Industrial multi-agent operation and maintenance decision-making method based on search enhancement and physical constraints

Abstract

This invention provides an industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints. The method includes receiving natural language operation and maintenance instructions, freezing collected industrial sensing data, and semantically deconstructing the instructions to generate structured parameters using equipment topology knowledge retrieved by a graph retrieval enhancement engine. It then invokes a deterministic computation engine in the physical constraint tool layer to pass the structured parameters to an external computing environment for physical and mathematical deduction, prohibiting the direct generation of physical numerical predictions. Based on the numerical results, a preliminary operation and maintenance strategy draft is generated, which is then verified through multiple rounds of debate in a multi-dimensional expert game layer to form the final optimized operation instructions. Finally, a rigid rule interceptor compares the data with equipment physical thresholds before executing or terminating the operation. This invention eliminates the risk of physical illusion in large models, ensures the absolute security of industrial operation and maintenance instructions, and improves the reliability of decisions under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and industrial automation control, and in particular to an industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints. Background Technology

[0002] Industrial operations and maintenance management is shifting from passive response based on fixed rules to proactive decision-making based on data.

[0003] Existing technologies mainly fall into two categories. One category consists of traditional operation and maintenance diagnostic systems based on expert rules and single-mechanism models. These systems rely on hard-coded thresholds and fixed rule trees, making them unsuitable for the complex, multi-variable coupled conditions of industrial environments. They exhibit extremely poor generalization capabilities and cannot provide effective operation and maintenance decisions in the face of unknown fault scenarios. The other category comprises industrial auxiliary decision-making systems based on large language models and retrieval-enhanced generation techniques. These systems rely on the natural language understanding capabilities of large language models to generate operation and maintenance decisions. However, large language models are essentially autoregressive lexical prediction engines based on probability distributions. When processing physical numerical calculations and constraint judgments in industrial scenarios, they are prone to generating illusory instructions that violate physical common sense, posing serious safety risks. Furthermore, these systems employ a unidirectional linear decision-making process, lacking multi-dimensional cross-validation mechanisms and independent rigid safety interception mechanisms. This results in unreliable decision-making results with zero fault tolerance, failing to meet the zero-error rigid requirements of closed-loop control in industrial environments and hindering the direct issuance of operation and maintenance control commands.

[0004] Therefore, how to eliminate the risk of physical illusion in large models, ensure the absolute security of industrial operation and maintenance instructions, and improve the reliability of decision-making under complex working conditions are technical problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides an industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints, which eliminates the risk of physical illusion in large models, ensures the absolute security of industrial operation and maintenance instructions, and improves the reliability of decision-making under complex working conditions.

[0006] On one hand, this invention provides an industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints, which includes: The system receives natural language operation and maintenance instructions, freezes and collects industrial sensing data, and uses a graph retrieval enhancement engine to retrieve equipment topology knowledge. It then performs semantic deconstruction on the natural language operation and maintenance instructions, extracts equipment entities and target parameters, and generates structured parameters. Based on the structured parameters, the deterministic computation engine in the physical constraint tool layer is invoked through the application programming interface to pass the structured parameters to the external computing environment; during the invocation process, direct generation of physical numerical predictions is prohibited. The physical and mathematical deductions are performed in the external computing environment, and deterministic numerical calculation results are output. Receive the numerical calculation results and use natural language generation capabilities to convert them into a preliminary draft of operation and maintenance strategy in natural language form; The preliminary operation and maintenance strategy draft is broadcast to the multi-dimensional expert game layer, where multiple expert agents conduct multiple rounds of debate and verification on the preliminary operation and maintenance strategy draft, and perform logical game based on the review protocol until a consensus is reached, forming the final optimized operation instructions. Before the final optimized operation instruction is issued, the operation parameters are extracted by a rigid rule interceptor and compared with the preset device hardware physical threshold. The operation is then executed or terminated based on the comparison result.

[0007] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints.

[0008] The present invention provides an industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints. After receiving natural language operation and maintenance instructions, the method freezes collected industrial sensing data and combines this with equipment topology knowledge retrieved by a graph retrieval enhancement engine to semantically deconstruct the operation and maintenance instructions, generating structured parameters. Through an application programming interface, it calls a deterministic computation engine in the physical constraint tool layer to pass the structured parameters to an external computing environment for physical and mathematical deduction. It prohibits large language models from directly generating physical numerical predictions. Based on the numerical calculation results obtained from the deduction, a preliminary draft of the operation and maintenance strategy is generated, which is then verified through multiple rounds of debate by multi-expert agents in a multi-dimensional expert game layer. This complete technical solution achieves consensus to form the final optimized operation command. Finally, an independent rigid rule interceptor performs a physical threshold comparison on the final optimized operation command before executing or terminating the operation. This decouples the large language model from deterministic physical computation, eliminates the physical illusion risk of the large language model in industrial operation and maintenance decision-making, and constructs a full-process security control system from intent parsing to execution fallback. It ensures 100% physical compliance and absolute security of the final issued operation and maintenance control command, while improving the reliability and fault tolerance of operation and maintenance decision-making under complex and multi-constrained industrial conditions, and meets the rigid requirement of zero error in closed-loop control in industrial sites. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in this 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 this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the industrial multi-agent operation and maintenance decision-making system based on retrieval enhancement and physical constraints provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0012] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0013] Figure 1 This is a flowchart illustrating the industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints provided in an embodiment of the present invention.

[0014] like Figure 1 As shown, the industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints provided in this embodiment of the invention mainly includes the following steps: 101. Receive natural language operation and maintenance instructions, freeze and collect industrial sensing data, and use a graph retrieval enhancement engine to retrieve equipment topology knowledge, perform semantic deconstruction on the natural language operation and maintenance instructions, extract equipment entities and target parameters, and generate structured parameters. In a specific implementation process, the deployment and configuration of each level can be completed in advance. Figure 2This is a schematic diagram of the structure of the industrial multi-agent operation and maintenance decision-making system based on retrieval enhancement and physical constraints provided in an embodiment of the present invention, as shown below. Figure 2 As shown, an interaction and perception module 21, a main control routing module 22, a physical constraint tool layer 23, a multi-dimensional expert game layer 24, and a fallback and execution module 25 can be built. Data interaction between each layer / module is completed through a pre-built high-performance application programming interface.

