A hot rolling decision control method based on process condition change

Abstract

The application discloses a hot rolling decision control method based on process condition change. The method uses historical rolling performance to construct a condition deviation prediction model cluster, combines current roll performance and setting information to predict the influence of current condition on finishing temperature, and adjusts the control parameters of the temperature model to complete dynamic correction of finishing temperature control. The application improves the adaptability of process control to process conditions and enhances control stability.

Description

Technical Field

[0001] This invention relates to control technology, and more specifically, to a hot rolling decision control method based on changes in process conditions. Background Technology

[0002] During the manufacturing process, it is difficult to maintain stable operating conditions such as equipment status and environmental conditions. The product structure or production process is also constantly changing. Traditional process control models cannot effectively capture the patterns of these characteristic changes. Traditional model self-learning methods are also prone to causing poor control stability when rolling products of the same specifications at different times.

[0003] Taking the final rolling temperature control in hot rolling process control as an example, when strip steel of the same specification is rolled at different times, there will be significant differences in process conditions such as slab heating uniformity, descaling process, cooling water flow rate, planned product structure, size change pattern, and rolling rhythm. These differences will cause significant interference to the final rolling temperature control, and these interferences are not easy to quantify, ultimately affecting the stability of the final rolling temperature control.

[0004] Traditional final rolling temperature control models start with the given exit temperature of the heating furnace and calculate the heat exchange and work-heat conversion processes of air cooling, water cooling, and deformation zone step by step according to the production process. Finally, they estimate the control information such as the threading speed and the cooling water flow rate between stands required to reach the target final rolling temperature. When the given exit temperature of the heating furnace is unstable, the rolling rhythm is unstable, the cooling water control is unstable, or other process changes occur, the control model cannot capture the impact of these changes on the temperature, which ultimately leads to deviations in the final rolling temperature control. Summary of the Invention

[0005] In view of the deficiencies in the existing technology, the purpose of this invention is to provide a hot rolling decision control method based on changes in process conditions, so as to improve the adaptability of process control to process conditions and enhance control stability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A hot rolling decision control method based on changes in process conditions;

[0008] By utilizing historical rolling records, a cluster of working condition deviation prediction models is constructed. Combining current coil records with set information, the influence of current working conditions on the final rolling temperature is predicted. By adjusting the control parameters of the temperature model, the dynamic correction of the final rolling temperature control is achieved.

[0009] Preferably, the hot rolling decision control method specifically includes the following steps:

[0010] S1, Product Structure Analysis;

[0011] S2, Establish process condition characteristic engineering;

[0012] S3, Sample set optimization;

[0013] S4, complete the multi-task network architecture working condition model;

[0014] S5, model cluster training;

[0015] S6, Prediction of deviations due to operating conditions;

[0016] S7, Final rolling temperature control setting correction;

[0017] S8, Feedback Analysis;

[0018] S9, sample reinforcement retraining;

[0019] S10, Feedforward Retraining of Roll Changing Mechanism.

[0020] Preferably, in step S1, the products are classified according to steel type, rolling pass combination, hot coil usage status, surface grade requirements, material properties, and number of descaling groups;

[0021] The steel grades are classified into carbon steel, high manganese steel, chromium stainless steel, nickel stainless steel, low grade silicon steel, high grade silicon steel, oriented silicon steel, pure iron, and titanium plate.

[0022] The rolling pass combinations are distinguished according to the number of rolling passes in roughing and finishing rolling;

[0023] The hot roll operation status is divided into operation and non-operation;

[0024] The surface grade requirements are divided into seven levels according to product design requirements, namely levels 0 to 6;

[0025] The material properties are classified into ironing roller material, ironing roller width transition material, planned main material, and inter-variety transition material;

[0026] The number of descaling groups is combined and differentiated according to the deployment strategy of three types of descaling equipment: roughing mill descaling box, roughing mill last pass descaling, and finishing mill descaling box.

[0027] Preferably, step S2 specifically includes:

[0028] Based on the input features of the mechanism model, operating conditions and process features that are strongly correlated with the prediction deviation of the mechanism model are introduced by using a big data-based correlation coefficient calculation method.

[0029] The feature set of the mechanism model includes composition, furnace exit temperature distribution, roughing descaling strategy, roughing dimensional change history, roughing speed diagram, roughing rolling power, roughing pass entry and exit descaling, finishing descaling strategy, finishing speed specification, finishing water strategy, and finishing load distribution information.

