Multi-agent decision-making and task scheduling method and system for blast furnace ironmaking multi-objective collaboration

Abstract

The present application relates to a kind of industrial multi-agent decision and task arrangement method and system for blast furnace iron-making multi-objective cooperation, belong to process industry intelligent decision and industrial artificial intelligence technical field.The method mainly includes: collecting the multi-source data of blast furnace site, and based on the preset working condition window, shared furnace condition state memory is constructed;Based on the shared furnace condition state memory, local suggestion object is generated in parallel by multiple blast furnace role agent;Structural conflict detection is performed on the local suggestion object, and conflict relationship diagram is constructed;Based on process rule, equipment state, safety boundary and current production target, the conflict relationship diagram is arbitrated, and target action set is determined;Task arrangement is carried out on the target action set, and structured collaborative execution suggestion is generated.The present application realizes the unified state perception, local suggestion generation, conflict detection, process constraint arbitration and task arrangement of the cooperative decision-making closed loop for blast furnace continuous production scene, and improves the decision efficiency.

Description

Technical Field

[0001] This invention relates to the fields of intelligent decision-making in process industries and industrial artificial intelligence, and in particular to an industrial multi-agent decision-making and task scheduling method and system for multi-objective collaboration in blast furnace ironmaking. Background Technology

[0002] Blast furnace ironmaking is a continuous production process with strong coupling of multiple variables. The gas flow distribution, thermal regime and hearth state are deeply related. Operating parameters such as air volume, air temperature, pulverized coal injection and charging are mutually constrained. Dynamic balance needs to be achieved among multiple objectives such as stable permeability, thermal regime control, output guarantee, fuel consumption and equipment safety, which constitutes a typical multi-constraint collaborative optimization problem.

[0003] Multi-agent collaborative technology, by constructing intelligent entities with specialized capabilities and establishing collaborative mechanisms, enables distributed decision-making and comprehensive optimization of complex systems, and has become an important technological path for the intelligent upgrading of process industries. However, directly applying a general multi-agent framework to blast furnace scenarios has significant adaptation defects. Existing solutions lack a role-based division for blast furnace-specific process objectives, making it difficult to address unique decision-making functions such as thermal regime, permeability, airflow distribution, and equipment safety. Each module typically acquires data, models, and makes judgments independently, lacking a unified and shared basis for furnace conditions, leading to inconsistent understanding. Furthermore, existing technologies generally lack structured detection and arbitration mechanisms for parameter conflicts, objective conflicts, timing conflicts, resource conflicts, and safety conflicts among multi-source suggestions, making it difficult to transform heterogeneous suggestions into comparable and arbitrable structured objects. In addition, existing solutions lack the ability to orchestrate tasks with execution order, observation windows, manual confirmation conditions, and rollback strategies, making it difficult to generate executable task flows on-site. They also fail to effectively integrate memory information such as tasks to be observed across shifts, incomplete tasks, prohibited actions, and handover instructions, failing to construct a closed-loop decision-making system that integrates state understanding, conflict resolution, task organization, execution control, and result feedback. They still rely on manual experience for conflict selection and execution arrangements, making it difficult to meet the actual needs of continuous industrial production for real-time multi-objective collaborative optimization. Summary of the Invention

[0004] The purpose of this invention is to provide an industrial multi-agent decision-making and task scheduling method and system for multi-objective collaboration in blast furnace ironmaking, in order to solve the problems of lack of shared state, difficulty in unifying and organizing local suggestions, difficulty in identifying and arbitrating conflicts, difficulty in directly executing output results, and difficulty in continuing shift memory in existing intelligent decision-making schemes for blast furnaces.

[0005] The technical means employed in this invention are as follows: On the one hand, this invention provides an industrial multi-agent decision-making and task orchestration method for multi-objective collaboration in blast furnace ironmaking, comprising the following steps: Collect multi-source data from the blast furnace site and construct a shared furnace status memory based on a preset operating condition window; Based on the shared furnace condition memory, local suggestion objects are generated in parallel by multiple blast furnace role intelligent agents; Perform structured conflict detection on the local proposal objects and construct a conflict relationship graph; wherein, the nodes of the conflict relationship graph represent the local proposal objects and the edges represent conflict relationships; Arbitrate the conflict relationship diagram based on process rules, equipment status, safety boundaries, and current production goals to determine the target action set; The target action set is task orchestrated to generate structured collaborative execution suggestions; the structured collaborative execution suggestions include at least the execution order, observation window, and rollback conditions.

