A multi-agent collaboration-based cross-layer dynamic regulation method and system

Abstract

This invention discloses a cross-level dynamic control method and system based on multi-agent collaboration, belonging to the field of industrial intelligent control technology. The invention uses a lightweight edge model to extract features and perform semantic summarization on low-level high-frequency sensor signals, assembling them into a hybrid intent data package of "structured snapshot + natural language summary," thus solving the token bottleneck and digital illusion problems of large language models processing industrial time-series signals. A central coordinating agent dynamically wakes up specialized agents such as supply chain, production scheduling, and process control based on abnormal intents, performing asynchronous parallel deduction and multi-objective conflict resolution on a shared blackboard. Finally, a cross-level control command with a mandatory physical safety threshold is output through a secure sandbox protocol, and after human-machine collaborative approval, it is sent to the heterogeneous MES and DCS systems for execution. This invention achieves cross-level globally optimal control, possessing both high robustness and high security, and is suitable for industrial scenarios such as continuous smelting of non-ferrous metals.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial intelligent control, and particularly relates to a cross-level dynamic regulation method and system based on multi-agent collaboration. Background Art

[0002] In the process of continuous smelting of non-ferrous metals (such as zinc smelting and other process industries), the production chain is long and the material flow and energy flow are strongly coupled. The typical information system architecture of a smelter includes three levels: the enterprise resource planning (ERP) level, the manufacturing execution system (MES) level, and the distributed control system (DCS) level. Supply chain management, production planning and scheduling, and bottom-layer process control are respectively carried out by the systems at the above different levels.

[0003] The existing technologies have the following core defects:

[0004] (I) Cross-level information silos lead to global sub-optimality

[0005] The supply chain ERP, production planning and scheduling MES, and bottom-layer process control DCS systems are fragmented and operate independently. When sudden disturbances occur during the production process (such as fluctuations in raw material grade and equipment anomalies), it is difficult to automatically achieve a global联动 response of "changing materials, reducing speed, and modifying process parameters". The systems at each level cannot form collaborative decisions, resulting in the existing regulation schemes only considering partial rather than overall optimal solutions.

[0006] (II) Vulnerability of traditional mathematical programming and mechanism models

[0007] Traditional operations research production scheduling algorithms and mechanism control models highly rely on static assumptions and accurate mathematical models. Facing the sudden changes in complex working conditions, these deterministic solution models are extremely prone to divergence or failure and cannot cope with high-dimensional multi-source uncertainties.

[0008] (III) Inherent defects of large language models in processing high-frequency time series signals

[0009] Real industrial disturbances mostly come from bottom-layer sensors (such as high-frequency time series data of temperature, pressure, etc.). Due to the limitation of the context window length (Token limitation) and the "hallucination" problem of large language models (LLMs) for dense numbers, they cannot directly access and process the massive and high-frequency DCS / SCADA industrial data streams. This is the core bottleneck restricting the application of large models in the field of industrial control.

[0010] (IV) Safety hazards of directly controlling physical devices by large models

[0011] Pure end-to-end large models lack a priori constraint mechanisms for the physical limits and safety red lines of industrial equipment. If they are directly used to generate industrial bottom-layer control instructions, there are uncontrollable safety risks.

[0012] Chinese invention patent CN116088439B discloses an optimized scheduling method for heating furnace groups in steel enterprises based on multi-agent simulation, which optimizes the scheduling of heating furnaces through a multi-agent simulation model. However, this scheme is limited to the scheduling optimization of a single section, does not involve cross-level system-level collaborative control, and does not solve the problem of large models handling high-frequency industrial data streams.

[0013] In conclusion, there is an urgent need for a new control scheme that can break down information silos across different levels, fully utilize the reasoning capabilities of large models, and ensure industrial security. Summary of the Invention

[0014] To address the aforementioned issues, this invention discloses a cross-level dynamic control method and system based on multi-agent collaboration. It constructs a novel control mode characterized by "collaborative division of labor between large and small models, hierarchical multi-agent control, and guaranteed security." A lightweight model semantically aligns and encapsulates structured monitoring signals to form a hybrid intent data packet that can be directly inferred by the large model. This drives multiple agents to perform parallel inference and conflict resolution on a shared blackboard. Finally, a secure sandbox protocol outputs control instructions with physical red lines, achieving highly secure and highly interpretable collaborative control in the process industry.

