A networking method and device for steel production if-then rule system

Abstract

The application discloses a steel production IF-THEN rule system network method and device, the method comprises the following steps: the IF-THEN rule in the field of steel production is disassembled into explicit rule features and original data features; a fusion network containing explicit rule paths and implicit data paths is constructed to maintain rule logic and learning data implicit rules respectively; a loss function containing rule constraint terms is used to train the network, so that it fits the original conclusion on the rule trigger sample and fits the actual control result on all samples; real-time reasoning and decision explanation are carried out using the trained network; and after network deployment, autonomous optimization is carried out through an incremental learning mechanism. The application converts the traditional rule system into a trainable, interpretable and autonomous evolution network model, effectively solves the problems of rule expansion and conflict, poor generalization ability and low iteration efficiency, and significantly improves the intelligent level, self-adaptive ability and decision reliability of steel production control.

Description

Technical Field

[0001] This application relates to the field of steel production control systems, and in particular to a networked method and apparatus for a steel production IF-THEN rule system. Background Technology

[0002] In various process systems of steel production, IF-THEN rules, as a traditional means of achieving automated decision-making, are constructed by domain experts based on experience and have the advantages of clear logic and strong interpretability. However, with the increasing complexity of production processes, the application of traditional rule systems in the steel production field faces severe challenges. First, the number of rules continues to grow with changes in operating conditions and the accumulation of experience, easily leading to mutual exclusion or coverage gaps between rules, resulting in control conflicts and process fluctuations. Second, existing rule systems can only handle preset operating conditions and lack adaptability to undefined new scenarios such as fluctuations in raw material composition and equipment aging, exhibiting poor generalization. Furthermore, rule updates heavily rely on manual maintenance, cannot utilize massive amounts of historical data for automatic optimization, struggle to cope with dynamic production environments, and suffer from low iteration efficiency.

[0003] In existing technologies, although some studies have attempted to combine rule systems with data-driven models, they are mostly limited to specific non-industrial fields (such as advertising recommendations) or rely on specific network architectures. They have failed to fully consider the characteristics of steel production processes, such as strong coupling of multiple parameters, strong temporality, and high security constraints, and lack a general, architecture-independent network mapping method for rule systems.

[0004] Therefore, there is an urgent need for a networked rule system that can effectively integrate explicit rule knowledge and implicit data patterns, has adaptive learning capabilities, and maintains interpretability, in order to solve problems such as rule expansion and conflict, poor generalization ability, and low iteration efficiency in steel production, and improve the level of intelligent production control. Summary of the Invention

[0005] This application provides a networked method and apparatus for an IF-THEN rule system in steel production, which can overcome the problems of rule expansion and conflict, insufficient generalization ability and low iteration efficiency faced by existing IF-THEN rule systems in steel production.

[0006] The first aspect of this application provides a networked method for an IF-THEN rule system in steel production, comprising: The IF-THEN type rules in the steel production field are decomposed into steel production condition parameters, comparison operators, thresholds and control conclusion actions, and explicit rule features and original data features are generated. Construct a fusion network that includes explicit rule paths and implicit data paths. The explicit rule paths are used to process explicit rule features to maintain the rule triggering logic, while the implicit data paths are used to process raw data features to learn the implicit patterns in the production parameters. Based on a historical dataset containing both rule-triggered and non-triggered samples, a loss function containing rule constraints is used to train the fusion network, enabling the fusion network to fit the original rule conclusions on rule-triggered samples and fit the actual control results on all samples. Real-time inference is performed using a trained fusion network. The final control decision is generated based on the outputs of explicit rule paths and implicit data paths through a weighted fusion mechanism or attention allocation mechanism. Explanatory information associated with the final control decision is output. The explanatory information includes at least one or more of the following: the original steel production rules that were triggered, the key features that affect the control decision and their contribution, and the matching degree information with similar scenarios in the history of steel production. After the converged network is put into use, it autonomously optimizes the network parameters based on the newly added production data through an incremental learning mechanism.

[0007] Optionally, explicit rule features include Boolean features and cross features generated by the combination of steel production rule conditions, used to characterize the explicit triggering logic of the rules; raw data features include the raw values ​​of continuous parameters of steel production and the embedded vectors of discrete parameters, used to characterize the state of the unprocessed production system.