[0006] Furthermore, this invention also provides an industrial multi-agent decision-making and task orchestration system for multi-objective collaboration in blast furnace ironmaking, comprising: The shared furnace condition status construction module is used to collect multi-source data from the blast furnace site and construct a shared furnace condition status memory. The role-based intelligent agent execution module is used to generate local suggestion objects in parallel through multiple blast furnace role-based intelligent agents based on the shared furnace condition state memory. The conflict detection module is used to perform structured conflict detection on the local suggestion objects and construct a conflict relationship graph; wherein, the nodes of the conflict relationship graph represent the local suggestion objects and the edges represent conflict relationships; The arbitration module is used to arbitrate the conflict relationship diagram based on process rules, equipment status, safety boundaries, and current production goals to determine the target action set; The task orchestration module is used to orchestrate the target action set and generate structured collaborative execution suggestions. Compared with the prior art, the present invention has the following advantages: This invention constructs a shared furnace condition state memory to uniformly organize multi-source data, enabling each role's intelligent agent to generate local suggestions based on the same factual basis, thereby eliminating inconsistencies in state interpretation. Furthermore, the outputs of each role are uniformly expressed as local suggestion objects, and a conflict relationship graph is constructed, allowing explicit detection and recording of parameter, objective, timing, resource, and safety conflicts. Based on this, the conflict relationship graph is arbitrated according to process rules, equipment status, safety boundaries, and production goals, generating retained actions, disabled actions, alternative actions, and priority results, transforming conflict suggestions into a set of usable actions that satisfy constraints. Subsequently, the retained and alternative actions are task orchestrated, and their execution order, observation windows, manual confirmation requirements, and rollback conditions are configured, forming an executable and traceable set. The system proposes structured collaborative execution suggestions; then, through gating execution and rollback control, it differentiates between automatic execution, execution after manual confirmation, and non-execution modes based on risk level and equipment status. Rollback is triggered within the observation window based on changes in key variables to ensure the safety and robustness of collaborative decision-making in blast furnace field applications. Finally, execution feedback and shift handover information are written back to the shared furnace status memory and shift memory and imported into the next round of decision-making. This enables the handover between shifts of tasks to be observed, incomplete tasks, executed actions, and currently prohibited actions, thereby adapting to continuous blast furnace production scenarios. This forms a closed-loop collaborative decision-making mechanism from unified understanding, conflict resolution, task scheduling to safe execution and memory inheritance, significantly improving the coordination, executability, and continuity of multi-role collaboration. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart of an industrial multi-agent decision-making and task orchestration method for multi-objective collaboration in blast furnace ironmaking, as described in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the shared furnace status memory structure in an embodiment of the present invention.

[0010] Figure 3 This is a schematic diagram illustrating the parallel execution of the blast furnace role intelligent agent in an embodiment of the present invention.

[0011] Figure 4 This is a flowchart of local suggestion conflict detection and process constraint arbitration in an embodiment of the present invention.

[0012] Figure 5 This is a schematic diagram of task orchestration, gating execution, and rollback control in an embodiment of the present invention.

[0013] Figure 6 This is a closed-loop flowchart for field application in an embodiment of the present invention.

[0014] Figure 7 This is a schematic diagram illustrating the continuity of memory across shifts in an embodiment of the present invention.

[0015] Figure 8 This is a schematic diagram of the overall architecture of an industrial multi-agent decision-making and task orchestration system for multi-objective collaboration in blast furnace ironmaking, as described in an embodiment of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] Example 1 like Figure 1 As shown in the figure, this embodiment provides an industrial multi-agent decision-making and task orchestration method for multi-objective collaboration in blast furnace ironmaking. The method specifically includes the following steps: S1. Collect multi-source data from the blast furnace site and construct a shared furnace status memory.

[0018] Multi-source data includes at least three of the following: continuous process data, equipment status data, laboratory test data, operational event data, and text record data. The shared furnace status memory includes at least the current status of major process variables, trends of key variables, equipment availability, recent operational events, anomaly markers, current production target constraints, shift memories, and a list of tasks to be observed. Continuous process data extracts features such as recent values, mean, rate of change, fluctuation amplitude, extreme values, and duration of exceeding limits. Laboratory test data records the most recent valid test value, test time, sampling interval, and time-sensitive markers. Equipment status data records availability, operating mode, fault markers, alarm levels, and status change times. Operational event data records event type, time, affected object, adjustment direction, and magnitude. Text record data uses entity extraction, keyword extraction, or template matching to identify various entities. The processed results are uniformly organized and written to the shared furnace status memory, providing unified status input to the task decomposition stage and the role-based intelligent agent stage. The shared furnace status memory can be implemented using a blackboard-style shared memory, a unified status object, or a versioned status database. When using version numbers, the latest version is read in each round of decision-making, and a new version is generated after updates to avoid read / write chaos.

[0019] S2. Execute task decomposition based on shared furnace status memory and current production targets.

[0020] Task decomposition is dynamically generated by combining current production goals and the main risks in the shared furnace status memory. Specifically, priority goals are first determined from the current set of production goals, and then the main task is identified based on anomaly markers, key variable trends, equipment availability status, recent operation events, and shift memories. Sub-tasks are generated around the main task and a task tree or task graph is constructed through dependencies, prerequisites, and mutual exclusions. In the task tree or task graph, nodes represent main tasks or sub-tasks, and edges represent dependencies, prerequisites, or mutual exclusions.

