114] S55. Through repeated iterations, the required mapping space of input and output is finally obtained for subsequent use in online calculations.

[0115] S6. Embed the third model constructed in S5 into the on-site production model, collect parameters such as steel type, rolling force, temperature, etc. in production in real time, and generate prediction results of local yield strength and tensile strength of heavy rails through the third model.

[0116] S7. During the use of the model, random testing can be arranged to verify the accuracy of the third model constructed by S5. If inaccurate predictions of certain extreme working conditions are encountered, the calculation accuracy of the model can be improved by correcting the first database established by S4.

[0117] The method establishes a finite element model for the heavy rail rolling and air cooling process to obtain the temperature, stress and strain rate change curves of the local position of the heavy rail that needs to be detected, and then calculates the local grain size information of the heavy rail through a cellular automaton model including dynamic recrystallization and static softening modules. After comparing the local grain size information with the original microstructure photograph and correcting the relevant model parameters of the cellular automaton, a high-precision offline calculation model is obtained. Since the production process window of the heavy rail is relatively fixed, the orthogonal test method is used to perform batch offline calculations on the production parameter changes that may be encountered in conventional production. The calculation results are trained through a deep learning model, and finally an online prediction model for the local strength of heavy rails that can be used for online calculations is obtained. By inputting the production data such as the steel type, rolling force, and temperature of the heavy rail production entrance into the model, the grain size, local yield strength, and tensile strength of the conventional local position to be detected of the heavy rail can be calculated. Since the prediction method adopts a cellular automaton model including dynamic recrystallization and static softening modules, it has high calculation accuracy and can replace the offline detection of heavy rails to a certain extent, especially when the production process is abnormal, it can provide early warning for heavy rail products that may have problems and reduce the generation of problem products; it completely solves the problem of the increase in time exponent after the micro-calculation accuracy of cellular automata is improved, and the cellular automaton calculation model and the finite element calculation model are set as offline calculation models. The orthogonal batch calculation method is used to obtain sufficient data points for offline cellular automaton calculations, and the deep learning model is used as a medium to convert it into a model that can be fully used for online real-time prediction.

[0118] The following examples are implemented on the premise of the technical solution of the present invention, and provide detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to the following examples. The methods used in the following examples are conventional methods unless otherwise specified.

[0119] [Implementation Example] Taking the production of hot-rolled heavy rails in a steel plant as an example, the focus is on the local tensile strength and yield strength of the heavy rail tread, and the relevant values ​​need to be calculated online;

[0120] Step 1: Use the Geleeble thermal simulation test machine to obtain the stress-strain curve of a certain specification of heavy rail, and inversely deduce its deformation resistance curve;

[0121] Step 2: Use ABAQUS software to build the entire process of heavy rail rolling of a certain specification, use Python language for batch pre-processing and post-processing, reduce model construction errors, improve information processing efficiency, and use the deformation resistance curve in step 1 as the input of ABAQUS software to calculate the strain rate curve of the focus point;

[0122] Step 3: Perform a constant temperature and constant strain rate hot compression experiment on a Gleeble thermal simulation test machine to study the dynamic recrystallization process, and perform a constant temperature and constant strain rate hot compression + heat preservation experiment on a Gleeble thermal simulation test machine to study the static recrystallization process. Cut and etch the test samples after the experiment to obtain the average grain size and distribution of the original austenite, as well as the average grain size and distribution after simulated rolling;

[0123] Step 4: construct a cellular automaton model, use the strain rate curve calculated in step 2 and the original grain distribution obtained in step 3 as input conditions, calculate the final average grain size and distribution after rolling, and compare them with the sample results after the experiment in step 3, and correct the cellular automaton model parameters to make them consistent with the sample results after the experiment in step 3;

[0124] Step 5: construct a database based on the upper and lower limits of the on-site production process fluctuations, use the cellular automaton model constructed in step 4 to complete the calculation of all working conditions, and store the results in the form of a database;

[0125] Step 6: Build a suitable machine learning model, fit the database obtained in step 5, and obtain an online calculation model with an error less than 5%;

[0126] Step 7: embed the online calculation model obtained in step 6 into the online production system to complete real-time production data collection and calculate the yield strength and tensile strength values ​​of the local area of ​​the heavy rail;

[0127] Step eight, using an offline statistical method, calibrate the calculation accuracy of the online model in step seven, and gradually improve its calculation accuracy so that it can completely replace manual sampling measurement.