[0030] Preferably, step S3 specifically includes:

[0031] The samples are grouped according to product structure, and the sample set is optimized in two dimensions: first, to keep the sample set fresh in terms of time series; and second, to ensure that the sample set is similar to and uniform with the planned products in terms of feature dimension.

[0032] The homogenization strategy for the sample set includes the following steps:

[0033] S31, Divide the samples into subclasses according to the specified features, with the number of subclasses being G;

[0034] S32, 100 samples are drawn from each subcategory at a fixed ratio;

[0035] S33, when G*100 is greater than 2000, the sample size is reduced proportionally; when G*100 is less than 1000, the sample size is increased proportionally.

[0036] Preferably, step S4 specifically includes:

[0037] Features are grouped according to input feature categories, including component-related characteristic models, deformation-related characteristic models, and water-related characteristic models, to complete a multi-task neural network working condition prediction model.

[0038] Preferably, step S5 specifically includes:

[0039] Establish a model training service cluster to conduct concurrent training only for models related to the product structure in the next plan.

[0040] Preferably, step S6 specifically includes:

[0041] Organize the model input feature information, call the working condition prediction model, and predict the influence of the current working condition on the final rolling temperature ΔT.

[0042] Preferably, in step S7, the correction for the deviation in operating condition is achieved by updating the work-heat conversion compensation coefficient of the deformation zone, as calculated below:

[0043]

[0044] In Equation (1), Δd(i) is the heat and power compensation coefficient of the deformation zone of the i-th frame after correction, ΔDeform is the heat and power conversion compensation coefficient, ΔDesc is the influence of the finishing mill descaling compensation coefficient on temperature, ΔIsc is the influence of the cooling water compensation coefficient on temperature, ΔAir is the influence of the radiation compensation coefficient on temperature, i is the frame number, and N is the number of active frames.

[0045] The final rolling temperature setting is calculated using the new power-to-heat conversion compensation coefficient Δd.

[0046] Preferably, in step S8, after the finishing rolling is completed, multiple points of actual strip temperature are collected to complete sample collection, and the deviation ΔT between the actual final rolling temperature and the control set temperature is calculated according to formula (2). corr Determine whether feedback training is needed;

[0047]

[0048] In equation (2), L represents the number of locations where actual performance samples were collected, and T represents the number of locations where actual performance samples were collected. act T represents the actual final rolling temperature at the data collection location. set Set the final rolling temperature for the data acquisition location under the current operating conditions;

[0049] When ΔT corr >θ or ΔT cirr When <-θ, the feedback function of the current volume needs to be triggered, where θ is the threshold for determining whether to trigger feedback.

[0050] Preferably, step S9 specifically includes:

[0051] When feedback training is triggered, according to ΔT corr The current volume samples are dynamically augmented by increasing their size. A new sample set is updated by copying samples, and training is restarted to generate a new model. The calculation is as follows:

[0052] x=|ΔT corr |-θ+1 (3)

[0053] In equation (3), the integer part of x represents the number of times the sample is copied.

[0054] Preferably, step S10 specifically includes:

[0055] The roll change event triggers the analysis of the rolling plan product structure, leading to the reconstruction of the product classification model.

[0056] The hot rolling decision control method based on process condition changes provided by this invention makes full use of historical production data to extract relevant features of working conditions and processes, constructs a working condition prediction model, improves the deviation of finishing rolling temperature setting caused by working condition changes, and effectively improves the setting accuracy and control stability of finishing rolling temperature in large-scale production. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the hot rolling decision control method of the present invention;

[0058] Figure 2 This is a schematic diagram illustrating the principle of the hot rolling decision control method of the present invention;

[0059] Figure 3This is a schematic diagram of step S5 in the hot rolling decision control method of the present invention. Detailed Implementation

[0060] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0061] This invention utilizes data from the entire hot rolling process, including settings and actual data such as heating furnace, roughing descaling process, large side pressure process, roughing rolling rhythm, roughing pass reduction, rolling speed, descaling at pass inlet and outlet, roughing exit temperature measurement, intermediate roll table insulation cover status, intermediate roll table air cooling time, finishing rolling descaling box process, finishing pass reduction, and rolling speed. Based on big data and deep learning technology, it constructs a working condition, process characteristic engineering, and working condition influence prediction model. Combining model clustering, feedforward adaptive training, and feedback adaptive training techniques, it achieves dynamic correction of the final rolling temperature control setting, improves the control model's adaptability to changes in working conditions and processes, and enhances the stability of the final rolling temperature control.