[0021] The main task identification can be achieved by combining a task mapping table with task trigger scores. Specifically, for each candidate main task m, a correspondence between it and a set of key variables is pre-established. For example, the task of restoring stable air permeability corresponds to one or more of the following: pressure difference, air permeability index, air volume, air pressure, material line, fabric regime switching event, and top temperature. The task trigger score F_m is calculated as follows: F_m = α1D_m + α2T_m + α3E_m + α4G_m + α5M_m; In this context, D_m, T_m, E_m, G_m, and M_m are all normalized to values ​​between 0 and 1. D_m represents the degree of deviation of a key variable related to the candidate primary task m, determined by the degree of deviation from a preset normal range. If the variable is within the preset normal range, its deviation value is 0; if it exceeds the preset normal range, the deviation value is determined by the ratio between the excess and the preset deviation range, and the deviation value is limited to between 0 and 1. D_m can be the weighted average, maximum, or a combination of the deviation values ​​of its corresponding key variables. The preset normal range and preset deviation range can be determined by blast furnace process specifications, historical stable operation data statistics, or expert configuration. T_m represents the degree of trend anomaly of the relevant key variable, determined based on its changing trend within a preset operating condition window. Trend anomalies can be determined based on the difference between the variable's current value and the window's initial value, the sliding window slope, the rate of change, or the duration of continuous increase / decrease. E_m represents the correlation between recent operational events and candidate primary tasks m. A pre-established event correlation mapping table between operational events and candidate primary tasks is used. Operational events include one or more of the following: air volume adjustment, air temperature adjustment, oxygen enrichment rate adjustment, pulverized coal injection rate adjustment, material distribution system switching, material line adjustment, iron tapping, ventilation shutdown, ventilation restart, and alarm handling. For operational events occurring within a preset backtracking time window, the event contribution value is determined based on the event type, the object affected by the event, the adjustment direction, the adjustment magnitude, and the time interval between the event occurrence and the current decision time. G_m represents the priority weight of the current production target to candidate primary task m. The current production target may include one or more of the following: restoring stable permeability, controlling thermal regime deviation, maintaining output, reducing fuel ratio, limiting equipment load, and ensuring safe operation. The current production target is divided into high priority, medium priority, low priority, and irrelevant targets, mapped to 1, 0.6, 0.3, and 0 respectively; alternatively, a corresponding normalized priority value can be set according to the production strategy. M_m represents the triggering degree of the observed or incomplete task in the shift memory on the candidate main task m, determined based on the correlation between various tasks and the candidate main task m. α1 to α5 are preset weights. These can be pre-configured by blast furnace process experts based on experience; set according to the company's current production strategy; calibrated offline based on historical blast furnace operation samples and manual decision-making records; or selected through offline verification to achieve the highest consistency between the task identification result and the historical expert handling result. When historical calibration samples are lacking, α1 to α5 can be initially set to 0.30, 0.25, 0.15, 0.20, and 0.10 respectively; during field application, the weights can be updated based on task identification accuracy, execution feedback, and expert correction results. If F_m is greater than the preset task triggering threshold, or if F_m ranks highest among all candidate main tasks, then the candidate main task m is identified as the current main task.

[0022] The task mapping table records the correspondence between different production goals, anomaly types, and candidate tasks. After determining the main task, the corresponding set of subtasks is selected according to the task mapping table, and a task tree or task graph is constructed based on the process sequence and safety verification requirements. For example, diagnostic subtasks precede adjustment subtasks, safety verification subtasks serve as preceding nodes of adjustment subtasks, and observation subtasks serve as successor nodes after adjustment actions. Mutually exclusive adjustment actions are represented by mutually exclusive edges.

[0023] For example, when the shared furnace status memory indicates that the differential pressure is continuously increasing, the air permeability index is decreasing, and there is a recent material feeding regime switching event, "Restore air permeability stability" can be identified as the main task, and further sub-tasks such as "Diagnose the source of air permeability abnormality", "Assess material feeding regime adjustment", "Assess material line adjustment", "Assess heat regime related actions" and "Verify equipment safety conditions" can be generated.

[0024] S3. Input the shared furnace condition status memory and task tree or task graph into multiple blast furnace role agents, and have each blast furnace role agent generate local suggestion objects in parallel.

[0025] Local suggestions should include at least a role identifier, corresponding task, suggested action or judgment, target object, suggested parameters, risk level, and observation window, and be written to the associated area of ​​the shared furnace status memory. For example... Figure 3As shown, the multiple blast furnace role agents include at least diagnostic role agents, regulation role agents, and safety monitoring role agents. Specifically, the diagnostic role agents determine the current dominant risk or source of anomaly. Inputs include the current status of major process variables, the changing trend of key variables, anomaly markers, recent operational events, and shift memories. Differential pressure, permeability index, top temperature, gas utilization rate, material level, molten iron composition, recent operational events, and alarm status are selected as diagnostic features. The judgment results are generated through anomaly classification models, causal association rules, process knowledge graphs, or retrieval-enhanced reasoning methods. The output includes a local suggested object containing role identifier, corresponding task, anomaly type, anomaly source judgment, confidence level, risk level, and observation window. The regulatory role intelligent agent proposes regulatory actions based on thermal regime, air permeability, airflow distribution, fabric distribution, material line, energy consumption or fuel ratio. The inputs include the current production target, task tree or task graph, trend of key variable changes, equipment availability, safety boundary and currently prohibited actions. Based on the candidate action library, action effect prediction model, process regulation rules or historical similar working condition search results, candidate regulation actions and their expected effects are generated. The output includes role identifier, corresponding task, suggested action, target object, suggested parameters, expected effect, risk level and observation window of the local suggested object. Safety monitoring roles verify the current equipment status, safety boundaries, and process prohibitions. Inputs include equipment status data, safety boundaries, alarm status, maintenance status, current production targets, and candidate actions output by various regulation roles. The status of hot blast stove, pulverized coal injection system, material distribution system, air supply system, key valve position, actuator status, and alarm status are selected as safety verification features. Based on the safety boundary table, equipment availability table, process prohibition rules, and linkage constraint rules, the feasibility of candidate actions is verified. The output includes a local suggested object containing role identifier, corresponding task, prohibited actions, restricted actions, executable actions, safety risk level, and conflict reason.

