Nuclear power maintenance decision-making system and method based on multi-Agent cooperation

Abstract

The invention belongs to the technical field of nuclear power station maintenance management, and particularly relates to a nuclear power maintenance decision-making system and method based on multi-Agent cooperation. The system comprises an input layer, a multi-Agent cooperation layer, a knowledge support layer, a decision processing layer, an output and interaction layer and a feedback learning layer. The input layer receives initial work order information including equipment basic information and fault description; the multi-Agent collaboration layer allocates sub-tasks to predefined six types of professional Agents according to work order information and completes information interaction; the knowledge support layer comprises a nuclear power professional knowledge base, a historical maintenance database and a rule and regulation library; the decision processing layer performs conflict detection, negotiation and decision fusion on Agent output; the output and interaction layer provides a decision result display and man-machine interaction interface; and the feedback learning layer realizes comprehensive improvement of nuclear power maintenance decision-making efficiency and safety. The method has the beneficial effects that a multi-professional Agent collaborative network technical means is adopted, and the technical effects of maintenance decision cross-professional information instant sharing and efficient collaboration are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of nuclear power plant maintenance management, and particularly relates to a nuclear power maintenance decision system and method based on multi-agent cooperation. BACKGROUND

[0002] The traditional nuclear power engineering maintenance decision process has great limitations. At present, nuclear power plants mainly rely on single professional perspective and experience judgment in the process of equipment maintenance, and lack systematic multi-professional collaborative decision mechanism. In the existing technical solutions, when the user reports the equipment, only the fault phenomenon and the cause are simply described. Although the nuclear power plant has generally adopted a digital EAM work order system to support online operation, the maintenance personnel still need to do a lot of preparation work after receiving the work order. The maintenance personnel need to manually complete several dozen kinds of field information such as unit, plant room, sensitive area, system version, equipment classification, nuclear safety plan, quality plan, defect analysis, work materials, work steps, etc. according to the reported basic work order information (such as equipment code, equipment name, etc.), and then cross-check manually to confirm the completeness of the work order information and operation steps before going to the scene for maintenance. The whole maintenance preparation process is time-consuming and long, and often takes more than 2 hours to fill in the relevant information for a work order.

[0003] There are some improved solutions in the industry at present, such as "a nuclear power plant equipment fault diagnosis method based on deep learning", which proposes a single intelligent agent fault diagnosis method, but this method only focuses on fault identification and does not involve multi-professional collaboration in maintenance decision; such as "nuclear power plant equipment maintenance management system", which provides a digital platform for maintenance management, but lacks intelligent decision assistance function.

[0004] These existing technical solutions have the following main problems:

[0005] 1. The decision-making process highly depends on the experience and judgment of single professional personnel, and lacks systematic integration of multi-professional perspectives;

[0006] 2. There is information asymmetry among the parties involved in maintenance decision-making, low collaboration efficiency, and professional barriers;

[0007] 3. It is difficult to simultaneously consider safety risk assessment, resource optimization and cost control in complex maintenance scenarios;

[0008] 4. A large amount of manual operation is required in the work order preparation process, which is time-consuming and has the risk of human error;

[0009] 5. It is difficult to effectively accumulate and utilize historical maintenance experience, and there is a risk of knowledge transmission gap.

[0010] Although some systems attempt to improve efficiency through intelligent work order filling, such as some implementations can shorten the work order filling preparation process to 5 minutes, these systems usually only focus on the automation of information filling and fail to solve the problem of multi-specialty collaboration from the decision-making level.

[0011] Therefore, there is an urgent need for a system that can simulate multi-specialty collaboration and implement intelligent maintenance decision-making to improve the scientificity, safety and efficiency of nuclear power plant maintenance decision-making. SUMMARY

[0012] The purpose of the present application is to provide a nuclear power maintenance decision-making system and method based on multi-agent collaboration, which can solve the technical problems of insufficient multi-specialty collaboration, information silos and low decision-making efficiency in the process of nuclear power maintenance decision-making.