[0062] This invention, based on traditional control principles, adds quantitative calculations of the impact of changes in process conditions on the final rolling temperature, and optimizes the final rolling temperature control setting function: When the strip reaches the FET temperature measurement point at the finishing mill inlet, the control system activates the process deviation prediction model and uses the prediction results of the process condition model to correct the final rolling temperature setting parameters; when the strip reaches the FDT (Front-to-Distance Temperature) point, feedback analysis and feedback training are initiated; upon receiving a finishing mill roll change event, rolling plan product structure analysis is initiated, feature engineering is re-established according to the product structure, and feedforward training is performed. This system uses a combination of feedforward and feedback methods to perform clustered training of small-sample models, improving the prediction accuracy of the process condition model and enhancing the control stability and accuracy of the final rolling temperature control model.

[0063] This invention provides a hot rolling decision control method based on changes in process conditions. It utilizes historical rolling performance to construct a cluster of condition deviation prediction models, combines current coil performance with set information, predicts the impact of current operating conditions on the final rolling temperature, and dynamically corrects the final rolling temperature control by adjusting the control parameters of the temperature model.

[0064] Combination Figure 1 and Figure 2 As shown, the hot rolling decision control method of the present invention specifically includes the following steps:

[0065] S1, Product Structure Analysis

[0066] Based on control principles, products with significant differences in production processes are classified, mainly according to steel type, rolling pass combination, hot coil usage status, surface grade requirements, material properties, and number of descaling groups.

[0067] 1) The major categories of steel are carbon steel, high manganese steel, chromium stainless steel, nickel stainless steel, low grade silicon steel, high grade silicon steel, oriented silicon steel, pure iron, titanium plate, and others.

[0068] 2) The rolling pass combination is distinguished according to the number of rolling passes of roughing and finishing rolling; the roughing rolling is divided into two stands, R1 can perform 1 or 3 passes, R2 can perform 1, 3, 5 or 7 passes, and the finishing rolling can perform 5, 6 or 7 passes.

[0069] 3) The hot-rolled coil usage status is divided into usage (1) and non-use (0);

[0070] 4) Surface grade requirements are divided into seven levels according to product design requirements, namely levels 0 to 6;

[0071] 5) Material properties are classified into ironing roller material, ironing roller width transition material, planned main material, and inter-variety transition material;

[0072] 6) The number of descaling groups is combined and differentiated according to the deployment strategies of three types of descaling equipment: roughing mill descaling boxes, roughing mill last pass descaling boxes, and finishing mill descaling boxes. Roughing mill descaling boxes include two groups: inlet and outlet. Roughing mill last pass descaling boxes also include two groups: inlet and outlet. The deployment strategies can be divided into:

[0073] Rough rolling descaling box: No opening (0), one set opening (1), two sets opening (2);

[0074] Descaling in the final pass of rough rolling: No opening (00), inlet group (10), outlet group (01), two groups (11);

[0075] Fine rolling descaling box: No opening (0), one set opening (1), two sets opening (2).

[0076] S2, Establish process condition characteristic engineering

[0077] Based on the input features of the mechanism model, operating conditions and process features that are strongly correlated with the prediction deviation of the mechanism model are introduced by using a big data-based correlation coefficient calculation method.

[0078] The feature set of the mechanism model includes composition, furnace exit temperature distribution, roughing descaling strategy, roughing dimensional change history, roughing speed diagram, roughing rolling power, roughing pass entry and exit descaling, finishing descaling strategy, finishing speed specification, finishing water strategy, and finishing load distribution information.

[0079] S3, Sample Set Optimization

[0080] The samples are grouped according to product structure, and the sample set is optimized in two dimensions: first, to keep the sample set fresh in terms of time series; and second, to ensure that the sample set is similar to and uniform with the planned products in terms of feature dimension.

[0081] Sample freshness strategy: Prioritize retaining the most recently produced samples among similar samples.

[0082] The homogenization strategy for the sample set includes the following steps:

[0083] S31, Divide the samples into subclasses according to the specified features, with the number of subclasses being G;

[0084] S32, 100 samples are drawn from each subcategory at a fixed ratio;

[0085] S33, when G*100 is greater than 2000, the sample size is reduced proportionally; when G*100 is less than 1000, the sample size is increased proportionally.