[0015] When on-site operators input natural language operation and maintenance commands, the system synchronously collects all industrial sensing data from the current industrial site through the interaction and perception module 21. It then freezes the collected industrial sensing data, locking a snapshot of the equipment's current operating status. Simultaneously, the main control routing module 22 activates the graph retrieval enhancement engine to search pre-built equipment topology knowledge and obtain the equipment topology information associated with the operation and maintenance command. The main control routing module 22, combined with the retrieved equipment topology knowledge, performs semantic deconstruction on the received natural language operation and maintenance command, identifying and extracting the equipment entities and target operation and maintenance parameters contained in the command, and generating structured parameters according to a preset format specification.

[0016] The process of collecting and freezing industrial sensing data may include: Structured sensor time-series data are synchronously acquired via an industrial fieldbus. The structured sensor time-series data includes temperature, pressure, rotational speed, and vibration amplitude. Meanwhile, unstructured visual data is collected by a visual inspection model deployed at key workstations. The visual inspection model uses an instance segmentation algorithm to extract the defect state vector of the equipment surface from real-time images. The defect state vector includes crack length, wear area, or the location and severity of leaks. The structured sensor time series data and the unstructured visual data are aligned and fused under a unified timestamp to form a multimodal state snapshot at the current moment, thereby freezing the industrial sensing data.

[0017] In detail, the system establishes synchronous communication connections with all sensors and detection devices in the field through an industrial fieldbus. It synchronously collects structured sensor time-series data of the field devices at a unified sampling frequency. The collected structured sensor time-series data includes, but is not limited to, structured physical parameters that reflect the real-time operating status of the equipment, such as temperature, pressure, speed, and vibration amplitude.

[0018] While collecting structured sensor time-series data, the system acquires real-time images through image acquisition devices deployed at key workstations on-site. It then calls a pre-trained visual detection model to collect corresponding unstructured visual data. The visual detection model uses an instance segmentation algorithm to segment the acquired real-time images and extract defect state vectors from the real-time images. The extracted defect state vectors include, but are not limited to, unstructured data that can reflect the health status of the equipment, such as crack length, wear area, and the location and severity of leaks.

[0019] After acquiring structured sensor time-series data and unstructured visual data, the system aligns the two types of data under a unified timestamp. It then performs feature fusion on the aligned data to form a complete multimodal state snapshot at the current moment. By locking this multimodal state snapshot, the system freezes the industrial sensing data, ensuring that all subsequent semantic deconstruction, physical calculation, and strategy generation processes are executed based on the device state at this frozen moment, thus avoiding decision-making biases caused by dynamic data changes.

[0020] In one specific implementation, the graph retrieval enhancement engine constructs a knowledge graph based on the equipment building information model and the physical topological connections between equipment; The semantic deconstruction process includes: Natural language processing technology is used to identify the device entity name and physical parameter keywords in the natural language operation and maintenance instructions. Graph retrieval technology is used to locate the core node of the corresponding device entity in the knowledge graph and automatically associate the core node with the upstream device, downstream device and parallel redundant pipeline in the topology. Based on the device entity name, physical parameter keywords, and the topology relationship, the natural language operation and maintenance instructions are converted into structured parameters containing the device's unique identifier, the associated topology path, and the set of affected parameters.

[0021] The system collects building information model data of all equipment on site, extracts the basic attributes, hardware parameters and operating boundary information of each equipment, and sorts out the physical topology connection relationship between all equipment, including pipeline connection, electrical connection, signal connection relationship between equipment, as well as upstream and downstream linkage relationship and parallel redundancy configuration relationship. Based on the above data, a complete equipment knowledge graph is constructed and the initial configuration of the graph retrieval enhancement engine is completed.

[0022] In detail, during the semantic deconstruction process, the main control routing module 22 first uses natural language processing technology to segment, identify entities, and extract keywords from the received natural language operation and maintenance instructions. This accurately identifies the device entity names and physical parameter keywords related to the operation and maintenance contained in the natural language operation and maintenance instructions. After the identification and extraction are completed, the graph retrieval enhancement engine starts graph retrieval technology to locate the core node of the device entity corresponding to the identified device entity name in the pre-built knowledge graph. After the core node is located, the topological connection relationship in the knowledge graph is automatically traversed to associate all upstream and downstream devices corresponding to the core node, as well as parallel redundant pipelines associated with the device, thus completely sorting out the fully associated topological network of the device entity.

[0023] Based on the identified device entity name, physical parameter keywords, and the fully associated topology obtained by the graph retrieval enhancement engine, the main control routing module 22 performs structured transformation of natural language operation and maintenance instructions, generating standardized structured parameters. The structured parameters clearly include the unique identifier of the corresponding device, the complete topology path associated with the device, and the set of affected parameters corresponding to this operation and maintenance operation, ensuring that the structured parameters can fully cover the entire intent and associated impact of the operation and maintenance instructions.

[0024] 102. Based on the structured parameters, the deterministic computation engine in the physical constraint tool layer 23 is called through the application programming interface to pass the structured parameters to the external computing environment; wherein, the direct generation of physical numerical predictions is prohibited during the calling process; In a specific implementation process, the master routing module 22 can call the deterministic computing engine in the physical constraint tool layer 23 through a predefined application interface based on the generated structured parameters, and completely transfer the structured parameters to the external computing environment isolated from the large language model's running environment. Throughout the entire calling process, the large language model on which the master routing module 22 depends is forcibly prohibited from directly generating any physical numerical prediction content, and only undertakes the responsibilities of parameter routing and interface calling.

[0025] The deterministic computing engine includes a sensitivity analysis engine, a value stream evaluation engine, or a mechanism simulation solver.

[0026] The sensitivity analysis engine is used to calculate the sensitivity of each physical operation variable to the current global optimization objective function value and output the sensitivity value. The value stream evaluation engine is used to calculate the estimated economic benefits under different operating strategies based on real-time electricity prices, material costs, or production efficiency data, and output the value stream evaluation value. The mechanism simulation solver is used to perform numerical simulations of physical processes based on thermodynamic, fluid mechanics, or structural mechanics equations, and outputs predicted values ​​of the changing trends of physical variables.

[0027] In detail, during the deployment phase, the system pre-computes the independent encapsulation and configuration of multiple deterministic computing engines in the physical constraint tool layer 23. Each deterministic computing engine runs in an external computing environment isolated from the large language model, and each engine is configured with an independent application programming interface and standardized input / output specifications.

[0028] The deterministic computing engine configured in the physical constraint tool layer 23 includes a sensitivity analysis engine, a value stream evaluation engine, and a mechanism simulation solver. It can call one or more corresponding deterministic computing engines according to the type of operation and maintenance instructions and the global optimization objective.

[0029] When the core objective of operation and maintenance is to determine the optimal adjustment variables, the sensitivity analysis engine is invoked. Based on the received structured parameters, real-time equipment operation data, and global optimization objective, the sensitivity analysis engine calculates the degree of influence of each adjustable physical operation variable on the current global optimization objective function value, and finally outputs the sensitivity value corresponding to each physical operation variable to guide the selection of operation variables to be adjusted first.

