Metallurgy control method and system, electronic equipment and storage medium

Abstract

The invention relates to a metallurgy control method and system, electronic equipment and a storage medium. The method comprises the steps that a metallurgy optimization task is received, the metallurgy optimization task comprises a task optimization target, a knowledge graph reasoning path of the metallurgy optimization task is determined through a pre-constructed metallurgy knowledge graph, the knowledge graph reasoning path comprises at least one metallurgy execution device, and the at least one metallurgy execution device is used for executing the metallurgy optimization task; based on the knowledge graph reasoning path and a pre-constructed equipment digital twin network, multi-modal operation information associated with the metallurgy execution equipment is obtained, the multi-modal operation information is analyzed according to the task optimization target, a metallurgy control strategy is generated, and the metallurgy execution equipment is subjected to task optimization. And controlling the metallurgy execution equipment to execute corresponding control operation according to the metallurgy control strategy so as to complete the metallurgy optimization task. According to the scheme provided by the invention, the intelligent level and the production efficiency of metallurgy can be greatly improved.

Description

Technical Field

[0001] This application relates to the field of intelligent metallurgical production technology, and in particular to metallurgical control methods, systems, electronic devices and storage media. Background Technology

[0002] With the continuous advancement of science and technology, the metallurgical industry is facing a major opportunity to deeply transform from traditional manufacturing to intelligent manufacturing. The intelligent upgrading of the metallurgical industry will not only help to improve production efficiency and reduce energy consumption costs, but also ensure the stability of product quality.

[0003] From the perspective of metallurgical processes, the iron and steel metallurgical production chain encompasses key processes such as raw material preparation, ironmaking, steelmaking, continuous casting, and rolling, exhibiting typical process industry characteristics such as strong coupling, nonlinear dynamics, and high energy consumption. For example, the composition of molten iron in the blast furnace is affected by the interaction of more than 20 parameters, the carbon control at the converter endpoint needs to achieve a high precision of ±0.02%, and the energy consumption and speed control in the rolling stage are directly related to product quality and cost.

[0004] Although steel companies have made some progress in metallurgical automation in recent years, achieving basic automatic control of key processes in the metallurgical process by integrating detection instruments, programmable controllers and other equipment with computers as the core, frequent manual intervention is still required in the metallurgical automation process, which restricts the response speed of the metallurgical system and makes it difficult to improve the level of intelligence. Summary of the Invention

[0005] To address or partially address the problems existing in related technologies, this application provides a metallurgical control method, system, electronic device, and storage medium, which can significantly improve the level of intelligence and production efficiency in metallurgy.

[0006] The first aspect of this application provides a metallurgical control method, comprising: Receive a metallurgical optimization task, wherein the metallurgical optimization task includes a task optimization objective; The knowledge graph reasoning path of the metallurgical optimization task is determined by a pre-constructed metallurgical knowledge graph, and the knowledge graph reasoning path includes at least one metallurgical execution device. Based on the knowledge graph reasoning path and the pre-built device digital twin network, multimodal operation information associated with the metallurgical execution equipment is obtained; To optimize the task objectives, the multimodal operational information is analyzed to generate a metallurgical control strategy. The metallurgical execution equipment is controlled to perform corresponding control operations according to the metallurgical control strategy in order to complete the metallurgical optimization task.

[0007] In one example, the construction of the metallurgical knowledge graph includes: acquiring historical multi-source heterogeneous data; extracting the equipment name and attribute information of the metallurgical execution equipment from the historical multi-source heterogeneous data; using the equipment name as a metallurgical node, and establishing edge relationships between each metallurgical node based on the attribute information; constructing a semantic network containing the metallurgical nodes and the edge relationships, and using the semantic network as the metallurgical knowledge graph; wherein, the metallurgical knowledge graph is used to output a corresponding knowledge graph reasoning path based on the association relationship between each metallurgical execution equipment when the metallurgical optimization task is received, and the knowledge graph reasoning path is used to represent the causal chain of the metallurgical execution equipment executing the metallurgical optimization task.

[0008] In one example, the construction of the equipment digital twin network includes: acquiring historical multi-source heterogeneous data; extracting the operating parameters of the metallurgical execution equipment and the topological relationships between the various metallurgical execution equipment from the historical multi-source heterogeneous data; using the metallurgical execution equipment as a twin object, and constructing an equipment digital twin of the twin object based on the operating parameters; connecting the equipment digital twins according to the topological relationships to generate an equipment digital twin network, which is used to simulate the operating state of the twin object.

[0009] In one instance, obtaining multimodal operational information associated with the metallurgical execution equipment based on the knowledge graph reasoning path and a pre-built equipment digital twin network includes: determining, one by one, the metallurgical execution equipment used to perform the metallurgical optimization task based on the knowledge graph reasoning path; obtaining production rhythm data and energy consumption data of the metallurgical execution equipment; collecting sensor data of the equipment digital twin corresponding to the metallurgical execution equipment in the equipment digital twin network in real time; and fusing the sensor data, production rhythm data, and energy consumption data to generate the multimodal operational information.

[0010] In one example, the step of analyzing the multimodal operational information to generate a metallurgical control strategy for the task optimization objective includes: when the task optimization objective is defect detection, extracting surface image data of the metallurgical execution equipment from the multimodal operational information; performing feature extraction on the surface image data to obtain visual features and defect features of the surface image data; mapping the visual features and defect features to the same feature space, performing a defect classification operation to generate defect information of the metallurgical execution equipment; inputting the defect information into a pre-trained metallurgical control model, and having the metallurgical control model output the metallurgical control strategy.

[0011] In one instance, the step of analyzing the multimodal operational information to generate a metallurgical control strategy for the task optimization objective includes: when the task optimization objective is equipment health detection, extracting sensor time-series data of the metallurgical execution equipment from the multimodal operational information; extracting degradation features from the sensor time-series data through a multi-head self-attention mechanism; inputting the degradation features into a pre-trained metallurgical control model, and outputting the metallurgical control strategy from the metallurgical control model.

[0012] In one instance, the step of analyzing the multimodal operating information to generate a metallurgical control strategy for the task optimization objective includes: when the task optimization objective is energy efficiency optimization, extracting energy consumption data and operating parameters of the metallurgical execution equipment from the multimodal operating information; inputting the energy consumption data and operating parameters into a pre-trained metallurgical control model, and having the metallurgical control model output the metallurgical control strategy.

[0013] A second aspect of this application provides a metallurgical control system, comprising: The task receiving module is used to receive metallurgical optimization tasks, wherein the metallurgical optimization tasks include task optimization objectives; The reasoning path determination module is used to determine the knowledge graph reasoning path of the metallurgical optimization task through a pre-constructed metallurgical knowledge graph, wherein the knowledge graph reasoning path includes at least one metallurgical execution device. The multimodal operation information acquisition module is used to acquire multimodal operation information associated with the metallurgical execution equipment based on the knowledge graph reasoning path and the pre-built equipment digital twin network; The metallurgical control strategy generation module is used to analyze the multimodal operation information and generate a metallurgical control strategy in response to the task optimization objectives. The optimization execution module is used to control the metallurgical execution equipment to perform corresponding control operations according to the metallurgical control strategy, so as to complete the metallurgical optimization task.

[0014] A third aspect of this application provides an electronic device, comprising: Processor; and A memory that stores executable code, which, when executed by the processor, causes the processor to perform the method described above.

[0015] A fourth aspect of this application provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method described above.

[0016] The fifth aspect of this application provides a computer program product comprising computer instructions that, when executed by a processor, implement the method described above.

[0017] The technical solution provided in this application may include the following beneficial results: This application is capable of receiving metallurgical optimization tasks, which include task optimization objectives. First, a knowledge graph reasoning path for the metallurgical optimization task is determined using a pre-constructed metallurgical knowledge graph. The knowledge graph reasoning path includes at least one metallurgical execution device. Then, based on the knowledge graph reasoning path and a pre-constructed device digital twin network, multimodal operating information associated with the metallurgical execution device is obtained. Subsequently, the multimodal operating information is analyzed for the task optimization objective to generate a metallurgical control strategy. This strategy is then used to control the metallurgical execution device to perform corresponding control operations, thereby completing the metallurgical optimization task.