[0026] Each role-based intelligent agent can construct its own knowledge base using historical blast furnace operation data, operation records, anomaly handling records, shift reports, expert-annotated samples, and process rules. When using machine learning models, it can train anomaly classification models, action effect prediction models, or risk level prediction models. When using large-scale model inference, it can build a retrieval-enhanced knowledge base and efficiently fine-tune its role-based reasoning capabilities through prompt templates, few-sample examples, or parameters. When using a rule-based approach, it can form a role-based knowledge base based on process rules, equipment interlocking conditions, safety boundaries, and expert experience. These methods can be used individually or in combination. Each role-based intelligent agent reads the same shared blast furnace status memory and task tree or task graph, generates local suggestion objects with a unified field format, and writes them to the associated area of ​​the shared blast furnace status memory for direct comparison and processing in the subsequent conflict detection stage.

[0027] For example, an agent focused on air permeability or airflow distribution can output the following local recommendations: The role is identified as an agent focused on air permeability or airflow distribution; the corresponding task is to restore stable air permeability; the recommended action is to restore the previous fabric distribution regime; the target is the fabric distribution regime; the recommended parameter is the number of the previous stable regime; the risk level is medium; the observation window is 30 minutes; and the expected effect is a decrease in pressure differential and improved air permeability. An agent focused on thermal regime regulation can output: The role is identified as an agent focused on thermal regime regulation; the corresponding task is to suppress the risk of excessively hot thermal regimes; the recommended action is to reduce the oxygen enrichment rate by 0.2%; the target is the oxygen enrichment rate; the risk level is medium; the observation window is 20 minutes; and the expected effect is a stabilization of the top temperature. An agent focused on safety monitoring can output: The role is identified as an agent focused on equipment and safety monitoring; the corresponding task is to verify safety conditions; the recommended judgment is that the current pulverized coal injection system is fluctuating significantly, and it is prohibited to simultaneously increase the pulverized coal injection rate and the oxygen enrichment rate.

[0028] S4. Read local suggestion objects in a unified manner, perform structured conflict detection, and construct a conflict relationship graph.

[0029] like Figure 4 As shown, the fields of each local suggestion object are first normalized to form a standardized suggestion object. A standardized suggestion object includes at least a role identifier, corresponding task, suggested action, target object, suggested direction, suggested parameter range, target label, execution resources, preconditions, observation window, risk level, and safety constraint markers. The suggested direction indicates whether a target object is adjusted upwards, downwards, maintained, restored, prohibited, or observed. The suggested parameter range indicates the range of parameter changes corresponding to the suggested action. The target label indicates that the suggested action primarily serves one or more of the following objectives: restoring stable permeability, controlling thermal regime deviation, maintaining production, reducing fuel ratio, limiting equipment load, or ensuring safe operation. Execution resources indicate the actuators, adjustment channels, or operating time windows required for the action.

[0030] Subsequently, conflicts are detected based on the target object, target label, execution resources, and observation window pairing. For two standardized suggestion objects with the same target object or a coupling relationship, if the suggestion direction is opposite or the suggestion parameter range is incompatible, it is identified as a parameter conflict. For example, one suggestion requires increasing the oxygen enrichment rate, while the other requires decreasing it; or one suggestion requires increasing the pulverized coal injection rate, while the other prohibits increasing the pulverized coal injection rate. For two standardized suggestion objects with different target labels, target conflict is detected based on the current production target priority and the direction of the action effect. If one suggested action is beneficial to maintaining output but detrimental to restoring stable permeability, while the other suggested action is beneficial to restoring stable permeability but may reduce output, it is identified as a target conflict. Two standardized suggestion objects with incompatible preconditions, observation windows, or execution times are identified as timing conflicts; two standardized suggestion objects occupying the same actuator, adjustment channel, or time window and cannot be parallelized are identified as resource conflicts; and suggested actions corresponding to standardized suggestion objects that violate equipment constraints, safety boundaries, or process prohibitions are identified as safety conflicts. A conflict relationship graph is constructed based on the detection results. The nodes are standardized suggested objects, and the edges record the conflict type, cause, and intensity. This can be implemented using an adjacency list, adjacency matrix, or relationship table.

[0031] Conflict intensity is determined using a weighted scoring method. For any conflicting edge e, calculate the conflict intensity S_e: S_e = β1C_e + β2R_e + β3P_e + β4O_e; Wherein, C_e is the base value for conflict type, R_e is the risk level value, P_e is the production target priority correlation value, and O_e is the overlap value; β1 to β4 are weight values; C_e is mapped to different base values ​​according to safety, parameters, resources, timing, and target conflict type (e.g., mapped to 1.0, 0.8, 0.6, 0.5, and 0.4 respectively), and the maximum value or weighted combination is taken when the same conflict edge corresponds to multiple types; R_e is mapped according to low, medium, and high action risk levels (can be mapped to 0.2, 0.6, and 1.0 respectively), and the maximum risk value is taken if there is a conflict between the two suggested objects connected by conflict edge e, and the risk value of that action is taken if the suggested action corresponding to conflict edge e conflicts with the safety boundary; P_e is mapped according to the priority relationship between the production target and the current target as high. The values ​​are: low, medium, or irrelevant; for multiple objectives, the maximum or weighted average is used. O_e is determined by the overlap of the target, execution resources, and time window (if the target of two standardized recommendations is exactly the same, O_e is 1; if the target is different but there is a pre-defined coupling relationship, O_e is between 0.5 and 0.8; if the target is irrelevant, the overlap value is 0). β1 to β4 are non-negative normalized weights, which can be calibrated based on security strategies, process experience, or historical samples. For security-priority scenarios, β1 and β2 are increased; for multi-objective collaboration scenarios, β3 is increased; and for resource coordination scenarios, β4 is increased. When historical calibration samples are lacking, β1, β2, β3, and β4 can be initially set to 0.35, 0.30, 0.20, and 0.15, respectively. S_e is divided into three levels: low, medium, and high. High-intensity conflicts prioritize arbitration; high-intensity conflicts involving security can directly trigger disabling or manual confirmation.

