A smelting process intelligent optimization method and system based on multi-source data fusion

Abstract

The application relates to a smelting process intelligent optimization method and system based on multi-source data fusion, which comprises the following steps: obtaining target process parameters, real-time sensor data and production process data; dividing thermodynamic and chemical component parameter sets according to characteristics, constructing corresponding sub-atlas and fusing into a process knowledge atlas. Time series are preprocessed and aligned, and a comprehensive smelting state space model is constructed. Key state vectors are extracted to input a hierarchical reinforcement learning framework, the framework coordinates double-network decoupling decisions according to parameter priority and coupling relationship, and generates a collaborative optimization action set. The knowledge atlas is called to solve multi-objective constraints and conflict resolution, and an executable instruction set is generated. The instructions are executed, feedback data is collected, a differential reward value is calculated, and the framework parameters and atlas dynamic weight edges are updated. Through multi-source data fusion and hierarchical reinforcement learning collaborative optimization, dynamic adjustment and precise control of smelting process parameters are realized.

Description

Technical Field

[0001] This application relates to the field of intelligent optimization control technology for industrial processes, and in particular to an intelligent optimization method and system for smelting processes based on multi-source data fusion. Background Technology

[0002] With the continuous development of intelligent manufacturing technology, the steel industry, as an important foundation of the national economy, is accelerating its transformation towards green, efficient, and intelligent manufacturing. In this process, achieving precise control and optimization of the core smelting process is crucial for improving product quality, reducing production costs, and ensuring production safety and stability.

[0003] In recent years, sensing and connectivity technologies, such as online detection sensors, the Internet of Things, and big data, have provided the hardware foundation for real-time acquisition of key parameters such as molten pool temperature, oxygen activity, and chemical composition. Meanwhile, artificial intelligence technologies, represented by machine learning and optimization algorithms, have provided new theoretical tools for mining knowledge from massive amounts of data and finding optimal process paths. The development of these technologies collectively constitutes the external conditions driving the evolution of smelting processes from experience-driven to model- and data-driven approaches.

[0004] However, in practical applications, the smelting process is a dynamic system with multiple variables, strong coupling, large time delays, and complex physicochemical reactions. Traditional optimization models based on a single mechanism or single data-driven approach often struggle to coordinate the intricate interactions between thermodynamic state control and chemical composition control, neglecting the deep coupling mechanisms between the two. This leads to situations where one aspect is prioritized at the expense of another, making it difficult to achieve global optimum. Furthermore, the optimization suggestions generated by existing data-driven optimization models lack interpretability and cannot strictly guarantee compliance with process safety regulations, limiting their practical application in industrial settings with high safety requirements. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for intelligent optimization of smelting processes based on multi-source data fusion.

[0006] Firstly, this application provides an intelligent optimization method for smelting processes based on multi-source data fusion, employing the following technical solution:

[0007] Acquire the target process parameters of the steel grade to be smelted, as well as real-time sensor data and production process data during the smelting process;

[0008] Based on the parameter characteristics, the target process parameters are divided into a thermodynamic parameter set and a chemical composition parameter set, and thermodynamic constraint sub-maps and chemical composition correlation sub-maps are constructed respectively, and then fused to generate a process knowledge graph.

[0009] The real-time sensor data and production process data are preprocessed and time-series aligned to construct a comprehensive smelting state space model for the current furnace batch.

[0010] Key state vectors are extracted from the integrated smelting state-space model and input into a pre-defined hierarchical reinforcement learning optimization framework; wherein, the hierarchical reinforcement learning optimization framework includes a thermodynamic policy network and a chemical composition policy network;

[0011] The hierarchical reinforcement learning optimization framework, based on the parameter priority and coupling relationship in the target process parameters, coordinates the thermodynamic strategy network and the chemical composition strategy network to make decoupled decisions and generate a set of collaborative optimization actions.

[0012] The process knowledge graph is invoked to perform multi-objective constraint solving and conflict resolution on the collaborative optimization action set, generating an executable optimization instruction set.

[0013] The executable optimized instruction set is executed, and multi-dimensional feedback detection data is collected simultaneously during the execution process;

[0014] Based on the multi-dimensional feedback detection data, the differential reward value for each parameter dimension is calculated, and the differential reward value is used to update the network parameters of the hierarchical reinforcement learning optimization framework and the dynamic weight edges in the process knowledge graph.

[0015] By adopting the above technical solutions, a process map and real-time state model integrating domain knowledge were constructed. A hierarchical reinforcement learning framework was designed for decoupled decision-making and collaborative optimization, achieving closed-loop intelligent control of the entire smelting process. Deeply combining the data-driven advantages of artificial intelligence with the mechanistic models of the metallurgical field not only enables the handling of complex smelting processes with multiple variables, strong coupling, and large time delays, providing precise optimization instructions in real time, but also achieves dual self-evolution of system parameters and the knowledge base through closed-loop feedback. This improves the hit rate of smelting endpoints, the accuracy of composition control, and process stability, thereby shortening smelting time, reducing alloy consumption, and achieving cost reduction and efficiency improvement.

[0016] Secondly, this application provides an intelligent optimization system for smelting processes based on multi-source data fusion, employing the following technical solution:

[0017] The multi-source data acquisition module is used to acquire the target process parameters of the steel to be smelted, as well as real-time sensor data and production process data during the smelting process.

[0018] The process knowledge graph construction module is used to divide the target process parameters into a thermodynamic parameter set and a chemical composition parameter set according to the parameter characteristics, and to construct a thermodynamic constraint sub-graph and a chemical composition association sub-graph respectively, and then merge them to generate a process knowledge graph.

[0019] The smelting state space modeling module is used to preprocess and time-series align the real-time sensor data and production process data to construct a comprehensive smelting state space model for the current furnace.

[0020] The strategy network driving module is used to extract key state vectors from the integrated smelting state space model and input them into a pre-set hierarchical reinforcement learning optimization framework; wherein, the hierarchical reinforcement learning optimization framework includes a thermodynamic strategy network and a chemical composition strategy network.

[0021] The multi-strategy collaborative decision-making module is used to control the hierarchical reinforcement learning optimization framework to coordinate the thermodynamic strategy network and the chemical composition strategy network to make decoupled decisions based on the parameter priority and coupling relationship in the target process parameters, and generate a set of collaborative optimization actions.

[0022] The multi-objective constraint solving module is used to call the process knowledge graph to perform multi-objective constraint solving and conflict resolution on the collaborative optimization action set, and generate an executable optimization instruction set.

[0023] The execution feedback module is used to execute the executable optimized instruction set and synchronously collect multi-dimensional feedback detection data during the execution process;

[0024] The optimization and update module is used to calculate the differential reward value of each parameter dimension based on the multi-dimensional feedback detection data, and to update the network parameters of the hierarchical reinforcement learning optimization framework and the dynamic weight edges in the process knowledge graph using the differential reward value.

[0025] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution:

[0026] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect.

[0027] In summary, this application includes at least one of the following beneficial technical effects: By constructing a multi-domain coupled process knowledge graph and designing a hierarchical reinforcement learning framework, decoupled decision-making and collaborative optimization of thermodynamic and chemical composition control in the smelting process are achieved, effectively solving the dynamic control problem of multi-variable strong coupling and large time delay in complex smelting systems; at the same time, the constraint solution mechanism of the knowledge graph ensures the process safety and interpretability of the optimization instructions, and the closed-loop feedback mechanism drives the dual self-evolution of model parameters and knowledge base, improving the smelting endpoint hit rate, composition control accuracy and production stability, realizing the intelligent leap of smelting process from "experience-dependent" to "data and knowledge dual-driven", which has important practical application value for promoting cost reduction and efficiency improvement and intelligent manufacturing upgrade in the steel industry. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the first process of a smart optimization method for smelting process based on multi-source data fusion, which is one embodiment of this application.

[0029] Figure 2 This is a schematic diagram of the second process of a smart optimization method for smelting process based on multi-source data fusion, which is one embodiment of this application.

[0030] Figure 3 This is a schematic diagram of the third process of a smelting process intelligent optimization method based on multi-source data fusion according to one embodiment of this application.

[0031] Figure 4 This is a schematic diagram of the fourth process of a smart optimization method for smelting process based on multi-source data fusion, according to one embodiment of this application.

[0032] Figure 5 This is a schematic diagram of the fifth process of a smart optimization method for smelting process based on multi-source data fusion, according to one embodiment of this application.

[0033] Figure 6 This is a schematic diagram of the sixth process of a smelting process intelligent optimization method based on multi-source data fusion, which is one embodiment of this application.

[0034] Figure 7 This is a schematic diagram of the seventh process of a smart optimization method for smelting process based on multi-source data fusion, according to one embodiment of this application.

[0035] Figure 8 This is a schematic diagram of the eighth process of a smelting process intelligent optimization method based on multi-source data fusion according to one embodiment of this application. Detailed Implementation

[0036] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-8 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0037] This application discloses an intelligent optimization method for smelting processes based on multi-source data fusion.

[0038] Reference Figure 1 A method for intelligent optimization of smelting processes based on multi-source data fusion, specifically including:

[0039] Step S101: Obtain the target process parameters of the steel to be smelted, as well as real-time sensor data and production process data during the smelting process.

[0040] Among them, the target process parameters clearly define the various technical indicators that the steel to be smelted must ultimately achieve, such as the tapping temperature range, specific oxygen content, and precise composition of alloying elements such as carbon, silicon, and manganese. These parameters are the ultimate goals that the optimization process needs to approach.

[0041] Simultaneously, the system also needs to acquire real-time sensor data streams from the production site, including instantaneous information such as steel temperature and oxygen activity measured by devices like oxygen constant probes and temperature probes. This information reflects the rapidly changing physical and chemical state within the smelting furnace. Furthermore, production process data, such as the type and quantity of molten iron, scrap steel, and alloy materials input, oxygen blowing volume, and energizing time, provide the background and input conditions for the smelting operation.

