A Deep Learning-Based Method and System for Controlling the Solidification of Steel Billets During Rapid Cooling of Molten Steel

By using deep learning technology, a rapid cooling control system for molten steel was constructed, which solved the problem of the difficulty in uniformly modeling the freezing trajectory of the internal structure of steel billets in existing technologies. It realized the online representation and dynamic guidance of the freezing trajectory of the internal structure of steel billets, and improved the coordinated adjustment capability of the rapid cooling process.

CN123102955APending Publication Date: 2026-11-17XUZHOU HUAHONG SPECIAL STEEL CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611620181.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing rapid cooling control technologies for continuously cast steel billets are unable to unify the modeling of steel composition, molten steel superheat, and microstructure evolution during the cooling and solidification process. They cannot characterize the freezing trajectory of the internal microstructure of the billet online, resulting in a disconnect between the rapid cooling process, the solidification process, and the microstructure forming process. It is difficult to coordinate the regulation of phase transformation advancement, microstructure homogenization, segregation suppression, and thermal stress release.

Method used

A deep learning approach is employed to preprocess the rapid cooling control data of molten steel, generate a standardized cooling and solidification state sequence, construct a microstructure freezing driving sequence, and use a microstructure freezing trajectory prediction network to extract the current microstructure freezing trajectory. Combined with steel composition and superheat data, a target microstructure freezing trajectory is generated, and matching and deviation characterization are performed to generate adjustment commands to achieve trajectory regression control. The system is continuously updated to guide the formation of the billet microstructure.

Benefits of technology

It enables online characterization and dynamic guidance of the freezing trajectory of the internal structure of steel billets, improves the synergistic regulation of phase transformation advancement, microstructure homogenization, segregation suppression and thermal stress release during rapid cooling, and enhances the pertinence and continuity of microstructure forming control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123102955A_ABST
    Figure CN123102955A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for controlling the rapid cooling and solidification of molten steel billets based on deep learning, relating to the field of intelligent control in continuous steel casting. The method includes the following steps: acquiring rapid cooling control data of molten steel during the rapid cooling and solidification process and performing preprocessing; constructing a microstructure freezing driving sequence; extracting the current microstructure freezing trajectory through a microstructure freezing trajectory prediction network; constructing a target microstructure freezing trajectory; matching the current microstructure freezing trajectory with the target microstructure freezing trajectory to generate a microstructure freezing trajectory deviation representation; generating adjustment instructions to form a trajectory regression control sequence; executing the rapid cooling and solidification of molten steel billets and performing cyclic updates. This invention uses deep learning to achieve microstructure freezing trajectory control of billets, possessing precise regulation, strong segregation suppression, and high continuity.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent control in continuous steel casting, and in particular to a method and system for controlling the rapid cooling and solidification of molten steel billets based on deep learning. Background Technology

[0002] In existing rapid cooling control of continuously cast steel billets, process adjustments are typically made around the heat flux of the crystallizer, the spray intensity of the secondary cooling zone, the casting speed, and the surface temperature of the billet. These adjustments mainly employ empirical settings, rule corrections, or feedback control methods targeting single thermal parameters to maintain stable surface temperature and solidification progression. While this technology can regulate the external cooling process, the control is primarily based on measurable surface parameters.

[0003] The shortcomings of the above-mentioned technologies are that it is difficult to model the steel composition, molten steel superheat and microstructure evolution during the cooling and solidification process in a unified manner, it is impossible to characterize the freezing trajectory of the internal microstructure of the billet online, and it is also difficult to dynamically guide and control the freezing trajectory of the target microstructure. This results in the rapid cooling process, solidification process and microstructure forming process being isolated from each other, and thus it is difficult to simultaneously take into account the coordinated regulation of phase transformation promotion, microstructure homogenization, segregation suppression and thermal stress release. Summary of the Invention

[0004] One objective of this invention is to propose a method and system for controlling the rapid cooling and solidification of molten steel billets based on deep learning. This invention uses deep learning to achieve control of the freezing trajectory of the billet microstructure, which has the advantages of precise regulation, strong segregation suppression and high continuity.

[0005] The deep learning-based method for controlling the rapid cooling and solidification of molten steel billets according to embodiments of the present invention includes the following steps:

[0006] Acquire rapid cooling control data of molten steel during the rapid cooling and solidification process, perform preprocessing on the rapid cooling control data of molten steel, and generate a standardized cooling and solidification state sequence;

[0007] Based on the standardized cooling and condensation state sequence, a tissue freezing driving sequence was constructed;

[0008] The tissue freeze-driving sequence is input into the tissue freeze trajectory prediction network to extract the current tissue freeze trajectory.

[0009] Based on steel composition data, molten steel superheat data, and preset billet microstructure indicators, target microstructure forming constraints are applied, and target microstructure freezing trajectory is generated.

[0010] The current tissue freezing trajectory is matched with the target tissue freezing trajectory to determine the deviations in phase change propulsion, tissue homogenization, segregation suppression, and thermal stress release, thereby generating a tissue freezing trajectory deviation characterization.

[0011] Based on the characteristics of the tissue freezing trajectory deviation and the standardized cooling and condensation state sequence, adjustment instructions are generated to form a trajectory regression control sequence.

[0012] The control of rapid cooling and solidification of molten steel billets is executed according to the trajectory regression control sequence, and the updated rapid cooling control data of molten steel is continuously collected and updated in a cycle to complete the billet microstructure forming guidance control.

[0013] Optionally, the rapid cooling control data for molten steel includes steel composition data, molten steel superheat data, crystallizer heat flow data, secondary cooling zone spray intensity data, billet drawing speed data, and billet surface temperature data.

[0014] Optionally, the preprocessing includes time synchronization, anomaly removal, missing data completion, and normalization.

[0015] Optionally, the generation of the tissue freeze-driven sequence specifically includes:

[0016] Based on the superheat data of molten steel, heat flow data of the crystallizer, spray intensity data of the secondary cooling zone, and surface temperature data of the billet in the standardized cooling and solidification state sequence, the changes in heat flow, surface temperature, spray intensity, and superheat of molten steel between adjacent sampling times are extracted and correlated with the structure to generate a heat flow migration characterization.

[0017] By temporally mapping the heat flow migration characterization with the billet speed data, the solidification propulsion cycle value, solidification propulsion continuous deviation value, and solidification propulsion position migration amount are extracted to generate the solidification propulsion characterization.

[0018] Based on solidification propagation characterization, crystallizer heat flow data and billet surface temperature data, the solidification coverage, surface cooling capacity and solidification expansion of each sampling section are calculated to generate shell growth characterization.

[0019] Based on the spray intensity data of the second cooling zone, the surface temperature data of the billet, and the shell growth characterization, the cooling effect transfer amount, cooling response duration, and cooling cycle matching value between adjacent cooling zones are calculated to generate a cooling bearing characterization.

[0020] The heat flow migration characterization, solidification propulsion characterization, shell growth characterization, and cooling reception characterization are arranged sequentially according to time index and correlated to generate a tissue freezing drive sequence.

[0021] Optionally, the generation of the current tissue freeze trajectory specifically includes:

[0022] The frozen drive sequence is divided into continuous sliding windows, and the thermal migration, solidification propulsion, shell growth and cooling reception characteristics of each window are organized to generate a set of windowed drive subsequences.

[0023] Input the windowed-driven subsequence set into the tissue freeze trajectory prediction network;

[0024] The tissue freezing trajectory prediction network includes an input mapping layer, a thermal history embedding layer, a stage transition layer, a tissue evolution mapping layer, and a trajectory output layer.

[0025] In the input mapping layer, vectorized encoding and splicing are performed on the heat flow migration characterization, solidification propulsion characterization, shell growth characterization and cooling acceptance characterization in each window to generate a driving embedding sequence.

[0026] The driving embedding sequence is input into the thermal history embedding layer to extract thermal history embedding features from multiple time periods.

[0027] The multi-period thermal history embedding features and driving embedding sequences are input into the stage bearing layer to calculate the solidification propulsion connection value, shell growth connection value and cooling bearing connection value, and generate solidification stage bearing features.