[0008] Optionally, the IF-THEN type rule in the steel production field can be decomposed into steel production condition parameters, comparison operators, thresholds, and control conclusion actions, including: When the condition part of an IF-THEN rule contains "OR" logic, in the parsing step, the IF-THEN rule is decomposed into multiple sub-rules, the condition of each sub-rule is "AND" logic, and corresponding cross features are generated for each sub-rule.

[0009] Optionally, a fusion network comprising explicit rule paths and implicit data paths is constructed, including: The explicit rule path receives explicit rule features as input and processes them using a linear layer, factorization machine layer, cross network layer, or rule node embedding layer to adapt to the rule logic of steel production. The implicit data path receives raw data features as input and processes them using deep fully connected networks, self-attention layers, or feature map networks to adapt to the nonlinear relationships between steel production parameters.

[0010] Optionally, the architecture of the fusion network can be selected from WIDE&DEEP networks, deep factorization machines, deep cross networks, Transformers, or graph neural networks.

[0011] Optionally, the loss function includes a basic loss term and a rule constraint loss term; The basic loss term is used to measure the deviation between the model's predicted output and the actual control result. Mean squared error is used for regression tasks, and cross-entropy loss is used for classification tasks. The rule constraint loss term is used to penalize the deviation between the model output and the original rule conclusion on rule-triggered samples; The loss function adjusts the weight of the rule constraint loss term in the total loss through the rule constraint coefficient.

[0012] Optionally, after the converged network is put into use, the network parameters are optimized online using an incremental learning mechanism based on the newly added production data, including: Continuously collect real-time input characteristics, decision outputs of the fusion network, and corresponding actual process results or product quality feedback during the steel production process; A sliding window mechanism is used to periodically fine-tune the parameters of the fusion network using newly added production data; When the deviation between the decision output of the fusion network and the manual intervention instructions or process quality indicators exceeds the safety threshold, an emergency update process is triggered to adjust the parameters of the fusion network using data from the relevant scenario.

[0013] The second aspect of this application provides a networked device for an IF-THEN rule system in steel production, comprising: The rule parsing unit is used to decompose IF-THEN type rules in the steel production field into steel production condition parameters, comparison operators, thresholds and control conclusion actions, and generate explicit rule features and original data features. Network building units are used to construct a fusion network that includes explicit rule paths and implicit data paths. The explicit rule paths are used to process explicit rule features to maintain the rule triggering logic, while the implicit data paths are used to process raw data features to learn the implicit patterns in the production parameters. The training unit is used to train the fusion network based on a historical dataset containing rule-triggered and non-triggered samples, using a loss function that includes rule constraints, so that the fusion network fits the original rule conclusions on rule-triggered samples and fits the actual control results on all samples. The reasoning and decision-making unit is used to perform real-time reasoning using the trained fusion network. It generates the final control decision based on the output of the explicit rule path and the implicit data path through a weighted fusion mechanism or attention allocation mechanism, and outputs explanatory information associated with the final control decision. The explanatory information includes at least one or more of the following: the original steel production rules that were triggered, the key features that affect the control decision and their contribution, and the matching degree information with similar scenarios in the history of steel production. The autonomous optimization unit is used to autonomously optimize the parameters of the fusion network based on newly added production data through an incremental learning mechanism after the fusion network is put into use.

[0014] A third aspect of this application provides a networked device for an IF-THEN rule system in steel production, comprising: One or more processors; A memory on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the networked method for the IF-THEN rule system of steel production as described in any of the preceding claims.

[0015] The fourth aspect of this application provides a computer storage medium for storing a program, which, when executed, is used to implement the networked method for the IF-THEN rule system of steel production as described in any of the preceding claims.