[0032] S5. Based on process rules, equipment status, safety boundaries, and current production goals, arbitrate the conflict relationship diagram to obtain retained actions, prohibited actions, alternative actions, and priority results.

[0033] By constructing a conflict relationship graph, one-to-many, many-to-one, and chain-like conflict relationships among multiple local advisory objects can be uniformly organized, and conflict type, conflict cause, and conflict intensity can be stored as edge attributes. In this way, the arbitration stage can make unified decisions within the group based on the conflict-connected subgraph, rather than making local judgments on individual conflict pairs, thereby avoiding the omission of indirect conflicts.

[0034] Among them, reserved actions are those currently recommended for priority execution; alternative actions are those that can be executed after reserved actions fail to achieve the expected results or specific conditions are met; observation actions are those that continuously monitor key variables after adjustment actions; rollback actions are those used to restore or limit subsequent operations when rollback conditions are triggered; and disabled actions do not enter the execution sequence but are written into the structured collaborative execution recommendations as prohibition constraints. Reserved actions, alternative actions, observation actions, and rollback actions are converted into task nodes, and disabled actions are converted into prohibition constraints.

[0035] In this embodiment, to ensure the enforceability of the arbitration result, a layered arbitration logic is preferably adopted. For example... Figure 5 As shown, the process begins with safety screening, marking suggested actions that violate equipment availability, safety boundaries, or process prohibitions as prohibited actions. For example, when the pulverized coal injection system is in an abnormally fluctuating state, suggesting actions that simultaneously increase the pulverized coal injection rate and oxygen enrichment rate are marked as prohibited actions. Then, based on the edge connections in the conflict relationship graph data structure, nodes with conflicting relationships are divided into conflict-connected subgraphs. Each conflict-connected subgraph represents a group of local suggested objects with direct or indirect conflicts. Isolated nodes without conflicting edges and that do not violate safety rules directly enter the candidate set of retained actions. Finally, action priority scores are calculated for each candidate action in each conflict-connected subgraph. Based on the score ranking and conflict intensity, retained actions, prohibited actions, and alternative actions are determined to form the final arbitration result.

[0036] Then, an action priority score is calculated for each candidate action in the conflicting connected subgraph. In one implementation, the action priority score Q_a for action a can be expressed as: Q_a = γ1G_a + γ2B_a + γ3H_a + γ4M_a - γ5R_a; Wherein, G_a represents the degree of matching between the target label of the action and the current production target priority. The higher the current production target priority, the higher G_a. The current production target priority can be jointly determined by the production plan, shift strategy, abnormal status, and safety constraints. B_a represents the expected improvement effect, which is determined by the action effect prediction model, historical similar working condition retrieval results, process rules, or expert experience. H_a represents the execution effect of historical similar working conditions, with a higher value for a higher effective improvement ratio. M_a represents the shift memory support, which is not yet complete. The task continuation value is high, and the value is 0 if it conflicts with prohibited actions; R_a is the risk level, with low, medium and high mapped to 0.2, 0.6 and 1.0 respectively, and direct disabling if the safety boundary is violated; γ1 to γ5 are non-negative weights, which can be preset by blast furnace process experts according to production strategies, or can be calibrated offline based on historical samples. When historical calibration samples are lacking, the weights are 0.25, 0.30, 0.15, 0.10 and 0.20 respectively. γ5 is increased when there are anomalies, and γ1 and γ2 are increased when recovery is emphasized; when there is sufficient historical data on similar working conditions, γ3 can be increased.

[0037] Subsequently, based on action priority scores, conflict intensity, and conflict type, retained actions, disabled actions, and alternative actions are determined. Actions that violate safety boundaries due to conflict are identified as disabled actions; for actions with opposing parameter conflict directions, the action with the higher priority score is retained, and the other action is identified as disabled or an alternative; for actions with conflicting objectives, the action that aligns with the current priority production objective is retained, and the other action is designated as an alternative or pending observation action; for actions with conflicting timing, the execution order of the actions is determined; for actions with conflicting resources, the action that provides stronger mitigation of the current major risk and has a lower risk level is retained, and the other action can be used as an alternative after resource release; the results of retained, disabled, and alternative actions and their priorities are output for task orchestration.

[0038] For example, in a scenario where the permeability is abnormal and the pulverized coal injection system fluctuates significantly, if the permeability or airflow distribution agent suggests restoring the previous material distribution regime, the material distribution or material line agent suggests lowering the material line by 0.15m, and the thermal regime adjustment agent suggests increasing the oxygen enrichment rate, while the safety monitoring agent determines that the pulverized coal injection system fluctuates significantly and prohibits simultaneously increasing the pulverized coal injection rate and the oxygen enrichment rate, then during the arbitration phase, the increase in the oxygen enrichment rate can be marked as a prohibited action. Then, the timing conflict between "restoring the previous material distribution regime" and "lowering the material line by 0.15m" can be handled, with the former being designated as a reserved action and the latter as a substitute action.