[0028] The multi-period thermal history embedding features and solidification stage inheritance features are spliced ​​together to generate tissue evolution input vector, which is then input into the tissue evolution mapping layer to generate tissue evolution mapping features.

[0029] The tissue evolution mapping features are input into the trajectory output layer to generate the current tissue frozen trajectory.

[0030] Optionally, the generation of the target tissue freeze trajectory specifically includes:

[0031] Read the steel composition data and molten steel superheat data corresponding to the current casting, and calculate the solidification effect value of the composition and the initial heat reserve value of superheat.

[0032] Based on the solidification effect value of the composition, the initial heat reserve value of overheating, and the preset billet structure index, the target phase transformation advancement value, the target structure homogenization value, the target segregation suppression value, and the target thermal stress release value are generated.

[0033] The target phase change advancement value, target microstructure homogenization value, target segregation suppression value, and target thermal stress release value are assigned to each sampling segment to generate target phase change advancement component sequence, target microstructure homogenization component sequence, target segregation suppression component sequence, and target thermal stress release component sequence. These are then mapped to the unified time index and billet pulling direction position index corresponding to the standardized cooling and condensation state sequence to complete the position alignment.

[0034] The target phase change propulsion component sequence, target tissue homogenization component sequence, target segregation suppression component sequence, and target thermal stress release component sequence, after being aligned in position, are associated with the tissue to generate the target tissue freezing trajectory.

[0035] Optionally, the generation of the tissue freeze trajectory deviation characterization specifically includes:

[0036] By matching the current tissue freezing trajectory with the target tissue freezing trajectory segment by segment, the current phase change propulsion component trajectory, the current tissue homogenization component trajectory, the current segregation suppression component trajectory, the current thermal stress release component trajectory, and the target corresponding component trajectory are extracted to generate a trajectory alignment sequence.

[0037] Based on the current phase transition propulsion component and the target phase transition propulsion component, and the current tissue homogenization component and the target tissue homogenization component of each sampling segment in the trajectory alignment sequence, a phase transition propulsion difference sequence and a tissue homogenization difference sequence are generated. A moving average is then performed on the phase transition propulsion difference and tissue homogenization difference of adjacent sampling segments to generate phase transition propulsion deviation and tissue homogenization deviation.

[0038] Based on the current segregation suppression component and the target segregation suppression component, and the current thermal stress release component and the target thermal stress release component of each sampling segment in the trajectory alignment sequence, a segregation suppression difference sequence and a thermal stress release difference sequence are generated. A moving average is then performed on the segregation suppression difference and thermal stress release difference of adjacent sampling segments to generate segregation suppression deviation and thermal stress release deviation.

[0039] By associating phase transformation propulsion deviation, microstructure homogenization deviation, segregation suppression deviation, and thermal stress release deviation with the microstructure, a microstructure freezing trajectory deviation characterization is generated.

[0040] Optionally, the generation of the trajectory regression control sequence specifically includes:

[0041] Read the phase transition propagation deviation, tissue homogenization deviation, segregation suppression deviation and thermal stress release deviation in the tissue freezing trajectory deviation characterization, and generate a control decision sequence by corresponding each sampled segment of the standardized cooling and condensation state sequence;

[0042] Based on the phase change propulsion deviation, segregation suppression deviation and corresponding secondary cooling zone spray intensity data of each sampling segment in the control decision sequence, the spray adjustment amount of each cooling segment is calculated and the zone spray adjustment command is generated.

[0043] Based on the phase change propagation deviation, thermal stress release deviation and corresponding billet pulling speed data of each sampling section, the billet pulling speed correction amount of each sampling section is calculated, and the billet pulling rhythm adjustment command is generated.

[0044] Based on the deviation of microstructure homogenization, thermal stress release deviation and corresponding billet surface temperature data of each sampling section, the stage switching correction amount of each sampling section is calculated, and solidification stage switching adjustment command is generated.

[0045] The zone spraying adjustment command, the billet pulling rhythm adjustment command, and the solidification stage switching adjustment command are linked together to generate a trajectory regression control sequence.

[0046] Optionally, the cyclical update specifically includes:

[0047] The zone spraying adjustment command, the billet pulling rhythm adjustment command, and the solidification stage switching adjustment command in the trajectory regression control sequence are respectively sent to the corresponding cooling control unit, billet pulling execution unit, and stage switching execution unit.

[0048] During the control execution process, the rapid cooling control data of molten steel is continuously collected, and the standardized cooling solidification state sequence, the microstructure freezing drive sequence, and the current microstructure freezing trajectory are updated.

[0049] The target microstructure freezing trajectory is regenerated, and the microstructure freezing trajectory deviation characterization is updated to obtain the updated trajectory regression control sequence, thus completing the billet microstructure forming guidance control.

[0050] A deep learning-based control system for rapid cooling and solidification of molten steel billets according to an embodiment of the present invention includes:

[0051] The data acquisition module is used to acquire rapid cooling control data of molten steel during the rapid cooling and solidification process, perform preprocessing, and generate a standardized cooling and solidification state sequence.

[0052] The driving sequence construction module is used to construct tissue freezing driving sequences based on standardized cooling and condensation state sequences.

[0053] The current trajectory prediction module is used to input the tissue freeze driving sequence into the tissue freeze trajectory prediction network and extract the current tissue freeze trajectory.

[0054] The target trajectory generation module is used to constrain the formation of the target structure based on steel composition data, molten steel superheat data and preset billet structure indicators, and generate the target structure freezing trajectory.

[0055] The deviation characterization generation module is used to match the current tissue freezing trajectory with the target tissue freezing trajectory, determine the phase change propulsion deviation, tissue homogenization deviation, segregation suppression deviation, and thermal stress release deviation, and generate a tissue freezing trajectory deviation characterization.

[0056] The control sequence generation module is used to generate adjustment instructions based on the tissue freezing trajectory deviation characterization and standardized cooling and condensation state sequence, forming a trajectory regression control sequence;

[0057] The control update module is used to execute the control of rapid cooling of molten steel to solidify billet according to the trajectory regression control sequence, and continuously collect updated rapid cooling control data of molten steel for cyclic update to complete the billet microstructure forming guidance control.

[0058] The beneficial effects of this invention are:

[0059] This invention is based on rapid cooling control data of molten steel. After preprocessing, a standardized cooling and solidification state sequence is formed. A microstructure freezing driving sequence is then constructed, and the current microstructure freezing trajectory is extracted through a microstructure freezing trajectory prediction network. Simultaneously, by combining steel composition data, molten steel superheat data, and preset billet microstructure indicators, a target microstructure freezing trajectory is generated. This elevates the existing technology's external process adjustments, which focus on surface temperature, spray intensity, and billet drawing speed, to online characterization and guidance control of the internal microstructure evolution path of the billet. Consequently, it can simultaneously reflect the dynamic changes in phase transformation advancement, microstructure homogenization, segregation suppression, and thermal stress release during the rapid cooling and solidification process of molten steel, establishing a unified correspondence between the rapid cooling process, the solidification advancement process, and the microstructure forming process.

[0060] Furthermore, this invention generates a deviation characterization of the microstructure freezing trajectory by matching the current microstructure freezing trajectory with the target microstructure freezing trajectory. Based on this deviation characterization and a standardized cooling and solidification state sequence, it generates zoned spray adjustment commands, billet pulling rhythm adjustment commands, and solidification stage switching adjustment commands, forming a trajectory regression control sequence. This sequence is then cyclically updated using continuously collected data. Therefore, it can continuously correct the microstructure forming path during the control process, avoiding the disconnect between different cooling sections and solidification stages, and enhancing the coordination between billet microstructure evolution and cooling rhythm. Compared with existing technologies, this invention not only improves the perception of the internal microstructure freezing state but also enhances the targeting and continuity of microstructure forming control, making it more conducive to balancing microstructure equilibrium, segregation control, and thermal stress release coordination under rapid cooling conditions. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a flowchart of the deep learning-based method for controlling the rapid cooling and solidification of molten steel billets proposed in this invention.

[0063] Figure 2 This is a schematic diagram of the microstructure freezing trajectory prediction network of the deep learning-based method for controlling the solidification of molten steel billets during rapid cooling.