[0086] S4, complete the multi-task network architecture working condition model

[0087] Features are grouped according to input feature categories, including component-related characteristic models, deformation-related characteristic models, and water-related characteristic models, to complete a multi-task neural network working condition prediction model.

[0088] These three characteristic models operate relatively independently. Therefore, features are grouped according to input feature categories to construct sub-models for each operating condition. Based on these sub-models, a multi-task neural network operating condition prediction model is then built.

[0089] Sub-model A is composed of air-cooled related features, f_A(X_A), where X_A ​​represents the component-related features;

[0090] Sub-model B is composed of water-cooling related features, f_B(X_B), where X_B represents water-related features;

[0091] The sub-model C is composed of deformation-related features, f_C(X_C), where X_C represents the deformation-related features;

[0092] The multi-task neural network condition prediction model D consists of the outputs of A, B, and C and the feature X_O, where X_O represents other feature combinations.

[0093] S5, Model Cluster Training

[0094] To improve model training efficiency, a model training service cluster will be established, allowing concurrent training only on models related to the product structure within the next planned phase. This will improve training efficiency while reducing the load on the model training server. Figure 3 As shown.

[0095] S6, Prediction of Deviation Due to Operating Conditions

[0096] Organize the model input feature information, call the working condition prediction model, and predict the influence of the current working condition on the final rolling temperature ΔT.

[0097] S7, Final rolling temperature control setting correction

[0098] In the calculation process of the final rolling temperature control model, the prediction deviation is mainly corrected by adjusting the water cooling heat exchange, the work-heat conversion in the deformation zone, and the air cooling radiation coefficient. The deviation of the working condition is corrected by updating the work-heat conversion compensation coefficient in the deformation zone. The calculation is as follows:

[0099]

[0100] In Equation (1), Δd(i) is the heat and power compensation coefficient of the deformation zone of the i-th frame after correction, ΔDeform is the heat and power conversion compensation coefficient, ΔDesc is the influence of the finishing mill descaling compensation coefficient on temperature, ΔIsc is the influence of the cooling water compensation coefficient on temperature, ΔAir is the influence of the radiation compensation coefficient on temperature, i is the frame number, and N is the number of active frames.

[0101] The final rolling temperature setting is calculated using the new power-to-heat conversion compensation coefficient Δd.

[0102] S8, Feedback Analysis

[0103] After the finishing rolling is completed, multiple actual measurements of the strip are collected to complete the sample collection. The deviation ΔT between the actual final rolling temperature and the control set temperature is calculated according to formula (2). corr Determine whether feedback training is needed;

[0104]

[0105] In equation (2), L represents the number of locations where actual performance samples were collected, and T represents the number of locations where actual performance samples were collected. act T represents the actual final rolling temperature at the data collection location. set Set the final rolling temperature for the data acquisition location under the current operating conditions;

[0106] When ΔT corr >θ or ΔT cirr When <-θ, the feedback function of the current volume needs to be triggered, where θ is the threshold for determining whether to trigger feedback.

[0107] S9, Sample reinforcement and retraining

[0108] When feedback training is triggered, according to ΔT corr The current volume samples are dynamically augmented by increasing their size. A new sample set is updated by copying samples, and training is restarted to generate a new model. The calculation is as follows:

[0109] x=|ΔT corr |-θ+1 (3)

[0110] In equation (3), the integer part of x represents the number of times the sample is copied.

[0111] S10, Feedforward Retraining of Roll Changing Mechanism

[0112] The roll change event triggers the analysis of the rolling plan product structure, leading to the reconstruction of the product classification model.

[0113] Example

[0114] This embodiment provides a hot rolling decision control method based on changes in process conditions, which specifically includes the following steps:

[0115] S1, Product Structure Analysis

[0116] Currently, the rolling mill is changing rolls in the finishing mill. The planned coils are categorized based on several dimensions: steel grade, rolling pass combination, hot-rolled coil usage status, surface grade requirements, material properties, and number of descaling groups. The resulting categories are as follows:

[0117]

[0118] S2, Establish process condition characteristic engineering

[0119] The feature engineering is constructed by collecting relevant actual data on furnace area, steel grade, roughing rolling rhythm, roughing rolling water cooling, roughing rolling heat and energy, roughing rolling deformation, finishing rolling rhythm, finishing rolling water cooling, finishing rolling heat and energy, and finishing rolling deformation.