[0039] S6. Based on priority results and dependencies between actions, orchestrate retained and alternative actions to generate structured collaborative execution suggestions.

[0040] Dependencies between actions include prerequisite dependencies, observation dependencies, condition dependencies, mutual exclusion dependencies, resource dependencies, and manual confirmation dependencies. Prerequisite dependencies indicate that one action must be completed before another; observation dependencies indicate that subsequent actions must wait for the observation window to end and determine whether to execute based on changes in key variables; condition dependencies indicate that an action is executed only when specific conditions are met; mutual exclusion dependencies indicate that two actions cannot be executed simultaneously; resource dependencies indicate that two actions occupying the same actuator, control channel, or operation time window need to be scheduled sequentially; and manual confirmation dependencies indicate that an action must be confirmed by the operator before it can be executed.

[0041] like Figure 6 As shown, the specific steps of task orchestration include: converting various actions into task nodes and configuring relevant attributes; generating dependency edges based on process prerequisites, equipment availability, safety verification results, observation windows, execution resource usage, and manual confirmation requirements; sorting task nodes or constructing a task graph with conditional branches based on priority results and dependency edges; inserting observation nodes after each adjustment-type task node and binding them to target variables, risk variables, observation time, and rollback conditions; and finally generating structured collaborative execution suggestions.

[0042] Structured collaborative execution recommendations should include at least the execution sequence, observation window, manual confirmation requirements, and rollback conditions. They should also include at least three of the following: main task, sub-task, recommended action, alternative action, prohibited action or forbidden item, execution mode, risk level, and shift continuation marker. For example, in a scenario with abnormal air permeability, the main task is to restore stable air permeability; the recommended action is to restore the previous fabric configuration; the alternative action is to lower the material line by 0.15m; the execution sequence is to first restore the fabric configuration and then observe for 30 minutes; the manual confirmation requirement is for the on-duty operator to confirm execution; the forbidden item is to prohibit simultaneously increasing the oxygen enrichment rate; and the rollback condition is to switch to the alternative action if the pressure differential does not improve within 30 minutes.

[0043] S7. Based on the risk level and equipment status, gate the structured collaborative execution suggestions to determine whether to execute automatically, execute after manual confirmation, or simply prompt not to execute.

[0044] The automatic execution mode is suitable for low-risk actions with small parameter changes, normal equipment status, and those that have passed safety verification. The manual confirmation-based execution mode is suitable for actions involving multiple parameter linkages that may cause significant changes in permeability or thermal regime. The prompt-only-not-execute mode is suitable for high-risk actions, actions with abnormal equipment status, or those where multiple high-intensity conflicts have not yet been resolved. The rollback condition is linked to changes in key variables within the observation window. If the target variable does not show the expected improvement, key risk variables continue to deteriorate, or protection variables trigger boundaries, the corresponding rollback action is executed, and the system re-enters the arbitration phase or freezes subsequent related actions.

[0045] S8. Collect execution feedback and shift handover information, and write it back to the shared furnace status memory and shift memory for input in the next round of decision-making.

[0046] like Figure 8 As shown, the execution feedback includes at least the changes in key variables after execution, manual confirmation results, action execution results, anomaly handling effects, whether rollback was triggered, tasks to be observed, incomplete tasks, status of executed actions, and currently prohibited actions. Shift handover information includes at least handover instructions and matters requiring continued attention in the next shift. All of the above information is uniformly written back to the shared furnace status memory and shift memory. Shift memory is organized by shift; at the end of the current shift, the current shift's memory item is written; at the beginning of the next shift, the previous shift's memory item is read to form the status inheritance result, including the input of tasks to be observed, incomplete tasks, status of executed actions, status of currently prohibited actions, and handover instructions. This is then written into the new round of shared furnace status memory to participate in subsequent task decomposition, local suggestion object generation, and arbitration processes, achieving a natural transition between shift-related matters.

[0047] The following specific application examples will further illustrate the solution and effects of the present invention.

[0048] In Application Example 1, a 2500m³ blast furnace experienced an anomaly at 02:10, characterized by a continuously increasing differential pressure and decreased permeability. The system collects data on air volume, air pressure, oxygen enrichment rate, pulverized coal injection rate, top temperature, differential pressure, material level, charging regime, and the status of the hot blast stove and pulverized coal injection system from the previous two hours. Simultaneously, it reads the most recent charging regime switching event and the shift report, constructs a shared furnace status memory, and sets the current production target as "prioritizing the restoration of stable permeability, followed by maintaining output."

[0049] The task decomposition phase identifies the primary task as restoring furnace stability. Sub-tasks include diagnosing the source of permeability anomalies, assessing adjustments to the charging regime, assessing material line adjustments, assessing actions related to the thermal regime, and verifying equipment safety conditions. During the role-based intelligent agent execution phase, local suggestion objects are generated as follows: Diagnostic intelligent agents determine that the current anomaly is primarily due to permeability risk; permeability or airflow distribution intelligent agents generate local suggestion objects for "restoring the previous charging regime"; charging or material line intelligent agents generate local suggestion objects for "lowering the material line by 0.15m"; thermal regime adjustment intelligent agents generate local suggestion objects for "temporarily suspending the increase in oxygen enrichment rate"; and safety monitoring intelligent agents generate local suggestion objects for "significant fluctuations in the pulverized coal injection system, prohibiting simultaneous increases in pulverized coal injection rate and oxygen enrichment rate."