[0018] Compared with related technologies, the technical solution of this application has the following advantages: First, by using a pre-constructed metallurgical knowledge graph, the knowledge graph reasoning path of the metallurgical optimization task is automatically determined, which helps to analyze the causal relationship of each metallurgical execution device in the metallurgical optimization task from an objective perspective, avoiding the one-sidedness and subjectivity caused by manual judgment. At the same time, combined with a pre-constructed equipment digital twin network, the multimodal operation information associated with the metallurgical execution devices can be obtained quickly and accurately, significantly improving the efficiency and accuracy of data acquisition, thereby improving the overall efficiency of the metallurgical production process. Second, after obtaining the multimodal operation information, the multimodal operation information can be differentiated for different task optimization objectives, generating a metallurgical control strategy that is highly matched with the task optimization objective. Based on the metallurgical control strategy, the metallurgical execution devices are precisely driven to perform corresponding control operations, realizing intelligent regulation of the metallurgical production process, thereby improving the level of intelligence of metallurgical control and meeting complex production needs.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] The above and other objects, features and advantages of this application will become more apparent from the following description of exemplary embodiments of this application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components.

[0021] Figure 1 This is a schematic flowchart illustrating a metallurgical control method according to an embodiment of this application; Figure 2 This is another schematic flowchart illustrating a metallurgical control method according to an embodiment of this application; Figure 3 This is a schematic diagram of the data platform of a metallurgical control system shown in an embodiment of this application; Figure 4 This is a schematic diagram of the intelligent central hub of a metallurgical control system shown in an embodiment of this application; Figure 5 This is a schematic diagram of a digital twin network of equipment in a metallurgical control system, as shown in an embodiment of this application. Figure 6 This is a schematic diagram of the structure of a metallurgical control system shown in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application. Detailed Implementation

[0022] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0023] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0024] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0025] With the continuous advancement of science and technology, the metallurgical industry is facing a major opportunity to deeply transform from traditional manufacturing to intelligent manufacturing. The intelligent upgrading of the metallurgical industry will not only help to improve production efficiency and reduce energy consumption costs, but also ensure the stability of product quality.

[0026] From the perspective of metallurgical processes, the iron and steel metallurgical production chain encompasses key processes such as raw material preparation, ironmaking, steelmaking, continuous casting, and rolling, exhibiting typical process industry characteristics such as strong coupling, nonlinear dynamics, and high energy consumption. For example, the composition of molten iron in the blast furnace is affected by the interaction of more than 20 parameters, the carbon control at the converter endpoint needs to achieve a high precision of ±0.02%, and the energy consumption and speed control in the rolling stage are directly related to product quality and cost.

[0027] Although steel companies have made some progress in metallurgical automation in recent years, achieving basic automatic control of key processes in the metallurgical process by integrating detection instruments, programmable controllers and other equipment with computers as the core, frequent manual intervention is still required in the metallurgical automation process, which restricts the response speed of the metallurgical system and makes it difficult to improve the level of intelligence.

[0028] In related technologies, there are problems such as low response efficiency and incomplete information acquisition, which lead to low production efficiency and difficulty in improving the level of intelligence.

[0029] To address the aforementioned problems, this application provides a metallurgical control method that can significantly improve the level of intelligence and production efficiency in metallurgy.

[0030] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0031] Figure 1 This is a schematic flowchart illustrating a metallurgical control method according to an embodiment of this application. See also... Figure 1 This method, applied to metallurgical control systems, includes at least the following steps: Step 101: Receive the metallurgical optimization task, which includes the task optimization objective.

[0032] In actual production, typical steel metallurgy processes include at least key steps such as raw material preparation, ironmaking, steelmaking, continuous casting, and rolling.

[0033] The main objective of raw material preparation is to process raw materials such as iron ore and coking coal into furnace feed that meets the requirements of blast furnace smelting. For example, iron ore is sintered or pelletized to form sintered ore or pellets with suitable particle size and uniform composition, while coking coal is dry-distilled at high temperature in a coke oven to produce coke, while recoverable resources such as coal gas and coal tar are produced as byproducts. The process stability at this stage directly affects the quality of the furnace feed and the subsequent blast furnace smelting efficiency.

[0034] Ironmaking typically employs a blast furnace method, in which coke, sinter, flux, and other materials are added to the blast furnace in proportion and subjected to a reduction reaction at a high temperature of approximately 1500°C. This process ultimately produces molten iron containing about 4% carbon and byproduct slag, which is then transported to the steelmaking process.

[0035] Steelmaking is carried out through converter smelting, which means that after high-temperature molten iron is poured into the converter, decarburization, desulfurization and other reactions are carried out by top blowing or bottom blowing of high-purity oxygen, so as to control the content of various elements in the molten steel and achieve precise control of composition.

[0036] Continuous casting is the process of cooling and solidifying refined molten steel through a continuous casting machine to form slabs or billets. The casting speed is usually between 1 m / min and 2 m / min, depending on the steel grade and cross-sectional dimensions.

[0037] Rolling processes can be divided into hot rolling and cold rolling based on the state of the billet. Hot rolling uses high temperatures to roll the billet into steel products of the target thickness and width, such as hot-rolled coils and rails. Cold rolling, on the other hand, involves further processing at room temperature to obtain finished steel products with higher surface quality and more precise dimensions, such as cold-rolled sheets and galvanized sheets.

[0038] In the embodiments of this application, the metallurgical control system can be applied to the intelligent upgrading of the metallurgical industry, such as smelting process control, intelligent regulation of continuous casting and rolling, predictive maintenance, surface defect detection, real-time composition prediction, carbon footprint optimization, waste slag resource utilization, raw material procurement optimization, intelligent production scheduling system and alloy design acceleration, etc.

[0039] It is worth noting that the metallurgical control system maintains communication connections with the MES (Manufacturing Execution System) and EMS (Energy Management System).

[0040] The MES system is an information system that connects the factory workshop with the upper management system. It is responsible for tasks such as production plan execution, workshop scheduling, equipment status monitoring, production progress tracking, and quality management. It can collect data such as production rhythm, process parameters, and personnel operation in real time to ensure that production proceeds efficiently according to plan.

[0041] EMS systems are mainly used for energy monitoring and management in factories or enterprises, including the metering, energy consumption data collection, analysis and optimization of various energy sources such as electricity, gas, water and steam. They can help reduce energy consumption, improve energy utilization efficiency and achieve energy conservation and emission reduction goals.

[0042] Optionally, the metallurgical control system can receive metallurgical optimization tasks.

[0043] The metallurgical optimization tasks can be manually input by relevant technical personnel or automatically triggered by the metallurgical control system based on metallurgical needs. Each metallurgical optimization task includes a clear optimization objective, enabling the metallurgical control system to configure relevant information based on the optimization objective and efficiently complete the metallurgical optimization task.

[0044] In this application, the task optimization objectives include, but are not limited to, defect detection, equipment health monitoring, and energy efficiency optimization. For example, if the task optimization objective is "reducing hot rolling energy consumption," the metallurgical control system can automatically identify the metallurgical equipment performing this task optimization objective.

[0045] As an example, when a blast furnace experiences abnormal operating conditions such as abnormal temperature or increased vibration, the metallurgical control system can identify abnormal indicators in real time and automatically trigger metallurgical optimization tasks with "equipment health detection" as the optimization objective, thereby improving the timeliness of response and the level of system automation.

[0046] In addition, the metallurgical control system supports natural language interaction. When a user enters the query "What are the main causes of plate and coil defects?", the metallurgical control system can automatically parse the user's intent represented by the query and trigger the corresponding metallurgical optimization task with "defect detection" as the optimization objective.

[0047] Step 102: Determine the knowledge graph reasoning path for the metallurgical optimization task using a pre-constructed metallurgical knowledge graph. The knowledge graph reasoning path includes at least one metallurgical execution device.

[0048] In this embodiment of the application, after receiving a specific metallurgical optimization task, the metallurgical control system automatically identifies the causal relationship between the metallurgical execution equipment and its upstream and downstream processes that are directly related to the optimization target of the task from the complex process chain by utilizing a pre-built metallurgical knowledge graph, and constructs one or more knowledge graph reasoning paths.

[0049] Optionally, a metallurgical knowledge graph refers to a network obtained by structurally modeling relevant data such as process flow, control parameters, raw material characteristics, equipment types, and performance indicators in the metallurgical field. When a metallurgical optimization task is received, it outputs the corresponding knowledge graph reasoning path based on the relationships between various metallurgical execution equipment. A metallurgical knowledge graph includes several metallurgical nodes and their edge relationships. Each node can represent a metallurgical execution equipment unit (such as a blast furnace, converter, or continuous casting machine) or a key parameter (such as temperature, composition, or speed), while the edge relationships between nodes reflect the causal relationships between their upstream and downstream processes.

[0050] A knowledge graph reasoning path refers to the path of nodes that conform to causal relationships, obtained by traversing the metallurgical knowledge graph under the constraint of the metallurgical optimization task objective. It represents the causal chain of metallurgical execution equipment performing the metallurgical optimization task. The knowledge graph reasoning path starts with the relevant metallurgical execution equipment and ends with the task optimization objective.