[0013] The technical solution of the present application is as follows: a nuclear power maintenance decision-making system based on multi-agent collaboration, comprising an input layer, a multi-agent collaboration layer, a knowledge support layer, a decision-making processing layer, an output and interaction layer, and a feedback learning layer; the input layer receives initial work order information, including device basic information and fault description; the multi-agent collaboration layer assigns sub-tasks to six types of pre-defined professional agents according to the work order information through task decomposition and fitness algorithm, and completes information interaction with the help of DAG communication protocol; the knowledge support layer includes a nuclear power professional knowledge base, a historical maintenance database and a regulations and rules database; the decision-making processing layer performs conflict detection, negotiation and decision-making fusion on the output of the agent; the output and interaction layer provides a decision-making result display and a human-computer interaction interface; the feedback learning layer updates knowledge and rules based on execution results and expert modifications to achieve comprehensive improvement of nuclear power maintenance decision-making efficiency and safety.

[0014] It includes a task decomposition and distribution module, an agent communication protocol module, a knowledge sharing module, a multi-agent decision-making generation module, a conflict detection and negotiation module, a decision-making fusion and scoring module, and a human-computer hybrid decision-making module.

[0015] The task decomposition and distribution module inputs: maintenance work order basic information W, and outputs: sub-task set T and distribution result A;

[0016] The agent communication protocol module is based on a directed acyclic graph communication model and a communication protocol implementation method, inputs: agent set A and communication dependency graph G, and outputs: agent state set S after communication is completed;

[0017] The knowledge sharing module realizes knowledge sharing between agents using a shared semantic space model, inputs: agent professional knowledge K_i and decision-making context C, and outputs: shared knowledge space S;

[0018] The multi-Agent decision generation module inputs: maintenance work order W, Agent set A, knowledge base K; outputs: decision scheme set D;

[0019] The conflict detection and negotiation module inputs: decision scheme set D, negotiation rule set R; outputs: negotiated decision scheme D';

[0020] The decision fusion and scoring module inputs: negotiated decision scheme D', scoring rule E; outputs: final decision scheme F;

[0021] The human-machine hybrid decision framework inputs: system decision F, human expert E, interactive interface I, outputs: final execution scheme P.

[0022] The task decomposition and distribution module has the following processing process:

[0023] Step 1: Analyze the work order W and extract the key information set I = {equipment code, fault description, task type...};

[0024] Step 2: Initialize the subtask set T = {};

[0025] Step 3: Generate subtasks according to the preset task decomposition template M;

[0026] For each template m e M, if m. Condition (I) == True, T = T U m. Generated subtasks (I);

[0027] Step 4: For each subtask t e T:

[0028] Calculate the fitness score of each Agent: S(t, a) = f, t is the knowledge requirement, a is the professional field, a is the historical performance, select the Agent with the highest score to assign the task, return the subtask set T and the assignment result A.

[0029] The Agent communication protocol module has the following processing process:

[0030] Step 1: Initialize the state of all Agents to "wait";

[0031] Step 2: Find the Agent set H with no in-degree nodes in the graph G;

[0032] Step 3: Set the state of all Agents in H to "ready";

[0033] Step 4: When there are "ready" state Agents: execute all "ready" state Agent tasks in parallel;

[0034] For each finished task Agent a: send message to all dependent a Agents, update a state to "finished", update dependency graph G, recalculate in-degree node set H, set Agents in H with state "waiting" to "ready";

[0035] Step 5: when all Agent states are "finished", end the communication process;

[0036] Return Agent state set S.

[0037] The knowledge sharing module, the specific implementation method is as follows:

[0038] Step 1: initialize shared knowledge space S = {};

[0039] Step 2: for each Agent i, knowledge projection: P_i = projection function (K_i, C), knowledge conversion: T_i = conversion function (P_i), S = S U T_i;

[0040] Step 3: consistency verification is performed on the knowledge elements in S:

[0041] Detect conflict knowledge C_k, for each conflict C_k, apply conflict resolution rule R;

[0042] Step 4: optimize knowledge space S, delete redundant knowledge;

[0043] Step 5: index and access mechanism is established for shared knowledge

[0044] Return shared knowledge space S.