[0015] To achieve the above objectives, the present invention utilizes the following five progressively advancing technical modules:

[0016] 1. Multimodal perception and semantic alignment for complex working conditions based on large and small model collaboration

[0017] A semantic encapsulation layer is set up between the underlying physical gateway and the central intelligent agent. The system receives high-frequency structured monitoring signals from the underlying layer in real time (such as OPC-UA data), and uses a lightweight anomaly detection model at the edge to capture time-series data snapshots, change rates, and associated device statuses before and after anomaly points.

[0018] The lightweight anomaly detection model can be implemented using various existing algorithms, such as: statistical process control (SPC) threshold determination based on sliding window, trend fitting slope calculation based on least squares, outlier detection in Isolation Forest, or K-Means clustering analysis.

[0019] This invention does not limit the specific algorithm of the lightweight anomaly detection model at the edge; the key is the next step:

[0020] The system assembles the purely digital features output by the lightweight model (including time series data windows before and after the abnormal trigger point, signal change rate, and associated device state variables) with the automatically generated natural language description into a hybrid intent data packet consisting of "structured snapshot + natural language summary" and sends it to the large model hub.

[0021] The function of this module is to compress high-frequency time-series data that originally required many tokens to describe into lightweight data packets containing only a few key features and a summary sentence, achieving a data compression ratio of 95-99%. This not only completely avoids the 'context loss' and 'numerical illusion' problems inherent in large language models when processing ultra-long and dense numerical sequences, but also reduces the computing power overhead and network latency of a single control decision to the millisecond level, greatly meeting the stringent requirements of the process industry for real-time performance and reliability.

[0022] 2. Agent Behavior Boundaries and Modular Design

[0023] The system of the present invention includes a central coordinating intelligent agent and several specialized intelligent agents, including but not limited to: a supply chain management intelligent agent, a planning and scheduling intelligent agent, and a process control expert intelligent agent.

[0024] To constrain the illusion of large models, each specialized intelligent agent adopts a standardized "three-tier microservice" architecture design:

[0025] (1) Memory component: Stores the context of the current exception event being processed, as well as the agent’s historical proposals in the current session.

[0026] (2) Reflection Component: Solidifies industry constraints for specific domains, rather than relying solely on prompts for flexible constraints. For example, the constraint rule for the scheduling agent is "daily output fluctuations must not exceed 10%"; the constraint rule for the process agent is "core equipment load must not exceed the rated upper limit". When these constraints are exceeded, the reflection component directly prevents the proposal from entering the blackboard at the logical level, rather than simply providing a prompt through prompts.

[0027] (3) External tool components (Tools / API): Limit the capability boundaries of intelligent agents. For example, the process control expert intelligent agent is only authorized to call the mechanism model prediction interface, the SOP knowledge base retrieval interface and the DCS real-time data reading interface. Each intelligent agent is strictly isolated at the physical and logical levels, fundamentally preventing unauthorized decision-making.

[0028] It adopts a hierarchical multi-agent architecture of "centralized routing - expert distributed" and designs four core agents around the metallurgical production chain.

[0029] Intelligent agent role definition:

[0030] A: Coordinator Agent

[0031] Role: Chief Dispatcher of the workshop. Responsible for natural language interaction, multi-source abnormal intent recognition, task breakdown and distribution, and final solution assembly.

[0032] Core capabilities (Tools): Anomaly level determination, sub-Agent wake-up, and scheme conflict detection.

[0033] Example of a Prompt setting: "You are the chief intelligent scheduler of a zinc smelter. When you receive an anomaly alarm, you need to analyze whether it will affect the supply chain, production schedule, or underlying processes, and assign the task to the corresponding expert agent. If the sub-agents' plans conflict, you need to coordinate them to regenerate."

[0034] B: Supply Chain Agent

[0035] Role: Purchasing and warehousing expert.