[0051] As an example, when a blast furnace experiences abnormal operating conditions such as abnormal temperature or increased vibration, the metallurgical control system can automatically associate information such as molten iron composition, equipment vibration signals, and historical process operation logs based on a pre-built metallurgical knowledge graph. This allows it to identify the relevant metallurgical execution equipment and deduce a complete knowledge graph reasoning path: "Blast furnace burden composition → Molten iron composition → Slag viscosity → Hearth heat flow distribution → Cooling wall temperature." Through this knowledge graph reasoning path, the metallurgical control system can analyze the causes one by one and formulate appropriate metallurgical control strategies.

[0052] Step 103: Based on the knowledge graph reasoning path and the pre-built digital twin network of the equipment, obtain multimodal operation information associated with the metallurgical execution equipment.

[0053] In this embodiment of the application, the metallurgical control system determines the relevant metallurgical execution equipment through the knowledge graph reasoning path determined in step 102. Subsequently, based on the pre-built equipment digital twin network, it collects and integrates various types of operating data associated with the metallurgical execution equipment in real time to generate multimodal operating information.

[0054] Optionally, multimodal operation information refers to operation data fused from different data sources and different data types, which can comprehensively reflect the real-time operation of metallurgical execution equipment. Multimodal operation information includes, but is not limited to, sensor data, production rhythm data, and energy consumption data.

[0055] Step 104: Analyze the multimodal operation information to generate a metallurgical control strategy based on the task optimization objectives.

[0056] In this embodiment of the application, after receiving the multimodal operation information obtained in step 103, the metallurgical control system performs in-depth analysis and processing of the multimodal operation information based on the current task optimization objective, thereby generating a metallurgical control strategy that is highly matched with the metallurgical optimization task.

[0057] Optionally, the process of analyzing multimodal operation information includes at least the following: first, cleaning and extracting features from the multimodal data; then, using a pre-trained metallurgical control model, reasoning or predicting the multimodal operation information and outputting a metallurgical control strategy.

[0058] Step 105: Control the metallurgical execution equipment to perform corresponding control operations according to the metallurgical control strategy in order to complete the metallurgical optimization task.

[0059] In this embodiment of the application, the metallurgical control system can issue control commands to the relevant metallurgical execution equipment according to the metallurgical control strategy generated in step 104, thereby controlling the metallurgical execution equipment to perform corresponding control operations to complete the metallurgical optimization task.

[0060] As an example, after analyzing multimodal operating information, the metallurgical control system discovered that in a certain hot rolling mill section, large fluctuations in rolling speed caused local thickness deviations, affecting the final product quality. The metallurgical control system ultimately generated the following metallurgical control strategy: increase the rolling speed to speed ①, gradually increase it over 3 seconds, and keep the fluctuation less than ± speed ②.

[0061] Accordingly, the generated control command can be: set the target speed to speed ①, and adjust the time by 3 seconds. The relevant metallurgical actuators will then automatically adjust their speed.

[0062] In addition, the metallurgical control system monitors the control operations performed by the metallurgical equipment in real time to ensure stable rolling speed, thereby ensuring thickness consistency and completing the metallurgical optimization task.

[0063] This application is capable of receiving metallurgical optimization tasks, which include task optimization objectives. First, a knowledge graph reasoning path for the metallurgical optimization task is determined using a pre-constructed metallurgical knowledge graph. The knowledge graph reasoning path includes at least one metallurgical execution device. Then, based on the knowledge graph reasoning path and a pre-constructed device digital twin network, multimodal operating information associated with the metallurgical execution device is obtained. Subsequently, the multimodal operating information is analyzed for the task optimization objective to generate a metallurgical control strategy. This strategy is then used to control the metallurgical execution device to perform corresponding control operations, thereby completing the metallurgical optimization task.

[0064] Compared with related technologies, the technical solution of this application has the following advantages: First, by using a pre-constructed metallurgical knowledge graph, the knowledge graph reasoning path of the metallurgical optimization task is automatically determined, which helps to analyze the causal relationship of each metallurgical execution device in the metallurgical optimization task from an objective perspective, avoiding the one-sidedness and subjectivity caused by manual judgment. At the same time, combined with a pre-constructed equipment digital twin network, the multimodal operation information associated with the metallurgical execution devices can be obtained quickly and accurately, significantly improving the efficiency and accuracy of data acquisition, thereby improving the overall efficiency of the metallurgical production process. Second, after obtaining the multimodal operation information, the multimodal operation information can be differentiated for different task optimization objectives, generating a metallurgical control strategy that is highly matched with the task optimization objective. Based on the metallurgical control strategy, the metallurgical execution devices are precisely driven to perform corresponding control operations, realizing intelligent regulation of the metallurgical production process, thereby improving the level of intelligence of metallurgical control and meeting complex production needs.

[0065] Figure 2 This is another schematic flowchart illustrating a metallurgical control method in an embodiment of this application. Figure 2 relatively Figure 1 The technical solutions of the embodiments of this application are described in more detail.

[0066] See Figure 2 This method, applied to metallurgical control systems, includes at least the following steps: Step 201: Construct a metallurgical knowledge graph in the field of metallurgy.

[0067] In this application, the construction method of the metallurgical knowledge graph includes: acquiring historical multi-source heterogeneous data, extracting the equipment name and attribute information of metallurgical execution equipment from the historical multi-source heterogeneous data, using the equipment name as a metallurgical node, and establishing edge relationships between various metallurgical nodes based on the attribute information, constructing a semantic network containing metallurgical nodes and edge relationships, and using the semantic network as a metallurgical knowledge graph.

[0068] Historical multi-source heterogeneous data refers to a large amount of data related to the metallurgical field stored in metallurgical databases. This data accurately records process parameters, equipment operating status, causes of equipment failures, optimization schemes, and other information relevant to the metallurgical field, and can be used to construct professional metallurgical knowledge graphs. Due to the different data sources, inconsistencies often exist in data structure, data type, and data format, making it a typical example of multi-source heterogeneous data.

[0069] In response to the aforementioned historical multi-source heterogeneous data, the metallurgical control system first performs data preprocessing operations such as data cleaning, missing value completion, outlier removal, and format standardization on the historical multi-source heterogeneous data.

[0070] Subsequently, using structured parsing and natural language processing techniques, the equipment names and attribute information of the metallurgical equipment were extracted from the aforementioned data. This included equipment names such as blast furnaces, converters, continuous casting machines, roughing mills, and finishing mills, as well as attribute information such as equipment type, type of material processed, location number, and operating parameters.

[0071] Next, the equipment name of each metallurgical execution device is treated as an independent metallurgical node, and edge relationships between the metallurgical nodes are established based on attribute information. For example, upstream and downstream relationships such as "blast furnace → converter" and "roughing mill → finishing mill", and causal relationships such as "cooling water temperature → slab crack risk".

[0072] Finally, a semantic network containing metallurgical nodes and edge relationships is constructed, thus forming a metallurgical knowledge graph stored in graph structure, which is managed and maintained through the Neo4j graph database.

[0073] As an example, the process of constructing a metallurgical knowledge graph in the field of metallurgy includes at least the following steps: Step a) Data acquisition: Integrate diverse data such as process manuals, equipment drawings, and expert experience (unstructured text).

[0074] Step b) Knowledge extraction: Adopt the domain-adaptive BERT model (Bidirectional Encoder Representations from Transformers) to extract process rules such as "blast furnace molten iron [Si] content → fuel ratio" from text.

[0075] Step c) Graph generation: Use Neo4j to build a semantic network containing 370,000 nodes. Step d) Dynamic update: Real-time embedding of new standards is achieved through LoRA (Low-Rank Adaptation) adapters. For example, when national recommended standards are updated, the synchronization time is less than 10 minutes.

[0076] Step 202: Construct a digital twin network for the devices.

[0077] In this application, the construction method of the equipment digital twin network includes: acquiring historical multi-source heterogeneous data; extracting the operating condition parameters of the metallurgical execution equipment and the topological relationships between the various metallurgical execution equipment from the historical multi-source heterogeneous data; treating the metallurgical execution equipment as twin objects; and constructing equipment digital twins of the twin objects based on the operating condition parameters. Connecting the equipment digital twins according to the topological relationships generates an equipment digital twin network, which is used to simulate the operating state of the twin objects.

[0078] Similarly, after performing data preprocessing on historical multi-source heterogeneous data, the metallurgical control system extracts the operating parameters of the metallurgical actuators and the topological relationships between the various metallurgical actuators from the data using structured parsing and natural language processing techniques.

