A parameter recommendation method for controlling abnormal fluctuation of a crystallizer liquid level

Abstract

The application provides a crystallizer liquid level abnormal fluctuation control parameter recommendation method, relates to the technical field of steel metallurgy continuous casting production control, and comprises the following steps: multi-source heterogeneous data acquisition and feature engineering; multi-flow coupling state perception based on a space-time graph convolution network; multi-agent state decoupling based on an attention mechanism; multi-agent collaborative decision making based on a multi-agent deep deterministic policy gradient algorithm (MADDPG) framework; hierarchical and hierarchical parameter recommendation and execution. The multi-flow collaborative intelligent control scheme provided by the application realizes overall modeling and accurate collaborative control of a multi-flow system of a caster by combining feature capture coupled with a graph neural network and multi-agent reinforcement learning collaborative decision making technology, effectively suppresses multi-flow neighborhood interference, and greatly reduces overall liquid level fluctuation.

Description

Technical Field

[0001] This invention relates to the technical field of continuous casting production control in iron and steel metallurgy, and more particularly to a parameter recommendation method for controlling abnormal fluctuations in the liquid level of the crystallizer. Background Technology

[0002] Continuous casting technology is hailed as the heart of the steel industry, and the crystallizer is the heart of the continuous casting machine. The stability of the molten steel level within the crystallizer is one of the key indicators for evaluating the operational level of continuous casting. As continuous casting technology develops towards higher casting speeds and higher quality, the requirements for crystallizer molten steel level fluctuations are becoming increasingly stringent. Typically, the fluctuation range of the crystallizer molten steel level needs to be controlled within a certain range. Within acceptable limits. Abnormal fluctuations in the liquid level can lead to serious quality and safety incidents: violent fluctuations can entrain the protective slag from the mold surface into the molten steel, forming large inclusions, resulting in incomplete billet inspection or cracking of the rolled product. Fluctuations disrupt the slag film structure at the meniscus, causing uneven lubrication between the billet shell and the mold wall, increasing drawing resistance, and even causing sticking leaks. Frequent liquid level fluctuations cause frequent changes in the meniscus position, resulting in uneven initial billet shell thickness; weak points are highly susceptible to cracking under thermal stress or leaks.

[0003] Existing control technologies and patents for liquid level fluctuations in crystallizers have the following limitations: Single-flow islanded control ignores inter-flow coupling: Most existing technologies treat each crystallizer flow as an independent control object. However, in actual production, a single machine with eight flows shares the same tundish. When one flow experiences an anomaly (such as nozzle blockage or a sudden drop in casting speed), it causes a rebalancing of the molten steel distribution within the tundish, which in turn induces surface fluctuations in adjacent flows. Existing single-flow control models cannot detect this neighboring interference, often leading to alternating alarms from multiple flows, exponentially increasing the control difficulty.

[0004] The control logic is rigid and lacks self-learning ability: traditional PID control or simple expert rule bases are inadequate when faced with complex fluctuation patterns. Although some existing patents mention intelligent recognition capabilities, most are based on shallow machine learning (such as SVM and decision trees), which have limited feature extraction capabilities, cannot handle high-dimensional time series data, and cannot achieve online evolution and adaptive optimization of control strategies.

[0005] Parameter recommendations lack dynamic coordination: In multi-stream scenarios, adjusting one parameter (such as the common pull speed) will simultaneously affect all streams. Current technology cannot balance this complex relationship where a change in one part affects the whole, and the recommended parameters are often locally optimal solutions, or even suboptimal solutions globally.