[0016] This invention transforms a traditional static rule system into a trainable, interpretable, and autonomously optimizable dual-path fusion network model. This invention not only fully preserves the interpretability and reliability of expert rule knowledge but also significantly enhances the system's generalization and adaptability to unknown operating conditions. It achieves complementary advantages between rule logic and data patterns, effectively manages rule conflicts, reduces maintenance costs, and achieves autonomous evolution through continuous learning of the production environment. Therefore, it comprehensively improves the intelligent control level, decision-making quality, and operational efficiency of the steel production process. Specifically, compared to existing technologies, it has the following beneficial effects: 1. Achieving the networking and learnability of the steel production rule system: By analyzing the steel production rules in a structured manner as learnable features, the traditional static rule system is transformed into a trainable neural network model, solving the technical problem that it cannot be optimized autonomously; 2. Strong adaptability to steel production: Stripping away the generalized domain characteristics of rules, focusing on steel production process parameters and control logic, and providing a rule-feature-network mapping framework specifically adapted to blast furnace ironmaking, converter steelmaking and other processes; 3. Network architecture independence: The core innovation lies in the dual-path feature fusion mechanism, rather than a specific network structure, which enhances the technical flexibility and scalability for different processes in steel production; 4. Balancing rule knowledge retention and generalization ability: Explicit rule paths ensure the accurate execution of known rules in steel production, while implicit data paths enhance adaptability to uncovered scenarios such as raw material fluctuations and equipment aging. 5. Strong autonomous iteration capability: The model can be dynamically optimized through online incremental learning, reducing the maintenance cost of steel production rules and adapting to the temporal changes of the production system. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a networked method for an IF-THEN rule system in steel production, provided as an embodiment of this application; Figure 2 A schematic diagram of the structure of a networked device for an IF-THEN rule system in steel production, provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a networked device for an IF-THEN rule system in steel production, provided as an embodiment of this application. Detailed Implementation

[0019] This application provides a networked method and apparatus for an IF-THEN rule system in steel production, which can overcome the problems of rule expansion and conflict, insufficient generalization ability and low iteration efficiency faced by existing IF-THEN rule systems in steel production.

[0020] To facilitate understanding, the application scenarios of the embodiments of this application will be introduced first.

[0021] With the rapid development of the steel industry and the increasing complexity of production processes, traditional IF-THEN rule-based control systems have been widely used in steel production. These rule-based systems are typically developed by domain experts based on experience, existing in the form of conditional judgments and control actions to provide decision support for the steel production process. However, with the expansion of production scale and the increase in process complexity, traditional rule-based systems face numerous challenges.

[0022] In steel production, rule-based systems need to handle multi-source heterogeneous data, including continuous time-series data, discrete state data, and various process parameters. CN111832703B discloses a method for modeling irregular sampling dynamic sequences in process manufacturing industries. This method can handle irregular sampling data and nonlinear dynamic characteristics in process manufacturing industries, improving the applicability and accuracy of soft measurement models. However, this method mainly focuses on processing single data streams and fails to effectively solve the problem of collaborative analysis of multi-source data in steel production.

[0023] With the development of artificial intelligence technology, deep learning methods have shown great potential in the field of industrial production. CN112150384A proposes a method and system based on the fusion of residual networks and dynamic convolutional networks, which improves model performance by fusing multiple training networks. This network fusion idea provides a new approach to solving decision-making problems in complex industrial scenarios, but this method is mainly aimed at the field of image processing and fails to fully consider the special characteristics of rule-based systems in steel production.

[0024] In the area of ​​digital transformation of steel production, CN117930786B describes an intelligent digital twin simulation system for steel production processes. This system improves the intelligence and adaptability of steel production processes by integrating advanced modeling, simulation, optimization, and decision support technologies. The system can more accurately predict changes in the production process, adjust model parameters in real time, and effectively cope with fluctuations in raw material quality and energy costs. However, the system fails to fully utilize existing expert rule knowledge, resulting in insufficient model interpretability.

[0025] At the production management level, CN120338430B discloses a deep learning-based digital collaborative management method and system for MES (Manufacturing Execution System). This method analyzes the dynamic relationships between data items through node feature embedding and edge relationship learning, generating a production line operation dependency graph. This method has advantages in handling multi-source heterogeneous data, but it fails to be specifically optimized for rule-based systems in steel production.

[0026] Furthermore, CN111932044A proposes a machine learning-based steel product price forecasting system and method. This method uses multi-source data and leverages machine learning models for modeling and training, providing price forecasts that are highly applicable, accurate, and stable. This indicates that machine learning methods have broad application prospects in the steel industry, but this method primarily focuses on price forecasting rather than production control.

[0027] The following problems still exist in the application of existing technologies to the IF-THEN rule system in the steel production field: First, traditional rule systems face the problems of rule expansion and conflict. Steel production processes are complex, involving multiple interconnected stages and parameter coupling. As production conditions change and experience accumulates, the number of rules continues to grow, easily leading to mutually exclusive rules or overlapping gaps, resulting in control contradictions or process fluctuations. Current technologies lack effective mechanisms to manage and coordinate the relationships between a large number of complex rules.