[0064] Figure 3 This is the network training convergence diagram of the deep learning-based method for controlling the rapid cooling and solidification of molten steel billets proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 and Figure 2 A deep learning-based method for controlling the solidification of molten steel billets during rapid cooling includes the following steps:

[0067] Acquire rapid cooling control data of molten steel during the rapid cooling and solidification process, perform preprocessing on the rapid cooling control data of molten steel, and generate a standardized cooling and solidification state sequence;

[0068] Based on the standardized cooling and condensation state sequence, a tissue freezing driving sequence was constructed;

[0069] The tissue freeze-driving sequence is input into the tissue freeze trajectory prediction network to extract the current tissue freeze trajectory.

[0070] Based on steel composition data, molten steel superheat data, and preset billet microstructure indicators, target microstructure forming constraints are applied, and target microstructure freezing trajectory is generated.

[0071] The current tissue freezing trajectory is matched with the target tissue freezing trajectory to determine the deviations in phase change propulsion, tissue homogenization, segregation suppression, and thermal stress release, thereby generating a tissue freezing trajectory deviation characterization.

[0072] Based on the characteristics of the tissue freezing trajectory deviation and the standardized cooling and condensation state sequence, adjustment instructions are generated to form a trajectory regression control sequence.

[0073] The control of rapid cooling and solidification of molten steel billets is executed according to the trajectory regression control sequence, and the updated rapid cooling control data of molten steel is continuously collected and updated in a cycle to complete the billet microstructure forming guidance control.

[0074] In this embodiment, the rapid cooling control data for molten steel includes steel composition data, molten steel superheat data, crystallizer heat flow data, secondary cooling zone spray intensity data, billet drawing speed data, and billet surface temperature data. The steel composition data characterizes the current chemical composition of the molten steel, reflecting the proportions of carbon, manganese, silicon, phosphorus, sulfur, and alloying elements. The molten steel superheat data is the difference between the actual pouring temperature and the liquidus temperature of the steel grade, used to characterize the initial heat reserve state of the molten steel before entering the solidification stage. The crystallizer heat flow data reflects the temperature difference between the crystallizer wall and the molten steel or the initial billet shell during continuous casting. Data on heat exchange intensity is used to characterize the heat transfer state and billet shell formation rhythm during the initial solidification stage; data on the spray intensity of the secondary cooling zone is the cooling effect intensity applied to the billet surface by each spray area in the secondary cooling zone, used to characterize the zoned cooling level experienced by the billet after leaving the crystallizer; data on the billet pulling speed is the speed data of the billet continuously moving along the pulling direction during continuous casting, used to characterize the billet's dwell rhythm and solidification propagation speed in the cooling zone; data on the billet surface temperature is the real-time temperature distribution data of the billet's outer surface during the rapid cooling and solidification process, used to characterize the billet's surface thermal state and its changes over time.

[0075] In this embodiment, preprocessing includes time synchronization, anomaly removal, missing data completion, and normalization.

[0076] In this embodiment, the generation of the tissue freeze-driven sequence specifically includes:

[0077] Based on the superheat data of molten steel, heat flow data of crystallizer, spray intensity data of secondary cooling zone and surface temperature data of billet in the standardized cooling and solidification state sequence, the heat flow change, surface temperature change, spray intensity change and molten steel superheat change between adjacent sampling times are extracted according to the unified time index and the billet pulling direction position index, and the heat flow migration characterization is generated by associating and organizing them according to the unified time index.

[0078] The billet pulling direction position index is positional information used to mark the order of the positions of each sampled data point along the billet pulling direction;

[0079] By temporally mapping the heat flow migration characterization with the billet speed data, the solidification propulsion cycle value, solidification propulsion continuous deviation value, and solidification propulsion position migration amount are extracted to generate the solidification propulsion characterization.

[0080] The generation of solidification propulsion characterization specifically includes: first, binding the heat flow migration characterization of each sampling moment with the billet pulling speed data of the same sampling moment according to a unified time index; then, multiplying the billet pulling speed data of each sampling moment by the time interval between adjacent sampling moments to calculate the segment propulsion displacement value of each sampling segment, and arranging them in the order of the billet pulling direction position index to generate a segment propulsion displacement sequence; then, selecting sampling segments whose segment propulsion displacement value is greater than the preset minimum effective displacement threshold, and whose heat flow change, surface temperature change, spray intensity change, and molten steel superheat change of adjacent sampling segments are in the same direction, counting the number of sampling segments that meet the conditions per unit time, and dividing by the corresponding statistical duration to generate the solidification propulsion cycle value; finally, subtracting the segment propulsion displacement values ​​of adjacent sampling segments to generate the segment propulsion displacement difference. The values ​​are then calculated; the differences in heat flow, surface temperature, spray intensity, and molten steel superheat are calculated for adjacent sampling sections; the differences in section advance displacement, heat flow, surface temperature, spray intensity, and molten steel superheat are normalized, and the normalized differences are weighted and summed to generate a solidification advance continuous deviation value; the section advance displacement values ​​of each sampling section are accumulated segment by segment according to the order of the billet direction position index to generate the cumulative advance position corresponding to each sampling time; the cumulative advance position of the previous sampling time is subtracted from the cumulative advance position of the later sampling time to generate the solidification advance position migration; the solidification advance rhythm value, solidification advance continuous deviation value, and solidification advance position migration are associated and organized according to a unified time index to generate a solidification advance characterization;

[0081] The preset minimum effective displacement threshold is set to the average value of the section advancement displacement in historical stable castings under the same steel grade and the same billet pulling conditions.

[0082] Based on solidification propagation characterization, crystallizer heat flow data and billet surface temperature data, the solidification coverage, surface cooling capacity and solidification expansion of each sampling section are calculated to generate shell growth characterization.

[0083] The generation of the shell growth characterization specifically includes: mapping the solidification propulsion cycle value, solidification propulsion continuous deviation value, and solidification propulsion position migration amount in the solidification propulsion characterization to the crystallizer heat flow data and billet surface temperature data of the same sampling section according to a unified time index; then multiplying the crystallizer heat flow data of each sampling section with the solidification propulsion cycle value, and performing subtraction correction with the solidification propulsion continuous deviation value to generate solidification coverage; then subtracting the billet surface temperature data of adjacent sampling sections to generate surface temperature difference value, and performing weighted summation after normalizing the surface temperature difference value and the solidification propulsion cycle value to generate surface cooling absorption amount; subsequently, multiplying the solidification propulsion position migration amount with the solidification coverage amount, and performing weighted summation with the surface cooling absorption amount to generate solidification expansion amount; and associating the solidification coverage amount, surface cooling absorption amount, and solidification expansion amount according to a unified time index to generate the shell growth characterization.

[0084] Based on the spray intensity data of the second cooling zone, the surface temperature data of the billet, and the shell growth characterization, the cooling effect transfer amount, cooling response duration, and cooling cycle matching value between adjacent cooling zones are calculated to generate a cooling bearing characterization.

[0085] The generation of cooling bearing characteristics specifically includes: mapping the spray intensity data of the secondary cooling zone, billet speed data, billet surface temperature data, and shell growth characteristics of adjacent cooling sections to the order of cooling sections according to a unified time index; multiplying the time interval between adjacent sampling moments with the spray intensity data of the secondary cooling zone to calculate the cooling input of each cooling section, and subtracting the cooling input of adjacent cooling sections to generate the cooling effect transfer amount; subtracting the billet surface temperature data of adjacent cooling sections to generate the surface temperature bearing difference value, and then normalizing the surface temperature bearing difference value with the solidification coverage amount and surface cooling bearing amount in the shell growth characteristics and performing a weighted sum to generate the cooling response duration amount; normalizing the changes in cooling input amount and solidification expansion amount of adjacent cooling sections and performing a weighted sum to generate the cooling cycle time matching value; and associating the cooling effect transfer amount, cooling response duration amount, and cooling cycle time matching value according to a unified time index to generate the cooling bearing characteristics.