[0006] Therefore, there is an urgent need for a parameter recommendation method that can sense the state of multi-flow coupling and has dynamic decision-making and self-learning capabilities to solve the problem of controlling abnormal fluctuations in the liquid level of the crystallizer in a one-machine-eight-flow scenario. Summary of the Invention

[0007] Addressing the technical problems mentioned in the background section, such as the difficulty in controlling crystallizer liquid level fluctuations, lag in regulation, and even erroneous regulation due to neglecting inter-flow coupling, rigid control models, and the inability to dynamically and collaboratively optimize, this invention provides a parameter recommendation method for controlling abnormal fluctuations in the crystallizer liquid level. This method constructs a digital twin environment for the multi-flow continuous casting process and utilizes advanced artificial intelligence algorithms—fusion of temporal graph neural networks and multi-agent reinforcement learning—to achieve accurate perception of multi-flow fluctuation states, in-depth mining of coupling relationships, and autonomous decision-making and recommendation of collaborative control strategies. This fundamentally improves the overall suppression capability of an 8-flow continuous casting machine for abnormal liquid level fluctuations.

[0008] The technical means employed in this invention are as follows: A parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer includes the following steps: Step 1: Multi-source heterogeneous data acquisition and feature engineering; collect flow state data and common state data from the primary PLC and secondary server to construct a structured dataset. ; for the structured dataset Preprocessing is performed, and the preprocessed structured dataset is used as a basis. Three derived features are constructed: first-order difference of liquid surface deviation, short-time energy of liquid surface, and correlation coefficient of inter-flow liquid surface fluctuation. Historical time-series data of fixed duration are extracted to generate the dimension. The time series samples, where, Indicates the number of historical moments. Indicates feature dimension; Step 2: Multi-flow coupling state perception based on spatiotemporal graph convolutional network; model the 8 crystallizer flows as a dynamic topological graph structure, with crystallizer flows as nodes and inter-flow molten steel coupling relationships as edges, and construct a dynamic adjacency matrix by combining the physical mechanism of continuous casting and learnable parameters; input the time series samples generated in Step 1 and the dynamic adjacency matrix into the stacked spatiotemporal graph convolutional block, aggregate node coupling information through spatial graph convolution, capture temporal evolution features through gated temporal convolution, and encode the global coupled state tensor that integrates spatiotemporal coupling information; Step 3: Multi-agent state decoupling based on attention mechanism; Define an independent agent for each crystallizer flow, obtain the corresponding intrinsic state of the flow through each agent, and dynamically aggregate the coupling states of the other crystallizer flows from the global coupling state tensor through multi-head attention mechanism, and splice the intrinsic state and coupling state to obtain the local observation value of each agent; Step 4: Multi-agent collaborative decision-making based on the MADDPG (Multi-Agent Deep Deterministic Policy Gradient Algorithm) framework; setting the continuous action space of each agent and a global reward function including fluctuation penalty, action penalty, equilibrium penalty, and constraint penalty; building a distributed executor policy network and a centralized evaluator network based on the MADDPG algorithm, and completing offline model training according to the experience replay pool to obtain the trained executor policy network; adjusting actions by agents through parallel output of process parameters based on local observations according to the trained executor policy network. Step 5: Layered Parameter Recommendation and Execution; A three-layered hierarchical structure is adopted to output macro-adjustment suggestions and single-stream parameter fine-tuning amounts. Threshold and equipment status safety checks are performed on the output macro-adjustment suggestions and single-stream parameter fine-tuning amounts, and the final recommended parameters are output.

[0009] Furthermore, in step 1, the flow state data includes: liquid level deviation. , pulling speed Argon flow rate Stopper rod position and opening Submersible water inlet insertion depth Crystallizer heat flux density Cooling water flow rate of crystallizer and the standard deviation of liquid level fluctuation The common status data includes: molten steel level in the tundish. 1. Tundish molten steel temperature Total argon pressure and the grade of cast steel .

[0010] Furthermore, in step 1, the structured dataset... The preprocessing includes the following steps: using the median absolute deviation method, outlier data exceeding 3 times the median absolute deviation (MAD) within the sliding window are removed, and missing data are filled in by linear interpolation; a Kalman filter is constructed for the liquid level deviation signal, using random walk as the state equation and combining sensor noise to construct the observation equation to complete signal filtering; and Z-score standardization is performed on all continuous feature variables.