[0120]

[0121] S3, Sample Set Optimization

[0122] Optimize the sample sets for each product category, divide them into subcategories according to the planned finished product thickness range h, and count the quantities:

[0123]

[0124] The final sample size composition for each product structure is as follows:

[0125]

[0126] S4, Multi-task Network Architecture Working Model

[0127] Based on the input feature categories, a component-related characteristic model, a deformation-related characteristic model, and a water-related characteristic model are used to complete a multi-task neural network working condition prediction model.

[0128]

[0129] The f_A sub-model is constructed using the X_A ​​feature set, the f_B sub-model is constructed using the X_B feature set, and the f_C sub-model is constructed using the X_C feature set. The outputs of the three sub-models are then combined with other features from feature engineering to construct the f_D working condition prediction model.

[0130] S5, Model Cluster Training

[0131] For the product structure currently in the plan, initiate concurrent training of relevant models within the cluster to improve model training efficiency while reducing the load on the model training server.

[0132] S6, Prediction of Deviation Due to Operating Conditions

[0133] Collect all data related to the feature engineering of the working condition model:

[0134]

[0135]

[0136]

[0137] Calling the operating condition prediction function f_D returns ΔT = 28℃.

[0138] This indicates that under the current operating conditions, the final rolling temperature will have a prediction deviation of +28℃ from the baseline.

[0139] S7, Final rolling temperature control setting correction

[0140] The work-heat conversion coefficient in the deformation zone of the final rolling temperature setting model is corrected to compensate for the deviation in the final rolling temperature setting.

[0141]

[0142] The finishing mill uses 7 stands, N=7. The corrected work-to-heat conversion coefficient is as follows:

[0143]

[0144] The final rolling temperature control setting is performed using the corrected work-heat conversion compensation coefficient.

[0145] S8, Feedback Analysis

[0146] After the finishing rolling is completed, three actual records (L=3) of the strip head are collected to complete the sample collection. The deviation between the actual final rolling temperature and the predicted control set temperature is calculated, and it is determined whether feedback training is needed.

[0147]

[0148]

[0149] The feedback threshold θ is set to 3, ΔT corr <3, therefore no feedback training is required.

[0150] In this embodiment, the codes and abbreviations are explained as follows:

[0151] R1: First horizontal rolling mill in the roughing zone

[0152] R2: The second horizontal rolling mill in the roughing zone

[0153] F1~F7: The 7 stands in the finishing rolling area are numbered sequentially from front to back.

[0154] RDT: Temperature measuring instrument at the exit of the roughing mill

[0155] FET: Temperature measuring instrument at the inlet of the finishing mill

[0156] FDT: Temperature measuring instrument at the exit of the finishing mill

[0157] In summary, by using the working condition prediction model, it was determined that the current working condition has a deviation of 22 degrees on the final rolling temperature. After correcting the heat conversion coefficient of finishing rolling, the actual final rolling temperature is close to the set temperature, thus improving the setting accuracy of the final rolling temperature.

[0158] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A hot rolling decision control method based on changes in process conditions, characterized in that: By utilizing historical rolling records, a cluster of working condition deviation prediction models is constructed. Combining current coil records with set information, the influence of current working conditions on the final rolling temperature is predicted. By adjusting the control parameters of the temperature model, the dynamic correction of the final rolling temperature control is achieved.

2. The hot rolling decision control method based on changes in process conditions according to claim 1, characterized in that, The hot rolling decision control method is specifically... Includes the following steps: S1, Product Structure Analysis; S2, Establish process condition characteristic engineering; S3, Sample set optimization; S4, complete the multi-task network architecture working condition model; S5, model cluster training; S6, Prediction of deviations due to operating conditions; S7, Final rolling temperature control setting correction; S8, Feedback Analysis; S9, sample reinforcement retraining; S10, Feedforward Retraining of Roll Changing Mechanism.

3. The hot rolling decision control method based on changes in process conditions according to claim 2, characterized in that, In step S1, the products are classified according to steel type, rolling pass combination, hot-rolled coil usage status, surface grade requirements, material properties, and number of descaling groups. The steel grades are classified into carbon steel, high manganese steel, chromium stainless steel, nickel stainless steel, low grade silicon steel, high grade silicon steel, oriented silicon steel, pure iron, and titanium plate. The rolling pass combinations are distinguished according to the number of rolling passes in roughing and finishing rolling; The hot roll operation status is divided into operation and non-operation; The surface grade requirements are divided into seven levels according to product design requirements, namely levels 0 to 6; The material properties are classified into ironing roller material, ironing roller width transition material, planned main material, and inter-variety transition material; The number of descaling groups is combined and differentiated according to the deployment strategy of three types of descaling equipment: roughing mill descaling box, roughing mill last pass descaling, and finishing mill descaling box.