[0086] The heat flow migration characterization, solidification propulsion characterization, shell growth characterization, and cooling reception characterization are arranged sequentially according to time index and correlated to generate a tissue freezing drive sequence.

[0087] The structure freezing drive sequence is a time-series data sequence organized in a unified time index order to characterize the continuous driving state of the billet structure freezing process. The heat flow migration characterization is used to characterize the heat transfer changes in the time dimension and the billet pulling direction during the rapid cooling and solidification process of molten steel, reflecting the synergistic migration characteristics of heat flow release, surface cooling, spraying effect, and changes in molten steel superheat. The solidification advancement characterization is used to characterize the advancement rhythm, advancement continuity, and positional migration characteristics of the solidification process along the billet pulling direction, reflecting the dynamic situation of the solidification front advancing downstream over time. The shell growth characterization is used to characterize the coverage growth, surface cooling reception, and expansion characteristics of the primary billet shell under continuous cooling conditions, reflecting the billet shell formation and thickening process. The cooling reception characterization is used to characterize the transmission of cooling effect between adjacent cooling sections, the continuation of temperature response, and the matching characteristics between cooling input and solidification expansion, reflecting the connection between each cooling section in the continuous cooling process.

[0088] In this embodiment, the generation of the current tissue freeze trajectory specifically includes:

[0089] The frozen drive sequence is divided into continuous sliding windows according to a unified time index and a pulling direction position index. The heat flow migration characterization, solidification propulsion characterization, shell growth characterization and cooling acceptance characterization in each window are organized to generate a set of windowed drive subsequences.

[0090] The windowed-driven subsequence set is a collection of multiple local temporal subsequences formed by dividing the tissue freeze-driven sequence into consecutive sliding windows;

[0091] Input the windowed-driven subsequence set into the tissue freeze trajectory prediction network;

[0092] The tissue freezing trajectory prediction network includes an input mapping layer, a thermal history embedding layer, a stage transition layer, a tissue evolution mapping layer, and a trajectory output layer.

[0093] The training of the microstructure freezing trajectory prediction network specifically includes: collecting rapid cooling control data of molten steel corresponding to historical castings, and merging them according to the same casting identifier, the same time index, and the same billet pulling direction position index to form a historical training sample set; performing preprocessing on the historical training sample set consistent with the current rapid cooling control data of molten steel to generate a historical standardized cooling and solidification state sequence, and generating a historical microstructure freezing driving sequence based on the historical standardized cooling and solidification state sequence; performing continuous sliding window division on the historical microstructure freezing driving sequence according to the unified time index and billet pulling direction position index to generate a historical windowed driving sub-sequence set; constructing a historical target microstructure freezing trajectory based on the steel composition data, molten steel superheat data, and preset billet microstructure indicators corresponding to each historical casting, and dividing the historical target microstructure freezing trajectory according to the time window consistent with the historical windowed driving sub-sequence set to form historical trajectory label samples; and inputting the historical windowed driving sub-sequence set into the microstructure freezing trajectory prediction network. The network is tested to output the predicted historical current tissue freezing trajectory. The predicted historical current tissue freezing trajectory vector is mapped to the corresponding vector of the historical trajectory label sample using a unified time index and pulling direction position index. The squared difference between the predicted value and the label value in each dimension is calculated, and the average of the squared differences in each dimension is calculated to generate the mean squared error loss value. The mean squared error loss value corresponding to each training sample is then averaged batch by batch to generate the training loss value. The Adam optimizer is used to iteratively update the parameters of the tissue freezing trajectory prediction network. The initial learning rate is set to 0.001, and the batch size is set to 32. After each training epoch, the validation loss value is calculated on the validation sample set. When the decrease in the validation loss value is less than 0.0001 for 10 consecutive training epochs, or when the training epochs reach 100, the tissue freezing trajectory prediction network is considered to have reached the convergence condition and training is stopped. The network parameters corresponding to the minimum validation loss value are saved, resulting in the trained tissue freezing trajectory prediction network.

[0094] The frozen trajectory prediction network employs a hierarchical structure that is primarily serial with local cross-layer fusion. Specifically, a windowed set of driving sub-sequences first enters the input mapping layer to generate a driving embedding sequence. This driving embedding sequence is then input into the thermal history embedding layer to generate multi-time-period thermal history embedding features. The stage-bearing layer simultaneously receives both the multi-time-period thermal history embedding features and the driving embedding sequence, generating solidification stage-bearing features. These features are then concatenated to form a tissue evolution input vector, which is input into the tissue evolution mapping layer to generate tissue evolution mapping features. Finally, the trajectory output layer performs regression analysis on these features to generate the current tissue frozen trajectory. The network backbone is a serial connection, with the input mapping layer, thermal history embedding layer, stage-bearing layer, tissue evolution mapping layer, and trajectory output layer proceeding sequentially. The stage-bearing layer and the tissue evolution mapping layer jointly invoke previous features.

[0095] The microstructure freezing trajectory prediction network adopts a hierarchical structure based on a serial backbone and introducing cross-layer feature fusion between key layers. This enables the originally dispersed heat flow migration information, solidification advancement information, shell growth information, and cooling reception information during the rapid cooling and solidification process of molten steel to be transformed layer by layer into the current microstructure freezing trajectory along a feature extraction path from shallow to deep. Specifically, the input mapping layer completes the unified encoding of multi-source driving information, the thermal history embedding layer is used to retain the cumulative influence of the previous cooling stage on the formation of subsequent microstructures, the stage reception layer is used to characterize the continuous connection relationship between adjacent solidification stages, and the microstructure evolution mapping layer is used to further map the cumulative thermal history information and stage reception information to the coupled evolution space of phase transformation advancement, microstructure homogenization, segregation suppression, and thermal stress release. The trajectory output layer is used to regress the aforementioned high-level microstructure evolution characteristics into the current microstructure freezing trajectory. The advantage of this structure is that it no longer relies on external process parameters such as surface temperature, spray intensity, and billet drawing speed for control. Instead, through a layered and progressive approach and cross-layer fusion, it gradually maps externally measurable data into continuous trajectory data of the internal microstructure freezing process of the billet, unifying the rapid cooling process, solidification process, and microstructure forming process onto the same predictive mainline. It elevates the control object in traditional thermal control from temperature field and cooling parameters to the microstructure freezing trajectory, which cannot be directly observed. Through a hierarchical network design of thermal history embedding, stage succession, and microstructure evolution mapping, it achieves online characterization, prediction, and subsequent guidance control of the internal microstructure formation path of the billet.

[0096] In the input mapping layer, vectorized encoding and splicing are performed on the heat flow migration characterization, solidification propulsion characterization, shell growth characterization and cooling acceptance characterization in each window to generate a driving embedding sequence.

[0097] The driving embedding sequence is input into the thermal history embedding layer to extract thermal history embedding features from multiple time periods.

[0098] The thermal history embedding layer uses a gated recurrent unit network to perform window-by-window recursive encoding on the driving embedding sequence along a unified time index, extracting multi-time thermal history embedding features. These multi-time thermal history embedding features are used to characterize the cumulative evolution of heat release, surface cooling and cooling reception during the rapid cooling and solidification process of the billet.

[0099] The multi-period thermal history embedding features and driving embedding sequences are input into the stage bearing layer to calculate the solidification propulsion connection value, shell growth connection value and cooling bearing connection value, and generate solidification stage bearing features.

[0100] The generation of solidification stage transition features specifically includes: pairing the multi-period thermal history embedding features and driving embedding sequences corresponding to adjacent windows according to a unified time index; calculating the difference between the encoding vectors corresponding to the solidification advancement representations of adjacent windows to generate a solidification advancement difference vector; concatenating the solidification advancement difference vector with the multi-period thermal history embedding features corresponding to the previous and next windows to form a solidification advancement combination vector; and then performing a weighted sum on the solidification advancement combination vector to generate a solidification advancement transition value; further, calculating the difference between the encoding vectors corresponding to the shell growth representations of adjacent windows to generate a shell growth difference vector; and concatenating the shell growth difference vector with the encoding vectors of the previous and next windows. The multi-period thermal history embedding features corresponding to the next window are concatenated to form a shell growth combination vector. Then, a weighted sum is performed on the shell growth combination vector to generate a shell growth connection value. Next, the difference between the corresponding encoding vectors of the cooling bearing representation of adjacent windows is calculated to generate a cooling bearing difference vector. The cooling bearing difference vector is then concatenated with the multi-period thermal history embedding features corresponding to the previous and next windows to form a cooling bearing combination vector. Then, a weighted sum is performed on the cooling bearing combination vector to generate a cooling bearing connection value. The solidification advancement connection value, shell growth connection value, and cooling bearing connection value are associated and organized according to the window order to generate solidification stage bearing features.