[0079] It is worth noting that topology refers to the connection method of various metallurgical execution equipment in terms of physical structure, control logic, and process flow. For example, the temperature curve of the blast furnace, the gas composition detection data of the converter, and the operating parameters such as the speed and load of the rolling mill, are all part of the topology of "blast furnace → converter → continuous casting machine → rolling mill".

[0080] The difference between topological relationships and edge relationships mentioned above is that topological relationships emphasize the actual physical connection structure between devices, while topological relationships emphasize the extensive connections between metallurgical execution devices. They may include multiple semantic levels such as edge relationships, attribute relationships, and attribution relationships, and are more focused on the semantic level of process. Edge relationships emphasize the semantic connection in the graph structure and are used to describe the semantic or causal relationship between two metallurgical nodes. Essentially, they are the "edges" in the graph.

[0081] Each metallurgical actuator is treated as a twin object, and a corresponding digital twin of the actuator is constructed based on its corresponding operating parameters. The digital twin of the actuator is a mirror image of the metallurgical actuator in virtual space, used to map the operating status of the actuator in real time, or to reflect its performance and potential trends.

[0082] All device digital twins are logically connected according to their topological relationships to construct a device digital twin network with a clear structure.

[0083] In practical applications, equipment digital twin networks can simulate the linkage relationship and overall operating status between equipment in the entire metallurgical production line, which is the key to realizing data visualization and real-time user interaction.

[0084] Step 203: Receive the metallurgical optimization task, which includes the task optimization objective.

[0085] In this application, the metallurgical control system can receive multiple metallurgical optimization tasks, and different metallurgical optimization tasks include different task optimization objectives.

[0086] In practical applications, metallurgical optimization tasks can be processed one by one according to the order in which the metallurgical control system receives the tasks, or they can be processed in order according to a pre-set priority.

[0087] As an example, users can set the priority of "Device Anomaly Handling Task" to be higher than that of "Daily Energy Consumption Optimization Task", that is, the former will be scheduled and executed first.

[0088] Step 204: Determine the knowledge graph reasoning path for the metallurgical optimization task using a pre-constructed metallurgical knowledge graph. The knowledge graph reasoning path includes at least one metallurgical execution device.

[0089] In this application, the metallurgical control system starts from the task optimization objective and unfolds layer by layer according to the causal relationship between nodes in the graph, gradually deducing the path of the metallurgical execution equipment that is closely related to achieving the metallurgical optimization objective.

[0090] When there is only one knowledge graph reasoning path in the metallurgical knowledge graph that satisfies the task optimization objective, the metallurgical control system can directly select that knowledge graph reasoning path as the reasoning basis for this metallurgical optimization task.

[0091] When there are multiple knowledge graph reasoning paths in the metallurgical knowledge graph that satisfy the task optimization objective, the metallurgical control system can use a scoring mechanism to select the most suitable path from the multiple knowledge graph reasoning paths.

[0092] Step 205: Based on the knowledge graph reasoning path and the pre-built digital twin network of the equipment, obtain multimodal operation information associated with the metallurgical execution equipment.

[0093] In this application, the process of obtaining multimodal operation information associated with metallurgical execution equipment based on knowledge graph reasoning paths and pre-constructed equipment digital twin networks includes: determining the metallurgical execution equipment used to perform metallurgical optimization tasks one by one based on knowledge graph reasoning paths, obtaining production rhythm data and energy consumption data of the metallurgical execution equipment, collecting sensor data of the equipment digital twin corresponding to the metallurgical execution equipment in the equipment digital twin network in real time, and fusing the sensor data, production rhythm data and energy consumption data to generate multimodal operation information.

[0094] Sensor data refers to the real-time acquisition of operating parameters from the corresponding equipment's digital twin by the metallurgical control system through data interaction with a pre-built network of equipment digital twins. For example, operating parameters such as temperature, pressure, vibration, current, voltage, and gas composition collected by intelligent sensors deployed on metallurgical actuators reflect the real-time operating status of the equipment.

[0095] Production rhythm data refers to equipment operation-related information obtained by the metallurgical control system from the MES system. For example, information such as the production cycle, work schedule, and output rate of metallurgical execution equipment reflects the production progress and efficiency of the metallurgical execution equipment.

[0096] Energy consumption data refers to equipment energy consumption-related information obtained by the metallurgical control system from the EMS system. For example, information such as power consumption and fuel usage of metallurgical actuators reflects the energy consumption of these actuators.

[0097] Sensor data, production schedule data, and energy consumption data are fused together to form unified multimodal operation information.

[0098] As an example, the process of fusing sensor data, production rhythm data, and energy consumption data includes time alignment and spatial alignment.

[0099] Time alignment refers to aligning sensor data, production rhythm data, and energy consumption data in a time-sensitive manner by designing a dynamic sliding window mechanism that adapts the window size to the differences in data frequency.

[0100] Spatial alignment refers to a cross-modal coordinate system transformation algorithm based on a calibration board, which reduces the error between the infrared thermal imager (pixel coordinates) and the vibration sensor (device coordinates) from 3mm to 0.8mm.

[0101] Step 206: Analyze the multimodal operation information to generate a metallurgical control strategy based on the task optimization objectives.

[0102] In this application, the metallurgical control system can selectively extract relevant data from multimodal operation information for different task optimization objectives, and combine it with a pre-trained metallurgical control model to output a control strategy that meets the objectives.

[0103] The pre-trained metallurgical control model refers to a model trained using machine learning or deep learning methods. This model can learn the operating rules, fault modes, and energy consumption characteristics of metallurgical equipment, enabling accurate analysis and decision-making for different metallurgical optimization tasks. During pre-training, parameters can be continuously optimized using a large number of labeled samples or unsupervised methods to improve the reasoning ability on the input data, thereby outputting effective metallurgical control strategies.

[0104] As an example, the metallurgical control model can be a Transformer-based optimization model. The input layer of this model embeds a metallurgical knowledge graph. For instance, the edge weights for "molten iron Si → furnace temperature" serve as priors, while the output layer implements a multi-task head to simultaneously predict energy consumption, quality, and equipment wear.

[0105] The training strategy includes pre-training a base model on 5 million historical process data points and using reinforcement learning to optimize the generation efficiency and accuracy of metallurgical control strategies.

[0106] As an optional example of this application, when the task optimization objective is defect detection, surface image data of the metallurgical execution equipment is extracted from the multimodal operation information, feature extraction is performed on the surface image data to obtain visual features and defect features of the surface image data, the visual features and defect features are mapped to the same feature space, a defect classification operation is performed to generate defect information of the metallurgical execution equipment, the defect information is input into the pre-trained metallurgical control model, and the metallurgical control model outputs the metallurgical control strategy.

[0107] Among them, the surface image data is mainly image data collected in real time by vision sensors deployed near the metallurgical execution equipment, which can clearly record the surface of the metallurgical execution equipment.

[0108] Feature extraction refers to the use of relevant computer vision technology to extract features from surface image data, thereby obtaining visual features such as texture, color, and shape, as well as defect features such as cracks, pits, and wear that may exist on the surface of metallurgical equipment.

[0109] Defect classification is mainly the process of identifying and classifying various defect features detected on the surface of metallurgical equipment. For example, it can be done by calculating the similarity between visual features and defect features, and then determining the corresponding defect information based on the similarity results, or by directly applying a deep learning model to process visual features and defect features and outputting defect information.

[0110] Defect information refers to the types of defects present on the surface of metallurgical equipment and their related characteristics, including detailed information such as the type of defect, its location, size, shape, and severity.

[0111] The defect information is further input into a pre-trained metallurgical control model. The metallurgical control model combines historical data and expert experience to output targeted metallurgical control strategies, such as arranging maintenance or taking preventive measures, to improve the safety and production efficiency of equipment operation.

[0112] For example, after extracting images of the strip surface of metallurgical actuators from multimodal operating information, the metallurgical control system uses the ViT model (VisionTransformer) to process the images. The ViT model first divides the image into multiple small blocks, and then extracts local and global visual features simultaneously through a self-attention mechanism. This accurately captures subtle defect features on the strip surface, such as scratches and oxide spots, thereby achieving high-precision defect identification and classification.

[0113] Building upon this foundation, the CLIP architecture is further introduced, mapping visual features extracted from images to the same feature space along with textual descriptions of defects. By calculating the similarity between image and textual features, accurate defect type identification is achieved, supporting advanced detection capabilities including "zero-shot inference." Even defect types not present in the training set can be identified and classified. Ultimately, the system outputs classification results and annotation information containing both known and unseen defects.

[0114] As another optional example of this application, when the task optimization objective is equipment health detection, the sensor time series data of the metallurgical execution equipment is extracted from the multimodal operation information, the degradation features in the sensor time series data are extracted through a multi-head self-attention mechanism, the degradation features are input into the pre-trained metallurgical control model, and the metallurgical control model outputs the metallurgical control strategy.