[0028] Second, existing rule systems have poor generalization capabilities. Traditional IF-THEN rules can only handle preset operating conditions and lack adaptability to undefined new scenarios, such as fluctuations in raw material composition, equipment aging, and sudden process disturbances. Although machine learning methods have a certain generalization ability, existing technologies have failed to effectively combine rule knowledge with data-driven methods, making it difficult to improve system adaptability while maintaining rule interpretability.

[0029] Third, the rule system suffers from low iteration efficiency. Existing rule updates mainly rely on manual maintenance, making it impossible to automatically optimize using the massive historical data accumulated in steel production, and thus difficult to cope with dynamic changes in the production environment. Although some studies have proposed data-driven model update methods, they have failed to design effective online learning mechanisms tailored to the characteristics of the rule system.

[0030] Fourth, existing technologies lack methods to organically integrate explicit rule knowledge with implicit data patterns. Traditional rule systems and data-driven methods are often separated, failing to fully leverage their complementary advantages, resulting in systems that are either highly interpretable but poorly adaptable, or highly adaptable but poorly interpretable.

[0031] Therefore, in order to solve the technical problems of rule expansion and conflict, poor generalization ability and low iteration efficiency in the IF-THEN rule system in the steel production field, and to achieve technical effects such as network and learnability of the rule system, strong adaptability to steel production, network architecture independence, balance between rule knowledge preservation and generalization ability, and strong autonomous iteration ability, this invention provides a "rule-feature-network" mapping framework to transform the IF-THEN rule system in the steel production field into a trainable neural network model.

[0032] See Figure 1 This figure is a schematic flowchart of a networked method for an IF-THEN rule system in steel production, provided in an embodiment of this application. The networked method for an IF-THEN rule system in steel production provided in this embodiment can be implemented, for example, through the following steps 101-105.

[0033] S101: Decompose the IF-THEN type rules in the steel production field into steel production condition parameters, comparison operators, thresholds and control conclusion actions, and generate explicit rule features and original data features.

[0034] In this embodiment, the explicit rule features include Boolean features and cross features generated from the combination of steel production rule conditions, used to characterize the explicit triggering logic of the rule; the original data features include the original values ​​of continuous parameters of steel production and the embedded vectors of discrete parameters, used to characterize the state of the unprocessed production system. When the condition part of the IF-THEN rule contains "OR" logic, in the parsing step, the IF-THEN rule is decomposed into multiple sub-rules, the condition of each sub-rule is "AND" logic, and corresponding cross features are generated for each sub-rule.

[0035] Specifically, the IF-THEN type rules in steel production are decomposed into steel production condition parameters, comparison operators, thresholds, and control conclusion actions, generating explicit rule features and raw data features. Explicit rule features include Boolean features and cross features generated from combinations of steel production rule conditions, used to characterize the explicit triggering logic of the rules. Raw data features include the raw values ​​of continuous steel production parameters and the embedded vectors of discrete parameters, used to characterize the unprocessed production system state. In this step, steel production rules containing "OR" logic are decomposed into multiple sub-rules, with each sub-rule generating its own cross feature. For example, the rule "If temperature > 1200℃ or pressure > 50MPa, reduce heating power" is decomposed into two sub-rules: "If temperature > 1200℃, reduce heating power" and "If pressure > 50MPa, reduce heating power," each generating its own cross feature. For complex rules commonly found in steel production, such as those containing "OR" logic, this module can automatically decompose them into multiple sub-rules for processing, ensuring the integrity and accuracy of the rule logic.

[0036] S102: Construct a fusion network that includes explicit rule paths and implicit data paths.

[0037] In this embodiment, the explicit rule path processes explicit rule features to maintain the rule triggering logic, while the implicit data path processes raw data features to learn the implicit patterns in production parameters. The explicit rule path receives explicit rule features as input and processes them using linear layers, factorization machine layers, cross-network layers, or rule node embedding layers to adapt to the steel production rule logic. The implicit data path receives raw data features as input and processes them using deep fully connected networks, self-attention layers, or feature graph networks to adapt to the nonlinear relationships between steel production parameters. The architecture of the fusion network is selected from WIDE&DEEP networks, deep factorization machine networks, deep cross-networks, Transformers, or graph neural networks.