[0036] Core capabilities (Tools): Query real-time inventory, calculate raw material arrival time (ETA), and search for supplier alternatives.

[0037] Functional Boundaries: When raw material supply is interrupted or quality changes drastically, responsible for calculating how long existing inventory can sustain operations and providing emergency allocation recommendations.

[0038] C: Scheduling Agent

[0039] Role: Production Planner (PMC).

[0040] Core capabilities (Tools): Obtain the current production schedule Gantt chart, project capacity load, and rearrange material feeding batches.

[0041] Functional Boundaries: Responsible for dynamically adjusting the feeding speed and batch turnover time when equipment is unloaded or raw materials are insufficient.

[0042] D: Process Control Agent

[0043] Role: Senior Process Engineer.

[0044] Core capabilities (Tools): Retrieve historical similar operating procedures (SOPs), query process control constraint boundaries, and provide suggestions for fine-tuning parameter settings.

[0045] Functional Boundaries: In response to fluctuations in raw material impurities, provide advance suggestions for adjusting parameters in each process, such as temperature and feed rate.

[0046] 3. Dynamic wake-up and modular isolation of agents based on intent recognition

[0047] This system adopts a hierarchical multi-agent architecture of "centralized routing - expert distributed systems". Its core features include:

[0048] Physical isolation of permissions: Each specialized intelligent agent consists of three layers of components: memory, domain constraint reflection, and specific tool API interfaces. The toolsets of different intelligent agents do not overlap. For example, the process intelligent agent can only call the DCS read / write and mechanism model, but cannot call the scheduling system.

[0049] Dynamic wake-up mechanism: After receiving the mixed intent data packet, the central agent analyzes the anomaly impact domain and dynamically wakes up only the specialized agents related to the current anomaly to form a "temporary decision-making working group", instead of calling up all agents, thereby improving the response speed.

[0050] 4. Multi-agent collaborative communication mechanism based on "shared blackboard"

[0051] The system establishes an independent and isolated memory space for the temporary decision-making working group – the “shared blackboard”.

[0052] Parallel inference: The awakened specialized agents use their respective domain tools to conduct independent inferences and asynchronously read and write state analyses and initial control proposals for their respective regions on the blackboard.

[0053] Conflict Detection and Resolution: The central intelligence agent cross-validates the multi-dimensional proposals on the blackboard. If a conflict is found between multi-dimensional business constraints, for example, the production scheduling intelligence agent advocates not reducing speed while the process intelligence agent advocates reducing speed, the central intelligence agent will generate a feedback instruction with rejection reasons based on the global cost function and prioritizing the safety > quality > output. This instruction will be sent back to the conflicting node, forcing it to find an "alternative compromise solution" until a globally conflict-free joint control solution is generated.

[0054] When a rigid constraint deadlock is detected in the proposals of different professional agents, the central agent performs the following conflict resolution actions: (1) issues a prompt word carrying the reason for rejection and the global bottom line to the conflicting parties to force them to re-deductive; (2) triggers the external resource retrieval mechanism according to the deadlock type, dynamically wakes up other professional agents such as supply chain management agents that are currently in a dormant state to intervene in the shared blackboard, and generates alternative joint proposals by introducing external compensation variables of the system, such as calling high-grade spare raw materials for blending.

[0055] Through the above mechanism, the system can generate a globally conflict-free joint control scheme when faced with multiple conflicting constraints.

[0056] Example of collaborative simulation:

[0057] Phase 1 (Assessment): The central agent posted on the blackboard: "Current anomaly: The iron content of zinc concentrate from the No. 2 ore source exceeds the standard by 20%."

[0058] Phase 2 (Concurrent Thinking): The process agent calculates: "The purification time needs to be increased by 10 minutes"; the scheduling agent sees this and reports: "If it is extended by 10 minutes, the downstream electrolysis process will be shut down due to material shortage".

[0059] Phase 3 (Compromise Generation): The central agent detects the conflict, requests a recalculation, and finally arrives at a compromise solution (e.g., "increase the purification time by 5 minutes while reducing the electrolysis current density by 2%").

[0060] 5. Structured protocol conversion for human-machine collaboration with a secure sandbox.