[0030] When the core objective of operation and maintenance is to evaluate economic benefits and costs, the value stream evaluation engine is invoked. The value stream evaluation engine connects to external databases in real time to obtain real-time data on electricity prices, material costs, and production efficiency. Combined with the operation plans in the structured parameters, it calculates the input costs, expected outputs, and net economic benefits corresponding to different operation strategies, and finally outputs the corresponding value stream evaluation value to evaluate the economic rationality of the operation strategy.

[0031] When maintenance operations require verification of the changing trends and compliance of physical processes, a mechanism simulation solver is invoked. The mechanism simulation solver has built-in standard mechanism equations corresponding to thermodynamics, fluid mechanics, and structural mechanics. Based on the received structured parameters and the current operating status of the equipment, it performs numerical simulation of the physical process after the operation is executed, simulates the dynamic changes of physical variables, and finally outputs the predicted values ​​of the changing trends of physical variables to verify the physical feasibility of the operation strategy.

[0032] 103. Perform physical and mathematical deductions in the external computing environment and output deterministic numerical calculation results; In a specific implementation process, the deterministic computing engine, in an external computing environment, performs corresponding physical and mathematical deductions based on the received structured parameters, the real-time operating status of the device, and the preset physical mechanism rules. After completing the entire calculation process, it outputs deterministic numerical calculation results and sends the numerical calculation results back to the main control routing module 22.

[0033] When the deterministic computation engine is a sensitivity analysis engine, the physical and mathematical deduction process includes: The sensitivity values ​​of each physical operation variable are calculated based on the partial derivatives of the current global optimization objective function value with respect to each physical operation variable, the ratio of each physical operation variable to the current global optimization objective function value, and the exponential penalty term constructed based on the difference between the potential risk index and the safety red line threshold. The larger the sensitivity value, the more preferentially the corresponding physical operation variable is adjusted.

[0034] In detail, the sensitivity analysis engine first determines the global optimization objective function corresponding to the current operation and maintenance scenario based on the received structured parameters, obtains the current value of the global optimization objective function, and simultaneously identifies all adjustable physical operation variables for this operation and maintenance. For each physical operation variable, the sensitivity analysis engine first calculates the partial derivative of the global optimization objective function value with respect to that physical operation variable, obtaining the basic degree of influence of that physical operation variable on the global optimization objective; then it calculates the ratio of the current value of that physical operation variable to the current value of the global optimization objective function, completing the dimensionless processing and eliminating the differences between different physical dimensions.

[0035] Meanwhile, the sensitivity analysis engine obtains the potential risk index corresponding to the adjustment of each physical operation variable, as well as the system's preset insurmountable safety red line threshold. Based on the difference between the potential risk index and the safety red line threshold, an index penalty term is constructed. This index penalty term is used to penalize and correct operation variables that exceed the safety range.

[0036] The sensitivity analysis engine combines the calculation results of the basic impact level, the dimensionless processing results, and the exponential penalty term to calculate the comprehensive sensitivity value corresponding to each physical operation variable. After the calculation is completed, the physical operation variables are sorted from largest to smallest sensitivity value. It is clear that the physical operation variable with the larger the sensitivity value has a higher impact on the global optimization objective and should be the priority for adjustment.

[0037] Specifically, the formula for the coordinated penalty of multidimensional constraint sensitivity used in the physical and mathematical deduction process can be:

[0038] in, Indicates the first A physical manipulated variable (such as the opening degree of a valve) The overall sensitivity recommendation index after security optimization. The larger the value, the more priority the system should give to adjusting this variable.

[0039] This represents the current global optimization objective function value (such as the unit energy consumption or value stream evaluation value of the entire production line).

[0040] This represents the current global optimization objective function value. For variables The partial derivatives, i.e. the fundamental sensitivity in pure mathematical mechanisms (calculated by the external mechanism solver, not involved in the large model).

[0041] Dimensionless processing term is used to eliminate dimensional differences between different physical units (such as Celsius, Pascal, Ampere).

[0042] Indicates adjustment variable The potential risk index (quantified and assessed by expert agents during the debate).

[0043] This represents the inviolable security threshold hard-coded by the system.

[0044] This represents the penalty scaling factor.

[0045] This indicates an exponential penalty term.

[0046] 104. Receive the numerical calculation results and use natural language generation capabilities to convert them into a preliminary draft of operation and maintenance strategy in natural language form; The main control routing module 22 receives the numerical calculation results returned by the deterministic computing engine. Using its own natural language generation capabilities, it transforms the pure numerical calculation results into a preliminary draft of operation and maintenance strategy in natural language form that conforms to the industrial operation and maintenance operation specifications. The draft clearly includes the execution object, operation content, operation parameters and execution requirements of the operation and maintenance operation.

[0047] 105. Broadcast the preliminary operation and maintenance strategy draft to the multi-dimensional expert game layer 24. Multiple expert agents conduct multiple rounds of debate and verification on the preliminary operation and maintenance strategy draft, and perform logical game based on the review protocol until a consensus is reached, forming the final optimized operation instruction. The main control routing module 22 broadcasts the generated preliminary operation and maintenance strategy draft to the multi-dimensional expert game layer 24 through a message queue. Multiple independent expert agents deployed in the multi-dimensional expert game layer 24 conduct multiple rounds of debate and verification on the preliminary operation and maintenance strategy draft from their respective evaluation dimensions. All expert agents conduct logical games based on a preset review protocol until all expert agents reach a consensus on the operation and maintenance strategy and form the final optimized operation instructions.

[0048] The multi-dimensional expert game layer 24 includes a security expert agent, an energy efficiency expert agent, and a mechanism expert agent, each of which is configured with different system prompts and evaluation weights. The multi-round debate verification adopts a round-based review protocol, specifically including: Round 1: The agent simultaneously distributes the preliminary operation and maintenance strategy draft to each expert agent via a message queue; Round Two: Each expert agent, based on its role and responsibilities defined by the system prompts and in conjunction with preset evaluation weights, reviews the draft from its respective evaluation dimensions and outputs review opinions and reasons; among them, the energy efficiency expert agent and the mechanism expert agent output opinions of approval or disapproval with explanations; the safety expert agent outputs opinions of approval, disapproval, or rejection with explanations. Round 3: If the security expert agent outputs a rejection opinion, the main control agent will revise the preliminary operation and maintenance strategy draft based on the rejection reasons and regenerate the revised strategy draft; if the security expert agent does not output a rejection opinion but other expert agents output opposing opinions, the preliminary operation and maintenance strategy draft will be revised based on the combined opposing reasons. Repeat the second and third rounds as described above until all expert agents output their approval opinions and the security expert agents do not output their disapproval opinions, or until the preset maximum number of debate rounds is reached, thus forming the final optimized operation instruction.