[0038] Specifically, a fusion network comprising explicit rule paths and implicit data paths is constructed to adapt to the coupling and temporal characteristics of steel production parameters. The explicit rule path receives explicit rule features as input and processes them using linear layers, FM layers, Cross layers, or rule node embedding layers, adapting to the rule logic of steel production. The implicit data path receives raw data features as input and processes them using deep fully connected networks, self-attention layers, or feature graph networks, adapting to the nonlinear correlation of steel production parameters. The architecture of this fusion network can be selected from WIDE&DEEP, DeepFM, DCN, Transformer, or Graph Neural Network (GNN), but is not limited to these architectures, all adapting to the characteristics of the steel production scenario. The output of the fusion network generates the final control decision through a weighted fusion mechanism. The initial value of the fusion coefficient α is 0.7, which can be a fixed value or a learnable parameter, adapting to the control priorities of steel production.

[0039] S103: Based on the historical dataset containing rule-triggered and non-triggered samples, the fusion network is trained using a loss function that includes rule constraints.

[0040] In this embodiment, the fusion network fits the original rule conclusions on rule-triggered samples and fits the actual control results on all samples. The loss function includes a basic loss term and a rule constraint loss term; the basic loss term measures the deviation between the model's predicted output and the actual control results, with mean squared error used for regression tasks and cross-entropy loss used for classification tasks; the rule constraint loss term penalizes the deviation between the model output and the original rule conclusions on rule-triggered samples; the loss function adjusts the weight of the rule constraint loss term in the total loss through the rule constraint coefficient.

[0041] Specifically, based on historical datasets from the steel production field that include both rule-triggered and non-triggered samples, a loss function incorporating rule constraints is used to train the fusion network, ensuring that the model output simultaneously aligns with both the original rule conclusions and actual control results in steel production. Online optimization is then performed after the model is deployed. The loss function is: ; in, The basic loss is the mean squared error (MSE) for regression and the cross-entropy for classification. To penalize the deviation between the model output and the rule conclusion on the rule-triggered sample; This is the rule constraint coefficient, which is adjusted according to the degree of dependence of the domain on the rule.

[0042] S104: Use the trained fusion network for real-time inference, generate the final control decision based on the output of explicit rule paths and implicit data paths through a weighted fusion mechanism or attention allocation mechanism, and output explanatory information associated with the final control decision.

[0043] In the embodiments of this application, the explanatory information includes at least one or more of the following: the original rules of steel production that were triggered, the key features that affect the control decision and their contribution, and the matching degree information with similar scenarios in the history of steel production.

[0044] Specifically, the system outputs the original rules triggered by the steel production process, the key features influencing control decisions and their contributions, as well as matching information with similar historical steel production scenarios. This step is implemented through the decision output module of the fusion network, providing interpretable decision results. For example, when the system suggests adjusting the rolling speed, it simultaneously displays the triggered original rules (e.g., "Slow down when steel plate thickness deviation > 0.2mm and surface temperature < 850℃"), key influencing factors (e.g., "Thickness deviation contribution 65%, surface temperature contribution 35%), and similar historical cases (e.g., "92% similarity to April 15, 2023, 09:30, the measures taken at that time were..."). This information helps operators understand the basis of the system's decisions, increasing their trust in the AI ​​system. This transparent decision-making mechanism enables operators to understand the system's decision-making basis, increasing their trust in the system and facilitating its application in actual production. For example, in the hot rolling control subsystem, when the system suggests adjusting the rolling speed, it simultaneously displays the triggering rules, key influencing factors and their contributions, as well as similar historical cases, helping operators make more reasonable decisions.

[0045] S105: After the converged network is put into use, the network parameters are autonomously optimized based on the newly added production data through an incremental learning mechanism.

[0046] In this embodiment, real-time input features, decision outputs of the fusion network, and corresponding actual process results or product quality feedback during the steel production process are continuously collected. A sliding window mechanism is adopted to periodically fine-tune the parameters of the fusion network using newly added production data. When the deviation between the decision output of the fusion network and the manual intervention instructions or process quality indicators exceeds a safety threshold, an emergency update process is triggered to adjust the parameters of the fusion network using data from relevant scenarios.

[0047] Specifically, after the model is deployed in the steel production system, it uses an incremental learning mechanism to dynamically update network parameters using newly added production data, enabling the rule system to evolve autonomously—this is the online optimization mechanism. For example, when a rule is found to be no longer applicable under new operating conditions, the model can automatically adjust the rule's weight or modify its triggering conditions, thereby adapting to changes in steel production processes and fluctuations in equipment status, maintaining the stability and advancement of control effects.