[0042] By simultaneously acquiring three types of data—target (future expectations), current status (real-time status), and operation records (historical inputs)—essential raw materials are provided for building a digital model that can comprehensively and accurately describe the current smelting process.

[0043] Step S102: Based on the parameter characteristics, the target process parameters are divided into a thermodynamic parameter set and a chemical composition parameter set, and thermodynamic constraint sub-maps and chemical composition correlation sub-maps are constructed respectively, and then fused to generate a process knowledge graph.

[0044] In metallurgical physical chemistry, although thermodynamic parameters (such as temperature and oxygen activity) and chemical composition parameters (such as the content of alloying elements) influence each other, their control mechanisms, response rates, and constraints are fundamentally different. Thermodynamic parameters such as temperature and oxygen activity directly affect the driving force and rate of the reaction, and their constraints are often manifested as safety boundaries (such as the superheat range and the upper limit of oxygen activity to prevent over-oxidation) and dynamic response relationships.

[0045] In this embodiment, the thermodynamic constraint sub-graph is constructed by representing these parameters with nodes and defining the nonlinear constraint relationships between them with edges (connections), such as "increased temperature will exacerbate the oxidation and burn-off of certain elements". On the other hand, the chemical composition correlation sub-graph focuses on the complex interactions between alloying elements, such as the competitive oxidation of silicon and oxygen, and the combination of manganese and sulfur. Its edge weights can represent the yield influence coefficient or reaction coupling strength between elements.

[0046] Finally, these two sub-graphs are fused using the laws of conservation of matter and energy to generate a multi-domain coupled process knowledge graph. This graph is essentially a dynamic knowledge base containing rich domain rules and relationships, which makes subsequent optimization decisions no longer black-box computations, but rather a reasoning process based on process principles.

[0047] Step S103: Preprocess and time-series align the real-time sensor data and production process data to construct a comprehensive smelting state space model for the current furnace batch.

[0048] Raw data collected directly from the field often suffers from issues such as noise, inconsistent sampling frequencies of different sensors, and critical time lag characteristics. For example, after adding alloys or adjusting power, the uniform change in molten steel temperature requires a certain amount of time, resulting in a lag in the signals detected by the sensors.

[0049] Therefore, the purpose of preprocessing and time alignment is to unify all data onto the same accurate time base and compensate for known time delays through algorithms, thereby obtaining state information that can truly reflect the overall furnace condition at the current moment.

[0050] Next, a comprehensive smelting state-space model is constructed. This involves integrating the cleaned, aligned, and time-delay-compensated multi-dimensional data (such as compensated real-time temperature, oxygen activity, dynamically calculated heating rate, deoxidation rate, and known raw material ratios) into a high-dimensional, complete mathematical representation. This model is no longer an isolated data point, but a state vector that can describe the dynamic evolution of the smelting system. It provides accurate and reliable on-site situation reports for artificial intelligence decision-making networks.

[0051] Step S104: Extract key state vectors from the integrated smelting state space model and input them into a pre-set hierarchical reinforcement learning optimization framework; wherein, the hierarchical reinforcement learning optimization framework includes a thermodynamic policy network and a chemical composition policy network.

[0052] Specifically, extracting key state vectors from a complex state-space model is a process of feature dimensionality reduction and core information extraction. It filters out redundant and interfering information that could hinder the current decision-making process, while retaining the most relevant features, such as the deviation between the current and target temperatures, the real-time values ​​of key element components, and the activity index of the melt pool. This refined vector is then input into a specially designed hierarchical reinforcement learning optimization framework.

[0053] In this embodiment, the reinforcement learning optimization framework adopts a hierarchical structure, with a coordinator at the top and two parallel dedicated policy networks at the bottom: a thermodynamic policy network and a chemical composition policy network. This design stems from the inherent requirements of metallurgical process control—thermodynamic control (such as power supply and oxygen blowing) and chemical composition control (such as alloy addition) are two closely coupled tasks with different decision frequencies and objects. Allowing the two specialized networks to handle these tasks separately enables more efficient learning of optimal policies for their respective domains. The thermodynamic policy network focuses on learning how to adjust energy input to control temperature and oxygen level; the chemical composition policy network focuses on learning how to accurately calculate and add alloy materials to achieve the target composition.

[0054] Step S105: The hierarchical reinforcement learning optimization framework coordinates the thermodynamic strategy network and the chemical composition strategy network to make decoupled decisions based on the parameter priority and coupling relationship in the target process parameters, and generates a set of collaborative optimization actions.

[0055] In this process, the initial actions generated independently by the two strategy networks (such as "increase power by 100 kW" or "add 200 kg of ferrosilicon") may conflict. For example, the thermodynamic network might decide to significantly increase the temperature to meet process requirements, but this action could lead to a significant decrease in the yield of certain easily oxidized elements in the chemical composition network. Therefore, a coordination mechanism is needed to arbitrate and adjust based on the priority of target parameters (e.g., some steel grades have extremely high requirements for sulfur content, which takes precedence over temperature control accuracy) and known coupling relationships (i.e., the mutual influence rules defined in the knowledge graph).

[0056] Specifically, the purpose of decoupling decision-making is to minimize or compensate for the interference of one control action on another control loop. The coordinator evaluates the impact factors of thermodynamic actions on chemical composition control and adjusts chemical composition actions in reverse, or adjusts the timing of actions, ultimately outputting a globally coordinated set of co-optimized actions that can simultaneously approach the thermodynamic and chemical composition goals, thereby solving the multivariate coupling problem in traditional control.

[0057] Step S106: Call the process knowledge graph to solve multi-objective constraints and resolve conflicts in the collaborative optimization action set, and generate an executable optimization instruction set;

[0058] Even with coordination, the set of collaborative actions generated by a data-driven AI network may still violate basic process safety rules or physicochemical limits. In such cases, a process knowledge graph needs to be invoked for final verification and correction.

[0059] Specifically, the process knowledge graph will perform multi-objective constraint solving on the action set, checking whether it triggers the safety boundaries defined in the graph (such as the temperature exceeding the limit of the refractory material) or mutually exclusive reactions (such as the simultaneous addition of elements that will form harmful inclusions). If a conflict is found, the system will not simply reject it, but will search for an adjustment scheme that is closest to the original optimization intention and satisfies all hard constraints within the rule space defined by the knowledge graph, based on principles such as Pareto optimality. This process is conflict resolution. Finally, the output set of executable optimization instructions is a set of operation commands that possesses AI optimization intelligence and strictly adheres to process safety and scientific principles, and can be directly issued to the basic automation system for execution.

[0060] Step S107: Execute the executable optimized instruction set and simultaneously collect multi-dimensional feedback detection data during the execution process;

[0061] The optimized instruction set is sent to the actuators (such as motors, valves, and feeding systems) for implementation. Simultaneously, the system initiates high-frequency monitoring, using various online sensors to synchronously collect multi-dimensional feedback data of the molten steel after instruction execution. This includes new temperature, new oxygen activity, and new composition analysis values. This feedback data accurately records the actual smelting process's response to the system's actions, forming a complete "decision-execution-feedback" closed loop, which is the sole basis for evaluating the optimization effect and training the iterative evolution of the model.

[0062] Step S108: Calculate the differential reward value for each parameter dimension based on the multi-dimensional feedback detection data, and use the differential reward value to update the network parameters of the hierarchical reinforcement learning optimization framework and the dynamic weight edges in the process knowledge graph.

[0063] The system compares the feedback data with the target value and calculates the differential reward value for each key parameter dimension (such as temperature deviation and composition deviation). This reward value is a quantified signal; a positive value indicates that the action achieved good results (close to the target), while a negative value indicates poor results. In the reinforcement learning framework, this reward signal is used through the backpropagation algorithm to update the internal parameters of the thermodynamic policy network and the chemical composition policy network, respectively, so that their future decisions can obtain higher cumulative rewards, i.e., better smelting results.

[0064] Meanwhile, the optimized instance (including deviations in initial state, executed actions, and final result) was also used to update the dynamic weight edges in the process knowledge graph. For example, if, during this smelting process, under certain temperature conditions, the actual yield of a certain alloy differs significantly from the preset weight value in the graph, the system will fine-tune that weight, making the correlation between "temperature-element yield" in the graph more closely reflect the actual production patterns of the current factory. This allows the entire system not only for its AI model to evolve, but also for its built-in process knowledge base to continuously self-correct and improve with the accumulation of production data, becoming increasingly adaptable to specific production lines and steel grades.

[0065] In the above implementation, a process map and real-time state model integrating domain knowledge are constructed, and a hierarchical reinforcement learning framework is designed for decoupled decision-making and collaborative optimization, realizing closed-loop intelligent control of the entire smelting process. By deeply combining the data-driven advantages of artificial intelligence with the mechanistic models of the metallurgical field, it can not only handle complex smelting processes with multiple variables, strong coupling, and large time delays, providing precise optimization instructions in real time, but also achieve dual self-evolution of system parameters and the knowledge base through closed-loop feedback. This improves the hit rate of smelting endpoints, the accuracy of composition control, and process stability, thereby shortening smelting time, reducing alloy consumption, and achieving cost reduction and efficiency improvement.

[0066] Reference Figure 2As one implementation of step S102, the steps of dividing the target process parameters into a thermodynamic parameter set and a chemical composition parameter set according to the parameter characteristics, constructing a thermodynamic constraint sub-map and a chemical composition correlation sub-map respectively, and fusing them to generate a process knowledge graph include:

[0067] Step S201: Based on the physicochemical properties of the target process parameters, parameters involving changes in energy state are classified into a thermodynamic parameter set, and parameters involving changes in material composition are classified into a chemical composition parameter set.