[0045] The multi-Agent decision generation module, the implementation process is as follows:

[0046] Step 1: initialize decision space D = {};

[0047] Step 2: execute each Agent decision process in parallel;

[0048] For each Agent a e A, extract work order related information I_a = extraction function (W, a. professional), load professional knowledge K_a = loading function (K, a. professional), decision reasoning: d_a = reasoning function (I_a, K_a), D = D U d_a;

[0049] Step 3: preliminary integration is performed on the decision space D

[0050] Merge similar decisions, remove obvious error decisions, and return decision scheme set D.

[0051] The conflict detection and negotiation module, the implementation process is as follows:

[0052] Step 1: Construct the decision conflict graph G:

[0053] Nodes are the decisions of each Agent, edges e(i, j) represent that there is a conflict between decisions d_i and d_j, and edge weights w(i, j) represent the severity of the conflict;

[0054] Step 2: Conflict clustering:

[0055] Apply graph clustering algorithm to identify conflict clusters C = {c_1, c_2,... c_n};

[0056] Step 3: For each conflict cluster c_i:

[0057] Initialize negotiation round t = 0, when there is a conflict and t < maximum rounds T: each Agent proposes an adjustment scheme, evaluate the comprehensive impact of the adjustment scheme, select the optimal adjustment scheme, update the decisions of each Agent, re-detect conflicts, t = t + 1;

[0058] Step 4: For conflicts that cannot be resolved through negotiation:

[0059] Mark as a point of manual intervention, return: the decision scheme D' after negotiation.

[0060] The decision fusion and scoring module realizes the process as follows:

[0061] Step 1: Initialize the fusion decision scheme F = {};

[0062] Step 2: For each dimension m in the decision scheme D':

[0063] Collect the decisions of each Agent on dimension m {d_1, m, d_2, m,..., d_n, m}, calculate the decision credibility: C(d_i, m) = f(Agent expertise, decision basis strength, historical accuracy), weighted fusion:

[0064] F_m = ∑(C(d_i, m) × d_i, m) / ∑C(d_i, m), F = F ∪ F_m;

[0065] Step 3: Overall score for the fusion decision F:

[0066] Safety score S_s = safety scoring function(F), feasibility score S_f = feasibility scoring function(F), economy score S_e = economy scoring function(F), comprehensive score S = w_s × S_s + w_f × S_f + w_e × S_e;

[0067] Step 4: If S < threshold T:

[0068] Mark as needing manual review, return the final decision scheme F and score.

[0069] The human-machine hybrid decision-making framework comprises:

[0070] Step 1: The system generates an initial decision scheme F, including: maintenance scheme suggestions, safety measure suggestions, resource allocation suggestions, risk analysis reports;

[0071] Step 2: According to the decision risk level R:

[0072] If R < low risk threshold T_l: the system decision is automatically passed; otherwise, if R < medium risk threshold T_m: push to the supervising expert for quick review, otherwise: push to the expert team for detailed review;

[0073] Step 3: The expert reviews the decision through the interactive interface I:

[0074] Visualize the decision scheme, provide the basis for decision-making and reasoning process, and allow the expert to modify specific decision items;

[0075] Step 4: The system records the expert's modifications:

[0076] Capture the modification items δ = {δ_1, δ_2,..., δ_n}, analyze the modification mode and reason, and update the relevant Agent knowledge and rules;

[0077] Step 5: Generate the final execution scheme P:

[0078] Integrate the system decision F and the expert's modifications δ, automatically generate an execution document, and distribute it to the relevant execution system to return: the final execution scheme P.

[0079] A nuclear power maintenance decision-making method based on multi-Agent collaboration, comprising the following steps:

[0080] Task decomposition and distribution process:

[0081] Step 1: Analyze the work order W and extract the key information set I = {equipment code, fault description, task type...};

[0082] Step 2: Initialize the subtask set T = {};

[0083] Step 3: Generate subtasks according to the preset task decomposition template M;

[0084] For each template m ∈ M, if m. Condition (I) == True, T = T ∪ m. Generate subtasks (I);

[0085] Step 4: For each subtask t ∈ T:

[0086] Calculate the fitness score of each Agent: S(t, a) = f, t is the knowledge requirement, a is the professional field, a is the historical performance, assign the task to the Agent with the highest score, return the sub-task set T and the assignment result A;

[0087] Agent communication protocol, the process is as follows:

[0088] Step 1: Initialize all Agent states to "wait";

[0089] Step 2: Find the set of Agents H with no incoming degree in graph G;

[0090] Step 3: Set the state of all Agents in H to "ready";

[0091] Step 4: When there are Agents in "ready" state: execute the tasks of all "ready" state Agents in parallel;

[0092] For each Agent a that completes a task: send a message to all Agents that depend on a, update the state of a to "complete", update the dependency graph G, recalculate the set of nodes with no incoming degree H, set the Agents in H with state "wait" to "ready";

[0093] Step 5: When all Agents are in "complete" state, end the communication process;

[0094] Return the set of Agent states S;

[0095] Knowledge sharing, the implementation method is as follows:

[0096] Step 1: Initialize the shared knowledge space S = {};

[0097] Step 2: For each Agent i, knowledge projection: P_i = projection function (K_i, C), knowledge conversion: T_i = conversion function (P_i), S = S ∪ T_i;

[0098] Step 3: Verify the consistency of the knowledge elements in S:

[0099] Detect conflicting knowledge C_k, for each conflict C_k, apply conflict resolution rules R;

[0100] Step 4: Optimize the knowledge space S, delete redundant knowledge;

[0101] Step 5: Establish an index and access mechanism for shared knowledge

[0102] Return the shared knowledge space S;

[0103] Multi-Agent decision generation, the implementation process is as follows:

[0104] Step 1: Initialize decision space D = {};

[0105] Step 2: Execute each Agent decision process in parallel;

[0106] For each Agent a e A, extract ticket-related information I_a = extract function(W, a. specialty), load specialized knowledge K_a = load function(K, a. specialty), decision reasoning: d_a = inference function(I_a, K_a), D = D U d_a;

[0107] Step 3: Preliminary integration of decision space D

[0108] Merge similar decisions, remove obvious erroneous decisions, and return the decision scheme set D;

[0109] Conflict detection and negotiation, the implementation process is as follows:

[0110] Step 1: Construct decision conflict graph G:

[0111] Nodes are each Agent decision d_i, edges e(i, j) represent that decisions d_i and d_j have conflicts, and edge weights w(i, j) represent the severity of conflicts;

[0112] Step 2: Conflict clustering:

[0113] Apply graph clustering algorithm to identify conflict clusters C = {c_1, c_2,... c_n};

[0114] Step 3: For each conflict cluster c_i:

[0115] Initialize negotiation round t = 0, when there is conflict and t < maximum round T: each Agent proposes adjustment scheme, evaluate the comprehensive impact of adjustment scheme, select the optimal adjustment scheme, update each Agent decision, re-detect conflict, t = t + 1;

[0116] Step 4: For conflicts that cannot be resolved through negotiation:

[0117] Mark as a point of human intervention, return: negotiated decision scheme D';

[0118] Decision fusion and scoring, the implementation process is as follows:

[0119] Step 1: Initialize fusion decision scheme F = {};

[0120] Step 2: For each dimension m in the decision scheme D':

[0121] Collect the decisions of each Agent on dimension m {d_1, m, d_2, m,..., d_n, m}, calculate the decision credibility: C(d_i, m) = f(Agent expertise, decision basis strength, historical accuracy), weighted fusion:

[0122] F_m = ∑(C(d_i, m) × d_i, m) / ∑C(d_i, m), F = F ∪ F_m;

[0123] Step 3: Overall score of the fusion decision F:

[0124] Safety score S_s = safety score function (F), feasibility score S_f = feasibility score function (F), economy score S_e = economy score function (F), comprehensive score S = w_s × S_s + w_f × S_f + w_e × S_e;

[0125] Step 4: If S < threshold T:

[0126] Marked for manual review, return the final decision scheme F and score;

[0127] The human-machine hybrid decision-making framework includes:

[0128] Step 1: The system generates an initial decision scheme F, including: maintenance scheme suggestion, safety measure suggestion, resource allocation suggestion, risk analysis report;

[0129] Step 2: According to the decision risk level R:

[0130] If R < low risk threshold T_l: system decision automatically passes; otherwise if R < medium risk threshold T_m: push to supervising experts for quick review, otherwise: push to expert team for detailed review;

[0131] Step 3: Experts review the decision through the interactive interface I:

[0132] Visualize the decision scheme, provide decision basis and reasoning process, allow experts to modify specific decision items;

[0133] Step 4: The system records expert modifications:

[0134] Capture modification items δ = {δ_1, δ_2,..., δ_n}, analyze modification patterns and reasons, update related Agent knowledge and rules;

[0135] Step 5: Generate the final execution scheme P:

[0136] Integrate system decision F and expert modification δ, automatically generate execution document, distribute to related execution system return: final execution scheme P.