[0101] Solidification stage continuity characteristics are used to characterize the continuous connection between solidification progression, shell growth, and cooling continuity of steel billets between adjacent solidification stages.

[0102] The multi-period thermal history embedding features and solidification stage inheritance features are spliced ​​together to generate tissue evolution input vector, which is then input into the tissue evolution mapping layer to generate tissue evolution mapping features.

[0103] The tissue evolution mapping layer employs a multi-layer fully connected neural network to perform nonlinear mapping on the embedded features of multi-time period thermal history and the features inherited from the solidification stage, thereby generating tissue evolution mapping features.

[0104] The microstructure evolution input vector is a comprehensive feature vector formed by splicing together the embedded features of multi-time period thermal history and the features of solidification stage according to a unified time index. It is used to organize the cumulative thermal history information and the connection information of adjacent stages into the same input carrier, thereby characterizing the overall driving basis of the microstructure freezing process of the billet at the current time and adjacent time periods. The microstructure evolution mapping feature is a high-level feature formed by the microstructure evolution mapping layer performing nonlinear mapping on the microstructure evolution input vector. It is used to characterize the coupled evolution features between phase transformation advancement, microstructure homogenization, segregation suppression and thermal stress release during the microstructure freezing process of the billet, and serves as the direct feature basis for generating the current microstructure freezing trajectory.

[0105] The tissue evolution mapping features are input into the trajectory output layer to generate the current tissue frozen trajectory.

[0106] The trajectory output layer uses a multi-output fully connected regression network to regress the microstructure evolution mapping features and generate the current microstructure freezing trajectory. The current microstructure freezing trajectory is the time-series trajectory data output by the multi-output fully connected regression network, which is used to characterize the real-time evolution path of the internal microstructure of the billet along the time dimension and the position dimension of the billet pulling direction during the current rapid cooling and solidification process. It reflects the comprehensive evolution state of phase transformation advancement, microstructure homogenization, segregation suppression and thermal stress release under the current working conditions.

[0107] In this embodiment, the generation of the target tissue freeze trajectory specifically includes:

[0108] Read the steel composition data and molten steel superheat data corresponding to the current casting, and calculate the solidification effect value of the composition and the initial heat reserve value of superheat.

[0109] The solidification effect value is obtained by weighted summing the contents of each chemical component in the steel composition data after normalization. The initial heat reserve value is obtained by normalizing the superheat data of molten steel.

[0110] Based on the solidification effect value of the composition, the initial heat reserve value of overheating, and the preset billet structure index, the target phase transformation advancement value, the target structure homogenization value, the target segregation suppression value, and the target thermal stress release value are generated.

[0111] The solidification effect value of the composition is multiplied by the phase transformation advancement index, microstructure homogenization index, segregation suppression index and thermal stress release index in the preset billet microstructure index to generate four composition constraint components; then the initial heat reserve value of overheating is weighted and summed with the four composition constraint components to generate the target phase transformation advancement value, target microstructure homogenization value, target segregation suppression value and target thermal stress release value.

[0112] The preset billet microstructure index is a set of indices pre-set under the current steel grade and target billet quality requirements, used to limit the direction of target microstructure formation during the rapid cooling and solidification process of the billet. It includes phase transformation advancement index, microstructure homogenization index, segregation suppression index, and thermal stress release index. Among them, the phase transformation advancement index is used to limit the target advancement rhythm of solidification and phase transformation along the time dimension and the position dimension of the billet pulling direction; the microstructure homogenization index is used to limit the target equilibrium level of the internal microstructure distribution of the billet; the segregation suppression index is used to limit the target control level of component agglomeration; and the thermal stress release index is used to limit the target control level of the thermal stress release rhythm during the cooling and solidification process.

[0113] The phase transformation advancement index, microstructure homogenization index, segregation suppression index, and thermal stress release index are quantitatively set according to historical stable castings under the same steel grade, cross-sectional specifications, and casting conditions. Specifically, historical stable castings that meet both surface and internal quality requirements are first selected as benchmark samples. Then, for each benchmark sample, the four indices are calculated for each sampling segment. The phase transformation advancement index is set by normalizing the solidification advancement cycle value and solidification expansion amount, then performing a weighted sum, and finally averaging the weighted sum values ​​of each sampling segment in the entire casting. The microstructure homogenization index is set by first calculating the absolute value of the difference in billet surface temperature data between adjacent sampling segments, adding a preset small positive value to this absolute value, taking the reciprocal, and then performing a weighted sum with the normalized value of the surface cooling capacity. The average of each sampling section is calculated; the segregation suppression index is set by adding a preset small positive value to the solidification influence value of the composition and taking the reciprocal, then adding a preset small positive value to the solidification advancement continuous deviation value and taking the reciprocal, and then normalizing the two reciprocals before weighted summation, and then averaging the average of each sampling section in the entire casting cycle; the thermal stress release index is set by adding a preset small positive value to the normalized value of the cooling cycle matching value and the cooling response duration and the absolute value of the difference between the billet surface temperature data of adjacent sampling sections and taking the reciprocal, and then normalizing the values, and then weighted summation of the values ​​obtained, and then averaging the average of each sampling section in the entire casting cycle; the minimum and maximum values ​​of the above four averages are calculated separately in all benchmark samples of the same steel grade, and the four averages are normalized according to the minimum and maximum values ​​to obtain the preset billet structure index;

[0114] The target phase change advancement value, target microstructure homogenization value, target segregation suppression value, and target thermal stress release value are allocated to each sampling segment according to a unified time index and a billet pulling direction position index. This generates target phase change advancement component sequences, target microstructure homogenization component sequences, target segregation suppression component sequences, and target thermal stress release component sequences, which are then mapped to the unified time index and billet pulling direction position index corresponding to the standardized cooling and condensation state sequence, respectively, to complete the position alignment.

[0115] The target phase change propulsion component sequence, target microstructure homogenization component sequence, target segregation suppression component sequence, and target thermal stress release component sequence, after being aligned in position, are associated with the microstructure according to a unified time index and billet pulling direction position index to generate the target microstructure freezing trajectory.

[0116] The target microstructure freezing trajectory is generated based on steel composition data, molten steel superheat data, and preset billet microstructure indicators. It is expanded according to a unified time index and a billet pulling direction position index to form a time-series target trajectory data. It is used to characterize the target microstructure evolution path of the billet along the time dimension and the billet pulling direction position dimension during the rapid cooling and solidification process. It reflects the target distribution status and target advancement rhythm of phase transformation advancement, microstructure homogenization, segregation suppression, and thermal stress release in each sampling segment under the target billet quality requirements of the current steel grade.

[0117] In this embodiment, the generation of the tissue freeze trajectory deviation characterization specifically includes:

[0118] The current tissue freezing trajectory and the target tissue freezing trajectory are sampled segment by segment according to a unified time index and the position index of the billet pulling direction. The current phase change propulsion component trajectory, the current tissue homogenization component trajectory, the current segregation suppression component trajectory, the current thermal stress release component trajectory and the target corresponding component trajectory are extracted to generate a trajectory alignment sequence.

[0119] Based on the current phase transition propulsion component and the target phase transition propulsion component, and the current tissue homogenization component and the target tissue homogenization component of each sampling segment in the trajectory alignment sequence, a phase transition propulsion difference sequence and a tissue homogenization difference sequence are generated. A moving average is then performed on the phase transition propulsion difference and tissue homogenization difference of adjacent sampling segments to generate phase transition propulsion deviation and tissue homogenization deviation.