[0049] In detail, during the deployment phase, the system pre-deploys multiple expert agents in the multi-dimensional expert game layer 24. The instantiated expert agents include a security expert agent, an energy efficiency expert agent, and a mechanism expert agent. Each expert agent is configured with independent system prompts corresponding to its role and responsibilities, as well as preset evaluation weights. The evaluation weight of the security expert agent is set to the highest priority.

[0050] The multi-round debate and verification process for the preliminary operation and maintenance strategy draft adopts a pre-set round-based review protocol, and the specific execution process is as follows.

[0051] In the first round, the master control agent will simultaneously distribute the preliminary operation and maintenance strategy draft generated to the security expert agent, energy efficiency expert agent, and mechanism expert agent through a pre-built message queue, ensuring that all expert agents obtain the draft content to be reviewed at the same time.

[0052] In the second round, after receiving the initial draft of the operation and maintenance strategy, each expert agent, based on its role and responsibilities defined by the system prompts and in conjunction with preset evaluation weights, conducts a comprehensive review of the draft from its corresponding evaluation dimensions. Upon completion of the review, it outputs corresponding review opinions and complete explanations. Specifically, the energy efficiency expert agent and the mechanism expert agent can only output "for" or "against" opinions based on the review results, along with corresponding explanations; the safety expert agent can output "for," "against," or "reject" opinions based on the review results, along with corresponding explanations, with "reject" opinions having the highest priority and veto power.

[0053] In the third round, after collecting the review opinions of all expert agents, the master control agent first determines whether the security expert agent has issued a rejection opinion. If the security expert agent has issued a rejection opinion, the master control agent directly analyzes the corresponding rejection reasons from the rejection opinions, comprehensively revises the preliminary operation and maintenance strategy draft based on the rejection reasons, and regenerates the revised strategy draft. If the security expert agent has not issued a rejection opinion, but an energy efficiency expert agent or mechanism expert agent has issued an objection opinion, the master control agent collects the objection reasons corresponding to all objections, synthesizes all objection reasons, revises the preliminary operation and maintenance strategy draft based on all objection reasons, and generates the revised strategy draft.

[0054] After completing this round of revisions, repeat the review and revision process of the second and third rounds mentioned above until all expert agents output their approval opinions on the current strategy draft, and the security expert agents do not output their disapproval opinions, or the number of debate rounds executed reaches the maximum number of debate rounds preset by the system, then stop the debate process and determine the current strategy draft as the final optimized operation instruction.

[0055] Furthermore, the preliminary draft of the operation and maintenance strategy was revised based on the reasons for rejection, including: The physical boundary condition parameters that triggered the veto and the corresponding security risk types were extracted from the veto reasons. Based on the type of security risk, the main control agent determines whether the scope of operation in the current preliminary operation and maintenance strategy draft exceeds the safe operating range defined by the physical boundary condition parameters; If the limit is exceeded, the single large-scale operation instruction in the original draft will be decomposed into multiple step-like small-scale operation instructions. The operation range of each step is within the safe operating range, and a preset waiting delay will be inserted between adjacent steps to form a phased execution revision strategy draft. If the conditions are not exceeded but other potential risks exist, the physical constraint tool layer 23 is invoked to recalculate the alternative parameter combination that satisfies the physical constraint conditions indicated by the veto reason, and the corresponding operation parameters in the original draft are replaced with the alternative parameters to generate the revised strategy draft.

[0056] In detail, the main control agent first performs a structured analysis of the reasons for rejection, and accurately extracts the physical boundary condition parameters that triggered the rejection, as well as the type of safety risk corresponding to the rejection. The physical boundary condition parameters include the upper and lower safety limits and critical thresholds of the equipment operation, and the types of safety risks include overpressure risk, overtemperature risk, and mechanical stress over-limit risk.

[0057] After the analysis is completed, the main control agent checks the operation content in the preliminary operation and maintenance strategy draft based on the identified security risk type, and determines whether the operation range set in the draft exceeds the safe operation range limited by the physical boundary condition parameters.

[0058] If, upon verification, it is confirmed that the operational range in the draft exceeds the safe operating range, the main control agent will decompose the single large-scale operation instruction in the original draft into multiple step-like small-scale operation instructions. This ensures that the operational range corresponding to each step is within the safe operating range defined by the physical boundary condition parameters. At the same time, a preset waiting delay is inserted between two adjacent step operations for the stable transition and monitoring of the equipment status, ultimately forming a phased execution correction strategy draft.

[0059] If, upon verification, it is confirmed that the operational range in the draft does not exceed the safe operating range, but there are other potential security risks pointed out in the reasons for rejection, then the main control agent calls the deterministic computing engine in the physical constraint tool layer 23 through the application programming interface, takes the physical constraints indicated in the reasons for rejection as the computing boundary, recalculates the alternative parameter combinations that satisfy all physical constraints, and then replaces the corresponding operational parameters in the original draft with the calculated alternative parameters, finally generating a revised strategy draft that meets the security constraint requirements.

[0060] This embodiment achieves a balance between security constraints and operational efficiency by using targeted strategy modification methods to mitigate security veto risks while preserving the feasibility of operation and maintenance.

[0061] Furthermore, based on the reasons for opposition, the preliminary draft of the operation and maintenance strategy was revised, including: Collect objections from expert agents other than the security expert agent and their accompanying quantitative impact assessment values; All objections collected are sorted according to a preset constraint priority order, with safety-related constraints having the highest priority, followed by energy efficiency-related constraints, and then mechanism-related constraints. For the objection with the highest priority, adjust the corresponding operation parameters in the preliminary operation and maintenance strategy draft to meet the constraints required by the objection. The next priority objection is processed sequentially. If the adjusted operation parameters conflict with the currently processed objection, the conflict resolution rules are invoked to select a compromise parameter that partially satisfies both constraints. The compromise parameter is obtained by the main control agent by querying a pre-stored feasible parameter mapping table. After iterating through all the objections, the final parameter combination is encapsulated into a revised strategy draft.

[0062] In detail, the master control agent first collects the objections output by all expert agents other than the security expert agent, and extracts the quantitative impact assessment values ​​attached to each objection, thus completing the collection and organization of all objections and corresponding quantitative data.