[0137] The beneficial effects of the present application are that: the multi-specialized Agent collaboration network technology means is adopted, the technical effects of maintenance decision cross-specialty information instant sharing and efficient collaboration are realized. The Agent communication protocol technology means based on DAG is introduced, the technical effects of decision flow parallel execution without deadlock and overall time delay significantly shortened are realized. The shared semantic space knowledge fusion technology means is constructed, the technical effects of heterogeneous professional knowledge unified management and conflict automatic solution are realized. The decision fusion and multi-dimensional scoring technology means of credibility weighting are designed, the technical effects of comprehensive optimization of maintenance scheme in safety, feasibility and economy are realized. The man-machine hybrid closed-loop learning technology means is established, the technical effects of system adaptive iteration updating and work order preparation time from 2 hours to 15 minutes are realized. BRIEF DESCRIPTION OF DRAWINGS

[0138] Figure 1 The flow chart of the work flow of the nuclear power maintenance decision system based on multi-Agent collaboration provided by the present application is shown in the figure.

[0139] Figure 2 The detailed flow chart of Agent collaboration decision is shown in the figure. DETAILED DESCRIPTION

[0140] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0141] Specifically, the present application realizes the transformation from single professional decision to multi-specialty collaborative decision by constructing a plurality of specialized Agent collaboration networks to simulate the collaborative decision-making process of different post professionals in nuclear power plants. The system uses a specialized Agent role system to perform multi-dimensional analysis on maintenance tasks, realizes the ordered interaction between Agents through a collaboration protocol based on nuclear power safety specifications, and designs an automatic identification and negotiation solution mechanism for decision conflicts, as well as an interactive framework for collective intelligence and human expert hybrid decision-making, thereby significantly improving the scientificity, comprehensiveness and efficiency of nuclear power maintenance decision-making under the premise of safety.

[0142] Among them, the Agent includes six categories, namely maintenance engineer Agent, safety expert Agent, plan scheduling Agent, equipment expert Agent, operation expert Agent, quality supervision Agent

[0143] The present application also makes full use of multi-specialty knowledge base and historical maintenance data, through the multi-Agent collaboration mechanism, after receiving the basic work order information, it can automatically complete equipment diagnosis, risk assessment, resource planning and work step generation and other professional tasks, shortening the traditional work order preparation process of more than 2 hours to within 15 minutes, while improving the decision-making quality and safety, effectively solving the multi-specialty collaboration problem in nuclear power maintenance decision-making. Specifically includes:

[0144] (1) Input layer: Receives initial work order information, including basic equipment information, fault description, etc.;

[0145] (2) Multi-Agent Collaboration Layer: A collaborative network composed of multiple specialized agents, which is the core of the system;

[0146] (3) Knowledge support layer: includes nuclear power professional knowledge base, historical maintenance database, rules and regulations database, etc.

[0147] (4) Decision Processing Layer: Responsible for integrating Agent opinions, conflict detection and negotiation, and decision generation;

[0148] (5) Output and Interaction Layer: Provides decision result display and human-computer interaction interface;

[0149] (6) Feedback learning layer: collects decision execution feedback and optimizes system performance.

[0150] A nuclear power plant maintenance decision-making system based on multi-agent collaboration includes a task decomposition and allocation module, an agent communication protocol module, a knowledge sharing module, a multi-agent decision generation module, a conflict detection and negotiation module, a decision fusion and scoring module, and a human-machine hybrid decision-making module.