[0068] Although the smelting process is a unified physicochemical system, its process parameters can be divided into two interrelated yet relatively independent categories based on their inherent properties and the dimensions of their impact on the system. "Thermodynamic parameters" are essentially macroscopic intensive properties describing the energy state of the system, with temperature and oxygen activity being the most typical and core representatives. Temperature directly characterizes the system's thermal energy level and is the driving factor for all chemical reaction rates and equilibrium; oxygen activity characterizes the redox potential of the molten pool, determining the trend and limit of oxidation or reduction of each element. These parameters are characterized by rapid dynamic changes, directly affecting process kinetics, and are often used as "environmental variables" or "boundary conditions" for process control. Classifying them into a set of thermodynamic parameters means that subsequent modeling will focus on characterizing their interrelationships (such as the nonlinear coupling of temperature and oxygen activity) and their constraints on the entire reaction process.

[0069] Conversely, "chemical composition parameters" describe the material composition of the system, namely the mass fraction or concentration of various alloying elements (such as C, Si, Mn, P, S, etc.). These are the material goals that the smelting process ultimately needs to achieve, and changes in these parameters are a direct result of chemical reactions and mass transport. By classifying them into a "set of chemical composition parameters," it means that subsequent modeling will focus on characterizing the chemical reactions (such as oxidation, reduction, and combination) that occur between different elements, as well as the coupling relationships between these reactions, such as competition, promotion, or inhibition.

[0070] Understandably, this division based on the essence of physicochemical processes achieves multi-dimensional decoupling of complex smelting processes, laying the foundation for the subsequent construction of specialized knowledge models focusing on "energy flow" and "material flow" respectively.

[0071] Step S202: Based on the thermodynamic parameter set, define the nonlinear constraint relationship between temperature and oxygen activity, and generate a thermodynamic constraint sub-map.

[0072] The core of this step lies in identifying and formally defining the nonlinear constraint relationship between the two core thermodynamic parameters of temperature and oxygen activity.

[0073] Specifically, in metallurgical physical chemistry, temperature and oxygen activity are not independent changes; they are closely related through the equilibrium constants of a series of chemical reactions (especially redox reactions of elements). For example, at a specific temperature, the oxygen activity (i.e., the deoxidizing capacity of that element) in equilibrium is a definite value. In the steelmaking process, operators need to control the molten steel within a suitable "temperature-oxygen activity" window: too low a temperature may cause the molten steel to become viscous or even solidify, while too high a temperature wastes energy and damages the furnace lining; excessively high oxygen activity leads to excessive burning of alloying elements and deterioration of steel quality, while too low an oxygen activity may hinder reactions such as dephosphorization.

[0074] In this embodiment, generating a thermodynamic constraint subgraph involves transforming these complex, nonlinear scientific laws and technological experiences into a computable network composed of nodes and edges. In this subgraph, nodes represent "temperature" and "oxygen activity," while the "edges" connecting them are assigned weights and rules to quantify the constraints between them. For example, "within the target temperature range T1-T2, the corresponding allowable oxygen activity range is a_O1-a_O2," or "when the temperature deviates from the baseline value ΔT, in order to maintain the same deoxidation effect, the minimum aluminum content required needs to be compensated according to a certain nonlinear function."

[0075] Step S203: Based on the chemical composition parameter set, define the reaction coupling relationship between alloying elements and generate a chemical composition correlation sub-map;

[0076] The core of this step lies in defining the reactive coupling relationships between alloying elements. Molten steel typically contains multiple alloying elements, which do not exist in isolation but rather within a complex chemical reaction network. These relationships mainly include:

[0077] 1. Competition: Multiple elements compete for the same reactant. The most typical example is the deoxidation process, where deoxidizing elements such as aluminum, silicon, and manganese react with oxygen simultaneously. The order and extent of the reaction are determined by the strength of their deoxidation abilities (i.e., their affinity for oxygen). Adding a strong deoxidizer (such as aluminum) can significantly change the oxidation loss of a weak deoxidizer (such as silicon).

[0078] 2. Chemical relationships: Elements combine with each other to form compounds. For example, manganese combines with sulfur to form MnS, which can fix the sulfur element and improve the hot brittleness of steel.

[0079] 3. Interrelationships: The presence of one element can alter the activity or reaction rate of another element. For example, a high carbon content in molten steel reduces the activity of oxygen; this is the effect of the "carbon-oxygen reaction" on the overall oxygen potential.

[0080] In this embodiment, generating a chemical composition correlation sub-map involves formally modeling these intricate chemical reaction networks. In this sub-map, nodes represent various alloying elements (C, Si, Mn, Al, O, etc.), and the "edges" connecting the nodes represent specific reaction coupling relationships between them. Each edge can be assigned attributes such as reaction type (competitive / combining), coupling strength coefficient, or yield influence factors based on stoichiometry and equilibrium calculations (e.g., how many kilograms of oxygen are consumed and how many kilograms of silicon might be "protected" from oxidation when 1 kilogram of aluminum is added for deoxidation). This sub-map allows the system to understand that adjusting the amount of any alloy added not only affects the content of that element itself but also ripples through this reaction network, impacting the final results of other elements.

[0081] Step S204: The thermodynamic constraint sub-graph and the chemical composition correlation sub-graph are connected by edges using the law of conservation of matter and energy, and then fused to generate a process knowledge graph.

[0082] Among them, the law of conservation of matter and energy (i.e., the law of conservation of mass and the law of conservation of energy) serves as the bridge and basis connecting these two sub-maps. Specifically, this connection is reflected in:

[0083] 1. Energy changes drive material transformation: Changes in temperature / oxygen activity (energy state) in the thermodynamic submap directly affect the equilibrium position and reaction rate of each reaction in the chemical submap. For example, by connecting edges, the event of "temperature increase" can be mapped to a change in the equilibrium constant of multiple oxidation reactions in the chemical submap (which is usually unfavorable for exothermic deoxygenation reactions), thereby predicting a general downward trend in elemental yield.

[0084] 2. Material transformation is accompanied by energy changes: Any alloy addition or elemental reaction (change in material composition) occurring in the chemical sub-map will cause a change in the system's enthalpy in the thermodynamic sub-map. For example, adding a large amount of cold alloy material will cause the molten pool temperature to drop (endothermic); the exothermic oxidation of aluminum will cause the temperature to rise.

[0085] In the embodiments of this application, such connections can be established quantitatively or qualitatively through the law of conservation of matter and energy. Generating a process knowledge graph involves establishing cross-domain connections between nodes of two sub-graphs based on these conservation laws and known physicochemical principles. For example, an edge can be established between the "temperature" node and the "aluminum element" node, with the rule that "the oxidation of aluminum is a strongly exothermic reaction; oxidizing 1 kg of aluminum will raise the temperature of molten steel by approximately X degrees."

[0086] Based on this, the two previously independent sub-graphs were merged into a multi-domain coupled process knowledge graph. This unified graph enables the system to perform complex associative reasoning. For example, when the system performs a power supply operation to increase temperature, it can automatically infer the possible changes in oxygen activity caused by this operation, and then deduce the potential impact on the yield of various elements through the chemical network, thus providing a comprehensive basis for subsequent intelligent decision-making that takes into account both "heat" and "quality".

[0087] In the above implementation, based on the parameter decoupling of the physicochemical essence, thermodynamic constraint and chemical correlation sub-graphs are constructed respectively, realizing the accurate characterization of the "energy flow" and "material flow" in the smelting process. Then, through the universal law of conservation of matter and energy, the two are deeply integrated to form a unified knowledge graph that can reflect the multivariable, strongly coupled, and nonlinear nature of the metallurgical process. This makes the system decision not only based on data-driven, but also rooted in the profound process mechanism, so as to more scientifically predict the multiple consequences of control actions and more effectively coordinate the contradiction between thermodynamic and chemical composition control. Ultimately, it lays the knowledge foundation for realizing high-precision, high-stability, and adaptive intelligent steelmaking.

[0088] Reference Figure 3 As a further implementation of the intelligent optimization method for smelting processes, after connecting the thermodynamic constraint sub-graph and the chemical composition correlation sub-graph through the law of conservation of matter and energy in step S204 to generate a process knowledge graph, the method further includes:

[0089] Step S301: Obtain a pre-set local expert domain knowledge base, including element reaction coupling relationships and security boundary parameters;

[0090] Among them, the local expert domain knowledge base is a structured knowledge collection that is built offline and stored locally in the factory. Its local attribute ensures the sovereignty and security of core process knowledge.

[0091] Specifically, the element-reaction coupling relationship is a formalized expression of the core laws of metallurgical physicochemistry. Examples include the deoxidation order (the order of the strength of each element's affinity for oxygen), the interaction coefficients between elements (such as the effect of aluminum's oxygen activity on the deoxidation effect of silicon), and harmful reactions that may occur under specific conditions (such as the effect of aluminum reacting with nitrogen to form AlN on steel properties). These relationships go beyond simple statistical correlations and reveal the essential connections and quantitative laws of chemical reactions.

[0092] Safety boundary parameters are rigid constraints defined from the perspectives of equipment safety, process safety, and product quality safety. Examples include the maximum withstand temperature of the converter lining, the ultimate pressure of the vacuum vessel, and the upper limits of phosphorus and sulfur content that must be controlled to prevent hot or cold brittleness in steel. Acquiring this knowledge base means that the system inherits valuable resources accumulated by experts from the outset, in the form of formulas, rules, and empirical values. This ensures that subsequent intelligent modeling and decision-making will not blindly explore without any prior knowledge, but will be constrained from the beginning within a scientifically sound and controllable domain.

[0093] Step S302: Map the element reaction coupling relationship to the association edge of the chemical composition association subgraph in the process knowledge graph, and map the safety boundary parameter to the constraint node of the thermodynamic constraint subgraph in the process knowledge graph.