[0063] The main control agent sorts all collected objections according to the system's preset constraint priority order. Objections related to safety constraints are given the highest priority, followed by those related to energy efficiency constraints, and then those related to mechanisms constraints, ensuring that the correction process prioritizes the fulfillment of high-priority constraints.

[0064] After sorting, the master control agent first adjusts the corresponding operation parameters in the preliminary operation and maintenance strategy draft for the objection with the highest priority. After the adjustment, it ensures that the operation parameters meet the constraints required by the objection.

[0065] The master control agent processes the objections of the next priority according to the sorting results. During the processing, if the operation parameters after the previous adjustment conflict with the constraints required by the objection being processed, the master control agent calls the preset conflict resolution rules and selects a compromise parameter that can simultaneously partially satisfy two conflicting constraints. This compromise parameter is obtained by the master control agent by querying the feasible parameter mapping table pre-stored in the system. The feasible parameter mapping table stores verified feasible parameter combinations that simultaneously satisfy multiple types of constraints.

[0066] After the master control agent has traversed all the sorted objections and completed all parameter adjustments, it encapsulates the final combination of operation parameters to generate a revised policy draft.

[0067] This embodiment uses a correction logic of priority sorting and conflict resolution to orderly resolve multi-dimensional constraint conflicts, ensuring that the revised strategy draft takes into account multiple industrial operation and maintenance goals.

[0068] In a specific implementation process, the security expert agent extracts the operational parameters from the draft, calculates the corresponding potential risk index, and compares it with the preset safety red line threshold and warning threshold: if the risk index exceeds the safety red line threshold, it outputs a rejection opinion and reasons for exceeding the risk limit; if it is between the warning threshold and the safety red line threshold, it outputs an objection opinion and reasons for risk warning; if it is below the warning threshold, it outputs a support opinion. The energy efficiency expert agent calculates the expected energy efficiency or economic benefits according to the draft's operating parameters. If the benefits are positive and exceed the preset threshold, it outputs a positive opinion; otherwise, it outputs a negative opinion and reasons for insufficient benefits. The mechanism expert agent calls the mechanism simulation solver to verify whether the draft's operating parameters meet the constraints of fluid mechanics, thermodynamics, or structural mechanics. If they do, it outputs a positive opinion; otherwise, it outputs a negative opinion and information indicating that the constraints are violated.

[0069] In detail, the security expert agent extracts the operational parameters from the preliminary operation and maintenance strategy draft, calculates the potential risk index corresponding to the operational parameters in combination with the real-time operating status of the equipment, compares the potential risk index with the preset safety red line threshold and warning threshold one by one, and outputs corresponding review opinions and complete reasons based on the comparison results. When the risk index exceeds the safety red line threshold, it outputs a rejection opinion and reasons for exceeding the risk limit; when the risk index is between the warning threshold and the safety red line threshold, it outputs an objection opinion and reasons for risk warning; and when the risk index is below the warning threshold, it outputs a support opinion.

[0070] Based on the operational parameters in the preliminary operation and maintenance strategy draft, and combined with on-site operational data such as real-time electricity prices, material costs, and production efficiency, the energy efficiency expert intelligent agent calculates the expected energy efficiency benefits or economic benefits corresponding to this operation and maintenance operation, and determines whether the benefits are positive and higher than the preset benefit threshold. If the requirements are met, it outputs a positive opinion; otherwise, it outputs a negative opinion and the specific reasons for insufficient benefits.

[0071] The mechanism expert agent retrieves the operation parameters from the preliminary operation and maintenance strategy draft, calls the mechanism simulation solver of the physical constraint tool layer 23, and verifies whether the operation parameters meet the physical constraint requirements of fluid mechanics, thermodynamics, or structural mechanics. If the verification passes, it outputs a positive opinion; if the verification fails, it outputs a negative opinion, while also marking the specific physical constraint type and boundary value that is violated.

[0072] Each expert agent will simultaneously feed back its review opinions and supporting reasons to the main control agent, providing a clear and actionable basis for the subsequent modification and optimization of operation and maintenance strategies.

[0073] 106. Before issuing the final optimized operation instruction, the operation parameters are extracted by a rigid rule interceptor and compared with the preset device hardware physical threshold. The issuance or termination operation is executed based on the comparison result.

[0074] In a specific implementation process, before the generated final optimized operation command is sent to the field control equipment, it must first pass through a rigid rule interceptor. The rigid rule interceptor extracts all operation parameters from the final optimized operation command, compares the extracted operation parameters with the pre-stored physical thresholds of the device hardware one by one, and executes the corresponding operation according to the comparison result. If the comparison result is in accordance with the rules, the operation is executed; if the comparison result exceeds the limit, the operation is terminated.

[0075] The rigid rule interceptor operates independently of the large language model and has a built-in immutable device factory physical threshold matrix. Before the final optimized operation instruction is issued, a rigid rule interceptor extracts the operation parameters and compares them with preset device hardware physical thresholds. Based on the comparison result, the issuance or termination operation is executed, including: At least one operation parameter is extracted from the final optimized operation instruction, the operation parameter including valve opening degree, spindle speed, temperature setpoint or discharge power; The operation parameters are compared with the corresponding absolute safety upper limit and absolute safety lower limit values ​​in the device's factory physical threshold matrix using hard-coded Boolean logic. If the operation parameter is between the absolute safety upper limit and the absolute safety lower limit, it is deemed compliant, and the final optimized operation instruction is compiled into executable control code and sent to the programmable logic controller or distributed control system. If the operation parameter exceeds the absolute safety upper limit or falls below the absolute safety lower limit, it is determined to be out of bounds, the veto right is exercised, the issuance is terminated, and an alarm is triggered.

[0076] In detail, during the deployment phase, the rigid rule interceptor is pre-deployed and configured independently. The rigid rule interceptor operates independently of the large language model and is not controlled or affected by the large language model. The rigid rule interceptor has an internal, tamper-proof device factory physical threshold matrix. The matrix stores the absolute safety upper limit and absolute safety lower limit values ​​corresponding to the hardware physical parameters of all devices on site. This threshold matrix only supports offline configuration and cannot be modified by software during operation.

[0077] Before the final optimized operation instructions are sent to the field control equipment, they must first undergo compliance verification through a rigid rule interceptor. The specific execution process is as follows.

[0078] The rigid rule interceptor first receives the final optimized operation instruction and extracts all operation parameters from the instruction. The extracted operation parameters include, but are not limited to, physical parameters that can directly control the operating status of the equipment, such as valve opening, spindle speed, temperature setpoint, and discharge power.

[0079] After the operation parameters are extracted, the rigid rule interceptor performs a hard-coded Boolean logic comparison between each operation parameter and the corresponding absolute safety upper limit and absolute safety lower limit values ​​in the device's factory physical threshold matrix. This comparison process is executed by the hardware logic circuit without any processing by the large language model.