[0011] Furthermore, in step 2, the formula for calculating the elements of the dynamic adjacency matrix is:

[0012] in, Indicates the distance between the two water outlets; Indicates the liquid level in the tundish; Indicates learnable parameters; All represent fixed hyperparameters; Indicates the reference intermediate package liquid level; matrix self-connecting elements .

[0013] Furthermore, the spatiotemporal graph convolutional block includes: a spatial graph convolutional layer and a gated temporal convolutional layer; the spatial graph convolutional layer is used to aggregate the spatial coupling features of adjacent crystallizer flow nodes, and the gated temporal convolutional layer uses GLU gated linear units to capture the temporal variation of parameters under single-flow and multi-flow coupling states; and the spatiotemporal graph convolutional network completes the initial parameter optimization through self-supervised pre-training.

[0014] Furthermore, in step 3, the calculation formula for the coupling state of each agent is as follows: ; in, The query vector, key vector, and value vector are respectively obtained by linear transformation of the deep embedding vector; Indicates the dimension of a vector; Indicates the first The coupling state characteristics corresponding to each agent.

[0015] Furthermore, in step 4, the action space of the intelligent agent includes: the pulling speed adjustment amount and the argon flow rate adjustment amount. For parameters that require manual intervention, the model outputs an abnormal confidence level. When the confidence level exceeds a preset threshold, a manual adjustment prompt is pushed to the operator.

[0016] Furthermore, in step 4, the global reward function expression is: ; in, Indicates the first The deviation of the liquid level in the flow is the first The difference between the actual liquid level and the target liquid level of molten steel in each crystallizer; Indicates the first Adjustment amount of flow pulling speed; Indicates the first Adjustment of argon flow rate, Indicates the penalty weight for liquid level deviation. Indicates the penalty weight for the magnitude of the action. Indicates the weight of the inter-flow balance penalty. Indicates the hard constraint penalty weight. This represents the absolute value of the variance of the fluctuation amplitude of each flow. This indicates a violation indicator function.

[0017] Furthermore, the violation indication function These include: absolute limit constraints, rate of change constraints, and process coupling constraints; The absolute limit constraint is: the flow rate of each stream. Argon flow rate Stopper rod opening And liquid level deviation ; The rate of change constraint is: the pulling speed adjustment amount within each inference cycle. Argon gas adjustment amount ; The process coupling constraints include: pull speed-heat flux density matching constraints, argon-stopper opening coordination constraints, pull speed-inter-flow consistency constraints, and pull speed limiting constraints under low liquid levels in the tundish.

[0018] Furthermore, in step 5, the three-layer hierarchical structure consists of: a macro trend recommendation layer, a micro fluctuation suppression layer, and a security watchdog.

[0019] Compared with the prior art, the present invention has the following advantages: 1. The multi-flow collaborative intelligent control scheme provided by this invention, by combining graph neural network coupled feature capture and multi-agent reinforcement learning collaborative decision-making technology, realizes the overall modeling and precise collaborative control of the multi-flow system of the casting machine, effectively suppresses multi-flow neighborhood interference, and significantly reduces the overall liquid level fluctuation.

[0020] 2. This invention takes the long-term cumulative reward of reinforcement learning as the optimization goal, which breaks the limitation of traditional control schemes that only focus on instantaneous control effects. It achieves forward-looking and global optimization of control decisions, takes into account both the current liquid level fluctuation suppression effect and the long-term stability of subsequent production conditions, and avoids the problem of short-term optimization and long-term imbalance of operating conditions.

[0021] 3. This invention is equipped with an intelligent model that has self-learning and iterative capabilities. It can adapt to the dynamic changes in equipment status such as copper plate wear and nozzle aging in real time, continuously iterate and optimize the control strategy, and has a strong adaptability to working conditions. It does not require frequent manual parameter adjustments and ensures stable control accuracy throughout the entire production process.