4. The hot rolling decision control method based on changes in process conditions according to claim 2, characterized in that, Step S2 specifically includes: Based on the input features of the mechanistic model, we introduce operating conditions and process features that are strongly correlated with the prediction deviation of the mechanistic model, which are identified by a correlation coefficient calculation method based on big data. The feature set of the mechanism model includes composition, furnace exit temperature distribution, roughing descaling strategy, roughing dimensional change history, roughing speed diagram, roughing rolling power, roughing pass entry and exit descaling, finishing descaling strategy, finishing speed specification, finishing water strategy, and finishing load distribution information.

5. The hot rolling decision control method based on changes in process conditions according to claim 2, characterized in that, Step S3 specifically includes: The samples are grouped according to product structure, and the sample set is optimized in two dimensions: first, maintaining the freshness of the sample set in terms of time series; and second, ensuring that the sample set is similar to and uniform in terms of features. The homogenization strategy for the sample set includes the following steps: S31, Divide the samples into subclasses according to the specified features, with the number of subclasses being G; S32, 100 samples are drawn from each subcategory at a fixed ratio; S33, when G*100 is greater than 2000, the sample size is reduced proportionally; when G*100 is less than 1000, the sample size is increased proportionally.

6. The hot rolling decision control method based on changes in process conditions according to claim 2, characterized in that, Step S4 specifically includes: Features are grouped according to input feature categories, including component-related characteristic models, deformation-related characteristic models, and water-related characteristic models, to complete a multi-task neural network working condition prediction model.

7. The hot rolling decision control method based on changes in process conditions according to claim 2, characterized in that, Step S5 specifically includes: Establish a model training service cluster to conduct concurrent training only for models related to the product structure in the next plan.

8. The hot rolling decision control method based on changes in process conditions according to claim 2, characterized in that, Step S6 specifically includes: Organize the model input feature information, call the working condition prediction model, and predict the influence of the current working condition on the final rolling temperature ΔT.

9. The hot rolling decision control method based on changes in process conditions according to claim 8, characterized in that, In step S7, the deviation of the working condition is corrected by updating the work-heat conversion compensation coefficient of the deformation zone, and the calculation is as follows: In Equation (1), Δd(i) is the heat and power compensation coefficient of the deformation zone of the i-th frame after correction, ΔDeform is the heat and power conversion compensation coefficient, ΔDesc is the influence of the finishing mill descaling compensation coefficient on temperature, ΔIsc is the influence of the cooling water compensation coefficient on temperature, ΔAir is the influence of the radiation compensation coefficient on temperature, i is the frame number, and N is the number of active frames. The final rolling temperature setting is calculated using the new power-to-heat conversion compensation coefficient Δd.

10. The hot rolling decision control method based on process condition changes according to claim 12, characterized in that, In step S8, after the finishing rolling is completed, multiple points of actual strip temperature are collected to complete the sample collection. The deviation ΔT between the actual final rolling temperature and the control set temperature is calculated according to formula (2). corr Determine whether feedback training is needed; In equation (2), L represents the number of locations where actual performance samples were collected, and T represents the number of locations where actual performance samples were collected. act T represents the actual final rolling temperature at the data collection location. set Set the final rolling temperature for the data acquisition location under the current operating conditions; When ΔT corr >θ or ΔT corr When <-θ, the feedback function of the current volume needs to be triggered, where θ is the threshold for determining whether to trigger feedback.

11. The hot rolling decision control method based on process condition changes according to claim 10, characterized in that, Step S9 specifically includes: When feedback training is triggered, according to ΔT corr The current volume samples are dynamically augmented by increasing their size. A new sample set is updated by copying samples, and training is restarted to generate a new model. The calculation is as follows: x=|ΔT corr |-θ+1 (3) In equation (3), the integer part of x represents the number of times the sample is copied.

12. The hot rolling decision control method based on changes in process conditions according to claim 2, characterized in that, Step S10 specifically includes: The roll change event triggers the analysis of the rolling plan product structure, leading to the reconstruction of the product classification model.

Images