[0115] Among them, sensor time-series data refers to the dynamic operating parameter sequence continuously collected by various intelligent sensors deployed on metallurgical equipment, such as temperature, pressure, current, voltage, vibration frequency and other data that change over time.

[0116] For sensor time-series data, the system employs a multi-head self-attention mechanism for deep feature extraction. This mechanism, a deep learning technique derived from the Transformer architecture, can capture the correlations and dependencies between time points in a time series in parallel across multiple attention subspaces, thereby more accurately identifying abnormal trends and critical state changes during device operation.

[0117] Among the extracted deep features, the system further identifies "degradation features" that may reflect equipment performance decline, aging, or potential failure risks. Degradation features can predict the deterioration trend of equipment health.

[0118] The metallurgical control system inputs the aforementioned degradation characteristics into a pre-trained metallurgical control model, which then outputs corresponding metallurgical control strategies, such as maintenance warnings, to achieve predictive maintenance and health management of the equipment.

[0119] For example, the metallurgical control system receives high-frequency sensor time-series data such as vibration signals, temperature changes, and current fluctuations from metallurgical actuators. The sampling frequency can be as high as 20kHz, which can capture minute changes in the operation of the metallurgical actuators.

[0120] A temporal Transformer model is used to model high-frequency time-series data, and its multi-head self-attention mechanism is used to extract key features from long-term series. This mechanism can focus on important patterns in the data in parallel at different time scales.

[0121] The metallurgical control model learns from the development patterns of equipment aging or failure in historical data, extracts potential time series information, estimates the remaining service life of the equipment, and ensures that the prediction error does not exceed 8 hours.

[0122] As another optional example of this application, when the task optimization objective is energy efficiency optimization, the energy consumption data and operating parameters of the metallurgical execution equipment are extracted from the multimodal operation information, and the energy consumption data and operating parameters are input into the pre-trained metallurgical control model, and the metallurgical control model outputs the metallurgical control strategy.

[0123] Among them, the metallurgical control system analyzes energy consumption data and operating parameters together to identify energy consumption patterns under different operating conditions, thereby reducing unit energy consumption while ensuring production efficiency, or rationally allocating resources among multiple devices to achieve optimal overall energy efficiency. The corresponding output metallurgical control strategy can adjust equipment operating parameters, optimize production scheduling, or adjust resource allocation schemes.

[0124] For example, energy efficiency optimization includes building a steel grade gene map model based on historical data of steel grades, and the metallurgical control system can accurately recommend alloy composition and process parameters under performance indicators.

[0125] The metallurgical control system first extracts data including alloy element content, process parameters, and energy consumption from multimodal operation information to construct a three-layer metallurgical knowledge graph with a "composition-process-performance" structure. Nodes in this network represent specific elements or process parameters, and edge weights are determined by calculating the relationships between nodes, characterizing the interactions between various factors, and ultimately dynamically recommending composition adjustment schemes.

[0126] For example, energy efficiency optimization includes steelmaking process optimization and anomaly tracing based on knowledge graphs.

[0127] The metallurgical control system collects multi-source information, including steelmaking process parameters, molten iron chemical composition, and equipment operating parameters, to construct a metallurgical knowledge graph. This graph can reach hundreds of thousands of nodes, covering information such as production, composition, equipment status, and operating rules, and can comprehensively reflect the multi-dimensional relationships in the steelmaking process.

[0128] During the reasoning process, based on the embedded rule logic (such as...) Automatic recommendations can be made to improve the response speed and intelligence level of process control.

[0129] Meanwhile, when the system detects anomalies such as fluctuations in molten steel temperature or deviations in elements, it infers potential fault points by matching the co-occurrence paths of anomalies between metallurgical nodes. For example, it identifies "thermocouple drift" as the root cause of temperature anomalies, thereby supporting rapid location and source tracing analysis. The final output includes suggestions for optimizing steelmaking process parameters and fault root cause diagnosis results.

[0130] For example, energy efficiency optimization includes rolling process optimization based on multi-agent reinforcement learning, and metallurgical control systems can achieve coordinated control of rolling process parameters.

[0131] The metallurgical control system can define multiple agents, such as agent 1 (rolling mill control agent): with the task optimization goal of minimizing rolling energy consumption, it intelligently adjusts the reduction rate of each stand; agent 2 (cooling system agent): with the task optimization goal of improving the plate shape qualification rate, it adjusts the water cooling intensity and cooling distribution in real time.

[0132] To achieve overall optimization, the metallurgical control system introduces the Nash equilibrium strategy solution mechanism, which solves the optimal balance point between rolling energy consumption and plate shape qualification rate without compromising their respective objectives, and outputs the combination of rolling process parameters.

[0133] Step 207: Control the metallurgical execution equipment to perform corresponding control operations according to the metallurgical control strategy in order to complete the metallurgical optimization task.

[0134] In this application, the metallurgical control system sends the metallurgical control strategy generated by the metallurgical control model to the corresponding metallurgical execution equipment or its local control unit through the industrial control bus. The metallurgical execution equipment that receives the control strategy instruction adjusts its working status in real time according to the metallurgical control strategy.

[0135] Control operations are guided by set task optimization goals, such as minimizing energy efficiency, maximizing product quality, maintaining equipment health, and minimizing abnormal risks, to ensure that the metallurgical process maintains optimal operating conditions.

[0136] As an example, a metallurgical control system can first input the metallurgical control strategy into the equipment's digital twin network, simulate the execution effect, and generate performance evaluation results. If the performance evaluation results meet the requirements, the strategy is officially sent to the actual equipment for execution. If the performance evaluation results do not meet the requirements, the metallurgical control strategy will be adjusted and optimized to form a closed-loop feedback, further improving the control effect and the quality of metallurgical task completion.

[0137] In addition, the metallurgical control system also supports real-time feedback and visual decision-making, forming a closed-loop intelligent control process of perception-decision-execution-feedback.

[0138] Real-time feedback includes closed-loop control, anomaly detection and alarm, and carbon emission optimization.

[0139] In terms of closed-loop control, the metallurgical control system employs a reinforcement learning mechanism based on the PPO (Proximal Policy Optimization) algorithm, as well as constraint optimization directions contained in the metallurgical knowledge graph, such as the relationship between manganese content and steel toughness, to adjust the alloy addition ratio in real time. In practical applications, the metallurgical control system can achieve a parameter update frequency of once every 30 seconds, with a final carbon content control accuracy of ±0.02%.

[0140] In terms of anomaly detection, the metallurgical control system supports multimodal anomaly diagnosis. It combines the Flink streaming processing framework to perform spatiotemporal fusion processing on vibration signals within a 1-second window and infrared images at 10 frames per second, and uses a temporal Transformer model and a ViT model together for early fault detection. At the same time, the metallurgical control system utilizes zero-sample inference to identify unknown anomaly patterns, such as novel electrode cracks, even without historical samples.

[0141] In terms of green and intelligent control, the metallurgical control system defines a state space that includes flue gas composition, energy consumption data, and production rhythm data, and designs a reward function as R=α·output-β·carbon emissions, where α=0.7 and β=0.3, to guide the agent to improve production efficiency while suppressing carbon emission levels.

[0142] Furthermore, to ensure the efficient and stable operation of real-time feedback functions such as closed-loop control, anomaly detection, and carbon emission optimization in the metallurgical control system, the system integrates several key technologies. First, a recompute mechanism is used to dynamically recompute intermediate tensors in the deep learning model, effectively reducing GPU memory usage by approximately 35%, enabling the successful training of long-sequence LSTM (Long Short-Term Memory) networks and improving the learning capability of reinforcement learning strategies in closed-loop control. Second, a Flink-based streaming architecture enables low-latency real-time processing of Kafka data streams (latency less than 200 milliseconds), and window aggregation of anomaly detection scores. Finally, distributed deployment using Kubernetes (a container orchestration system) enables hot model updates, with version switching taking less than 10 seconds, ensuring the continuous iteration and stable operation of the metallurgical control system in carbon emission optimization.

[0143] Among them, the visualization objects for visual decision-making include metallurgical knowledge graphs, equipment digital twin networks, and intelligent reports.

[0144] The metallurgical knowledge graph utilizes WebGL (Web Graphics Library) for accelerated rendering, displaying a real-time relationship network (up to 370,000 nodes) covering the entire metallurgical process. It employs the Barnes-Hut force-directed algorithm combined with semantic clustering to automatically cluster similar process parameters, maintaining the graph's readability and responsiveness. For example, a user can click on any process indicator, such as blast furnace molten iron temperature, to track over 20 related influencing factors, enabling source analysis of process parameters. The force-directed graph dynamically displays the fault propagation path: bearing vibration → mill shutdown → production line blockage, helping users quickly identify risk chains.