[0094] Specifically, while statistical correlations between elements can be obtained through data mining in chemical composition correlation sub-maps, the semantics and strength of these "edges" may be ambiguous. Mapping element-reaction coupling relationships onto these correlation edges injects explicit chemical meaning and quantitative parameters into them. For example, mapping the "silicon-oxygen" coupling relationship (including the reaction equation, equilibrium constant, and temperature relationship) onto the corresponding edge means that the edge no longer merely represents the covariance of the two values, but can support a miniature chemical reactor model, allowing calculation of the oxygen activity required to achieve the target silicon content at a specific temperature. Similarly, mapping safety boundary parameters onto the constraint nodes (such as temperature nodes) of thermodynamic constraint sub-maps adds an insurmountable threshold attribute to these nodes.

[0095] Step S303: Based on the mapping results, update the initial connection strength of the dynamic weight edges of the process knowledge graph to generate the updated process knowledge graph.

[0096] The initial dynamic weights may be based on general data or historical averages, which may not accurately reflect the true importance of specific process relationships for the current production line and steel grade. The provided expert rules serve as a highly reliable reference benchmark, allowing for the evaluation and adjustment of these weights.

[0097] For example, if expert knowledge strongly indicates that the precipitation of compounds of the microalloying element niobium and carbon is crucial to performance during the smelting of a certain high-strength steel (i.e., the weight of the "niobium-carbon" coupling relationship should be high), and the initial weight is low, the system will significantly increase the strength of that dynamic weight edge according to this rule. Conversely, the weight of some related edges that are irrelevant under the current smelting conditions may be decreased.

[0098] Ultimately, generating an updated process knowledge graph means that the overall topology and relationship confidence of the graph have been optimized in a targeted manner based on domain knowledge. This allows the graph to have an internal state that is closer to the actual needs of the current optimization task from the outset, shortening the exploration time required for subsequent pure data-driven learning and improving the decision quality and reliability of the system in the initial stage.

[0099] In the above implementation, the dispersed expert experience is systematically encoded and integrated into a computable knowledge graph, which realizes the effective injection of prior knowledge, enhances the graph's ability to represent complex coupling relationships and safety constraints in smelting, and enables the system's intelligent decision-making ability to combine the forward-looking nature of data-driven approaches with the reliability of mechanistic models, providing a highly reliable decision-making basis for subsequent intelligent optimization.

[0100] Reference Figure 4 As one implementation of step S103, the steps of preprocessing and aligning real-time sensor data and production process data to construct a comprehensive smelting state-space model for the current furnace batch include:

[0101] Step S401: Identify the acquisition timestamps of each detection parameter in the real-time sensor data, and establish a time delay compensation mapping relationship based on the response delay characteristics of the smelting equipment;

[0102] In the smelting process, the sampling frequency, installation location, and signal transmission and processing links of different sensors vary, resulting in accurate timestamps of data arrival at the system, but the physical events they reflect are not synchronized in real time. Furthermore, the smelting process inherently involves equipment response delays. For example, it takes several seconds for a temperature probe to be inserted into molten steel until its thermocouple reaches thermal equilibrium and outputs a stable reading; similarly, when the electric arc furnace power setting changes, there is a significant physical lag in the change of molten steel temperature. This means that the temperature data collected at time t may actually reflect the furnace state at time t-Δt.

[0103] Therefore, in this embodiment, the timestamps of each data stream are first identified. Then, based on prior knowledge or historical data learning of the physical characteristics of specific devices (such as specific models of sensors or heating systems), a time-delay compensation mapping relationship is established. This relationship can be a time offset (e.g., temperature data needs to be compensated forward by 3 seconds) or a dynamic transfer function model. The purpose of establishing this mapping is to compensate all observed data to the same time reference on which the theoretically represented real physical events occur. This is an absolute prerequisite for any meaningful multivariate correlation analysis and state estimation that follows.

[0104] Step S402: Align the temperature data stream and oxygen activity data stream with the time axis according to the time delay compensation mapping relationship to generate a synchronized sensor data sequence;

[0105] Temperature is the concentrated manifestation of energy state, and oxygen activity is the decisive factor in the direction and extent of metallurgical reactions. The two are the most critical dynamic variables describing the smelting process. Their synchronicity is crucial in metallurgical thermodynamics because any meaningful analysis (such as determining deoxidation equilibrium or calculating element activity) must be based on the same instantaneous temperature and oxygen activity values.

[0106] In this embodiment, the established time-delay compensation mapping relationship is used to perform precise time axis alignment on the two data streams. For example, the timestamps of the oxygen activity data are uniformly compensated to reflect the same physical moment as the temperature data. The resulting synchronized sensor data sequence ensures that each data point pair (temperature, oxygen activity) is strictly paired in the time dimension, truly representing the thermodynamic state of the molten pool at a specific moment.

[0107] Step S403: Extract dynamic process characteristic quantities from the synchronized sensor data sequence, including heating rate gradient, alloy element diffusion rate and deoxidation reaction efficiency.

[0108] Among them, the synchronized instantaneous data sequence (temperature, oxygen activity) only tells "what the current value is", while the dynamic process characteristics reveal "how the system is changing" and "how efficient the change is", which is the core of assessing process health, predicting future trends and carrying out fine control.

[0109] Specifically, the heating rate gradient is obtained by calculating the first or even second derivative of temperature with respect to time. It not only reflects the effectiveness of the current heating power, but its gradient change can also provide early warning of nonlinear phenomena such as "insufficient heating" or "overheating risk." The diffusion rate of alloying elements usually requires indirect measurement or model estimation (e.g., through time series analysis of compositional spectral data, or by back-calculation based on mass transfer models using parameters such as temperature and stirring energy). It characterizes how quickly the added alloy mixes in the molten pool and is directly related to the uniformity and hit rate of composition control. The deoxidation reaction efficiency is a comprehensive indicator that can be evaluated by calculating the decrease in oxygen activity per unit time and per unit deoxidizer consumption. It comprehensively reflects the kinetic efficiency under the combined effect of multiple factors such as deoxidizer type, molten pool temperature, and stirring conditions.

[0110] Step S404: Perform multidimensional tensor fusion on the synchronized sensor data sequence, dynamic process feature quantities, and raw material ratio information in the production process data;

[0111] This process organically integrates data from different dimensions, with varying physical meanings and dimensions, using a multidimensional tensor mathematical structure. Synchronized sensor data sequences provide dynamic observations with high temporal resolution; dynamic process characteristic quantities offer in-depth interpretations and derived information from the observed data; and raw material proportioning information in the production process data (such as the amount and composition of molten iron, scrap steel, and various alloys) defines the initial conditions and material basis of this smelting process, serving as the background and boundary for understanding all subsequent reactions.

[0112] Understandably, simple data splicing loses complex structural relationships, while tensors (which can be viewed as high-order matrices) are well-suited for representing such multi-dimensional, multi-feature data cubes. For example, the three dimensions of a three-dimensional tensor can be: time step, feature type (original observations and dynamic features), and feature channels (temperature, oxygen activity, various alloying elements, etc.). Through tensor fusion, the system obtains a complete, structured data object that simultaneously preserves temporal evolution, feature correlations, and batch initial conditions.

[0113] Step S405: Construct a smelting state space coordinate system containing time-varying characteristics based on the tensor dimensionality reduction algorithm, and map the feature vector of the multidimensional tensor in the smelting state space coordinate system.

[0114] This method utilizes tensor dimensionality reduction algorithms (such as high-order principal component analysis and tensor decomposition) to extract a few latent variables that best characterize the essential features of the smelting process from high-dimensional, complex fused tensors. Based on these latent variables, a low-dimensional smelting state space coordinate system is constructed. Each dimension of this coordinate system represents a comprehensive factor with clear physical or technological significance, such as "comprehensive thermodynamic potential," "chemical redox trend," and "degree of homogenization of the molten pool."

[0115] It should be noted that the time-varying nature of the smelting state space coordinate system is reflected in the fact that this coordinate system is not static; its basis (i.e., the principal component directions) can adapt to changes in process characteristics under different steel grades and raw material conditions through online learning. Projecting the multidimensional tensor data of the current heat into this coordinate system yields a low-dimensional feature vector. This feature vector, with its coordinate position in the state space, concisely encodes the comprehensive state of the current heat across all important dimensions.

[0116] Step S406: Generate the comprehensive smelting state space model for the current furnace based on the topological relationship of the feature vector distribution.

[0117] Specifically, the feature vector of the current furnace is placed within a "state point cloud" or "state map" composed of feature vectors from historical normal furnaces and various typical abnormal furnaces (such as over-oxidation, temperature anomalies, and out-of-range composition). By analyzing the position of the current feature vector within this historical state distribution topology (e.g., which typical "good smelting trajectory" cluster it is closer to, or whether it has begun to deviate from the normal region and move towards an abnormal region), a state model with relativity and interpretability is generated.

[0118] In this embodiment of the application, the model not only includes the absolute coordinates of the current state, but also the qualitative judgments of its "health level", "risk level" and "trend" relative to historical experience. For example, the model can be described as "the current state is in the excellent zone, but is moving towards the boundary of the high temperature and high oxygen zone at a relatively fast speed".

[0119] In the above implementation, the original isolated data stream is transformed into a synchronous, structured tensor rich in deep dynamic information and incorporating the initial conditions of the process. Then, through intelligent dimensionality reduction, a low-dimensional, interpretable, and time-varying state space that reflects the essence of the process is constructed. Finally, a comprehensive smelting state space model is generated that can accurately locate the current state and reveal its relative position and trend in the historical experience map, providing high-fidelity and high-information-density input for subsequent precise decision-making and control based on artificial intelligence.