[0061] The final solution, after conflict resolution, first generates a visual front-end work order for human expert approval (Human-in-the-loop). Once approved, the protocol conversion module forcibly encapsulates the decision logic into a standardized JSON format instruction with strict type validation before issuing it.

[0062] Before executing instructions, the underlying DCS system performs independent hardware-level verification on the absolute physical security threshold field carried in the instructions. If the set value exceeds the security threshold, the system will perform hardware-level interception to achieve physical fallback protection for AI control.

[0063] Key data structure design of this invention

[0064] A: Hybrid Event Payload (Hybrid Modal Abnormal Intent)

[0065] This data structure is the standard message that the edge perception gateway transmits to the upper-layer central intelligent agent. Its core function is to "feed the structured features calculated by the small model to the large model for inference," thus completely solving the problem of digital illusion in the large model.

[0066] { "event_id": "EVT-2026-002", "trigger_source": "DCS_Telemetry_Monitor", "urgency": "CRITICAL", "semantic_summary": "The temperature in Zone B of Calcination Furnace No. 1 suddenly increased, exceeding the steady-state range, and showed an accelerating upward trend.", "structured_telemetry": { "tag_id": "TC-ROAST-1B", "current_value": 875.5, "rate_of_change_per_min": "+12.5", "time_series_window": [852.1, 855.4, 860.2, 875.5], "correlated_signals": { "air_flow_valve_opening": "45%", "feed_rate": "15t / h"}}}

[0067] Field Definitions Table 1

[0068] event_id String Globally unique identifier for exception events It is used to provide end-to-end contextual binding throughout the entire multi-agent lifecycle, shared blackboard reading and writing, and final work order approval. trigger_source String The underlying data source module that triggered this exception This enables the central agent to identify the source of disturbance (such as the DCS system or LIMS laboratory) and thus decide which specialized agents should be activated first. urgency Enum An enumeration of the urgency level of the exception (e.g., CRITICAL, WARNING). Scheduling technology features: The central intelligent agent determines the priority of the task queue based on this field and forces a timeout for large model inference to ensure industrial real-time performance. semantic_summary String Natural language summaries automatically generated by small models or rule engines Multimodal features: Transform the tedious alarm thresholds into textual intent contexts that Large Language Models (LLMs) are best at understanding, reducing the difficulty of inference. structured_telemetry Object Structured telemetry data objects containing numerical snapshots and trends Collaborative features between large and small models: This allows the large model to avoid directly processing the original high-frequency measurement point flow, serving as a boundary isolation layer to address the conflict between token consumption and the computational power of the large model. tag_id String Physical sensor tag number that caused the anomaly Precisely anchor the physical location of production line equipment to provide an index for retrieving the mechanism model and SOP of specific equipment. current_value Float Transient values ​​when an alarm is triggered Provide the current accurate numerical deviation as a baseline input for algorithm deduction. rate of change per min String Rate of change / derivative characteristics of the signal Core dimensionality reduction feature: The slope (e.g., +12.5) is directly calculated by the underlying small model and provided to the large model, avoiding the large model having to calculate the derivative of the array itself and eliminating numerical illusion. time_series_window Array Simplified historical data sequence window before and after the anomaly trigger point This provides limited data observation basis for process expert agents to determine whether "the disturbance is an instantaneous pulse or a continuous drift". correlated_signals Object Snapshots of the status of other physically related equipment (such as valve opening and feed rate). Multivariate joint features: Provides a multi-condition cross-section for multiple agents to investigate the root cause of failures, avoiding local misjudgments.

[0069] Cross-level control command distribution structure (Dispatch Payload)

[0070] This data structure is a standard message sent to lower-level physical devices after multiple agents reach a consensus. Its core functions are "cross-layer penetration" and "mandatory physical security fallback".