[0145] The equipment digital twin network dynamically generates corresponding process digital twin views based on real-time operating data of metallurgical execution equipment, overlays key process indicators and knowledge graph reasoning paths, accurately reproduces changes in the physical state of the equipment, and synchronizes with actual process parameters in real time.

[0146] The intelligent reports are based on the DeepSeek-NLP (DeepSeek-Natural Language Processing) engine to convert structured data into natural language, automatically generating daily metallurgical reports and fault analysis reports. The intelligent reports cover key indicator trends, such as a 5% decrease in the converter's final carbon qualification rate today. They can also be correlated with cross-system data for analysis, with a root cause location accuracy of 89%, significantly improving decision-making efficiency.

[0147] In practical applications, the metallurgical control system is based on AR (Augmented Reality) technology to realize multiple real-time interaction methods, enabling immersive perception of the operating status of metallurgical equipment and intelligent assisted operation, and ensuring that the positional error between virtual content and real equipment is less than 2 millimeters.

[0148] Gesture interaction: Users can directly manipulate virtual models through natural movements. For example, they can use a grab gesture to retrieve and rotate a 3D model of a virtual blast furnace to view the spatial distribution of its internal temperature field, or use an air gesture to select equipment parts and retrieve a metallurgical knowledge graph.

[0149] Voice control: Supports natural language commands. For example, users can directly issue voice commands such as "Show the temperature rise curve of No. 3 tuyeres in the last hour" or "Show the heat load trend of No. 3 blast furnace last week," which will instantly overlay the corresponding historical curves or trend charts in the field of view to help determine whether there is a risk of overheating.

[0150] Intelligent prompts: When the system detects that the user's gaze is focused on a potentially faulty part of the equipment (such as a point of abnormal high-frequency vibration), it will automatically pop up a repair case video or operation manual related to that part.

[0151] The above-mentioned real-time interaction methods can effectively improve the intelligence level and response efficiency of metallurgical field operations.

[0152] To facilitate a deeper understanding of the metallurgical control method of the embodiments of this application, the metallurgical control system of the embodiments of this application will be described in detail below.

[0153] From the perspective of the core layer, the metallurgical control system includes a data simulation layer, a data processing and enhancement layer, an intelligent reasoning and decision-making layer, a real-time optimization and feedback layer, and a visualization decision support layer.

[0154] 1) Data simulation layer Time series data simulation: Based on the normal distribution and other statistical models, parameters such as temperature and pressure are generated as shown in Table 1. Gaussian noise is added to ensure that the error is controlled within 5%.

[0155] expression: .

[0156] Table 1

[0157] Defect image synthesis: Using OpenCV, images of metal surfaces with various defects such as scratches and pores are synthesized to enrich the training data.

[0158] 2) Data Processing and Enhancement Layer ① Multimodal data cleaning and noise reduction Time series data: Sliding window standardization (60 steps in length, 10 steps in step size) is used to normalize continuous time series data. At the same time, dynamic noise (standard deviation of 0.5% of the baseline value) is injected to enhance the robustness of the model. Key parameters are kept in FP32 (32-bit Floating Point) high precision to ensure that sensitive indicators such as temperature and carbon content are not distorted.

[0159] Visual data: Synthetic defect images (scratch length follows a Gamma distribution, and oxide spot radius follows a normal distribution) are generated using OpenCV (Open Source Computer Vision Library) to simulate real defects; nonlocal mean filtering is used for noise reduction to improve image quality.

[0160] Structured data: A metallurgical professional dictionary containing 32,000 terms was built, and an entity masking strategy was used to repair missing fields, thereby improving the integrity and accuracy of the structured data.

[0161] ② Multimodal data fusion Time alignment: Dynamic time warping (DTW) algorithm is used to match sensor data (such as vibration signals and temperature) with different sampling frequencies and asynchronous data, allowing for ±5% time deviation.

[0162] Spatial alignment: The coordinate system of the infrared camera and the equipment is calibrated using a calibration board (error less than 0.1mm) to achieve spatial correspondence between defect images and process parameters.

[0163] Fusion architecture: Temporal features are encoded using Transformer (e.g., rolling force sequence), and visual features are extracted by the ViT model (images are divided into 16×16 segments). These are then fused using a cross-attention mechanism, with dynamic gating to adjust the weights of different modalities. Structured process rules (e.g., vectorized IF-THEN statements) are injected as prior knowledge to enhance the model's understanding of the process flow.

[0164] ③Scarce data generation The GAN (Generative Adversarial Network) generates data using a U-Net architecture. It takes a noise vector and a working condition label (e.g., high-sulfur molten iron) as input and outputs synthesized spectral data with a wavelength range of 200-1000 nm. The discriminator is a convolutional neural network with spectral normalization, employing the Wasserstein distance loss function (weight λ=10) to augment the extreme working condition data, achieving an augmentation factor of up to 5 times.

[0165] Diffusion model generation: The forward process involves progressively adding noise in T=1000 steps, with cosine scheduling used for noise dispatching. The reverse process involves using Conditional U-Net, guided by process parameters, to generate physically plausible defect variants.

[0166] 3) Intelligent Reasoning and Decision-Making Layer ① Intelligent Defect Detection ViT Model + Zero-Shot Inference: High-precision classification of surface defects (scratches, oxide spots, etc.) on strip steel based on Vision Transformer. Combining a CLIP-like architecture (Contrastive Language–Image Pre-training) to align visual features with text descriptions (vertical scratches with a width > 2mm), zero-shot inference is achieved to detect defect types not covered by the training set.

[0167] ② Equipment health prediction Temporal Transformer Modeling: Input sensor sequences such as vibration and temperature (sampling frequency 20kHz), and capture long-cycle degradation features (such as bearing wear harmonics) through a multi-head attention mechanism. Output remaining service life prediction error ≤8 hours, supporting early maintenance decisions.

[0168] ③ Steel type gene map Construct a composition-process-performance correlation network (nodes = alloying elements / process parameters, edges = correlation coefficients) and dynamically recommend composition adjustment schemes (such as the effect of C content ±0.01% on strength).

[0169] ④ Knowledge graph optimization of steelmaking process Optimization of converter oxygen blowing volume and anomaly tracing based on DeepSeek-MetalKG (370,000 nodes).

[0170] Optimization of oxygen blowing rate in converter: Correlation between Si content in molten iron and endpoint carbon control rules ( ).

[0171] Anomaly tracing: Automatically match process deviations (such as temperature fluctuations) with equipment failure modes (such as thermocouple drift).

[0172] ⑤ Multi-agent reinforcement learning Rolling process optimization: Agent 1 (rolling mill): Adjusts the reduction rate with the goal of minimizing energy consumption.

[0173] Intelligent Agent 2 (Cooling System): Controls water cooling intensity with plate shape qualification rate as the target.

[0174] Dynamic game theory: Solving for the optimal parameter combination through Nash equilibrium.

[0175] ⑥ Visualized decision support Real-time generation of digital twin view of process, overlaying key indicators (such as blast furnace permeability index) and knowledge graph reasoning path.

[0176] 4) Real-time optimization and feedback layer ① Closed-loop control Dynamic steelmaking proportion optimization: Based on reinforcement learning (PPO algorithm), the alloy addition ratio is adjusted in real time, and the optimization direction is constrained by knowledge graph (such as the relationship between Mn content and toughness). The process parameters are updated every 30 seconds, and the final carbon content control accuracy reaches ±0.02%.

[0177] ② Anomaly detection and alarm Multimodal anomaly diagnosis: Streaming vibration signals (1-second Flink window) and infrared images (10fps) are used to detect early faults through temporal Transformer + ViT fusion. Zero-shot inference identifies unknown anomaly patterns (such as novel electrode cracks).

[0178] ③ Carbon emission optimization Reinforcement learning carbon footprint tracking: The state space includes flue gas composition ( Energy consumption data, production rhythm, and reward function R = α·output - β·carbon emissions (α = 0.7, β = 0.3).

[0179] 5) Visualized Decision Support Layer ①Dynamic knowledge graph visualization It renders a knowledge graph of the entire metallurgical process in real time (370,000 nodes), supports traceability of process parameters, allows users to track more than 20 related influencing factors by clicking on any indicator, and displays the fault propagation path. It presents the multi-level impact of equipment abnormalities in a force-directed graph, such as bearing vibration → mill shutdown → production line blockage.