[0048] This application provides a networked method for the IF-THEN rule system in steel production, which realizes the networking of the IF-THEN rule system in the steel production scenario. It retains the interpretability of the traditional rule system and introduces the adaptive capability of deep learning, thereby improving the intelligence level and efficiency of steel production process control.

[0049] Based on the methods provided in the above embodiments, this application also provides a networked device for the IF-THEN rule system in steel production. The following describes the networked device for the IF-THEN rule system in steel production in conjunction with the accompanying drawings.

[0050] See Figure 2 The figure is a schematic diagram of the structure of a networked device for an IF-THEN rule system in steel production provided in an embodiment of this application.

[0051] The steel production IF-THEN rule system network device 200 provided in this application embodiment includes: a rule parsing unit 201, a network construction unit 202, a training unit 203, a reasoning and decision-making unit 204, and an autonomous optimization unit 205. Each module works collaboratively to realize the intelligentization and networking of the steel production rule system.

[0052] The device adopts a modular design and can be integrated into the steel production control system in the form of software, firmware, or hardware, including the blast furnace ironmaking control subsystem, converter steelmaking control subsystem, continuous casting forming control subsystem, or hot rolling control subsystem. This integration method enables the device to seamlessly connect with existing steel production control systems, improving the overall intelligence level of the system.

[0053] The rule parsing unit 201 is used to decompose IF-THEN type rules in the steel production field into steel production condition parameters, comparison operators, thresholds and control conclusion actions, and generate explicit rule features and original data features. Network building unit 202 is used to build a fusion network containing explicit rule paths and implicit data paths, wherein the explicit rule paths are used to process explicit rule features to maintain the rule triggering logic, and the implicit data paths are used to process raw data features to learn the implicit patterns in the production parameters. Training unit 203 is used to train the fusion network based on a historical dataset containing rule-triggered samples and non-triggered samples, using a loss function containing rule constraint terms, so that the fusion network fits the original rule conclusions on rule-triggered samples and fits the actual control results on all samples. The reasoning and decision-making unit 204 is used to perform real-time reasoning using the trained fusion network. It generates the final control decision based on the output of the explicit rule path and the implicit data path through a weighted fusion mechanism or an attention allocation mechanism, and outputs explanatory information associated with the final control decision. The explanatory information includes at least one or more of the following: the original steel production rules that were triggered, the key features that affect the control decision and their contribution, and the matching degree information with similar scenarios in the history of steel production. The autonomous optimization unit 205 is used to autonomously optimize the parameters of the fusion network based on newly added production data through an incremental learning mechanism after the fusion network is put into use.

[0054] In one possible implementation, the explicit rule features include Boolean features and cross features generated by the combination of steel production rule conditions, used to characterize the explicit triggering logic of the rules; the raw data features include the raw values ​​of continuous parameters of steel production and the embedded vectors of discrete parameters, used to characterize the state of the unprocessed production system.

[0055] In one possible implementation, the rule parsing unit 201 is specifically used for: When the condition part of an IF-THEN rule contains "OR" logic, in the parsing step, the IF-THEN rule is decomposed into multiple sub-rules, the condition of each sub-rule is "AND" logic, and corresponding cross features are generated for each sub-rule.

[0056] In one possible implementation, the network building unit 202 has features for: The explicit rule path receives explicit rule features as input and processes them using a linear layer, factorization machine layer, cross network layer, or rule node embedding layer to adapt to the rule logic of steel production. The implicit data path receives raw data features as input and processes them using deep fully connected networks, self-attention layers, or feature map networks to adapt to the nonlinear relationships between steel production parameters.

[0057] In one possible implementation, the architecture of the fusion network is selected from WIDE&DEEP networks, deep factorization machine networks, deep cross networks, Transformers, or graph neural networks.

[0058] In one possible implementation, the loss function includes a basic loss term and a rule constraint loss term; The basic loss term is used to measure the deviation between the model's predicted output and the actual control result. Mean squared error is used for regression tasks, and cross-entropy loss is used for classification tasks. The rule constraint loss term is used to penalize the deviation between the model output and the original rule conclusion on rule-triggered samples; The loss function adjusts the weight of the rule constraint loss term in the total loss through the rule constraint coefficient.