[0120] Based on the current segregation suppression component and the target segregation suppression component, and the current thermal stress release component and the target thermal stress release component of each sampling segment in the trajectory alignment sequence, a segregation suppression difference sequence and a thermal stress release difference sequence are generated. A moving average is then performed on the segregation suppression difference and thermal stress release difference of adjacent sampling segments to generate segregation suppression deviation and thermal stress release deviation.

[0121] The phase transformation propulsion deviation, microstructure homogenization deviation, segregation suppression deviation, and thermal stress release deviation are associated with the microstructure using a unified time index and the position index of the billet pulling direction, thereby generating a microstructure freezing trajectory deviation characterization.

[0122] The tissue freezing trajectory deviation characterization is a set of deviation features formed by associating the phase change propagation deviation sequence, tissue homogenization deviation sequence, segregation suppression deviation sequence, and thermal stress release deviation sequence. It is used to characterize the comprehensive deviation of the current tissue freezing trajectory from the target tissue freezing trajectory in each sampling segment.

[0123] In this embodiment, the generation of the trajectory regression control sequence specifically includes:

[0124] Read the phase transition propagation deviation, microstructure homogenization deviation, segregation suppression deviation and thermal stress release deviation in the microstructure freezing trajectory deviation characterization, and correspond them with the standardized cooling and condensation state sequence by sampling segment according to a unified time index and the position index of the billet pulling direction to generate a control decision sequence;

[0125] Based on the phase change propulsion deviation, segregation suppression deviation and corresponding secondary cooling zone spray intensity data of each sampling segment in the control decision sequence, the spray adjustment amount of each cooling segment is calculated, and the zonal spray adjustment command is generated in the order of the cooling segments.

[0126] The generation of the zoned spray adjustment command specifically includes: mapping the phase change propulsion deviation value and segregation suppression deviation value of each sampling section to the corresponding secondary cooling zone spray intensity data item by item according to a unified time index and billet pulling direction position index; then normalizing the phase change propulsion deviation value and segregation suppression deviation value respectively, and performing a weighted sum of the two normalized deviation values ​​to generate a composite spray deviation value; using the composite spray deviation value as the spray intensity adjustment coefficient, multiplying it by the corresponding secondary cooling zone spray intensity data to generate the spray adjustment amount of each sampling section; then averaging the spray adjustment amounts of each sampling section within the same cooling zone to generate a zoned spray adjustment value; and finally associating and organizing the zoned spray adjustment values ​​according to the order of the cooling zones to generate the zoned spray adjustment command.

[0127] The cooling section is a series of continuous cooling segments within the secondary cooling zone, divided according to the cooling control unit.

[0128] Based on the phase change propagation deviation, thermal stress release deviation and corresponding billet pulling speed data of each sampling segment in the control decision sequence, the billet pulling speed correction amount of each sampling segment is calculated, and the billet pulling rhythm adjustment command is generated in a unified time index order.

[0129] The generation of the billet pulling rhythm adjustment command specifically includes: mapping the phase change propagation deviation value and thermal stress release deviation value of each sampling segment to the corresponding billet pulling speed data item by item according to a unified time index; then normalizing the phase change propagation deviation value and thermal stress release deviation value respectively, and performing a weighted sum on the two normalized deviation values ​​to generate a combined billet pulling deviation value; multiplying the combined billet pulling deviation value with the corresponding billet pulling speed data to generate the billet pulling speed correction amount for each sampling segment; then performing sequential arrangement and continuous association of the billet pulling speed correction amounts for each sampling segment according to a unified time index to generate a billet pulling speed correction sequence; and finally outputting the billet pulling speed correction sequence according to a unified time index to generate the billet pulling rhythm adjustment command.

[0130] Based on the deviation of homogenization of the structure, the deviation of thermal stress release and the corresponding surface temperature data of the billet in each sampling section of the control decision sequence, the stage switching correction amount of each sampling section is calculated, and the solidification stage switching adjustment command is generated according to the solidification progress sequence.

[0131] The stage switching correction amount is used to characterize the stage switching range that needs to be adjusted during the solidification process of the current sampling section. The stage switching range is the numerical value of the degree of switching that needs to be adjusted when the current sampling section switches from the current solidification stage to the next solidification stage.

[0132] The generation of solidification stage switching adjustment instructions specifically includes: matching the microstructure homogenization deviation value and thermal stress release deviation value of each sampling segment with the corresponding billet surface temperature data according to a unified time index and billet pulling direction position index; then normalizing the microstructure homogenization deviation value and thermal stress release deviation value respectively, and performing a weighted sum on the two normalized deviation values ​​to generate a stage deviation composite value; multiplying the stage deviation composite value with the corresponding billet surface temperature data to generate the stage switching correction amount for each sampling segment; then sequentially arranging and continuously associating the stage switching correction amounts of each sampling segment according to the solidification progress order to generate a stage switching correction sequence; and finally outputting the stage switching correction sequence according to the solidification progress order to generate a solidification stage switching adjustment instruction.

[0133] The zone spraying adjustment command, the billet pulling rhythm adjustment command, and the solidification stage switching adjustment command are linked and organized according to a unified time index and billet pulling direction position index to generate a trajectory regression control sequence.

[0134] In this embodiment, performing cyclic updates specifically includes:

[0135] The zone spraying adjustment command, the billet pulling rhythm adjustment command, and the solidification stage switching adjustment command in the trajectory regression control sequence are respectively sent to the corresponding cooling control unit, billet pulling execution unit, and stage switching execution unit.

[0136] During the control execution process, the rapid cooling control data of molten steel is continuously collected, and the standardized cooling solidification state sequence, the microstructure freezing drive sequence, and the current microstructure freezing trajectory are updated.

[0137] The target microstructure freezing trajectory is regenerated, and the microstructure freezing trajectory deviation characterization is updated to obtain the updated trajectory regression control sequence, thus completing the billet microstructure forming guidance control.

[0138] When the phase transformation propagation deviation, microstructure homogenization deviation, segregation suppression deviation, and thermal stress release deviation of each sampling segment in the current round of generated microstructure freezing trajectory deviation characterization are all lower than the corresponding preset deviation threshold, the update stops. The preset deviation threshold is set as the average of the corresponding deviation values ​​in the historical stable castings under the same steel grade, the same cross-sectional specification, and the same billet casting conditions.

[0139] A deep learning-based control system for rapid cooling and solidification of molten steel into billets includes:

[0140] The data acquisition module is used to acquire rapid cooling control data of molten steel during the rapid cooling and solidification process, perform preprocessing, and generate a standardized cooling and solidification state sequence.

[0141] The driving sequence construction module is used to construct tissue freezing driving sequences based on standardized cooling and condensation state sequences.

[0142] The current trajectory prediction module is used to input the tissue freeze driving sequence into the tissue freeze trajectory prediction network and extract the current tissue freeze trajectory.

[0143] The target trajectory generation module is used to constrain the formation of the target structure based on steel composition data, molten steel superheat data and preset billet structure indicators, and generate the target structure freezing trajectory.

[0144] The deviation characterization generation module is used to match the current tissue freezing trajectory with the target tissue freezing trajectory, determine the phase change propulsion deviation, tissue homogenization deviation, segregation suppression deviation, and thermal stress release deviation, and generate a tissue freezing trajectory deviation characterization.

[0145] The control sequence generation module is used to generate adjustment instructions based on the tissue freezing trajectory deviation characterization and standardized cooling and condensation state sequence, forming a trajectory regression control sequence;

[0146] The control update module is used to execute the control of rapid cooling of molten steel to solidify billet according to the trajectory regression control sequence, and continuously collect updated rapid cooling control data of molten steel for cyclic update to complete the billet microstructure forming guidance control.

[0147] refer to Figure 3 The horizontal axis represents the number of training epochs, indicating the number of times the organization-frozen trajectory prediction network iterates and updates its parameters during the training phase; the vertical axis represents the loss value. Figure 3 The blue curve represents the training loss, and the orange dashed line represents the validation loss. As the number of training rounds increases, both types of loss values ​​continue to decrease, indicating that the microstructure freezing trajectory prediction network gradually improves its ability to predict the microstructure freezing evolution path of steel billets by continuously learning the correlation between the thermal flow migration characterization, solidification propagation characterization, shell growth characterization, and cooling acceptance characterization during the historical rapid cooling and solidification process of molten steel.