[0022] 4. This invention, through precise multi-flow coordinated control, can ensure that the liquid levels of the eight flow lines of the casting machine remain stable within the standard target range, effectively unifying key indicators such as the subsurface quality and vibration mark depth of the casting billets in each flow, significantly improving the consistency of casting billet quality, and providing high-quality raw material support for subsequent rolling processes.

[0023] In summary, the technical solution of this invention solves the problems of existing multi-flow control technologies for casting machines, which are mostly single-flow independent control, do not consider inter-flow coupling interference, have large liquid level fluctuations, short-sighted control decisions, cannot adapt to aging equipment conditions, and have large differences in the quality of billets from different flows. By modeling a single multi-flow system, using graph neural network coupled perception, and multi-agent reinforcement learning for global collaborative decision-making, it achieves multi-flow interference suppression, forward-looking optimal control, adaptive optimization of working conditions, and unified management and control of billet quality.

[0024] Therefore, the technical solution of the present invention solves the problems of severe multi-flow coupling interference, poor liquid surface fluctuation control effect, solidified control strategy without adaptability, and poor consistency of multi-flow quality of billet in the prior art.

[0025] Based on the above reasons, this invention can be widely promoted in fields such as continuous casting of steel, collaborative control of multi-flow industrial equipment, and intelligent manufacturing in metallurgy. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0028] Figure 2 This is a diagram of the architecture of the spatiotemporal graph convolutional network ST-GCN and the multi-agent deep deterministic policy gradient (MADDPG) hybrid intelligent recommendation model of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] like Figures 1-2As shown, this invention provides a parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer, comprising the following steps: Step 1: Multi-source heterogeneous data acquisition and feature engineering. To achieve accurate state perception, a high-dimensional real-time data pool covering eight flows and including process parameters and state parameters needs to be constructed. Flow state data and common state data (including liquid level deviation) are collected from the primary PLC and secondary server. , pulling speed Argon flow rate Stopper rod position and opening Submersible water inlet insertion depth Crystallizer heat flux density Cooling water flow rate of crystallizer and the standard deviation of liquid level fluctuation The common status data includes: molten steel level in the tundish. 1. Tundish molten steel temperature Total argon pressure and the grade of cast steel ), build a structured dataset .

[0032] The raw data needs to be cleaned and feature-enhanced to improve the model's robustness and expressive power. This applies to the structured dataset. Preprocessing is performed on the structured dataset. The preprocessing includes the following steps: using the median absolute deviation method, outlier data exceeding 3 times the MAD (median absolute deviation) within the sliding window are removed, and missing data are filled in by linear interpolation; a Kalman filter is constructed for the liquid level deviation signal, using random walk as the state equation and combining sensor noise to construct the observation equation to complete signal filtering; and Z-score standardization is performed on all continuous feature variables.

[0033] To further enhance the model's ability to identify fluctuation patterns, the following additional derived features are constructed: First-order difference of liquid level deviation ; Short-time energy of liquid level deviation ; correlation coefficient of liquid surface fluctuation Calculate the flow within the sliding window With flow The Pearson correlation coefficient of the liquid level deviation, as a direct measure of coupling, is used as an auxiliary input for the subsequent graph adjacency matrix.

[0034] At the current moment Based on the baseline, take forward. At any moment (e.g.) The time series samples consist of 1 second of historical data, with each sample having a dimension of 1. ,in For feature dimensions.