[0059] In one possible implementation, the autonomous optimization unit 205 has the following functions: Continuously collect real-time input characteristics, decision outputs of the fusion network, and corresponding actual process results or product quality feedback during the steel production process; A sliding window mechanism is adopted to periodically fine-tune the parameters of the fusion network using newly added production data; When the deviation between the decision output of the fusion network and the manual intervention instructions or process quality indicators exceeds the safety threshold, an emergency update process is triggered to adjust the parameters of the fusion network using data from the relevant scenario.

[0060] In one possible implementation, the device is integrated into the blast furnace ironmaking control subsystem in the form of software. By collecting parameters such as blast furnace temperature, pressure, and gas composition in real time and combining them with an expert rule base, the device can intelligently control the blast furnace smelting process, effectively improving the stability of molten iron quality and energy utilization efficiency.

[0061] In one possible implementation, the device is integrated into the continuous casting forming control subsystem as firmware. By monitoring key parameters such as molten steel temperature, crystallizer level, and billet pulling speed, and in conjunction with continuous casting process rules, it achieves precise control of billet quality and stable operation of continuous casting production.

[0062] Since the device 200 is the same as the networked method for the IF-THEN rule system in steel production provided in the above method embodiments, the specific implementation of each unit of the device 200 is based on the same concept as in the above method embodiments. Therefore, for the specific implementation of each unit of the device 200, please refer to the description of the networked method for the IF-THEN rule system in steel production in the above method embodiments, and it will not be repeated here.

[0063] This application embodiment also provides a networked device for an IF-THEN rule system in steel production, the device comprising: a processor and a memory; The memory is used to store instructions; The processor is used to execute the instructions in the memory to perform the networked method for the IF-THEN rule system in steel production mentioned in the above embodiments.

[0064] It should be noted that the hardware structure of the networked equipment for the steel production IF-THEN rule system provided in this application embodiment can be as follows: Figure 3 The structure shown, Figure 3 This is a schematic diagram of the structure of a device provided in an embodiment of this application.

[0065] Please see Figure 3 As shown, device 300 includes: a processor 310, a communication interface 320, and a memory 330. The number of processors 310 in device 300 can be one or more. Figure 3 Taking a processor as an example, in this embodiment, the processor 310, communication interface 320, and memory 330 can be connected via a bus system or other means. Figure 3 Taking the connection between China and Israel via bus system 340 as an example.

[0066] Processor 310 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. Processor 310 may further include hardware chips. These hardware chips may be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), or combinations thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), generic array logic (GAL), or any combination thereof.

[0067] The memory 330 may include volatile memory, such as random-access memory (RAM); the memory 330 may also include non-volatile memory, such as flash memory, hard disk drive (HDD) or solid-state drive (SSD); the memory 330 may also include a combination of the above types of memory.

[0068] Optionally, the memory 330 stores an operating system and programs, executable modules, or data structures, or subsets thereof, or extended sets thereof. The programs may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic business processes and handling hardware-based tasks. The processor 310 can read the programs from the memory 330 to implement the networked method for the steel production IF-THEN rule system provided in this application embodiment.

[0069] The bus system 340 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus system 340 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0070] This application also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the networked method for the IF-THEN rule system in steel production mentioned in the above embodiments.

[0071] This application also provides a computer program product containing instructions that, when run on a computer, causes the computer to execute the networked method for the steel production IF-THEN rule system mentioned in the above embodiments.

[0072] Finally, it should be noted that the above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.

Claims

1. A networked method for an IF-THEN rule system in steel production, characterized in that, The method includes: The IF-THEN type rules in the steel production field are decomposed into steel production condition parameters, comparison operators, thresholds and control conclusion actions, and explicit rule features and original data features are generated. A fusion network comprising explicit rule paths and implicit data paths is constructed. The explicit rule paths process the explicit rule features to preserve the rule triggering logic, while the implicit data paths process the raw data features to learn the implicit patterns in production parameters. The explicit rule features include Boolean features and cross features generated from combinations of steel production rule conditions, used to characterize the explicit triggering logic of the rules. The raw data features include the raw values ​​of continuous steel production parameters and the embedding vectors of discrete parameters, used to characterize the state of the unprocessed production system. Based on a historical dataset containing both rule-triggered and non-triggered samples, the fusion network is trained using a loss function that includes rule constraints, so that the fusion network fits the original rule conclusions on rule-triggered samples and fits the actual control results on all samples. The trained fusion network is used for real-time inference. A final control decision is generated based on the outputs of the explicit rule path and the implicit data path through a weighted fusion mechanism or an attention allocation mechanism. Explanatory information associated with the final control decision is output. The explanatory information includes at least one or more of the following: the original steel production rule that was triggered, the key features that affect the control decision and their contribution, and the matching degree information with similar scenarios in the history of steel production. After the fusion network is put into use, the network parameters are autonomously optimized based on the newly added production data through an incremental learning mechanism.