[0120] Reference Figure 5 As one implementation of step S105, the hierarchical reinforcement learning optimization framework, based on the parameter priorities and coupling relationships in the target process parameters, coordinates the thermodynamic strategy network and the chemical composition strategy network to make decoupled decisions and generate a collaborative optimization action set, including the following steps:

[0121] Step S501: Analyze the priority weights of thermodynamic parameters and chemical composition parameters in the target process parameters, and generate a parameter decision weight matrix;

[0122] The smelting process is a complex system with multiple variables coupled. Different steel grades have significantly different sensitivities to temperature and composition control. For example, in the early stage of refining, temperature adjustment often takes precedence over composition fine-tuning to ensure the activity of the molten pool, while in the late stage of alloying, precise composition control becomes dominant.

[0123] In one embodiment of this application, dynamic priority weights can be assigned to different thermodynamic parameters (such as endpoint temperature and oxygen activity) and chemical composition parameters (such as C, Si, and Mn content) based on production needs and steel characteristics. These weights may stem from various considerations: when quality is prioritized, the precision weight of key components is highest; when cost is prioritized, the weight of reducing alloy consumption and shortening smelting time is increased; when safety is prioritized, the safety boundary constraints of temperature and oxygen potential have the greatest weight.

[0124] The system analyzes target process parameters and extracts preset priority weights for thermodynamic parameters (such as temperature and oxygen activity) and chemical composition parameters (such as alloy element content) to construct a parameter decision weight matrix. This matrix is ​​not merely a static set of weights, but a dynamic decision operator. It quantifies the relative importance of different process objectives within the same decision time slice, providing decision guidance for the generation of subsequent action sequences and ensuring that the optimization strategy allocates resources according to the primary and secondary contradictions of the current smelting stage.

[0125] Step S502: Prioritize the thermodynamic action sequence output by the thermodynamic strategy network according to the parameter decision weight matrix to determine the execution order of the basic power adjustment action and the basic deoxygenation action.

[0126] The action sequence generated by the thermodynamic strategy network based on the current state may include a variety of operations, such as power boosting, argon blowing and stirring, or deoxidizer addition. However, if these operations are executed randomly, they may interfere with each other.

[0127] Therefore, by introducing a parametric decision weight matrix, the system prioritizes the action sequence. This process is essentially a combinatorial optimization problem of finding the optimal action execution order. The prioritization criteria include the rate of impact of the operation on the physical state of the molten pool, the reversibility of the operation, and safety constraints, thereby determining the order of the basic power adjustment action and the basic deoxidation action. This ensures that the thermodynamic control layer can approach the target state with the most reasonable logical path, avoiding temperature fluctuations or energy waste caused by disordered operation sequences.

[0128] In some embodiments, if the current temperature is significantly low and the temperature weight is high, then all actions that can effectively and quickly raise the temperature (such as increasing power) will be given high priority; if the oxygen activity is close to the safe limit, then the priority of deoxygenation actions will immediately jump.

[0129] Step S503: Based on the element reaction coupling relationship in the chemical composition correlation sub-map, perform conflict labeling on the chemical composition action sequence output by the chemical composition strategy network to obtain the conflict labeling result;

[0130] Specifically, in complex alloying processes, there are differences in redox potential, density, and the possibility of mutual reactions among different alloying elements (for example, adding aluminum first for deoxidation may affect the yield of subsequently added silicon and manganese elements).

[0131] In this embodiment, the chemical composition correlation sub-map is a knowledge representation tool that structures the rules of chemical reactions, mutual exclusion relationships, and synergistic effects between elements. The system traverses this map, mapping the action sequences output by the chemical composition strategy network onto map nodes, and detects whether there are operation combinations that violate the element reaction coupling relationships (such as adding an easily oxidizable element under a specific oxygen potential). If a potential conflict is detected, the action is marked as conflicting. This process enables early warning and elimination of unreasonable composition control behaviors, preventing material waste or compositional deviations caused by chemical incompatibilities.

[0132] For example, the graph shows that elements A and B compete for oxygen in molten steel. If the action sequence simultaneously includes "adding a large amount of element A" and "adding a large amount of element B" to increase the content of A and B respectively, this could be a conflict. The element added first would consume a large amount of oxygen, severely impacting the yield of the element added later, leading to unmet goals and waste. This step identifies and flags such potential conflicts based on chemical reaction mechanisms, indicating that these action combinations may not produce the expected synergistic effect or even cause side effects. This provides early warning of operational risks due to insufficient understanding of the chemical network.

[0133] Step S504: Call the nonlinear constraint relationship in the thermodynamic constraint sub-map, calculate the interference influence factor of the priority-ranked thermodynamic action sequence on the action sequence of the chemical components with conflict, and perform feasibility verification by combining the interference influence factor with the conflict marking results.

[0134] In the smelting process, thermodynamic and chemical parameters are strongly coupled. For example, a large adjustment of power (thermodynamic action) will change the temperature of the molten pool and the oxygen potential, which will significantly affect the yield of alloying elements (chemical composition action).

[0135] In this embodiment, the thermodynamic constraint submap stores a nonlinear mapping model between temperature, oxygen potential, and elemental reaction kinetics. The system uses this model to calculate the interference factor of the thermodynamic action sequence on the chemical composition actions. This factor quantifies the degree to which thermodynamic operations disrupt chemical equilibrium. Combined with previous conflict labeling results, the system performs a comprehensive feasibility check to determine whether the current thermodynamic-chemical joint action set will lead to runaway process parameters, thereby identifying cross-domain interference risks before execution.

[0136] For example, the chemical composition of a "deep desulfurization" process (such as the addition of calcium carbide) is highly dependent on high temperature and high alkalinity for its reaction efficiency. If the rules of the thermodynamic constraint submap indicate that the efficiency of the desulfurization reaction is extremely low at the current temperature and oxygen availability (i.e., it does not meet the feasibility constraint), then the process is judged as "infeasible" or "inefficient" at that moment.

[0137] Step S505: If the interference impact factor exceeds the preset threshold or the feasibility verification fails, the output of the chemical component strategy network is adjusted using the attention mechanism to generate the verified chemical component action sequence.

[0138] When the verification finds that the thermodynamic action seriously interferes with the chemical control (the impact factor exceeds the threshold) or there is an infeasible conflict, the system does not directly reject the action, but activates the attention mechanism.

[0139] In this embodiment, the attention mechanism can focus on key feature dimensions that cause conflict (such as specific oxygen activity ranges or temperature-sensitive regions), and make weighted adjustments to the hidden layer states of the chemical composition strategy network, forcing the network to regenerate actions. By adjusting the network's focus, a modified chemical composition action sequence that can adapt to thermodynamic disturbances is generated, thus realizing dynamic decoupling and adaptive coordination between thermodynamic and chemical composition control.

[0140] In addition, if the interference impact factor does not exceed the threshold or the feasibility verification is passed, it can be directly output as the verified chemical component action sequence.

[0141] Step S506: Match the verified chemical composition action sequence with the priority-sorted thermodynamic action sequence in the spatiotemporal dimension to generate an initial cooperative action set;

[0142] Specifically, smelting operations have distinct spatiotemporal attributes, with different operations requiring different time windows and equipment resources (space). The system aligns the corrected chemical and thermodynamic actions on the time axis and matches them with their corresponding execution equipment units, constructing a multi-dimensional initial set of coordinated actions. This set contains the thermodynamic and chemical operation instructions that should be executed at each moment, initially forming a blueprint for coordinated control of the entire process, ensuring the self-consistency of various operations in logical timing.

[0143] In the embodiments of this application, the matching operation first needs to consider timing matching: certain chemical composition actions need to be carried out under specific thermodynamic conditions. For example, the addition of certain easily oxidized precious alloys usually needs to be carried out during a temperature plateau period with good deoxidation and low oxygen activity. Therefore, it needs to be embedded into the time window created by the thermodynamic action of "completing deoxidation and reaching a suitable temperature". Secondly, resource and space matching also needs to be considered: ensuring that the execution of actions does not conflict physically. For example, the use of the feeding chute needs to be staggered with the temperature measurement and sampling time.

[0144] Step S507: Based on the physical execution window constraints of the smelting process, the initial collaborative action set is rearranged in time to generate a collaborative optimized action set.

[0145] The initial set of collaborative actions is logically feasible, but it must comply with the physical limitations of the actual equipment, such as the feeding frequency and capacity limitations of the feeding system, the stabilization time and operating time required for temperature sampling, the mechanical response time for electrode lifting or oxygen lance movement, and the cycle required for switching between different workstations. Based on these hardware physical boundaries, the system fine-tunes and rearranges the timing of the initial action set, removing or shifting actions that exceed the physical window limits, and finally generating an executable set of collaboratively optimized actions. This ensures that the intelligent optimization decision is not only theoretically optimal but also engineering-executable.

[0146] For example, two closely spaced small batches of material can be combined into a single batch to reduce the number of hopper switching operations; or a temperature measurement operation can be slightly advanced to avoid the expected intense boiling period and ensure measurement accuracy. This process acts like a scheduler, further adapting the operation sequence to the rhythm and capacity of the actual production line while satisfying all process logic.

[0147] In the above implementation, a rigorous hierarchical reinforcement learning optimization logic was constructed, which effectively solved the problem of strong coupling interference between thermodynamic control and chemical composition control in the smelting process. It overcame the shortcomings of traditional methods that ignored element reaction conflicts and cross-domain influences, realized intelligent decoupling and precise coordination of multiple process parameters, and improved the scientific nature, safety and final product quality of smelting decisions.

[0148] Reference Figure 6 As a further implementation of the intelligent optimization method for smelting processes, before step S105, where the hierarchical reinforcement learning optimization framework coordinates the thermodynamic strategy network and the chemical composition strategy network to make decoupled decisions based on the parameter priorities and coupling relationships in the target process parameters, and generates a collaborative optimization action set, the method further includes:

[0149] Step S601: Input the key state vector into the preset working condition identification classifier to identify the smelting stage type of the current furnace and generate smelting stage type labels, including melting period, oxidation period and reduction period.