[0148] Depend on Figure 3 It can be seen that the tissue freezing trajectory prediction network exhibits good convergence characteristics during training. Neither the training loss nor the validation loss shows a continuous increase or significant divergence, indicating that the network parameter update process is stable and can effectively extract tissue evolution features during rapid cooling and solidification. When the training epochs reach the later stage, the loss value gradually stabilizes, indicating that the prediction network has a relatively stable tissue freezing trajectory prediction capability.

[0149] Example 1: To verify the feasibility of this invention in practice, it was applied to a scenario involving the rapid cooling and solidification of molten steel into billets in a continuous casting workshop of a steel company. In this scenario, the continuous casting machine faces the coupled effects of steel grade switching, fluctuations in molten steel superheat, changes in the heat transfer state of the crystallizer, and dynamic adjustments to the spray intensity in the secondary cooling zone during continuous casting. Existing control methods primarily rely on manual or rule-based adjustments based on surface temperature, spray intensity, and billet casting speed. While these methods can maintain a basic production rhythm, they struggle to incorporate the internal structure formation path of the billet into a single control framework. Consequently, imbalances easily arise between the solidification progress rhythm, the level of microstructure homogenization, the ability to suppress segregation, and the coordination of thermal stress release.

[0150] To address this issue, this embodiment first continuously collects data on steel composition, molten steel superheat, crystallizer heat flow, secondary cooling zone spray intensity, billet drawing speed, and billet surface temperature. The data undergoes time synchronization, anomaly removal, missing data completion, and normalization to generate a standardized cooling and solidification state sequence. Subsequently, a heat flow migration characterization is constructed based on changes in molten steel superheat, heat flow, surface temperature, and spray intensity. This is then combined with billet drawing speed data to construct a solidification propagation characterization. Furthermore, a shell growth characterization and a cooling reception characterization are generated, forming a microstructure freezing-driven sequence. The freezing sequence is then input into the freezing trajectory prediction network to extract the current freezing trajectory. Simultaneously, a target freezing trajectory is generated based on steel composition data, molten steel superheat data, and preset billet microstructure indicators. On this basis, the current freezing trajectory and the target freezing trajectory are matched segment by segment to form a deviation characterization of the freezing trajectory. Then, based on the deviation characterization and the standardized cooling and solidification state sequence, zone spray adjustment instructions, billet pulling rhythm adjustment instructions, and solidification stage switching adjustment instructions are generated to form a freezing trajectory regression control sequence, which is continuously updated in a loop during subsequent casting processes.

[0151] To verify the performance of the present invention, it was compared with the traditional method. The comparison results are shown in Table 1.

[0152] Table 1. Comparison of Production Data between the Invention and Traditional Thermal Feedback Control Methods

[0153] As shown in Table 1, the method of this invention first demonstrates stronger stability at the process control level. The average surface temperature fluctuation of the billet decreased from 18.7℃ to 11.4℃, a reduction of 39.0%. This indicates that the invention does not simply increase the spray intensity, but rather coordinates the adjustment commands for zoned spraying, billet pulling rhythm, and solidification stage switching by matching the deviation between the current and target microstructure freezing trajectories. This makes the actions of each cooling control unit more closely aligned with the internal microstructure evolution needs of the billet in both the time and pulling direction dimensions. Correspondingly, the average continuous deviation of solidification advancement decreased from 0.268 to 0.161, a reduction of 39.9%, indicating a smoother advancement rhythm at the solidification front and a more continuous connection between the preceding and following sampling sections. Considering both sets of data together, it is clear that this invention truly advances external thermal regulation to a control level oriented towards the internal microstructure evolution path, rather than remaining at the level of traditional surface temperature tracking.

[0154] Looking at the quality data, the method of this invention directly improves the center segregation index, surface crack incidence, and billet qualification rate. The center segregation index decreased from 1.24 to 0.96, indicating a more balanced composition distribution in the billet during solidification and stronger segregation suppression. The surface crack incidence rate decreased from 2.8% to 1.5%, a reduction of 46.4%, indicating a more reasonable match between the thermal stress release rhythm and the cooling cycle, effectively reducing stress concentration caused by excessively rapid local temperature drops and abrupt stage transitions. The billet qualification rate increased from 95.6% to 98.3%, an improvement of 2.7 percentage points, which has significant engineering value in continuous casting production. This invention unifies phase transformation propagation deviation, microstructure homogenization deviation, segregation suppression deviation, and thermal stress release deviation into a microstructure freezing trajectory deviation characterization. This deviation characterization then jointly drives the three control directions of spraying, billet pulling, and solidification stage transitions. Therefore, the improvement is continuous and synergistic, rather than localized.

[0155] From the perspective of performance improvement, traditional methods mainly rely on external process modifications based on crystallizer heat flow, secondary cooling zone spray intensity, billet drawing speed, and billet surface temperature. This approach can maintain a basic cooling rhythm, but it lacks direct characterization of the internal microstructure formation path of the billet. Therefore, it often only intervenes passively after surface temperature fluctuations occur, making it difficult to simultaneously consider the continuity of solidification, the level of microstructure homogenization, and the coordination of thermal stress release. The method of this invention first constructs a microstructure freezing driving sequence from a standardized cooling and solidification state sequence, then extracts the current microstructure freezing trajectory using a microstructure freezing trajectory prediction network, and simultaneously generates a target microstructure freezing trajectory based on steel composition data, molten steel superheat data, and preset billet microstructure indicators. Subsequently, a deviation characterization of the microstructure freezing trajectory is formed by matching the sampled segments of the two trajectories. In other words, the object controlled by this invention is no longer simply the surface temperature and spray intensity, but the internal microstructure freezing path of the billet itself. This is the fundamental reason why multiple indicators in the table can be improved simultaneously.

[0156] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A molten steel rapid cooling and solidification billet control method based on deep learning, characterized by, Includes the following steps: Acquire rapid cooling control data of molten steel during the rapid cooling and solidification process, perform preprocessing on the rapid cooling control data of molten steel, and generate a standardized cooling and solidification state sequence; Based on the standardized cooling and condensation state sequence, a tissue freezing driving sequence was constructed; The tissue freeze-driving sequence is input into the tissue freeze trajectory prediction network to extract the current tissue freeze trajectory. Based on steel composition data, molten steel superheat data, and preset billet microstructure indicators, target microstructure forming constraints are applied, and target microstructure freezing trajectory is generated. The current tissue freezing trajectory is matched with the target tissue freezing trajectory to determine the deviations in phase change propulsion, tissue homogenization, segregation suppression, and thermal stress release, thereby generating a tissue freezing trajectory deviation characterization. Based on the characteristics of the tissue freezing trajectory deviation and the standardized cooling and condensation state sequence, adjustment instructions are generated to form a trajectory regression control sequence. The control of rapid cooling and solidification of molten steel billets is executed according to the trajectory regression control sequence, and the updated rapid cooling control data of molten steel is continuously collected and updated in a cycle to complete the billet microstructure forming guidance control.

2. The deep learning-based molten steel rapid cooling billet control method of claim 1, wherein, The rapid cooling control data for molten steel includes steel composition data, molten steel superheat data, crystallizer heat flow data, secondary cooling zone spray intensity data, billet drawing speed data, and billet surface temperature data.

3. The method for controlling the rapid cooling and solidification of molten steel billets based on deep learning according to claim 1, characterized in that, The preprocessing includes time synchronization, anomaly removal, missing data completion, and normalization.