[0035] Preferably, step 2 involves multi-flow coupling state perception based on a spatiotemporal graph convolutional network; modeling the eight crystallizer flows as a dynamic topology graph structure. ,in, Indicates node association; Represents the set of edges; Representing a dynamic adjacency matrix; using the crystallizer flow as nodes and the inter-flow molten steel coupling relationship as edges, a dynamic adjacency matrix is ​​constructed by combining the physical mechanism of continuous casting and learnable parameters; the time-series samples generated in step 1 and the dynamic adjacency matrix are input into a stacked spatiotemporal graph convolution block, and node coupling information is aggregated through spatial graph convolution, and temporal evolution features are captured through gated temporal convolution, encoding a global coupled state tensor that integrates spatiotemporal coupling information; elements Quantified time flow convection The influence weights are jointly determined by physical mechanisms and data-driven factors. The formula for calculating the elements of the dynamic adjacency matrix is ​​as follows:

[0036] in, Indicates the distance between the two water outlets; Indicates the liquid level in the tundish; Indicates learnable parameters; All represent fixed hyperparameters; Indicates the reference intermediate package liquid level; matrix self-connecting elements .

[0037] In a preferred embodiment, the spatiotemporal graph convolutional block in this application includes: a spatial graph convolutional layer and a gated temporal convolutional layer; the spatial graph convolutional layer is used to aggregate the spatial coupling features of adjacent crystallizer flow nodes, and the gated temporal convolutional layer uses GLU gated linear units to capture the temporal variation of parameters under single-flow and multi-flow coupling states; and the spatiotemporal graph convolutional network completes the initial parameter optimization through self-supervised pre-training.

[0038] Step 3: Multi-agent state decoupling based on attention mechanism; An independent agent is defined for each crystallizer flow. Each agent obtains its corresponding intrinsic state, and the coupled states of the remaining crystallizer flows are dynamically aggregated from the global coupled state tensor using a multi-head attention mechanism. The eigenstates and coupled states are then concatenated to obtain the local observations of each agent. In this application, the local observations of each agent... It consists of two parts: Intrinsic state: recent history and current process parameters of the current flow; Coupled state: the state transition from global state using a multi-head attention mechanism. Information from other streams is selectively aggregated; the formula for calculating the coupling state of each agent is: ; in, The query vector, key vector, and value vector are respectively obtained by linear transformation of the deep embedding vector; Indicates the dimension of a vector; Indicates the first The coupling state characteristics corresponding to each agent.

[0039] Step 4: Multi-agent collaborative decision-making based on the MADDPG (Multi-Agent Deep Deterministic Policy Gradient Algorithm) framework; setting the continuous action space of each agent and a global reward function including fluctuation penalty, action penalty, equilibrium penalty, and constraint penalty; building a distributed Actor (executor) policy network and a centralized Critic (evaluator) network based on the MADDPG algorithm, and completing offline model training according to the experience replay pool to obtain the trained Actor policy network; according to the trained Actor policy network, the agents adjust the actions by outputting process parameters in parallel based on local observations; in Step 4, the action space of the agents includes: the pulling speed adjustment amount and the argon flow rate adjustment amount. For parameters that require manual intervention, the model outputs an abnormal confidence level. When the confidence level exceeds a preset threshold, a manual adjustment prompt is pushed to the operator.

[0040] In a preferred embodiment, the global reward function expression in this application is: ; in, Indicates the first The deviation of the liquid level in the flow is the first The difference between the actual liquid level and the target liquid level of molten steel in each crystallizer; Indicates the first Adjustment amount of flow pulling speed; Indicates the first Adjustment of argon flow rate, Indicates the penalty weight for liquid level deviation. Indicates the penalty weight for the magnitude of the action. Indicates the weight of the inter-flow balance penalty. Indicates the hard constraint penalty weight. This represents the absolute value of the variance of the fluctuation amplitude of each flow. This indicates a violation indicator function.

[0041] In this application, the violation indication function These include: absolute limit constraints, rate of change constraints, and process coupling constraints; The absolute limit constraint is: the flow rate of each stream. Argon flow rate Stopper rod opening And liquid level deviation ; The rate of change constraint is: the pulling speed adjustment amount within each inference cycle. Argon gas adjustment amount ; The process coupling constraints include: pull speed-heat flux density matching constraints, argon-stopper opening coordination constraints, pull speed-inter-flow consistency constraints, and pull speed limiting constraints under low liquid levels in the tundish.