[0150] The underlying logic of this step lies in addressing the impact of the time-varying and phased characteristics of the smelting process on the control strategy. The smelting process is not a static and homogeneous system. During the melting phase, the focus is mainly on the melting and heating of solid raw materials; during the oxidation phase, the focus is on decarburization and dephosphorization; and during the reduction phase, the focus is on deoxidation and alloying. The reaction mechanisms and control objectives of different stages are completely different.

[0151] In this embodiment, the key state vector includes real-time characterization information such as temperature and oxygen activity. The operating condition identification classifier (such as a classification model based on support vector machines or deep neural networks) can accurately determine the current process stage by learning the stage feature boundaries from a large amount of historical data. The generated smelting stage type labels provide a key index basis for subsequent knowledge retrieval, ensuring the pertinence and timeliness of the optimization strategy.

[0152] Step S602: Based on the smelting stage type label, retrieve the matching historical optimization case nodes in the process knowledge graph, obtain the historical state feature vector and historical optimization action sequence associated with the historical optimization case nodes, and form a set of historical successful optimization cases.

[0153] Specifically, the process knowledge graph stores a large amount of structured data from past successful smelting cases. Each case node is labeled with a corresponding stage and associated with its state characteristics at that time (such as specific temperature values ​​and composition content) and the sequence of optimization actions taken at that time (such as power adjustment range and alloy addition amount). By matching stage labels, the system quickly filters out irrelevant historical data and locks onto the set of historical successful cases at the current stage, thus realizing the logical process of quickly locating high-value experiential knowledge from massive amounts of data.

[0154] Step S603: Calculate the cosine similarity between the key state vector and the feature vectors of each historical state in the set of historical successful optimization cases, and select the historical optimization case node with the largest cosine similarity as the target reference case.

[0155] This step aims to address the problem of state similarity matching, seeking historical experiences that most closely resemble the current operating conditions. Although the stage is the same, specific micro-states such as temperature levels and oxygen activity may differ. Cosine similarity measures the directional consistency of two vectors by calculating the cosine of the angle between them in multidimensional space, effectively ignoring the influence of absolute numerical magnitude and focusing on the similarity of state patterns. The system calculates the similarity between the current state vector and each candidate historical state vector, identifying the node with the highest similarity as the target reference case. This is equivalent to finding the most successful historical precedent, providing the most intuitive reference sample for current decision-making.

[0156] Step S604: Extract the historical optimization action sequence from the target reference case and convert the historical optimization action sequence into a prior policy parameter matrix;

[0157] In this context, the historical optimization action sequence is a series of specific operational instructions (such as "increase power by 10%" or "add 5kg of aluminum granules"), while the reinforcement learning policy network requires probability distribution parameters or neural network weights. Through specific transformation algorithms (such as inverse reinforcement learning or policy parameterization mapping), these specific successful actions are transformed into a priori parameter matrix that the policy network can recognize and utilize. This process essentially compiles human experts' or historical best practices into a language that the machine model can understand, giving the policy network initial intuition and common sense.

[0158] Step S605: Configure the initial parameters of the policy network of the hierarchical reinforcement learning optimization framework using the prior policy parameter matrix to generate the initialized hierarchical reinforcement learning optimization framework.

[0159] Traditional reinforcement learning often requires lengthy periods of random exploration to converge to a better policy, which is unacceptable in industrial settings. By loading the transformed prior policy parameter matrix into the weights of the policy network, the network is essentially trained and makes decisions directly based on historical successes. The initialized framework is no longer a blank slate but an agent with a basic tendency to make correct decisions. This not only ensures the safety of the initial decisions but also significantly shortens the time required for subsequent online learning and optimization.

[0160] Step S606: Based on the initialized hierarchical reinforcement learning optimization framework, perform the step of coordinating the thermodynamic policy network and the chemical composition policy network to make decoupled decisions.

[0161] At this point, the strategy network has incorporated historical best practices at the network parameter level. When making decoupled decisions regarding thermodynamics and chemical composition, it can use historical successes as a benchmark and fine-tune and optimize based on current real-time state deviations. This approach avoids the process fluctuation risks caused by blind exploration, making the decision-making process more robust, efficient, and consistent with process logic, achieving a deep integration of data-driven and knowledge-guided approaches.

[0162] In the above implementation, a knowledge transfer path from historical experience to real-time decision-making is constructed, which effectively solves the problems of long cold start time and high randomness in early decision-making that exist in traditional reinforcement learning methods in industrial applications. It realizes the intelligent optimization of current complex smelting processes by using historical successful experience to guide them, which significantly improves the response speed, decision accuracy and operational stability of the optimization system, ensures the traceability and interpretability of smelting control strategies, and fully demonstrates the advantages of data-driven intelligent optimization under knowledge guidance.

[0163] Reference Figure 7 As one implementation of step S106, the step of calling the process knowledge graph to perform multi-objective constraint solving and conflict resolution on the collaborative optimization action set to generate an executable optimization instruction set includes:

[0164] Step S701: Extract the rule edges of the thermodynamic constraint subgraph and the association edges of the chemical composition association subgraph from the process knowledge graph to generate a multi-objective constraint rule set;

[0165] The process knowledge graph contains a wealth of domain rules, but these rules exist in the form of a graph structure (nodes and edges).

[0166] In this embodiment, the regular edges of the thermodynamic constraint sub-map encode the safety boundaries and dynamic constraints between parameters such as temperature and oxygen activity, for example, "the tapping temperature must be between T_min and T_max" or "after deoxidation, the rate of decrease in oxygen activity must not exceed Δ[O] / Δt_max to prevent slag entrapment." The correlation edges of the chemical composition correlation sub-map encode the reaction coupling and process requirements between elements, for example, "the final [S] content must be ≤0.010%" or "when adding Al for deoxidation, it is necessary to ensure that [Al]s is within the target range to control the morphology of inclusions."

[0167] By extracting these edges and their attributes, and formalizing them into computer-reasonable criteria such as "IF (condition) THEN (constraint)" or "action A and action B are mutually exclusive," a multi-objective constraint rule set is generated. This rule set integrates requirements from multiple dimensions, including quality (component accuracy), cost (yield), safety (temperature window), and efficiency (operation timing), providing a comprehensive, clear, and computable standard library for subsequent evaluation and verification.

[0168] Step S702: Based on the multi-objective constraint rule set, perform a feasibility assessment on each action unit in the collaborative optimization action set and identify conflicting action units;

[0169] Among them, the action units in the collaborative optimization action set (such as "increase the power to P1 at time t1" and "add M kg of alloy X at time t2") will be substituted into the rule set one by one or in groups for evaluation.

[0170] Specifically, the logic of feasibility assessment is to simulate the execution of these actions, predict the system state after execution, and determine whether the predicted state violates any constraint in the rule set. For example, assess whether the "increase power" action unit will cause the predicted temperature to exceed the constraint of "maximum safe temperature" in the rule set; assess whether the combination of the two action units "add alloy X" and "add alloy Y" triggers the mutually exclusive association of "X and Y will form a harmful phase at high temperature" in the rule set.

[0171] Next, conflicting action units are identified to find single actions or combinations of actions that would lead to rule violations. This is equivalent to a pre-emptive sand table simulation, which exposes in advance process risks or physical and chemical contradictions that may be overlooked from the perspective of optimization algorithms alone.

[0172] Step S703: Match conflict resolution strategies for conflict action units based on the dynamic weight edge priority in the process knowledge graph;

[0173] The dynamic weighted edges in the process knowledge graph record the relative importance or priority of each rule and association in the graph within the current production context. These weights can be dynamically adjusted based on production objectives (such as prioritizing quality in this batch), historical success rates, or real-time operating conditions. The priority of the dynamic weighted edges serves as the basis for conflict arbitration decisions.

[0174] In this embodiment, the principle of conflict resolution strategy matching is to select a preset resolution strategy based on the priority of the relevant rule edges for the identified conflict type (such as safety constraint conflict, quality constraint conflict, and mutually exclusive reaction conflict). For example, if the conflict is between "continuing heating to reach the target temperature" and "the upper limit of the safe temperature of the furnace lining," and the priority of the safety rule is set to the highest, the matching resolution strategy may be to "redefine" the heating action as "holding" or "reducing the heating rate." If the conflict is between two mutually exclusive alloy addition actions, and the quality rule corresponding to one alloy has a higher priority, the strategy may be to "reject" the addition of the other alloy or replace it with an equivalent but non-conflicting substitute material. This step ensures that the resolution of the conflict is not arbitrary, but rather a well-founded and intelligent arbitration that conforms to the current production value orientation.

[0175] Step S704: Apply a resolution strategy to redefine or eliminate conflicting action units to generate a conflict-resolved action sequence.

[0176] Specifically, the application of conflict resolution strategies involves two core operations: redefinition and elimination. Redefinition refers to modifying the parameters of conflicting action units to make them as close as possible to the original optimization intent while satisfying high-priority constraints. For example, redefining "increase power by 200kW" as "increase power by 100kW", or redefining "add all alloy at once" as "add in two batches". Elimination, on the other hand, involves directly removing an action unit when the conflict cannot be reconciled through parameter adjustment and the target priority corresponding to that action unit is relatively low.

[0177] After applying the resolution strategy, the original collaborative optimization action set is updated into a new action sequence that has passed all rule checks and is now conflict-resolved. This sequence logically eliminates the identified process conflicts and is an internally consistent and domain-knowledge-compliant feasible solution.

[0178] Step S705: Perform a compatibility check between the conflict-resolved action sequence and the physical execution capability of the smelting equipment to generate an executable optimized instruction set.

[0179] The principle of adaptability verification is to compare the details of each action with the physical execution capabilities of the underlying smelting equipment, including capability boundary verification, timing executability verification, and instruction granularity conversion.