[0071] { "dispatch_payload": { "auth_token": "xxx-human-signed-xxx", "cross_layer_commands": [ { "layer": "MES_Scheduling", "action": "update_blend_ratio", "params": {"bin_2": 0.5, "bin_3": 0.5}}, { "layer": "DCS_Process_Control", "action": "set_pid_target", "equipment": "roasting_furnace_1", "params": {"target_temp": 855.0, "max_limit": 860.0}} ]}}

[0072] Technical features description:

[0073] 1. Cross-layer penetration design: The cross_layer_commands array allows instructions from heterogeneous MES and DCS systems to be included simultaneously in a single action, enabling logical collaboration in a physically isolated environment.

[0074] 2. Safety Defense Design: The max_limit (absolute safety threshold) field forces large models to bind a safety upper limit when outputting a set value (target_temp). When the underlying PLC / DCS receives this instruction, it performs secondary authentication and verification. If the set value or subsequent drift exceeds this physical limit, it will automatically intercept, achieving "last resort protection".

[0075] Field Definition Table 2

[0076] auth_token String Human-in-the-loop authentication token Safety closed-loop feature: Proof that the instruction has been approved by the human operator's digital signature or authorization, ensuring that the large model cannot directly drive high-risk industrial equipment without human intervention. cross_layer_commands Array Cross-level action instruction object collection Architectural features: Allows operation instructions for multiple heterogeneous systems (such as MES and DCS) to be synchronously encapsulated in a single JSON message in the form of a transaction. layer String The execution target system level of this instruction (such as MES_Scheduling) Protocol routing identifier: The gateway uses this field to send instructions to the enterprise resource scheduling network or the underlying industrial control network, respectively. action String Specified operation method or API call name The output of large models must be mapped to a limited set of API interfaces pre-registered by the system to prevent large models from fabricating / illusory unknown illegal instructions. equipment String (Specific control device identifiers for the DCS layer) Directed addressing to the underlying physical controller or specific loop. params Object The dictionary of key-value pairs of parameters required to perform this action It carries specific business values ​​(such as mixing ratio and PID target setting value) derived by various professional intelligent agents. max_limit / min_limit Float Absolute physical safety threshold / engineering boundary line The strongest defense feature (safety sandbox): It mandates that large models bind safe extreme values ​​when outputting setpoints. The underlying PLC performs hardware-level interception before and during instruction execution, achieving physical fallback protection for AI control.

[0077] The beneficial effects of this invention are as follows:

[0078] 1. Breaking through the bottleneck of industrial perception in large-scale models (innovation in the perception layer):

[0079] A hybrid modal access mechanism of "extracting structured features from a low-level small model + semantic assembly" was invented. This completely solves the token bottleneck and numerical illusion problems encountered by LLM in processing high-frequency industrial time-series signals, enabling large models to truly "understand" industrial DCS monitoring dashboards.

[0080] 2. Overcoming the local limitations of a single system (architectural innovation):

[0081] A hierarchical multi-agent structure mapping the organizational structure of a physical factory was constructed. Complex continuous smelting control was decoupled into sub-tasks such as supply chain, scheduling, and control. The system's output simultaneously includes instructions to both the upper-level production scheduling system (MES) and the lower-level control system (DCS), breaking down the traditional silos of industrial software and achieving cross-level global optimization.

[0082] 3. Flexible resolution of multidimensional industrial constraint conflicts (algorithm mechanism innovation):

[0083] This system creatively combines "asynchronous deduction using a shared blackboard" with "dynamic game theory based on prompt feedback." Compared to the shortcomings of traditional static programming operations research, which crashes and throws errors when the limits are exceeded, this system can find alternative compromises like an "expert meeting" when faced with multiple conflicting constraints, greatly improving the system's robustness in harsh operating conditions.

[0084] 4. Uphold the red line of safety at the industrial level (engineering safety innovation):

[0085] The system features strict access control (limiting the boundaries of API tool calls by the Agent) and a structured protocol security sandbox. When issuing physical action commands, a mandatory security threshold (max_limit) is applied, along with a human-machine collaborative hard interrupt mechanism, ensuring the absolute security and feasibility of AI-driven decision-making in harsh industrial production environments. Attached Figure Description

[0086] Figure 1 This invention is based on a global architecture diagram of a multimodal perception and multi-agent collaborative system, demonstrating the entire chain from small model signal extraction to large model multi-agent inference.