[0042] Step 5, Hierarchical Parameter Recommendation and Execution: A three-layer hierarchical structure is adopted, outputting macro-level adjustment suggestions and single-stream parameter fine-tuning amounts. Threshold and device status safety checks are performed on the output macro-level adjustment suggestions and single-stream parameter fine-tuning amounts, and the final recommended parameters are output. The three-layer hierarchical structure consists of: a macro-level trend recommendation layer, a micro-level fluctuation suppression layer, and a safety watchdog.

[0043] Macro trend recommendation layer: Based on the long-term trends perceived by ST-GCN and information such as steel grade replacement, it provides macro adjustment suggestions for overall casting speed, argon gas, etc.

[0044] Micro-fluctuation suppression layer: The MADDPG model calculates the fine-tuning of each flow in real time, which directly affects the flow level control.

[0045] Safety watchdog: All recommended actions must pass safety checks (absolute value constraints, rate of change constraints, and device status checks) before being issued. If a violation occurs, the system will block and issue an alarm.

[0046] As a preferred implementation method, in this application, the experience generated during online operation is stored in a replay pool, and the model is periodically fine-tuned using new data. An offline reinforcement learning algorithm is employed to avoid the risks of online exploration. After fine-tuning, the model is verified through simulation before redeployment, and historical versions are retained for rollback support.

[0047] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the above embodiments of the present invention, the descriptions of each embodiment have their own emphasis; parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. It should be understood that the disclosed technical content in the several embodiments provided in this application can be implemented in other ways.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer, characterized in that, Includes the following steps: Step 1: Multi-source heterogeneous data acquisition and feature engineering; Stream status data and common status data are collected from the primary PLC and secondary server to construct a structured dataset. ; for the structured dataset Preprocessing is performed, and the preprocessed structured dataset is used as a basis. Three derived features are constructed: first-order difference of liquid surface deviation, short-time energy of liquid surface, and correlation coefficient of inter-flow liquid surface fluctuation. Historical time-series data of fixed duration are extracted to generate the dimension. The time series samples, where, Indicates the number of historical moments. Indicates feature dimension; Step 2: Multi-flow coupling state perception based on spatiotemporal graph convolutional network; model the 8 crystallizer flows as a dynamic topological graph structure, with crystallizer flows as nodes and inter-flow molten steel coupling relationships as edges, and construct a dynamic adjacency matrix by combining the physical mechanism of continuous casting and learnable parameters; input the time series samples generated in Step 1 and the dynamic adjacency matrix into the stacked spatiotemporal graph convolutional block, aggregate node coupling information through spatial graph convolution, capture temporal evolution features through gated temporal convolution, and encode the global coupled state tensor that integrates spatiotemporal coupling information; Step 3: Multi-agent state decoupling based on attention mechanism; Define an independent agent for each crystallizer flow, obtain the corresponding intrinsic state of the flow through each agent, and dynamically aggregate the coupling states of the other crystallizer flows from the global coupling state tensor through multi-head attention mechanism, and splice the intrinsic state and coupling state to obtain the local observation value of each agent; Step 4: Multi-agent collaborative decision-making based on the MADDPG (Multi-Agent Deep Deterministic Policy Gradient Algorithm) framework; setting the continuous action space of each agent and a global reward function including fluctuation penalty, action penalty, equilibrium penalty, and constraint penalty; building a distributed executor policy network and a centralized evaluator network based on the MADDPG algorithm, and completing offline model training according to the experience replay pool to obtain the trained executor policy network; adjusting actions by agents through parallel output of process parameters based on local observations according to the trained executor policy network. Step 5: Layered Parameter Recommendation and Execution; A three-layered hierarchical structure is adopted to output macro-adjustment suggestions and single-stream parameter fine-tuning amounts. Threshold and equipment status safety checks are performed on the output macro-adjustment suggestions and single-stream parameter fine-tuning amounts, and the final recommended parameters are output.

2. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 1, characterized in that, In step 1, the flow state data includes: liquid level deviation. , pulling speed Argon flow rate Stopper rod position and opening Submersible water inlet insertion depth Crystallizer heat flux density Cooling water flow rate of crystallizer and the standard deviation of liquid level fluctuation The common status data includes: molten steel level in the tundish.

1. Tundish molten steel temperature Total argon pressure and the grade of cast steel .

3. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 1, characterized in that, In step 1, the structured dataset The preprocessing includes the following steps: using the median absolute deviation method, outlier data exceeding 3 times the median absolute deviation (MAD) within the sliding window are removed, and missing data are filled in by linear interpolation; a Kalman filter is constructed for the liquid level deviation signal, using random walk as the state equation and combining sensor noise to construct the observation equation to complete signal filtering; and Z-score standardization is performed on all continuous feature variables.

4. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 1, characterized in that, In step 2, the formula for calculating the elements of the dynamic adjacency matrix is ​​as follows: in, Indicates the distance between the two water outlets; Indicates the liquid level in the tundish; Indicates learnable parameters; All represent fixed hyperparameters; Indicates the reference intermediate package liquid level; matrix self-connecting elements .

5. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 1, characterized in that, The spatiotemporal graph convolutional block includes a spatial graph convolutional layer and a gated temporal convolutional layer. The spatial graph convolutional layer is used to aggregate the spatial coupling features of adjacent crystallizer flow nodes. The gated temporal convolutional layer uses GLU gated linear units to capture the temporal variation of parameters under single-flow and multi-flow coupling states. The spatiotemporal graph convolutional network completes the initial parameter optimization through self-supervised pre-training.

6. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 1, characterized in that, In step 3, the formula for calculating the coupling state of each agent is as follows: ； in, The query vector, key vector, and value vector are respectively obtained by linear transformation of the deep embedding vector; Indicates the dimension of a vector; Indicates the first The coupling state characteristics corresponding to each agent.

7. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 1, characterized in that, In step 4, the action space of the intelligent agent includes: the pulling speed adjustment amount and the argon flow rate adjustment amount. For parameters that require manual intervention, the model outputs an abnormal confidence level. When the confidence level exceeds a preset threshold, a manual adjustment prompt is pushed to the operator.

8. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 1, characterized in that, In step 4, the global reward function expression is: ； in, Indicates the first The deviation of the liquid level in the flow is the first The difference between the actual liquid level and the target liquid level of molten steel in each crystallizer; Indicates the first Adjustment amount of flow pulling speed; Indicates the first Adjustment of argon flow rate, Indicates the penalty weight for liquid level deviation. Indicates the penalty weight for the magnitude of the action. Indicates the weight of the inter-flow balance penalty. Indicates the hard constraint penalty weight. This represents the absolute value of the variance of the fluctuation amplitude of each flow. This indicates a violation indicator function.

9. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 8, characterized in that, The violation indication function These include: absolute limit constraints, rate of change constraints, and process coupling constraints; The absolute limit constraint is: the flow rate of each stream. Argon flow rate Stopper rod opening And liquid level deviation ; The rate of change constraint is: the pulling speed adjustment amount within each inference cycle. Argon gas adjustment amount ; The process coupling constraints include: pull speed-heat flux density matching constraints, argon-stopper opening coordination constraints, pull speed-inter-flow consistency constraints, and pull speed limiting constraints under low liquid levels in the tundish.

10. The parameter recommendation method for controlling abnormal fluctuations in the liquid level of a crystallizer according to claim 1, characterized in that, In step 5, the three-layer hierarchical structure consists of: a macro trend recommendation layer, a micro fluctuation suppression layer, and a security watchdog.

Images