[0180] Specifically, capability boundary verification includes: verifying the maximum feeding rate of the feeding system, the maximum speed of electrode lifting, and the maximum flow rate of bottom-blowing gas, to ensure that the action commands do not exceed the hardware limits of the equipment. Timing executability verification includes: verifying whether the minimum time interval between two consecutive actions meets the requirements for equipment reset or reaction stabilization; and whether complex action combinations exceed the allowable processing time of the station. Command granularity conversion involves converting abstract actions generated by the optimization algorithm (e.g., "increase silicon content by 0.1%)" into specific commands that the equipment controller can understand (e.g., "release 150 kg of ferrosilicon from silo 3 and add it to chute 2 within 30 seconds").

[0181] Through this series of verifications and transformations, the final executable optimized instruction set is a final operating procedure that has been verified to be safe in terms of process, self-consistent in terms of logic, and can be accurately executed by production line equipment in terms of engineering. It can be directly issued to the basic automation system to drive the field equipment to complete the smelting.

[0182] In the above implementation, the process knowledge graph is visualized as a set of multi-objective constraint rules. A comprehensive, domain-mechanism-based compliance and security simulation of the initial optimization actions is performed, identifying and resolving potential process conflicts and physical contradictions in advance. By introducing a dynamic weighted edge priority mechanism, conflict resolution can flexibly adapt to the value orientation of different steel grades and production objectives, achieving a deep integration of intelligence and process principles. Finally, rigorous equipment compatibility verification ensures that all instructions fall within the physical capabilities of the equipment and are translated into a control language recognizable by the equipment, improving the reliability, safety, and operability of the smelting system.

[0183] Reference Figure 8 As one implementation of step S108, the steps of calculating the differential reward value for each parameter dimension based on multi-dimensional feedback detection data, and updating the network parameters of the hierarchical reinforcement learning optimization framework and the dynamic weight edges in the process knowledge graph using the differential reward value include:

[0184] Step S801: Extract the actual measured values ​​of each parameter dimension and the expected values ​​of the target process parameters from the multi-dimensional feedback detection data, and calculate the parameter deviation.

[0185] Among them, the multi-dimensional feedback detection data collected by online detection sensors (such as oxygen analyzers, temperature probes, and spectrometers) after the optimization instructions are executed is the real-world response of the smelting process to the system instructions.

[0186] In this embodiment, the response is precisely compared with the initial optimization objective, namely the expected value of the target process parameters (e.g., the expected endpoint temperature of 1750℃ and the expected carbon content of 0.20%). The parameter deviation is calculated, which quantifies the gap between the "ideal" and the "reality," for example, "actual temperature 1720℃, deviation -30℃," and "actual carbon content 0.22%, deviation +0.02%." These deviations are objective and quantitative performance signals, accurately indicating the successes and shortcomings of this optimization in various dimensions such as thermodynamic control and chemical composition control.

[0187] Step S802: Based on the parameter deviation and the preset reward function template, generate the differential reward value for each parameter dimension;

[0188] The pre-defined reward function template is a set of rule frameworks that map deviations to reward values. Its core logic is to define which deviations are acceptable and which require severe punishment. It can also unify the parameter deviations of different dimensions and importance into a comparable and synthesizable scalar signal.

[0189] For example, a simple template rule might be: a positive reward is given for temperature deviations within ±10℃, and a negative reward is given for deviations outside this range, with the intensity of the negative reward increasing with the square of the deviation; exceeding the sulfur content limit (even if the deviation is small) will be given a very large negative reward to reflect its extremely high quality importance. Generating differential reward values ​​for each parameter dimension involves applying these templates to calculate a specific reward score for each parameter such as temperature, carbon, and sulfur.

[0190] Understandably, the differential reward is calculated based on the difference between the current optimization result and the objective. It is an immediate evaluation of the effect of this specific decision. This reward value integrates multiple objective requirements such as quality, cost, and security, and is a unified and quantified value signal that drives the subsequent updates of the policy network and knowledge graph.

[0191] Step S803: Calculate the backpropagation gradient for the thermodynamic policy network and chemical composition policy network of the hierarchical reinforcement learning optimization framework based on the differential reward value, and update the network weight parameters.

[0192] Among them, the thermodynamic policy network and the chemical composition policy network in the hierarchical reinforcement learning optimization framework are composed of millions of connection weight parameters, which determine their thinking and decision-making methods.

[0193] In this embodiment of the application, the backpropagation gradient algorithm is used to use the calculated differential reward value as the target signal and propagate it back along the network from the output layer to the input layer to calculate the "contribution" or "responsibility" (i.e. gradient) of each weight parameter in the network to the final reward (or penalty).

[0194] Next, based on the principle of "the greater the contribution, the greater the adjustment," optimization algorithms such as gradient descent are used to update the network weight parameters. Specifically, the weights on decision paths leading to positive rewards are strengthened, while the weights on paths leading to negative rewards are weakened.

[0195] Through this process, the thermodynamic strategy network learns how to generate actions that yield higher temperature control rewards, while the chemical composition strategy network learns how to generate actions that yield higher composition hit rate rewards. This update allows both dedicated networks to continuously learn from historical successes and failures, making future decisions more likely to lead to better smelting results, thus achieving data-driven online strategy optimization.

[0196] Step S804: Based on the differential reward value, evaluate the importance of the dynamic weighted edges in the process knowledge graph and generate weight adjustment coefficients;

[0197] The dynamic weighted edges in the process knowledge graph encode key process parameters such as the coupling strength of reactions between elements and the influence coefficient of temperature on yield. However, the initial values ​​of these parameters may come from theory or historical experience and may not perfectly match the specific operating conditions of the current production line.

[0198] Therefore, the importance of dynamic weighted edges can be assessed by analyzing which weighted edges in the graph are most relevant to the current optimization result (reflected in the differential reward value). For example, if the actual silicon yield in this smelting process is much lower than the expected value of the "silicon yield weighted edge" in the graph, resulting in a large deviation in silicon composition and a negative reward, then the system will assess that this weighted edge is highly important, meaning its inaccuracy is significantly responsible for the poor result. Based on this assessment, the system will generate a weight adjustment coefficient, which is an adjustment amount used to quantify the correction of the specific value of this weighted edge. The magnitude and direction of the coefficient (whether it increases or decreases) are jointly determined by the sign of the reward value and the specific pattern of the deviation.

[0199] Step S805: Apply weight adjustment coefficients to update the connection strength of dynamic weight edges and generate an updated process knowledge graph.

[0200] The process of updating the connection strength of dynamic weighted edges using weight adjustment coefficients involves mathematically correcting the values ​​of corresponding edges in the graph. For example, the "temperature influence coefficient on aluminum yield" is revised from the initial 0.85 (theoretical value) to 0.78 (a value more consistent with the actual situation of this production line). This update process transforms a static, a priori knowledge base into a dynamic, self-evolving, and living knowledge system based on production data.

[0201] Next, an updated process knowledge graph is generated. In the next optimization decision, the process rules and parameters relied upon by the system will be closer to the actual factory conditions. This not only directly improves the accuracy of graph-based conflict verification and rule reasoning, but also provides a more reliable environmental model for the policy network to learn and simulate.

[0202] In the above implementation, a unified feedback signal based on multi-objective rewards is used to achieve collaborative optimization of a data-driven policy network and a mechanism model-driven knowledge graph. The parameter updates of the policy network enable the system to continuously explore and solidify better decision-making patterns within a massive state-action space; the updates of dynamic weighted edges in the knowledge graph provide continuous calibration of the inherent process rules and parameters, making them more accurately reflect the actual physicochemical laws of a specific production line. This closed-loop system of synchronous iteration between model parameters and knowledge parameters fundamentally solves the industry pain points of traditional imported systems or fixed models that struggle to adapt to complex changes in the field and lack deep self-learning capabilities.

[0203] This application also discloses an intelligent optimization system for smelting processes based on multi-source data fusion.

[0204] A smart optimization system for smelting processes based on multi-source data fusion, specifically comprising:

[0205] The multi-source data acquisition module is used to acquire the target process parameters of the steel to be smelted, as well as real-time sensor data and production process data during the smelting process.

[0206] The process knowledge graph construction module is used to divide the target process parameters into thermodynamic parameter sets and chemical composition parameter sets according to parameter characteristics, and construct thermodynamic constraint sub-graphs and chemical composition association sub-graphs respectively, and merge them to generate a process knowledge graph;

[0207] The smelting state space modeling module is used to preprocess and time-series align real-time sensor data and production process data to construct a comprehensive smelting state space model for the current furnace.

[0208] The policy network-driven module is used to extract key state vectors from the integrated smelting state space model and input them into a pre-defined hierarchical reinforcement learning optimization framework; wherein, the hierarchical reinforcement learning optimization framework includes a thermodynamic policy network and a chemical composition policy network.

[0209] The multi-strategy collaborative decision-making module is used to control the hierarchical reinforcement learning optimization framework to coordinate the thermodynamic strategy network and the chemical composition strategy network to make decoupled decisions based on the parameter priority and coupling relationship in the target process parameters, and generate a set of collaborative optimization actions.

[0210] The multi-objective constraint solving module is used to call the process knowledge graph to solve multi-objective constraints and resolve conflicts on the collaborative optimization action set, and generate an executable optimization instruction set.

[0211] The execution feedback module is used to execute the executable optimized instruction set and synchronously collect multi-dimensional feedback detection data during the execution process;

[0212] The optimization and update module is used to calculate the differential reward value of each parameter dimension based on multi-dimensional feedback detection data, and to update the network parameters of the hierarchical reinforcement learning optimization framework and the dynamic weight edges in the process knowledge graph using the differential reward value.

[0213] An intelligent optimization system for smelting processes based on multi-source data fusion, according to an embodiment of this application, can implement any of the above methods, and the specific working process of each module in the system can be referred to the corresponding process in the above method embodiments.

[0214] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0215] This application also discloses a computer-readable storage medium.