[0180] ② Intelligent report generation Based on the DeepSeek-NLP engine: Automatically generate daily metallurgical reports (structured data → natural language), including key indicator trends, such as a 5% decrease in the converter's final carbon qualification rate today.

[0181] Fault Analysis Report: By correlating data from multiple systems, the root cause was located with an accuracy rate of 89%.

[0182] ③ Multimodal expert interaction Real-world overlay: 3D model of the equipment + real-time sensor data (such as temperature field of thermal imager).

[0183] Gesture-based knowledge graph access: Select equipment components with an air gesture to instantly display the maintenance case library.

[0184] 6) Model Optimization ① Data preprocessing For timing process data: sliding window normalization (60 steps, step size 10), dynamic noise injection: Key parameters retain FP32 accuracy (such as temperature and carbon content). For text log data: Construct a metallurgical domain dictionary and employ a special entity masking strategy.

[0185] For image data: Defect generation parameters include at least scratches and oxide spots. Among them, the scratch length Gamma (2, 5), angle U (0, 180°), oxide spot radius N (10, 2), and density Poisson (λ=3).

[0186] ② Multi-task model architecture The pseudocode is: CLASS MetallurgyModel EXTENDS DeepSeek METHOD __init__() INIT quality_head: LinearLayer(1024 → 6) INIT fault_head: LinearLayer(1024 → 20) END METHOD METHOD forward(inputs) base_output = SUPER().forward(inputs) quality_pred = quality_head(base_output.last_hidden_state) fault_pred = fault_head(base_output.mean(dim=1)) RETURN {main:base_output, quality:quality_pred, fault:fault_pred} END METHOD END CLASS ③ Training strategies Optimizer configuration: AdamW (lr=5e-5, β1=0.9, β2=0.98), gradient pruning threshold: 1.0, learning rate scheduling: .

[0187] Mixed precision training: Global precision FP16, while key calculations such as the loss function retain FP32 precision.

[0188] ④ Model distillation scheme The data used in the teacher model training process was the full metallurgical dataset (5 million samples), the architecture was DeepSeek-7B, and the training time was 80 GPU-hours (A100×8).

[0189] The parameter compression ratio of the student model is 1 / 8 (7B→900M), and the {2nd, 5th, 8th, 11th, 14th}th Transformer layers are retained.

[0190] Distillation loss function: Ltotal=0.4LKD+0.3Ltask+0.3LattnLtotal=0.4LKD+0.3Ltask+0.3Lattn.

[0191] Among them, knowledge distillation loss: Attention distillation loss: .

[0192] ⑤ Hardware adaptation optimization Optimize as shown in Table 2: Table 2

[0193] ⑥ Deployment Plan Deploy as shown in Table 3: Table 3

[0194] ⑦ Continuous learning mechanism Incremental update protocol: The update frequency is triggered every 1000 new data entries, and the update scope is only to fine-tune the parameters of the last 3 layers + adapter.

[0195] Version rollback: Retain the 5 most recent model versions.

[0196] Automatic rollback will occur when the following conditions are met: key indicators decrease by more than 15% or process parameters exceed safety limits.

[0197] refer to Figure 3 , Figure 3 This is a schematic diagram of the data platform of a metallurgical control system shown in an embodiment of this application. In this application, the metallurgical control system constructs a unified data platform, deeply integrating more than ten previously dispersed business systems, including a flue gas analysis system, a rolling intelligent control system, a converter fine model control system, a factory intelligent scheduling system, a power fire prevention and detection system, an unmanned overhead crane system, and an equipment IoT platform. Upon entering the homepage, users can clearly see overview information and early warning information, helping them make quick decisions.

[0198] refer to Figure 4 , Figure 4 This is a schematic diagram of the intelligent hub of a metallurgical control system shown in an embodiment of this application. This application uses the improved DeepSeek industrial big data model as the intelligent hub, achieving intelligent association of cross-system data through a metallurgical knowledge graph (370,000 nodes), thereby accurately analyzing metallurgical data. For example, when a blast furnace malfunctions, the system can automatically associate the molten iron composition, vibration signals, and process logs to deduce the knowledge graph reasoning path of "high-silicon molten iron → increased slag viscosity → overheating of the cooling wall." Furthermore, the intelligent hub supports natural language interaction. For example, if a user inputs a query such as "cause of plate / coil defect" or "please provide hot-rolling temperature information for the past month," the intelligent hub can comprehensively analyze quality data and process parameters to generate a corresponding report, shortening cross-system response time and significantly improving the level of intelligence in metallurgical production.

[0199] In practical applications, this data platform demonstrates three core values: First, it achieves true global visualization, allowing managers to monitor the entire process from raw material intake to finished product delivery in real time through a single interface, including key indicators such as current production progress, equipment operating status, and energy consumption trends. Second, it establishes an intelligent early warning mechanism. By analyzing cross-system data correlations through machine learning algorithms, the platform can proactively identify complex issues such as "an increase in blast furnace cooling wall temperature will lead to abnormal rolling mill loads in the future."

[0200] Most importantly, the deep integration of DeepSeek's industrial big data model has completely changed the traditional human-computer interaction mode. Users can ask "Which processes are abnormal today?" in natural language, and the metallurgical control system can automatically retrieve relevant data and generate a comprehensive report containing charts and analysis conclusions.

[0201] refer to Figure 5 , Figure 5 This is a schematic diagram of a digital twin network of equipment in a metallurgical control system, as illustrated in an embodiment of this application. The platform constructs a digital twin network covering the entire plant, collecting over 20 operating parameters in real time, including vibration, temperature, and current, through intelligent sensors deployed on key equipment. After feature extraction and preliminary analysis, these real-time data streams are correlated with production rhythm data in the MES system and energy consumption data in the EMS system through multi-dimensional correlation analysis. This facilitates real-time updates of the status of key equipment, such as normal status, maintenance-required status, and fault status, and identifies them as online equipment, maintenance-required equipment, and faulty equipment, respectively.

[0202] The platform's established equipment lifecycle database uses machine learning to analyze historical operating data, which can predict the remaining service life of key components such as motors and gearboxes 200-500 hours in advance, and intelligently recommend the optimal replacement time in the equipment partition.

[0203] For example, when the reducer of a converter shows early signs of wear, the system will not only immediately highlight the maintenance status and location of the equipment in the 3D plant model, but also automatically push a handling plan that includes the equipment's historical maintenance records, spare parts inventory, and standard maintenance procedures in the equipment partition.

[0204] In the embodiments of this application, the metallurgical control system integrates cutting-edge technologies such as deep learning, knowledge graphs, reinforcement learning and multimodal data fusion to model and optimize key links in the metallurgical production process, thereby improving equipment management efficiency, reducing energy consumption, and enhancing the ability to finely control product quality.

[0205] Furthermore, the metallurgical control system incorporates the DeepSeek industrial model to construct five collaborative subsystems, including a data platform, knowledge engine, intelligent optimizer, AR interactive system, and digital twin network. This enables intelligent collaboration across the entire chain from perception and cognition to decision-making, overcoming bottlenecks such as fragmented traditional technologies and dispersed systems. It effectively supports the metallurgical industry in dynamic adjustment and intelligent decision-making under complex working conditions.

[0206] Corresponding to the aforementioned application function implementation method embodiments, this application also provides a metallurgical control system, electronic equipment, and corresponding embodiments.

[0207] Figure 6 This is a schematic diagram of the structure of a metallurgical control system shown in an embodiment of this application.

[0208] See Figure 6 A metallurgical control system, comprising: The task receiving module 601 is used to receive metallurgical optimization tasks, which include task optimization objectives. The reasoning path determination module 602 is used to determine the knowledge graph reasoning path of the metallurgical optimization task through a pre-constructed metallurgical knowledge graph. The knowledge graph reasoning path includes at least one metallurgical execution device. The multimodal operation information acquisition module 603 is used to acquire multimodal operation information associated with metallurgical execution equipment based on knowledge graph reasoning paths and a pre-built equipment digital twin network; The metallurgical control strategy generation module 604 is used to analyze multimodal operation information and generate metallurgical control strategies in response to task optimization objectives. The optimization execution module 605 is used to control the metallurgical execution equipment to perform corresponding control operations according to the metallurgical control strategy in order to complete the metallurgical optimization task.

[0209] As an optional example of this application, the construction method of the metallurgical knowledge graph includes: acquiring historical multi-source heterogeneous data; extracting the equipment name and attribute information of metallurgical execution equipment from the historical multi-source heterogeneous data; using the equipment name as a metallurgical node, and establishing edge relationships between each metallurgical node based on the attribute information; constructing a semantic network containing metallurgical nodes and edge relationships, and using the semantic network as the metallurgical knowledge graph; wherein, when a metallurgical optimization task is received, the metallurgical knowledge graph is used to output the corresponding knowledge graph reasoning path based on the association relationship between each metallurgical execution equipment, and the knowledge graph reasoning path is used to represent the causal chain of the metallurgical execution equipment performing the metallurgical optimization task.