2. The method according to claim 1, characterized in that, The decomposition of IF-THEN type rules in the steel production field into steel production condition parameters, comparison operators, thresholds, and control conclusion actions includes: When the condition part of the IF-THEN rule contains "OR" logic, in the parsing step, the IF-THEN rule is decomposed into multiple sub-rules, the condition of each sub-rule is "AND" logic, and corresponding cross features are generated for each sub-rule.

3. The method according to claim 1, characterized in that, The construction of the fused network, which includes explicit rule paths and implicit data paths, includes: The explicit rule path receives the explicit rule features as input and processes them using a linear layer, factorization machine layer, cross network layer, or rule node embedding layer to adapt to the rule logic of steel production. The implicit data path receives the original data features as input and processes them using a deep fully connected network, a self-attention layer, or a feature graph network to adapt to the nonlinear correlation between steel production parameters.

4. The method according to claim 1 or 3, characterized in that, The architecture of the fusion network is selected from WIDE&DEEP network, deep factorization machine network, deep cross network, Transformer or graph neural network.

5. The method according to claim 1, characterized in that, The loss function includes a basic loss term and a rule constraint loss term; The basic loss term is used to measure the deviation between the model's predicted output and the actual control result. The mean squared error is used for regression tasks, and the cross-entropy loss is used for classification tasks. The rule constraint loss term is used to penalize the deviation between the model output and the original rule conclusion on rule-triggered samples; The loss function adjusts the weight of the rule constraint loss term in the total loss through the rule constraint coefficient.

6. The method according to claim 1, characterized in that, After the fusion network is put into use, the network parameters are optimized online based on the newly added production data through an incremental learning mechanism, including: Continuously collect real-time input features during the steel production process, the decision output of the fusion network, and the corresponding actual process results or product quality feedback; A sliding window mechanism is used to periodically fine-tune the parameters of the fusion network using the newly added production data; When the deviation between the decision output of the fusion network and the manual intervention command or process quality indicator exceeds a safety threshold, an emergency update process is triggered to adjust the parameters of the fusion network using data from the relevant scenario.

7. A networked device for an IF-THEN rule system in steel production, characterized in that, include: The rule parsing unit is used to decompose IF-THEN type rules in the steel production field into steel production condition parameters, comparison operators, thresholds and control conclusion actions, and generate explicit rule features and original data features. A network construction unit is used to construct a fusion network containing explicit rule paths and implicit data paths. The explicit rule paths process the explicit rule features to maintain the rule triggering logic, while the implicit data paths process the raw data features to learn the implicit patterns in production parameters. The explicit rule features include Boolean features and cross features generated from combinations of steel production rule conditions, used to characterize the explicit triggering logic of the rules. The raw data features include the raw values ​​of continuous steel production parameters and the embedding vectors of discrete parameters, used to characterize the unprocessed state of the production system. The training unit is used to train the fusion network based on a historical dataset containing rule-triggered samples and non-triggered samples, using a loss function containing rule constraints, so that the fusion network fits the original rule conclusions on rule-triggered samples and fits the actual control results on all samples. The reasoning and decision-making unit is used to perform real-time reasoning using the trained fusion network, generate a final control decision based on the output of the explicit rule path and the implicit data path through a weighted fusion mechanism or attention allocation mechanism, and output explanatory information associated with the final control decision. The explanatory information includes at least one or more of the following: the original steel production rule that was triggered, the key features that affect the control decision and their contribution, and the matching degree information with similar scenarios in the history of steel production. The autonomous optimization unit is used to autonomously optimize the parameters of the fusion network based on newly added production data through an incremental learning mechanism after the fusion network is put into use.

8. An electronic device, characterized in that, The device includes: a processor and a memory; The memory is used to store instructions; The processor is configured to execute the instructions in the memory to perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Including instructions that, when run on a computer, cause the computer to perform the method described in any one of claims 1-6 above.

Images