[0050] The conflict detection phase identifies a timing conflict between "restoring the previous fabric deployment regime" and "reducing the material line by 0.15m," and also identifies a safety conflict between "increasing the oxygen enrichment rate" and the current safety status. The arbitration phase prioritizes "increasing the oxygen enrichment rate" as a prohibited action based on equipment status and safety boundaries, then prioritizes retaining "restoring the previous fabric deployment regime" based on current production targets, and lists "reducing the material line by 0.15m" as an alternative action. The orchestration phase generates the following structured collaborative execution suggestions: Step 1, restore the previous fabric deployment regime; Step 2, observe changes in pressure differential and air permeability within 30 minutes; Step 3, if air permeability is not restored, execute "reducing the material line by 0.15m"; Step 4, continue observing changes in air permeability and write this task into the shift memory. In this embodiment, restoring the previous fabric deployment regime corresponds to the direct action of "alleviating the current air permeability risk," the 30-minute observation window corresponds to the judgment phase of "verifying the action effect," and reducing the material line by 0.15m corresponds to "an alternative action when the retained action effect is insufficient," thus forming an executable task flow.

[0051] In Application Example 2, a blast furnace experienced increased top temperature fluctuations and higher Si levels in molten iron during the day shift. The system constructs a shared furnace status memory and identifies the current production objective as "suppressing the risk of excessive thermal regime overheating and maintaining stable production." After task decomposition, the main task is to suppress excessive thermal regime overheating, and the sub-tasks include determining the source of the excessive thermal regime overheating, assessing oxygen enrichment rate adjustments, assessing blast temperature adjustments, assessing the linkage between pulverized coal injection and the material line, and verifying the safety status of the hot blast stove and pulverized coal injection system.

[0052] Each role's intelligent agent generates local suggestion objects in parallel: the thermal regime adjustment intelligent agent generates a local suggestion object for "reducing the oxygen enrichment rate by 0.2%"; the energy consumption and fuel ratio intelligent agent generates a local suggestion object for "not changing the air volume for now"; the material distribution or material line intelligent agent generates a local suggestion object for "maintaining the current material distribution regime"; and the safety monitoring role's intelligent agent generates a local suggestion object for "the hot blast stove is in normal condition and the oxygen enrichment rate reduction can be executed". No obvious parameter conflicts were found during the conflict detection phase, but a target conflict was identified between "reducing the oxygen enrichment rate" and "maintaining production". During the arbitration phase, "suppressing the risk of excessive thermal regime heat" was identified as the priority target, and the oxygen enrichment rate reduction action was retained. The orchestration phase forms a structured collaborative execution suggestion: Step 1, reduce the oxygen enrichment rate by 0.2%; Step 2, observe the top temperature change within 20 minutes; Step 3, if the molten iron temperature drops too quickly, revert the oxygen enrichment rate to the pre-adjustment level. This embodiment illustrates that the present invention can organize potentially production-affecting adjustment actions into a monitored and reversible execution plan through target conflict identification and priority target arbitration.

[0053] Example 2 This embodiment, based on Embodiment 1, provides an industrial multi-agent decision-making and task orchestration system for multi-objective collaboration in blast furnace ironmaking, such as... Figure 8 As shown, it specifically includes a working condition status access and shared state construction module, a task decomposition and target modeling module, a role intelligent agent execution module, a conflict detection module, a process constraint arbitration module, an execution order arrangement module, a gating execution and rollback module, and a feedback write-back and shift memory update module; the above modules work together to execute the steps in Example 1.

[0054] In this embodiment, the system is deployed within the blast furnace production information platform of a steel enterprise and is connected to PLC, DCS, SCADA, laboratory testing system, equipment status system, and log system. The modules do not require physically independent deployment; they can also be implemented using different software modules on the same server or computing platform.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for multi-agent decision-making and task orchestration in industrial processes for multi-objective collaboration in blast furnace ironmaking, characterized in that, Includes the following steps: Collect multi-source data from the blast furnace site and construct a shared furnace status memory based on a preset operating condition window; Based on the shared furnace condition memory, local suggestion objects are generated in parallel by multiple blast furnace role intelligent agents; Perform structured conflict detection on the local proposal objects and construct a conflict relationship graph; wherein, the nodes of the conflict relationship graph represent the local proposal objects and the edges represent conflict relationships; Arbitrate the conflict relationship diagram based on process rules, equipment status, safety boundaries, and current production goals to determine the target action set; The target action set is task orchestrated to generate structured collaborative execution suggestions.

2. The industrial multi-agent decision-making and task scheduling method for multi-objective collaboration in blast furnace ironmaking according to claim 1, characterized in that, Based on the shared furnace condition memory, local suggestion objects are generated in parallel by multiple blast furnace role agents, including: Based on the shared furnace status memory, calculate the task trigger score of multiple candidate main tasks, and select the candidate main task whose task trigger score meets the preset conditions as the main task; determine at least one sub-task corresponding to the main task according to the preset task mapping table, and construct a task tree or task graph. The shared furnace condition state memory and the task tree or task graph are input into multiple blast furnace role agents, and each blast furnace role agent generates local suggestion objects in parallel; the multiple blast furnace role agents include at least diagnostic role agents, regulation role agents and safety monitoring role agents; the local suggestion objects include at least role identifier, corresponding task, suggested action or judgment, target, suggested parameters, risk level and observation window.