[0210] As an optional example of this application, the construction method of the equipment digital twin network includes: acquiring historical multi-source heterogeneous data; extracting the operating parameters of the metallurgical execution equipment and the topological relationship between each metallurgical execution equipment from the historical multi-source heterogeneous data; taking the metallurgical execution equipment as a twin object, and constructing the equipment digital twin of the twin object based on the operating parameters; connecting the equipment digital twins according to the topological relationship to generate the equipment digital twin network, which is used to simulate the operating state of the twin object.

[0211] As an optional example of this application, the multimodal operation information acquisition module 603 is used to: determine the metallurgical execution equipment used to perform metallurgical optimization tasks one by one based on the knowledge graph reasoning path; acquire the production rhythm data and energy consumption data of the metallurgical execution equipment; collect sensor data of the equipment digital twin corresponding to the metallurgical execution equipment in the equipment digital twin network in real time; and fuse the sensor data, production rhythm data and energy consumption data to generate multimodal operation information.

[0212] As an optional example of this application, the metallurgical control strategy generation module 604 includes: The defect detection strategy generation submodule is used to extract surface image data of metallurgical execution equipment from multimodal operation information when the task optimization objective is defect detection; perform feature extraction operation on the surface image data to obtain visual features and defect features of the surface image data; after mapping the visual features and defect features to the same feature space, perform defect classification operation to generate defect information of metallurgical execution equipment; input the defect information into the pre-trained metallurgical control model, and output the metallurgical control strategy from the metallurgical control model.

[0213] As an optional example of this application, the metallurgical control strategy generation module 604 includes: The equipment health detection generation submodule is used to extract sensor time-series data of metallurgical execution equipment from multimodal operation information when the task optimization objective is equipment health detection; extract degradation features from sensor time-series data through multi-head self-attention mechanism; input degradation features into pre-trained metallurgical control model, and output metallurgical control strategy from metallurgical control model.

[0214] As an optional example of this application, the metallurgical control strategy generation module 604 includes: The energy efficiency optimization strategy generation submodule is used to extract energy consumption data and operating parameters of metallurgical execution equipment from multimodal operation information when the task optimization objective is energy efficiency optimization; input the energy consumption data and operating parameters into the pre-trained metallurgical control model, and output the metallurgical control strategy from the metallurgical control model.

[0215] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0216] Figure 7 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application.

[0217] See Figure 7 The electronic device 700 includes a memory 710 and a processor 720.

[0218] The processor 720 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. Memory 710 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by the processor 720 or other modules of the computer. Permanent storage devices may be read-write storage devices. Permanent storage devices may be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use mass storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices may be removable storage devices (e.g., floppy disks, optical drives). System memory may be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory may store some or all of the instructions and data required by the processor during operation. Furthermore, memory 710 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (e.g., DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, memory 710 may include a removable storage device that is readable and / or writable, such as a laser disc (CD), a read-only digital versatile optical disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-high density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.

[0219] The memory 710 stores executable code, which, when processed by the processor 720, can cause the processor 720 to execute part or all of the methods described above.

[0220] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.

[0221] Alternatively, this application may be implemented as a computer-readable storage medium (or a non-transitory machine-readable storage medium or a machine-readable storage medium) storing executable code (or computer program or computer instruction code) that, when executed by a processor of an electronic device (or server, etc.), causes the processor to perform part or all of the steps of the methods described above according to this application.

[0222] This application also provides a computer program product, which includes computer instructions that, when executed by a processor, implement the method described above.

[0223] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A metallurgical control method, characterized in that, include: Receive a metallurgical optimization task, wherein the metallurgical optimization task includes a task optimization objective; The knowledge graph reasoning path of the metallurgical optimization task is determined by a pre-constructed metallurgical knowledge graph, and the knowledge graph reasoning path includes at least one metallurgical execution device. Based on the knowledge graph reasoning path and the pre-built device digital twin network, multimodal operation information associated with the metallurgical execution equipment is obtained; To optimize the task objectives, the multimodal operational information is analyzed to generate a metallurgical control strategy. The metallurgical execution equipment is controlled to perform corresponding control operations according to the metallurgical control strategy in order to complete the metallurgical optimization task.

2. The method according to claim 1, characterized in that, The construction methods of the metallurgical knowledge graph include: Acquire historical multi-source heterogeneous data; Extract the equipment name and attribute information of the metallurgical execution equipment from the historical multi-source heterogeneous data; The equipment name is used as a metallurgical node, and edge relationships between the metallurgical nodes are established based on the attribute information. Construct a semantic network containing the metallurgical nodes and the edge relationships, and use the semantic network as the metallurgical knowledge graph; The metallurgical knowledge graph is used to output a corresponding knowledge graph reasoning path based on the relationship between various metallurgical execution devices when the metallurgical optimization task is received. The knowledge graph reasoning path is used to represent the causal chain of the metallurgical execution devices executing the metallurgical optimization task.

3. The method according to claim 1, characterized in that, The construction methods of the device digital twin network include: Acquire historical multi-source heterogeneous data; The operating parameters of the metallurgical actuators and the topological relationships between the various metallurgical actuators are extracted from the historical multi-source heterogeneous data. The metallurgical execution equipment is used as a twin object, and a digital twin of the twin object is constructed based on the operating parameters. The device digital twins are connected according to the topology to generate a device digital twin network, which is used to simulate the operating state of the twin objects.

4. The method according to claim 1, characterized in that, The process of obtaining multimodal operational information associated with the metallurgical execution equipment based on the knowledge graph reasoning path and the pre-built device digital twin network includes: Based on the reasoning path of the knowledge graph, the metallurgical execution equipment used to perform the metallurgical optimization task is determined one by one; Obtain the production rhythm data and energy consumption data of the metallurgical execution equipment; Real-time acquisition of sensor data from the digital twin of the equipment corresponding to the metallurgical execution equipment in the equipment digital twin network; The sensor data, production rhythm data, and energy consumption data are fused to generate the multimodal operation information.

5. The method according to claim 1, characterized in that, The step of analyzing the multimodal operational information to generate a metallurgical control strategy, based on the task optimization objective, includes: When the task optimization objective is defect detection, the surface image data of the metallurgical execution equipment is extracted from the multimodal operation information; Perform feature extraction on the surface image data to obtain the visual features and defect features of the surface image data; After mapping the visual features and the defect features to the same feature space, a defect classification operation is performed to generate defect information of the metallurgical execution equipment. The defect information is input into a pre-trained metallurgical control model, which then outputs the metallurgical control strategy.

6. The method according to claim 1, characterized in that, The step of analyzing the multimodal operational information to generate a metallurgical control strategy, based on the task optimization objective, includes: When the task optimization objective is equipment health detection, the sensor timing data of the metallurgical execution equipment is extracted from the multimodal operation information; Degradation features in the sensor time-series data are extracted using a multi-head self-attention mechanism; The degradation features are input into a pre-trained metallurgical control model, and the metallurgical control model outputs the metallurgical control strategy.

7. The method according to claim 1, characterized in that, The step of analyzing the multimodal operational information to generate a metallurgical control strategy, based on the task optimization objective, includes: When the task optimization objective is energy efficiency optimization, the energy consumption data and operating parameters of the metallurgical execution equipment are extracted from the multimodal operation information; The energy consumption data and operating parameters are input into a pre-trained metallurgical control model, which then outputs a metallurgical control strategy.

8. A metallurgical control system, characterized in that, include: The task receiving module is used to receive metallurgical optimization tasks, wherein the metallurgical optimization tasks include task optimization objectives; The reasoning path determination module is used to determine the knowledge graph reasoning path of the metallurgical optimization task through a pre-constructed metallurgical knowledge graph, wherein the knowledge graph reasoning path includes at least one metallurgical execution device. The multimodal operation information acquisition module is used to acquire multimodal operation information associated with the metallurgical execution equipment based on the knowledge graph reasoning path and the pre-built equipment digital twin network; The metallurgical control strategy generation module is used to analyze the multimodal operation information and generate a metallurgical control strategy in response to the task optimization objectives. The optimization execution module is used to control the metallurgical execution equipment to perform corresponding control operations according to the metallurgical control strategy, so as to complete the metallurgical optimization task.

9. An electronic device, characterized in that, include: processor; as well as A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores executable code that, when executed by a processor of an electronic device, causes the processor to perform the method as described in any one of claims 1-7.

Images