4. The method for controlling the rapid cooling and solidification of molten steel billets based on deep learning according to claim 1, characterized in that, The generation of the tissue freeze-driven sequence specifically includes: Based on the superheat data of molten steel, heat flow data of the crystallizer, spray intensity data of the secondary cooling zone, and surface temperature data of the billet in the standardized cooling and solidification state sequence, the changes in heat flow, surface temperature, spray intensity, and superheat of molten steel between adjacent sampling times are extracted and correlated with the structure to generate a heat flow migration characterization. By temporally mapping the heat flow migration characterization with the billet speed data, the solidification propulsion cycle value, solidification propulsion continuous deviation value, and solidification propulsion position migration amount are extracted to generate the solidification propulsion characterization. Based on solidification propagation characterization, crystallizer heat flow data and billet surface temperature data, the solidification coverage, surface cooling capacity and solidification expansion of each sampling section are calculated to generate shell growth characterization. Based on the spray intensity data of the second cooling zone, the surface temperature data of the billet, and the shell growth characterization, the cooling effect transfer amount, cooling response duration, and cooling cycle matching value between adjacent cooling zones are calculated to generate a cooling bearing characterization. The heat flow migration characterization, solidification propulsion characterization, shell growth characterization, and cooling reception characterization are arranged sequentially according to time index and correlated to generate a tissue freezing drive sequence.

5. The method for controlling the rapid cooling and solidification of molten steel billets based on deep learning according to claim 1, characterized in that, The generation of the current organization freeze trajectory specifically includes: The frozen drive sequence is divided into continuous sliding windows, and the thermal migration, solidification propulsion, shell growth and cooling reception characteristics of each window are organized to generate a set of windowed drive subsequences. Input the windowed-driven subsequence set into the tissue freeze trajectory prediction network; The tissue freezing trajectory prediction network includes an input mapping layer, a thermal history embedding layer, a stage transition layer, a tissue evolution mapping layer, and a trajectory output layer. In the input mapping layer, vectorized encoding and splicing are performed on the heat flow migration characterization, solidification propulsion characterization, shell growth characterization and cooling acceptance characterization in each window to generate a driving embedding sequence. The driving embedding sequence is input into the thermal history embedding layer to extract thermal history embedding features from multiple time periods. The multi-period thermal history embedding features and driving embedding sequences are input into the stage bearing layer to calculate the solidification propulsion connection value, shell growth connection value and cooling bearing connection value, and generate solidification stage bearing features. The multi-period thermal history embedding features and solidification stage inheritance features are spliced ​​together to generate tissue evolution input vector, which is then input into the tissue evolution mapping layer to generate tissue evolution mapping features. The tissue evolution mapping features are input into the trajectory output layer to generate the current tissue frozen trajectory.

6. The method for controlling the rapid cooling and solidification of molten steel billets based on deep learning according to claim 1, characterized in that, The generation of the target tissue freeze trajectory specifically includes: Read the steel composition data and molten steel superheat data corresponding to the current casting, and calculate the solidification effect value of the composition and the initial heat reserve value of superheat. Based on the solidification effect value of the composition, the initial heat reserve value of overheating, and the preset billet structure index, the target phase transformation advancement value, the target structure homogenization value, the target segregation suppression value, and the target thermal stress release value are generated. The target phase change advancement value, target microstructure homogenization value, target segregation suppression value, and target thermal stress release value are assigned to each sampling segment to generate target phase change advancement component sequence, target microstructure homogenization component sequence, target segregation suppression component sequence, and target thermal stress release component sequence. These are then mapped to the unified time index and billet pulling direction position index corresponding to the standardized cooling and condensation state sequence to complete the position alignment. The target phase change propulsion component sequence, target tissue homogenization component sequence, target segregation suppression component sequence, and target thermal stress release component sequence, after being aligned in position, are associated with the tissue to generate the target tissue freezing trajectory.

7. The method for controlling the rapid cooling and solidification of molten steel billets based on deep learning according to claim 1, characterized in that, The generation of the tissue freezing trajectory deviation characterization specifically includes: By matching the current tissue freezing trajectory with the target tissue freezing trajectory segment by segment, the current phase change propulsion component trajectory, the current tissue homogenization component trajectory, the current segregation suppression component trajectory, the current thermal stress release component trajectory, and the target corresponding component trajectory are extracted to generate a trajectory alignment sequence. Based on the current phase transition propulsion component and the target phase transition propulsion component, and the current tissue homogenization component and the target tissue homogenization component of each sampling segment in the trajectory alignment sequence, a phase transition propulsion difference sequence and a tissue homogenization difference sequence are generated. A moving average is then performed on the phase transition propulsion difference and tissue homogenization difference of adjacent sampling segments to generate phase transition propulsion deviation and tissue homogenization deviation. Based on the current segregation suppression component and the target segregation suppression component, and the current thermal stress release component and the target thermal stress release component of each sampling segment in the trajectory alignment sequence, a segregation suppression difference sequence and a thermal stress release difference sequence are generated. A moving average is then performed on the segregation suppression difference and thermal stress release difference of adjacent sampling segments to generate segregation suppression deviation and thermal stress release deviation. By associating phase transformation propulsion deviation, microstructure homogenization deviation, segregation suppression deviation, and thermal stress release deviation with the microstructure, a microstructure freezing trajectory deviation characterization is generated.

8. The method for controlling the rapid cooling and solidification of molten steel billets based on deep learning according to claim 1, characterized in that, The generation of the trajectory regression control sequence specifically includes: Read the phase transition propagation deviation, tissue homogenization deviation, segregation suppression deviation and thermal stress release deviation in the tissue freezing trajectory deviation characterization, and generate a control decision sequence by corresponding each sampled segment of the standardized cooling and condensation state sequence; Based on the phase change propulsion deviation, segregation suppression deviation and corresponding secondary cooling zone spray intensity data of each sampling segment in the control decision sequence, the spray adjustment amount of each cooling segment is calculated and the zone spray adjustment command is generated. Based on the phase change propagation deviation, thermal stress release deviation and corresponding billet pulling speed data of each sampling section, the billet pulling speed correction amount of each sampling section is calculated, and the billet pulling rhythm adjustment command is generated. Based on the deviation of microstructure homogenization, thermal stress release deviation and corresponding billet surface temperature data of each sampling section, the stage switching correction amount of each sampling section is calculated, and solidification stage switching adjustment command is generated. The zone spraying adjustment command, the billet pulling rhythm adjustment command, and the solidification stage switching adjustment command are linked together to generate a trajectory regression control sequence.

9. The method for controlling the rapid cooling and solidification of molten steel billets based on deep learning according to claim 1, characterized in that, The cyclical update specifically includes: The zone spraying adjustment command, the billet pulling rhythm adjustment command, and the solidification stage switching adjustment command in the trajectory regression control sequence are respectively sent to the corresponding cooling control unit, billet pulling execution unit, and stage switching execution unit. During the control execution process, the rapid cooling control data of molten steel is continuously collected, and the standardized cooling solidification state sequence, the microstructure freezing drive sequence, and the current microstructure freezing trajectory are updated. The target microstructure freezing trajectory is regenerated, and the microstructure freezing trajectory deviation characterization is updated to obtain the updated trajectory regression control sequence, thus completing the billet microstructure forming guidance control.

10. A deep learning-based control system for rapid cooling and solidification of molten steel billets, executing the deep learning-based control method for rapid cooling and solidification of molten steel billets according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire rapid cooling control data of molten steel during the rapid cooling and solidification process, perform preprocessing, and generate a standardized cooling and solidification state sequence. The driving sequence construction module is used to construct tissue freezing driving sequences based on standardized cooling and condensation state sequences. The current trajectory prediction module is used to input the tissue freeze driving sequence into the tissue freeze trajectory prediction network and extract the current tissue freeze trajectory. The target trajectory generation module is used to constrain the formation of the target structure based on steel composition data, molten steel superheat data and preset billet structure indicators, and generate the target structure freezing trajectory. The deviation characterization generation module is used to match the current tissue freezing trajectory with the target tissue freezing trajectory, determine the phase change propulsion deviation, tissue homogenization deviation, segregation suppression deviation, and thermal stress release deviation, and generate a tissue freezing trajectory deviation characterization. The control sequence generation module is used to generate adjustment instructions based on the tissue freezing trajectory deviation characterization and standardized cooling and condensation state sequence, forming a trajectory regression control sequence; The control update module is used to execute the control of rapid cooling of molten steel to solidify billet according to the trajectory regression control sequence, and continuously collect updated rapid cooling control data of molten steel for cyclic update to complete the billet microstructure forming guidance control.