[0216] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the intelligent optimization methods for smelting processes based on multi-source data fusion.

[0217] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0218] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for intelligent optimization of smelting processes based on multi-source data fusion, characterized in that, The optimization method includes: Acquire the target process parameters of the steel grade to be smelted, as well as real-time sensor data and production process data during the smelting process; Based on the physicochemical properties of the target process parameters, parameters involving changes in energy state are classified into a thermodynamic parameter set, and parameters involving changes in material composition are classified into a chemical composition parameter set. Based on the set of thermodynamic parameters, a nonlinear constraint relationship between temperature and oxygen activity is defined, and a thermodynamic constraint sub-map is generated. Based on the set of chemical composition parameters, the reaction coupling relationship between alloying elements is defined, and a chemical composition correlation sub-map is generated. The thermodynamic constraint sub-map and the chemical composition correlation sub-map are connected by edges using the law of conservation of matter and energy, and then fused to generate a process knowledge graph. The real-time sensor data and production process data are preprocessed and time-series aligned to construct a comprehensive smelting state space model for the current furnace batch. Key state vectors are extracted from the integrated smelting state-space model and input into a pre-defined hierarchical reinforcement learning optimization framework; wherein, the hierarchical reinforcement learning optimization framework includes a thermodynamic policy network and a chemical composition policy network; The priority weights of thermodynamic parameters and chemical composition parameters in the target process parameters are analyzed to generate a parameter decision weight matrix; The thermodynamic action sequence output by the thermodynamic strategy network is prioritized according to the parameter decision weight matrix to determine the execution order of the basic power adjustment action and the basic deoxygenation action. Based on the element reaction coupling relationship in the chemical component correlation sub-map, conflict labeling is performed on the chemical component action sequence output by the chemical component strategy network to obtain conflict labeling results; The nonlinear constraint relationships in the thermodynamic constraint sub-map are invoked to calculate the interference influence factor of the priority-ranked thermodynamic action sequence on the action sequence of the marked conflicting chemical components, and the feasibility is verified by combining the interference influence factor with the conflict marking results. If the interference factor exceeds the preset threshold or the feasibility verification fails, the output of the chemical component strategy network is adjusted using an attention mechanism to generate a verified chemical component action sequence. The verified chemical component action sequence is matched with the priority-sorted thermodynamic action sequence in spatiotemporal dimensions to generate an initial cooperative action set. Based on the physical execution window constraints of the smelting process, the initial set of cooperative actions is rearranged in time to generate a set of cooperative optimized actions. The process knowledge graph is invoked to perform multi-objective constraint solving and conflict resolution on the collaborative optimization action set, generating an executable optimization instruction set. The executable optimized instruction set is executed, and multi-dimensional feedback detection data is collected simultaneously during the execution process; Based on the multi-dimensional feedback detection data, the differential reward value for each parameter dimension is calculated, and the differential reward value is used to update the network parameters of the hierarchical reinforcement learning optimization framework and the dynamic weight edges in the process knowledge graph.

2. The intelligent optimization method for smelting processes based on multi-source data fusion according to claim 1, characterized in that, After the step of fusing the thermodynamic constraint sub-map and the chemical composition correlation sub-map through the law of conservation of matter and energy to generate a process knowledge graph, the method further includes: Obtain a pre-built local expert domain knowledge base, including element reaction coupling relationships and safety boundary parameters; The element reaction coupling relationship is mapped to the association edge of the chemical composition association subgraph in the process knowledge graph, and the safety boundary parameter is mapped to the constraint node of the thermodynamic constraint subgraph in the process knowledge graph. Based on the mapping results, the initial connection strength of the dynamic weight edges of the process knowledge graph is updated to generate an updated process knowledge graph.

3. The intelligent optimization method for smelting processes based on multi-source data fusion according to claim 1, characterized in that, The steps for preprocessing and time-series aligning the real-time sensor data and production process data to construct the comprehensive smelting state-space model for the current furnace run include: Identify the acquisition timestamps of each detection parameter in the real-time sensor data, and establish a time delay compensation mapping relationship based on the response delay characteristics of the smelting equipment; The temperature data stream and oxygen activity data stream are aligned on the time axis according to the time delay compensation mapping relationship to generate a synchronized sensor data sequence. Dynamic process features, including heating rate gradient, alloy element diffusion rate, and deoxidation reaction efficiency, are extracted from the synchronized sensor data sequence. The synchronized sensor data sequence, the dynamic process feature quantity, and the raw material ratio information in the production process data are fused using multidimensional tensors. A smelting state space coordinate system with time-varying characteristics is constructed based on a tensor dimensionality reduction algorithm, and the feature vector of the multidimensional tensor is mapped in the smelting state space coordinate system. The integrated smelting state space model for the current furnace is generated based on the topological relationship of the distribution of the feature vectors.

4. The intelligent optimization method for smelting processes based on multi-source data fusion according to claim 1, characterized in that, Before the step of the hierarchical reinforcement learning optimization framework coordinating the thermodynamic strategy network and the chemical composition strategy network to decouple and generate a collaborative optimization action set based on the parameter priorities and coupling relationships in the target process parameters, the following steps are also included: The key state vector is input into a preset working condition identification classifier to identify the smelting stage type of the current furnace and generate smelting stage type labels, including melting period, oxidation period and reduction period. Based on the smelting stage type label, the matching historical optimization case nodes are retrieved in the process knowledge graph, and the historical state feature vectors and historical optimization action sequences associated with the historical optimization case nodes are obtained to form a set of historical successful optimization cases. Calculate the cosine similarity between the key state vector and the feature vectors of each historical state in the set of historical successful optimization cases, and select the historical optimization case node with the largest cosine similarity as the target reference case; Extract the historical optimization action sequence from the target reference case and transform the historical optimization action sequence into a prior policy parameter matrix; The prior policy parameter matrix is ​​used to configure the initial parameters of the policy network of the hierarchical reinforcement learning optimization framework, thereby generating the initialized hierarchical reinforcement learning optimization framework. Based on the initialized hierarchical reinforcement learning optimization framework, the step of coordinating the decoupling decision between the thermodynamic policy network and the chemical composition policy network is executed.

5. The intelligent optimization method for smelting processes based on multi-source data fusion according to claim 1, characterized in that, The steps of calling the process knowledge graph to perform multi-objective constraint solving and conflict resolution on the collaborative optimization action set to generate an executable optimization instruction set include: The rule edges of the thermodynamic constraint subgraph and the association edges of the chemical composition association subgraph are extracted from the process knowledge graph to generate a multi-objective constraint rule set. Based on the multi-objective constraint rule set, a feasibility assessment is performed on each action unit in the collaborative optimization action set to identify conflicting action units; Based on the dynamic weighted edge priority in the process knowledge graph, the conflict action unit is matched with a resolution strategy. The conflict resolution strategy is applied to redefine or eliminate the conflict action units to generate a conflict-resolved action sequence. The reconciled action sequence is then subjected to compatibility verification with the physical execution capabilities of the smelting equipment to generate an executable optimized instruction set.

6. The intelligent optimization method for smelting processes based on multi-source data fusion according to claim 1, characterized in that, The steps of calculating the differential reward value for each parameter dimension based on the multi-dimensional feedback detection data, and updating the network parameters of the hierarchical reinforcement learning optimization framework and the dynamic weight edges in the process knowledge graph using the differential reward value, include: Extract the actual measured values ​​of each parameter dimension and the expected values ​​of the target process parameters from the multi-dimensional feedback detection data, and calculate the parameter deviation. Based on the parameter deviation and the preset reward function template, the differential reward value for each parameter dimension is generated; Based on the differential reward value, backpropagation gradient calculation is performed on the thermodynamic policy network and chemical composition policy network of the hierarchical reinforcement learning optimization framework to update the network weight parameters; The importance of the dynamic weighted edges in the process knowledge graph is evaluated based on the differential reward value, and weight adjustment coefficients are generated. The connection strength of the dynamic weighted edges is updated by applying the weight adjustment coefficients to generate an updated process knowledge graph.

7. A smelting process intelligent optimization system based on multi-source data fusion, characterized in that, The system is used to execute the intelligent optimization method for smelting processes based on multi-source data fusion as described in any one of claims 1 to 6, wherein the optimization system comprises: The multi-source data acquisition module is used to acquire the target process parameters of the steel to be smelted, as well as real-time sensor data and production process data during the smelting process. The process knowledge graph construction module is used to divide the target process parameters into a thermodynamic parameter set and a chemical composition parameter set according to the parameter characteristics, and to construct a thermodynamic constraint sub-graph and a chemical composition association sub-graph respectively, and then merge them to generate a process knowledge graph. The smelting state space modeling module is used to preprocess and time-series align the real-time sensor data and production process data to construct a comprehensive smelting state space model for the current furnace. The strategy network driving module is used to extract key state vectors from the integrated smelting state space model and input them into a pre-set hierarchical reinforcement learning optimization framework; wherein, the hierarchical reinforcement learning optimization framework includes a thermodynamic strategy network and a chemical composition strategy network. The multi-strategy collaborative decision-making module is used to control the hierarchical reinforcement learning optimization framework to coordinate the thermodynamic strategy network and the chemical composition strategy network to make decoupled decisions based on the parameter priority and coupling relationship in the target process parameters, and generate a set of collaborative optimization actions. The multi-objective constraint solving module is used to call the process knowledge graph to perform multi-objective constraint solving and conflict resolution on the collaborative optimization action set, and generate an executable optimization instruction set. The execution feedback module is used to execute the executable optimized instruction set and synchronously collect multi-dimensional feedback detection data during the execution process; The optimization and update module is used to calculate the differential reward value of each parameter dimension based on the multi-dimensional feedback detection data, and to update the network parameters of the hierarchical reinforcement learning optimization framework and the dynamic weight edges in the process knowledge graph using the differential reward value.

8. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 6.

Images