Claims

1. A method for online prediction of local strength of heavy rail, characterized in that: The steps include: S1. Establishing a first model: establishing a multi-pass rolling model using the pass profiles of the heavy rails. The input conditions of the model include the heavy rail model, the material properties of the heavy rails and the incoming temperature of the heavy rails. The output information includes the final shape of the heavy rails and the stress and strain curves of each local position. S2. Establishing the first experiment: using a Geleeble thermal simulation test machine to simulate the rolling process of the first model established in S1, conducting a constant temperature and constant strain rate hot compression experiment on the Gleeble thermal simulation test machine to study the dynamic recrystallization process, and conducting a constant temperature and constant strain rate hot compression + heat preservation experiment on the Gleeble thermal simulation test machine to study the static recrystallization process; S3. Establish a second model: a cellular automaton model calculation model including dynamic recrystallization and static softening modules, simulate the first experimental process described in S2, use the austenite grain boundary obtained in S2 as an input condition, calculate the final grain size result, and use the result of the first experiment obtained in S2 to verify the result and correct the program, so as to obtain a cellular automaton model calculation model with a rolling process error of less than 5% covering the heavy rail steel grades to be studied; S4, establishing a first database: using an orthogonal test method, arranging and combining the production parameter changes encountered in production, and establishing a database of permutations and combinations suitable for calculation; selecting one of the combinations, using the stress, strain, and temperature curves obtained by the first model established by S1, and using the S3 method to process the samples collected on site before rolling as input conditions, and finally using the cellular automaton model calculation model developed by S3 to calculate the grain distribution and grain size under the corresponding working conditions; repeating the above process to obtain the grain distribution and grain size calculation results of each grouping, and placing them in the first database; S5. Establish the third model: transform the data in the database established in S4 into an input-output mapping space through machine learning or deep learning algorithms; S6, embed the third model constructed in S5 into the on-site production model, collect parameters such as steel type, rolling force, temperature, etc. in production in real time, and generate prediction results of local yield strength and tensile strength of heavy rails through the third model; S7. During the use of the model, when inaccurate predictions of certain extreme working conditions are encountered, the first database established in S4 is corrected.

2. The method for online prediction of local strength of heavy rail according to claim 1, characterized in that: The step S1 specifically includes the following steps: S11, select finite element software to construct the model, wherein the roller adopts a rigid body and the rolled piece adopts an elastic body; S12. Create a rolled piece model required for simulation based on actual production process, including temperature, material properties and heat transfer coefficient; S13, setting a driving force for a certain period of time behind the rolled piece, so as to force the rolled piece to enter the roll gap formed by the rollers through friction, and use the friction to complete the subsequent rolling simulation; S14, according to different parts of interest, set corresponding sets or paths during modeling, and use post-processing software to obtain calculation results of the corresponding sets or paths; S15, post-processing the calculation results obtained in S14 to obtain input conditions for subsequent calculation models.

3. The method for online prediction of local strength of heavy rail according to claim 1, characterized in that: The step S2 specifically includes the following steps: S21. Conduct a constant temperature and constant strain rate hot compression experiment on a Gleeble thermal simulation test machine to study the dynamic recrystallization process: select a cylindrical sample of φ8×12mm, and place graphite or tantalum sheets at both ends of the compression sample before hot compression; S22. After hot compression, measure the diameter of the middle bulging part and the end diameter of the sample in the height direction; S23, using thermocouples to track the actual temperature changes of the compressed samples in real time, and the selected hot compression temperature and strain rate cover the temperature and strain rate experienced by the heavy rail steel grades to be studied during continuous hot rolling; S24. In the compression test, the sample is heated to 1100-1200°C at a rate of 5-10°C / s and kept at this temperature for 3-5 minutes; then cooled to a deformation temperature of 900-1150°C at a rate of 5-10°C / s and kept at this temperature for 1-3 minutes; finally, the sample is heated at a temperature of 900-1150°C and a temperature of 0.001-10s. -1 The sample was compressed and deformed to a true strain of 0.7 to 0.9 at a strain rate, and quenched immediately after deformation; S25. A constant temperature and constant strain rate hot compression + heat preservation experiment was carried out on a Gleeble thermal simulation test machine to study the static recrystallization process. The sample size, lubrication conditions and temperature measurement requirements of the experiment were the same as those of the constant temperature and constant strain rate hot compression experiment. S26, heat the sample to 1100-1200℃ at a rate of 5-10℃ / s, keep it warm for 3-5min; then cool it to the deformation temperature at a rate of 3-5℃ / s, keep it warm for 1-3min; 900-1150℃ and 0.001-10s -1 The sample is compressed and deformed at a strain rate until the true strain is 0.15 to 0.3; then the deformation is stopped, and the sample is kept warm for a corresponding time of 1 to 30 seconds, and quenched immediately after the holding is completed; S27. After the experiment is completed, the sample is split longitudinally with a wire cutting machine, and the sample cross section is mechanically ground using 400#, 800#, 1500# and 2000# sandpaper in sequence, and then polished to a mirror state using a diamond polishing liquid with a particle size of 2.5μm. Finally, a saturated picric acid aqueous solution is used to corrode the original austenite grain boundary in a 50-70℃ water bath for 3-5 minutes until the sample surface color darkens, and then it is immediately cleaned with alcohol and blown dry with a hair dryer; S28. Place the corroded sample under an optical microscope for observation, and take photos under a 200x or 500x objective lens. Take at least 10 photos for each sample, and statistically analyze the obtained microstructure photos to obtain the average grain size of the original austenite.