3. The industrial multi-agent decision-making and task scheduling method for multi-objective collaboration in blast furnace ironmaking according to claim 1, characterized in that, Perform structured conflict detection on the local proposal objects and construct a conflict relationship graph, including: The fields of each local suggestion object are normalized to obtain a standardized suggestion object that includes at least the role identifier, corresponding task, suggested action, target object, suggested direction, suggested parameter range, target label, execution resources, preconditions, observation window, risk level and safety constraint mark. Candidate conflicts are matched based on the target object, target label, execution resources, and observation window. For two standardized suggestion objects with the same target object, the suggestion direction, suggestion parameter range, and allowable adjustment range are compared. If the suggestion direction is opposite or the parameter range is incompatible, it is identified as a parameter conflict. Compare the impact of two standardized recommendations on different production goals based on the target label and the current production goal priority. If the impacts are opposite, it is identified as a goal conflict. Based on the preconditions, observation window and execution time requirements, determine whether there is a relationship that cannot be executed simultaneously or must be executed sequentially. If such a relationship exists, it is identified as a timing conflict. Based on the execution resources, determine whether the same execution mechanism, adjustment channel, or operation time window is occupied and the conditions for parallel execution are not met. If so, it is identified as a resource conflict. Determine whether the recommended action violates safety constraints based on equipment status, safety boundaries, and process prohibition rules; if it does, it is identified as a safety conflict. Based on the identified conflict relationships, a conflict relationship graph data structure is constructed. Nodes represent standardized suggestion objects, edges represent conflict relationships, and edge attributes include conflict type, conflict cause, and conflict intensity.

4. The industrial multi-agent decision-making and task scheduling method for multi-objective collaboration in blast furnace ironmaking according to claim 1, characterized in that, Arbitrate the conflict relationship diagram based on process rules, equipment status, safety boundaries, and current production goals to determine the set of target actions, including: Read the nodes, edges, conflict types, conflict causes, and conflict intensities from the conflict graph data structure; Based on safety boundaries, equipment status, and process prohibition rules, actions corresponding to nodes that involve safety conflicts and do not meet the safety execution conditions are screened out for safety and marked as prohibited actions. Based on the edge connection relationships in the conflict relationship graph data structure, nodes with conflicting associations are divided into one or more conflicting connected subgraphs; Calculate the action priority score for each candidate action in each conflicting connected subgraph; In each conflict-connected subgraph, reserved actions, disabled actions, and alternative actions are determined based on action priority scores and conflict intensity. Among them, actions that directly conflict with the security boundary are determined as disabled actions; actions that have high-intensity parameter conflicts or resource conflicts with reserved actions and cannot be executed sequentially are determined as disabled actions; and actions that have temporal conflicts or target conflicts with reserved actions but can be executed when subsequent conditions are met are determined as alternative actions. Generate a set of target actions and priority results based on each retained action, disabled action, and alternative action.

5. The industrial multi-agent decision-making and task scheduling method for multi-objective collaboration in blast furnace ironmaking according to claim 4, characterized in that, The target action set is used for task orchestration to generate structured collaborative execution suggestions, including: The retained actions, alternative actions, observation actions, and rollback actions in the target action set are converted into task nodes, and the disabled actions are converted into prohibited constraints. The dependencies between the task nodes are determined based on the preconditions of the process, the availability of equipment, the results of safety verification, the observation window, the resource usage, and the requirements for manual confirmation. Based on the dependencies and priority results, the task nodes are sorted or a task graph with conditional branches is constructed. Insert observation nodes after adjustment task nodes, and bind the observation nodes to key variable changes and rollback conditions; Generate structured collaborative execution suggestions based on the sorting results or task graph.

6. The industrial multi-agent decision-making and task scheduling method for multi-objective collaboration in blast furnace ironmaking according to claim 1, characterized in that, The structured collaborative execution suggestion also includes a requirement for manual confirmation; the method also includes: The rollback condition is linked to the changes in key variables within the observation window; when the target variable in the observation window does not show the expected improvement or the key risk variable continues to deteriorate, the corresponding rollback action is executed.

7. The industrial multi-agent decision-making and task scheduling method for multi-objective collaboration in blast furnace ironmaking according to claim 6, characterized in that, Based on the risk level, equipment status, action attributes, and manual confirmation requirements in the structured collaborative execution proposal, the structured collaborative execution proposal is gated to determine whether to execute automatically, execute after manual confirmation, or only prompt not to execute.

8. The industrial multi-agent decision-making and task scheduling method for multi-objective collaboration in blast furnace ironmaking according to claim 1, characterized in that, The method further includes: Collect execution feedback and shift handover information, and write the changes in key variables after execution, action execution results, tasks to be observed, and currently prohibited actions back to the shared furnace status memory for import in the next round of decision-making.

9. An industrial multi-agent decision-making and task scheduling system for multi-objective collaboration in blast furnace ironmaking, characterized in that, include: The shared furnace condition status construction module is used to collect multi-source data from the blast furnace site and construct a shared furnace condition status memory. The role-based intelligent agent execution module is used to generate local suggestion objects in parallel through multiple blast furnace role-based intelligent agents based on the shared furnace condition state memory. The conflict detection module is used to perform structured conflict detection on the local suggestion objects and construct a conflict relationship graph; wherein, the nodes of the conflict relationship graph represent the local suggestion objects and the edges represent conflict relationships; The arbitration module is used to arbitrate the conflict relationship diagram based on process rules, equipment status, safety boundaries, and current production goals to determine the target action set; The task orchestration module is used to orchestrate the target action set and generate structured collaborative execution suggestions.

Images