[0080] If, after comparison, all operating parameters are within the corresponding absolute safety upper limit and absolute safety lower limit, the rigid rule interceptor determines that the final optimized operating instruction is compliant, compiles the final optimized operating instruction into control code that can be recognized and executed by the programmable logic controller or distributed control system, and sends it to the corresponding field control equipment for execution.

[0081] If, after comparison, any operational parameter exceeds the corresponding absolute safety upper limit or falls below the corresponding absolute safety lower limit, the rigid rule interceptor determines that the final optimized operation instruction is out of bounds, directly exercises its veto power, terminates the instruction issuance process, and triggers the corresponding out-of-bounds alarm to the on-site maintenance personnel, reporting the out-of-bounds parameters and corresponding threshold information.

[0082] This embodiment constructs the last line of defense for industrial operation and maintenance commands by using independent rigid rule interception and hard-coded threshold verification, preventing the issuance of dangerous commands that exceed the boundaries and ensuring the safe operation of industrial equipment.

[0083] The following is a specific example to illustrate: Example 1: Scenario of coordinated temperature control in chemical synthesis reactors: Scene composition: This embodiment is applied to the operation and maintenance scenario of reactors in high-risk chemical production lines. The system architecture fully matches the key core modules: the interactive perception layer connects to the reactor temperature and pressure sensors, the main control routing layer has a built-in graph retrieval enhancement engine, the physical constraint tool layer 23 is configured with a thermodynamic mechanism simulation solver, the multi-dimensional expert game layer 24 deploys a safety expert agent (highest weight) and a production efficiency expert agent, and the fallback execution layer has a built-in rigid rule interceptor with a hard-coded physical threshold matrix for the reactor.

[0084] Execution process: When an operator inputs the natural language operation and maintenance command "Reactor A is heating up too fast, how to handle it?", the system immediately freezes all industrial sensing data such as the reactor's internal temperature, jacket water temperature, and pressure at the current moment. It then uses a graph retrieval enhancement engine to search for topological knowledge of the reactor and its upstream and downstream pipelines, performs semantic deconstruction on the command, extracts equipment entities and target parameters, and generates structured parameters with unique equipment identifiers.

[0085] The main control agent calls the thermodynamic mechanism simulation solver of the physical constraint tool layer 23 through the API, and passes the structured parameters to the external computing environment. The large model is prohibited from directly generating physical numerical predictions throughout the process. The solver independently completes the thermodynamic deduction and outputs the deterministic numerical calculation result of "starting the backup cooling pump and adjusting the opening of the main circulating water valve to 80%".

[0086] The main control agent receives numerical results and transforms them into a preliminary operation strategy draft that conforms to the chemical operation and maintenance specifications through natural language generation capabilities, clarifying the operation objects, content, parameters and execution requirements.

[0087] The draft is broadcast to the multi-dimensional expert game layer 24, and verification is carried out using a round-based review protocol: After reviewing the production efficiency expert agent, it raises objections, believing that excessive valve opening will cause reactant crystallization and affect the yield; the safety expert agent, combined with the pressure rise rate, determines that failure to cool down immediately will trigger the risk of overpressure explosion, outputs a favorable opinion and exercises safety priority to suppress it, and finally reaches a consensus to form the final optimized operation instruction.

[0088] Before the final instruction is issued, it passes through a rigid rule interceptor to extract the valve opening parameters and compare them with the equipment's factory safety threshold. If it is confirmed that the 80% opening does not exceed the absolute safety range, it is deemed compliant and compiled into control code, which is then sent to the field programmable logic controller for execution.

[0089] Implementation results: This embodiment fully implements the core process of power control, deprives the large model of physical calculation authority, eliminates physical illusions through deterministic mechanism simulation, relies on multi-expert game with a safety veto power to safeguard the safety bottom line in high-risk scenarios, and finally completes the safety backup through rigid interception, avoiding the illegal decision of "gradually cooling down production rate" in traditional solutions, and eliminating major safety accidents such as explosions and overheating.

[0090] Example 2: Adaptive Degradation Operation Scenario for Heavy-Duty CNC Machine Tool Spindles: Scene composition: This embodiment is applied to the operation and maintenance scenario of CNC equipment in a precision machining workshop. The system architecture matches the key core modules: the interactive perception layer connects to the spindle vibration sensor and the key workstation visual inspection model; the main control routing layer has a built-in graph retrieval enhancement engine; the physical constraint tool layer 23 configures the dynamic stress calculation engine; the multi-dimensional expert game layer 24 deploys equipment life expert intelligent agents and mechanism expert intelligent agents; and the bottom-line execution layer has a built-in rigid rule interceptor with hard-coded thresholds for spindle speed and feed rate.

[0091] Execution process: The system identifies minor tool chipping using a visual inspection model and simultaneously collects abnormal spindle vibration frequency data. When the operator inputs a natural language maintenance command, "Handle the current tool abnormality to avoid workpiece scrapping and spindle damage," the system immediately freezes multimodal industrial sensing data such as spindle speed, vibration amplitude, and tool defect status. It then uses a graph retrieval enhancement engine to search for topological relationships between the spindle, tool, and feed system, semantically deconstructs the command, and generates structured parameters.

[0092] The main control agent calls the dynamic stress calculation engine of the physical constraint tool layer 23 through the API, passes the structured parameters to the external computing environment, prohibits the large model from directly generating numerical predictions, and the engine independently completes the mechanical deduction and outputs the deterministic numerical calculation result of "immediate stop and tool change".

[0093] The main control agent receives the numerical results and transforms them into a preliminary operation and maintenance strategy draft that conforms to the machining operation and maintenance specifications.

[0094] The draft is broadcast to the 24th layer of the multi-dimensional expert game for round-based review: the equipment lifespan expert agent reviews and votes in favor, believing that an emergency stop can protect the spindle from damage; the mechanism expert agent outputs opposing opinions, pointing out that the current stage is a critical surface finishing process, and a direct emergency stop would lead to workpiece scrap, suggesting "reducing the spindle speed by 30%, reducing the feed rate by 50%, and retracting the tool after completing the current cut." The main control agent integrates the opposing reasons according to the constraint priority, adjusts the operating parameters to resolve the conflict, and finally reaches a consensus to form the final optimized operating instructions for phased degraded operation.

[0095] Before the final instruction is issued, it passes through a rigid rule interceptor to extract the rotational speed and feed rate parameters and compare them with the equipment's factory safety thresholds. Once the parameters are confirmed to be within the safe range and deemed compliant, the instruction is sent to the machine tool control system for execution.