4. The method for online prediction of local strength of heavy rail according to claim 1, characterized in that: The step S3 specifically includes the following steps: S31. Establish a dislocation density evolution model in the dynamic recrystallization model: adopt the Kocks-Mecking model, that is, the KM dislocation density model, and the specific formula is as follows: Among them, ρ is the dislocation density, ε is the strain, k1 is the hardening index, and k2 is the softening index; S32, establish a dynamic recrystallization nucleation model in a dynamic recrystallization model, using the following model, the specific formula is as follows: in, is the nucleation rate, C DRX , m is the model constant, Q act is the dynamic recrystallization activation energy, is the strain rate, R is the ideal gas state constant, T is the deformation temperature, T m is the melting point used to calculate the shear modulus at a specific temperature; S33. Establishing a dynamic recrystallization grain growth model in a dynamic recrystallization model: v i =M i f i =M i [t(r m -r i )-c i k i ] (3) Among them, v i is the growth rate of dynamically recrystallized grains, M i is the grain boundary mobility, f i is the growth driving force, ρ m is the matrix dislocation density, ρ i is the dislocation density of the ith unit cell, γ i is the grain boundary energy of the i-th unit cell, κ i is the grain boundary curvature; S34. Establishing a static recovery model in a static softening model: The static reply SRV formula is: ρ(t1)=ρ(t0)exp(-k2ε res ) (4) Among them, ε res is the residual strain, ρ(t1) and ρ(t0) are the dislocation densities at the current and next moment, t0 = t1 + Δt, k2 is a constant; S35. Establishing a static recrystallization model in a static softening model: The nucleation rate of static recrystallization SRX is calculated as: in, is the nucleation probability, C SRX is the model constant, E is the residual deformation energy, and E min is the critical residual deformation energy, V Ω is the non-recrystallized volume fraction; S36. Establish a grain coarsening model in the static softening model: When the residual deformation energy in the material is completely consumed, the static recovery SRV, metadynamic recrystallization MDRX and static recrystallization SRX processes are completed, and the grains continue to grow under the single drive of the interface energy until the next deformation begins or the material is cooled. At this time, the calculation formula for the grain growth rate becomes: v i =M i f i =M i c i k i (6)。 5. The method for online prediction of local strength of heavy rail according to claim 1, characterized in that: The step S5 specifically includes the following steps: S51. Use machine learning or deep learning methods to build a multi-layer training model, and divide the input variables in the data in the database established by S4, including the stress-strain related data obtained by the Gleeble thermal simulation test machine, and the grain size parameters and output variables calculated by the offline calculation model of the cellular automaton constructed by S3: the original microstructure photo results obtained in S2 into a training set and a validation set by a random method to ensure that data grouping does not have an adverse effect on the training of the model; S52. Using the training set to train a machine learning or deep learning method; S53. Use the validation set to validate the machine learning or deep learning method; S54. Determine whether the verification result meets the production requirements. If not, adjust the machine learning or deep learning method. S55. Through repeated iterations, the required mapping space of input and output is finally obtained.