[0087] Figure 2 The present invention provides a centrally-led collaborative simulation and tool call sequence diagram, which details the workflow of the system in resolving the conflict between "process safety" and "production schedule" after an anomaly occurs. Detailed Implementation

[0088] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0089] The present invention describes a cross-level dynamic control method based on multi-agent collaboration, taking a sudden and severe 20% exceedance of iron impurities in zinc concentrate raw materials from a smelter as an example, combined with... Figure 1 and Figure 2 The specific embodiments of the present invention will be described in detail below.

[0090] Step S101: Signal perception and intent encapsulation

[0091] A lightweight anomaly detection model deployed at the edge (in this embodiment, a sliding window-based statistical process control algorithm) monitors zinc concentrate raw material composition data in real time. When the iron impurity content exceeds a preset value, the system automatically extracts features such as the time-series data window before and after the anomaly trigger point, calculates the rate of change, and the status of associated equipment. Using a semantic encapsulator, the system assembles these features with an automatically generated natural language summary (e.g., the raw material iron impurities exceed the standard by 20%) into a hybrid intent data packet, which is then sent to the central coordinating agent.

[0092] Step S201: Agent-oriented wake-up and task distribution

[0093] The central coordinating agent receives mixed intent data packets, parses out the anomaly impact domain, and dynamically wakes up the supply chain management agent, the planning and scheduling agent, and the process control expert agent to form a "temporary decision-making working group." These three specialized agents perform asynchronous parallel deductions on a shared blackboard.

[0094] Step S301: Asynchronous simulation and multi-objective conflict resolution

[0095] The intelligent supply chain management system queries real-time inventory, calculates raw material arrival times, and searches for supplier alternatives. When raw material supply is interrupted or its quality changes drastically, it is responsible for calculating how long the existing inventory can last and providing emergency allocation suggestions.

[0096] The planning and scheduling agent obtains the current production Gantt chart, calculates the capacity load, and rearranges the material feeding batches; when the equipment is unloaded or there is insufficient raw material, it dynamically adjusts the feeding speed and batch turnover time.

[0097] The process control expert agent retrieves historical similar operating procedures (SOPs) and queries process control constraint boundaries; it also provides advance parameter adjustment suggestions for each process in response to fluctuations in raw material impurities.

[0098] If the central coordinating agent reads conflicting proposals on the blackboard, such as one party requesting a production halt while the other requests continued production, the central coordinating agent, prioritizing safety over quality over output, determines that excessive iron impurities do not fall under the category of safety risks but rather represent a conflict between product quality and output, requiring scheduling to resolve.

[0099] The central coordinating agent issues a feedback instruction to the scheduling and process departments, including the reason for rejection: "Find an alternative solution that does not reduce speed"; at the same time, it triggers an external resource retrieval mechanism, dynamically waking up the supply chain management agent to intervene and requesting it to provide a "raw material alternative solution". After deduction, the supply chain management agent gives a new proposal: "Mix the concentrate from warehouse No. 4 with the concentrate from warehouse No. 2 at a 1:1 ratio".

[0100] Step S401: Safety Sandbox Constraints and Cross-Layer Command Issuance

[0101] After two rounds of negotiation, the above parties reached a global conflict-free solution: (1) Supply chain: the blending scheme was activated; (2) Scheduling: the order of feeding batches was adjusted; (3) Process: the parameters of the smelting process were adjusted.

[0102] The solution generates a visual work order and submits it to human experts for approval. Upon approval, the protocol conversion module encapsulates the approved solution into a standardized JSON instruction package and distributes it. The instruction package contains independent sub-instruction sets that simultaneously point to both the Manufacturing Execution System (MES) and the Control System (DCS). Furthermore, for physical operation instructions pointing to the DCS, in addition to the target setpoint field, the data structure of these instructions also forcibly configures and distributes an "absolute physical safety threshold (e.g., max_limit / min_limit)" field, allowing the underlying process controller to perform hardware-level extreme value interception and secondary verification.