[0096] Implementation results: This embodiment implements core elements such as multimodal data acquisition and multi-expert game conflict resolution. It ensures that decisions conform to physical laws through deterministic mechanical calculations and balances equipment safety and production yield by relying on multi-agent cross-validation. It avoids workpiece scrapping and economic losses caused by direct emergency stop and eliminates the equipment risk of spindle overload damage.

[0097] Example 3: Coordinated scheduling scenario of industrial microgrid and energy storage base station: Scene composition: This embodiment is applied to the operation and maintenance scenario of energy storage in a factory's own microgrid. The system architecture matches the key core modules: the interactive perception layer connects to the battery pack temperature and charge / discharge power sensors, the main control routing layer has a built-in graph retrieval enhancement engine, the physical constraint tool layer 23 configures the value stream evaluation engine and the battery charge / discharge thermal model simulation solver, the multi-dimensional expert game layer 24 deploys energy efficiency expert agents and safety expert agents, and the fallback execution layer has a built-in rigid rule interceptor with hard-coded thresholds for battery charge / discharge power and temperature.

[0098] Execution process: During peak summer electricity consumption periods, operators input natural language operation and maintenance commands such as "optimize the charging and discharging strategy of the energy storage power station to maximize revenue." The system immediately freezes industrial sensing data such as the current battery pack SOC, core temperature, and real-time electricity price. Through the graph retrieval enhancement engine, it retrieves the topological relationship knowledge of the energy storage unit, inverter, and load, performs semantic deconstruction on the command, and generates structured parameters.

[0099] The main control agent calls the value stream evaluation engine and thermal model simulation solver of the physical constraint tool layer 23 through API, and passes the structured parameters to the external computing environment. It prohibits large models from directly generating numerical predictions. The engine independently completes the deduction of economic benefits and thermal effects, and outputs the deterministic numerical calculation result of "energy storage battery fully discharged to obtain the maximum benefit of peak-valley price difference".

[0100] The main control agent receives the numerical results and transforms them into a preliminary operation and maintenance strategy draft that conforms to power operation and maintenance specifications.

[0101] The draft is broadcast to the multi-dimensional expert game layer 24 for round-based review: the energy efficiency expert agent reviews and votes in favor, confirming that the benefits of full-load discharge meet expectations; the safety expert agent reviews and outputs a rejection opinion, pointing out that the battery core temperature will exceed the safety red line after 20 minutes of full-load discharge, posing a risk of thermal runaway. Based on the rejection reasons, the main control agent breaks down the single full-load discharge command into a stepped discharge operation, re-calls the physical constraint tool layer 23 for calculation, and generates a revised strategy draft of "discharging at 70% rated power and simultaneously starting the liquid cooling system". After a second review, it is unanimously approved, forming the final optimized operation command.

[0102] Before the final instruction is issued, it passes through a rigid rule interceptor to extract the discharge power and temperature setpoints and compare them with the equipment's factory safety thresholds. After confirming that the parameters are within the safe range and determining compliance, the instruction is sent to the energy storage control system for execution.

[0103] Implementation results: This embodiment implements core elements such as the invocation of deterministic tools for implementing rights, strategy modification after security veto, and rigid interception. Through deterministic calculations of value stream and thermal effect, it balances economic benefits and equipment safety, avoids the risks of battery aging and thermal runaway caused by short-sighted full-load discharge, and achieves a long-term balance between economic benefits and equipment life cycle.

[0104] Figure 3This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340. The processor 310, communication interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions from the memory 330 to execute an industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints.

[0105] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints provided by the above methods.

[0107] It should be noted that all relevant information that may be involved in the various embodiments of the present invention is processed in strict accordance with the requirements of laws and regulations, following the principles of legality, legitimacy, and necessity, based on the reasonable purpose of the business scenario, and is information that users actively provide or generate during the use of the product / service, as well as information obtained with user authorization.

[0108] The information processed by this invention may vary depending on the specific product / service scenario and should be based on the specific scenario in which the user uses the product / service. This may involve user account information, device information, or other related information. This invention will treat the relevant information and its processing with the utmost diligence.

[0109] This invention places great emphasis on the security of relevant information and has adopted reasonable and feasible security protection measures that comply with industry standards to protect user information and prevent unauthorized access, public disclosure, use, modification, damage or loss of relevant information.

[0110] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0112] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints, characterized in that, include: The system receives natural language operation and maintenance instructions, freezes and collects industrial sensing data, and uses a graph retrieval enhancement engine to retrieve equipment topology knowledge. It then performs semantic deconstruction on the natural language operation and maintenance instructions, extracts equipment entities and target parameters, and generates structured parameters. Based on the structured parameters, the deterministic computation engine in the physical constraint tool layer is invoked through the application programming interface to pass the structured parameters to the external computing environment. During the invocation process, direct generation of physical numerical predictions is prohibited; The physical and mathematical deductions are performed in the external computing environment, and deterministic numerical calculation results are output. Receive the numerical calculation results and use natural language generation capabilities to convert them into a preliminary draft of operation and maintenance strategy in natural language form; The preliminary operation and maintenance strategy draft is broadcast to the multi-dimensional expert game layer, where multiple expert agents conduct multiple rounds of debate and verification on the preliminary operation and maintenance strategy draft, and perform logical game based on the review protocol until a consensus is reached, forming the final optimized operation instructions. Before the final optimized operation instruction is issued, the operation parameters are extracted by a rigid rule interceptor and compared with the preset device hardware physical threshold. The operation is then executed or terminated based on the comparison result.

2. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 1, characterized in that, The graph retrieval enhancement engine constructs a knowledge graph based on the equipment building information model and the physical topology connections between equipment. The semantic deconstruction process includes: Natural language processing technology is used to identify the device entity name and physical parameter keywords in the natural language operation and maintenance instructions. Graph retrieval technology is used to locate the core node of the corresponding device entity in the knowledge graph and automatically associate the core node with the upstream device, downstream device and parallel redundant pipeline in the topology. Based on the device entity name, physical parameter keywords, and the topology relationship, the natural language operation and maintenance instructions are converted into structured parameters containing the device's unique identifier, the associated topology path, and the set of affected parameters.

3. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 1, characterized in that, The deterministic computing engine includes a sensitivity analysis engine, a value stream evaluation engine, or a mechanism simulation solver. The sensitivity analysis engine is used to calculate the sensitivity of each physical operation variable to the current global optimization objective function value and output the sensitivity value. The value stream evaluation engine is used to calculate the estimated economic benefits under different operating strategies based on real-time electricity prices, material costs, or production efficiency data, and output the value stream evaluation value. The mechanism simulation solver is used to perform numerical simulations of physical processes based on thermodynamic, fluid mechanics, or structural mechanics equations, and outputs predicted values ​​of the changing trends of physical variables.

4. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 3, characterized in that, When the deterministic computation engine is a sensitivity analysis engine, the physical and mathematical deduction process includes: The sensitivity values ​​of each physical operation variable are calculated based on the partial derivatives of the current global optimization objective function value with respect to each physical operation variable, the ratio of each physical operation variable to the current global optimization objective function value, and the exponential penalty term constructed based on the difference between the potential risk index and the safety red line threshold. The larger the sensitivity value, the more preferentially the corresponding physical operation variable is adjusted.

5. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 1, characterized in that, The multidimensional expert game layer includes a security expert agent, an energy efficiency expert agent, and a mechanism expert agent. Each expert agent is configured with different system prompt words and evaluation weights. The multi-round debate verification adopts a round-based review protocol, specifically including: Round 1: The agent simultaneously distributes the preliminary operation and maintenance strategy draft to each expert agent via a message queue; Round Two: Each expert agent, based on its role and responsibilities defined by the system prompts and in conjunction with preset evaluation weights, reviews the draft from its respective evaluation dimensions and outputs review opinions and reasons; among them, the energy efficiency expert agent and the mechanism expert agent output opinions of approval or disapproval with explanations; the safety expert agent outputs opinions of approval, disapproval, or rejection with explanations. Round 3: If the security expert agent outputs a rejection opinion, the main control agent will revise the preliminary operation and maintenance strategy draft based on the rejection reasons and regenerate the revised strategy draft; if the security expert agent does not output a rejection opinion but other expert agents output opposing opinions, the preliminary operation and maintenance strategy draft will be revised based on the combined opposing reasons. Repeat the second and third rounds as described above until all expert agents output their approval opinions and the security expert agents do not output their disapproval opinions, or until the preset maximum number of debate rounds is reached, thus forming the final optimized operation instruction.

6. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 5, characterized in that, The preliminary operation and maintenance strategy draft was revised based on the reasons for rejection, including: The physical boundary condition parameters that triggered the veto and the corresponding security risk types were extracted from the veto reasons. Based on the type of security risk, the main control agent determines whether the scope of operation in the current preliminary operation and maintenance strategy draft exceeds the safe operating range defined by the physical boundary condition parameters; If the limit is exceeded, the single large-scale operation instruction in the original draft will be decomposed into multiple step-like small-scale operation instructions. The operation range of each step is within the safe operating range, and a preset waiting delay will be inserted between adjacent steps to form a phased execution revision strategy draft. If no risks are exceeded but other potential risks exist, the physical constraint tool layer is invoked to recalculate the alternative parameter combination that satisfies the physical constraint conditions indicated by the veto reason, and the corresponding operation parameters in the original draft are replaced with the alternative parameters to generate a revised strategy draft.

7. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 5, characterized in that, Based on the reasons for opposition, the preliminary draft of the operation and maintenance strategy was revised, including: Collect objections from expert agents other than the security expert agent and their accompanying quantitative impact assessment values; All objections collected are sorted according to a preset constraint priority order, with safety-related constraints having the highest priority, followed by energy efficiency-related constraints, and then mechanism-related constraints. For the objection with the highest priority, adjust the corresponding operation parameters in the preliminary operation and maintenance strategy draft to meet the constraints required by the objection. The next priority objection is processed sequentially. If the adjusted operation parameters conflict with the currently processed objection, the conflict resolution rules are invoked to select a compromise parameter that partially satisfies both constraints. The compromise parameter is obtained by the main control agent by querying a pre-stored feasible parameter mapping table. After iterating through all the objections, the final parameter combination is encapsulated into a revised strategy draft.

8. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 5, characterized in that, Each expert agent, based on its role and responsibilities defined by the system prompts and in conjunction with preset evaluation weights, reviews the draft from its respective evaluation dimensions and outputs review opinions and reasons, including: The security expert agent extracts the operational parameters from the draft, calculates the corresponding potential risk index, and compares it with the preset safety red line threshold and warning threshold: if the risk index exceeds the safety red line threshold, it outputs a rejection opinion and reasons for exceeding the risk limit; if it is between the warning threshold and the safety red line threshold, it outputs an objection opinion and reasons for risk warning; if it is below the warning threshold, it outputs a support opinion. The energy efficiency expert agent calculates the expected energy efficiency or economic benefits according to the draft's operating parameters. If the benefits are positive and exceed the preset threshold, it outputs a positive opinion; otherwise, it outputs a negative opinion and reasons for insufficient benefits. The mechanism expert agent calls the mechanism simulation solver to verify whether the draft's operating parameters meet the constraints of fluid mechanics, thermodynamics, or structural mechanics. If they do, it outputs a positive opinion; otherwise, it outputs a negative opinion and information indicating that the constraints are violated.

9. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 1, characterized in that, The rigid rule interceptor operates independently of the large language model and has a built-in, tamper-proof device factory physical threshold matrix. Before the final optimized operation instruction is issued, a rigid rule interceptor extracts the operation parameters and compares them with preset device hardware physical thresholds. Based on the comparison result, the issuance or termination operation is executed, including: At least one operation parameter is extracted from the final optimized operation instruction, the operation parameter including valve opening degree, spindle speed, temperature setpoint or discharge power; The operation parameters are compared with the corresponding absolute safety upper limit and absolute safety lower limit values ​​in the device's factory physical threshold matrix using hard-coded Boolean logic. If the operation parameter is between the absolute safety upper limit and the absolute safety lower limit, it is deemed compliant, and the final optimized operation instruction is compiled into executable control code and sent to the programmable logic controller or distributed control system. If the operation parameter exceeds the absolute safety upper limit or falls below the absolute safety lower limit, it is determined to be out of bounds, the veto right is exercised, the issuance is terminated, and an alarm is triggered.

10. The industrial multi-agent operation and maintenance decision-making method based on retrieval enhancement and physical constraints according to claim 1, characterized in that, Collect and freeze industrial sensing data, including: Structured sensor time-series data are synchronously acquired via an industrial fieldbus. The structured sensor time-series data includes temperature, pressure, rotational speed, and vibration amplitude. Meanwhile, unstructured visual data is collected by a visual inspection model deployed at key workstations. The visual inspection model uses an instance segmentation algorithm to extract the defect state vector of the equipment surface from real-time images. The defect state vector includes crack length, wear area, or the location and severity of leaks. The structured sensor time series data and the unstructured visual data are aligned and fused under a unified timestamp to form a multimodal state snapshot at the current moment, thereby freezing the industrial sensing data.

Images