[0103] This invention also introduces a cross-level dynamic control system based on multi-agent collaboration for executing the above-described method. The system comprises: an edge-aware gateway module, a central coordination module, a passive shared blackboard memory module, a specialized agent cluster module, and a secure sandbox protocol conversion module; each specialized agent in the specialized agent cluster module includes mutually isolated memory components, reflection components, and tool components.

[0104] The present invention also describes a computer-readable storage medium storing a computer program, wherein the computer program is executed by a processor using the methods described above.

[0105] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A cross-level dynamic control method based on multi-agent collaboration, characterized in that, Includes the following steps: S1: Receive structured time-series signals from underlying industrial equipment in real time, extract time-series features through a lightweight anomaly detection model at the edge, generate natural language summaries, and concatenate the features and summaries into a hybrid intent data packet containing structured telemetry snapshots and natural language summaries; S2: The central coordinating agent receives the hybrid intent data packet, parses the abnormal influence domain, and based on the influence domain, selectively wakes up the affected professional agent combination from the preset agent pool. S3: The awakened professional intelligent agents combine to call their respective restricted domain-specific tools to perform asynchronous parallel deduction and proposal reading and writing in the passive shared memory area; the central coordinating intelligent agent performs cross-domain constraint cross-verification on multi-party proposals based on a preset global cost function, and issues feedback instructions containing rejection reasons and constraints to the conflicting intelligent agents based on the preset global cost function and conflict type, so as to drive them to re-perform asynchronous deduction and form a globally conflict-free collaborative solution; S4: Transform the conflict-free collaborative scheme into a cross-level control instruction package containing data structures, and after forcibly attaching an absolute physical security threshold and human-machine collaborative verification, issue the execution to the underlying heterogeneous system.

2. The method according to claim 1, characterized in that, In step S1, when the edge-end lightweight anomaly detection model detects that the signal deviates from the preset steady-state range, it extracts the time series data window before and after the anomaly trigger point, the signal change rate characteristics, and the state variables of the associated devices, and encapsulates them with the automatically generated natural language description in the same hybrid data structure and sends them to the central coordinating agent.

3. The method according to claim 1, characterized in that, In step S2, the professional intelligent agent combination includes a supply chain management intelligent agent, a planning and scheduling intelligent agent, and a process control expert intelligent agent; each of the professional intelligent agents is configured as a mutually isolated three-layer functional structure, including: a memory layer for storing event context, a reflection layer for solidifying domain safety red lines, and a tool layer for limiting permissions; the tool layer strictly restricts each professional intelligent agent, both physically and logically, to only call the application programming interface of its corresponding domain.

4. The method according to claim 3, characterized in that, The safety red lines fixed in the reflection layer include: the daily output fluctuation of the scheduling agent does not exceed 10%, and the core equipment load of the process agent does not exceed the rated upper limit; when the proposal generated by the large model exceeds the safety red lines, the reflection component directly prevents the proposal from entering the shared blackboard.

5. The method according to claim 1, characterized in that, In step S3, when the central coordinating agent detects a rigid constraint deadlock in the proposals of different professional agents, the conflict resolution actions performed by the central coordinating agent include: not only issuing prompts to the conflicting parties with the reasons for rejection and the global bottom line to force them to re-deductively, but also triggering an external resource retrieval mechanism according to the deadlock type, dynamically waking up other professional agents that are currently in a dormant state to intervene in the shared memory area, and generating alternative joint proposals by introducing external compensation variables of the system.

6. The method according to claim 1, characterized in that, In step S4, the output cross-level control instruction package contains an independent set of sub-instructions that simultaneously point to the production planning and scheduling layer and the process control layer; and for the physical operation instructions pointing to the process control layer, in addition to the target set value field in its data structure, an "absolute physical safety threshold" field is also forcibly configured and issued for the underlying process controller to perform hardware-level extreme value interception and secondary verification.

7. A cross-level dynamic control system based on multi-agent collaboration, characterized in that, The system comprises: an edge-aware gateway module, a central coordination module, a passive shared blackboard memory module, a specialized intelligent agent cluster module, and a secure sandbox protocol conversion module; each specialized intelligent agent in the specialized intelligent agent cluster module comprises mutually isolated memory components, reflection components, and tool components; the system is used to execute the method of any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images