[0151] The task decomposition and allocation module takes as input: basic maintenance work order information W, and outputs: a set of subtasks T and an allocation result A. The specific operation is as follows:

[0152] Step 1: Parse work order W and extract key information set I = {equipment code, fault description, task type...};

[0153] Step 2: Initialize the subtask set T = {};

[0154] Step 3: Generate subtasks based on the preset task decomposition template M:

[0155] For each template m∈M, if m.condition(I)==True, T=T∪m.generatingsubtask(I);

[0156] Step 4: For each subtask t∈T:

[0157] Calculate the suitability score for each Agent: S(t, a) = f(t. knowledge requirement, a. professional field, a. historical performance), select the Agent with the highest score to assign tasks, and return the subtask set T and the assignment result A;

[0158] The working principle of this part is explained in detail below:

[0159] (1) Work order information parsing:

[0160] The system first receives a maintenance work order W, from which it extracts key information such as equipment code, fault description, task type, priority, etc., to form a structured information set I.

[0161] (2) Task template matching:

[0162] The system maintains a set of preset task decomposition templates M, each template corresponding to a typical maintenance task decomposition method.

[0163] For example: equipment maintenance templates, periodic testing templates, fault diagnosis templates, and emergency repair templates.

[0164] (3) Subtask generation:

[0165] The system matches the work order information I with the template conditions. When a certain template condition is met, the system generates the corresponding subtask. For example, when "pump valve failure" is identified: - Generate the "valve status check" subtask - Generate the "sealing test" subtask - Generate the "safety isolation" subtask.

[0166] (4) Suitability assessment and allocation:

[0167] For each generated subtask t, the system evaluates the suitability score S(t, a) of each agent. This score is based on: the matching degree between the professional knowledge required for the task and the agent's professional domain, the agent's historical performance in handling similar tasks, and the agent's current workload. Finally, the system assigns the subtask to the agent with the highest score for processing.

[0168] Taking a "cooling pump failure" repair work order as an example:

[0169] Enter work order information:

[0170] Equipment Code: RCP-P-001, Equipment Name: Main Cooling Pump, Fault Description: Excessive pump vibration, abnormal operating noise, Task Type: Fault Repair, Priority: High, Key Information Extracted:

[0171] I = {Equipment Type: "Cooling Pump", Fault Phenomenon: "Excessive Vibration", System: "Main Cooling System", Priority: "High"}

[0172] Template matching and subtask generation: The template "rotating equipment failure" is matched, and a set of subtasks T is generated.

[0173] T = {t1: "Vibration Analysis and Cause Diagnosis", t2: "Safety Risk Assessment", t3: "Maintenance Resource Planning", t4: "System Impact Analysis", t5: "Maintenance Procedure Design", t6: "Quality Inspection Plan"};

[0174] Task assignment:

[0175] The allocation result A after calculating the suitability scores of each Agent:

[0176] A = {t1 → Maintenance Engineer Agent (Score: 0.92), t2 → Safety Expert Agent (Score: 0.88), t3 → Planning and Scheduling Agent (Score: 0.95), t4 → Operations Expert Agent (Score: 0.91), t5 → Maintenance Engineer Agent (Score: 0.87), t6 → Quality Supervision Agent (Score: 0.93)}.

[0177] The Agent communication protocol module is based on a Directed Acyclic Graph (DAG) communication model. The communication protocol implementation method takes the following inputs: Agent set A, communication dependency graph G, and output: Agent state set S after communication is complete, including the following:

[0178] Step 1: Initialize all Agents to "Waiting" state;

[0179] Step 2: Find the set H of Agents in graph G that have no in-degree nodes;

[0180] Step 3: Set the status of all Agents in H to "Ready";

[0181] Step 4: When there are Agents in the "ready" state: Execute all Agent tasks in the "ready" state in parallel;

[0182] For each Agent a that completes the task: send a message to all Agents that depend on a, update the status of a to "complete", update the dependency graph G, recalculate the set of no-in-degree nodes H, and set the Agents in H with the status of "waiting" to "ready".

[0183] Step 5: When all Agents are in the "Completed" status, end the communication process.

[0184] Returns: Agent state set S.

[0185] The agent communication protocol is based on a directed acyclic graph (DAG) communication model, ensuring that information flows in the correct order in a multi-agent system and avoiding deadlocks and conflicts. This protocol manages dependencies and communication timing between agents, ensuring efficient collaboration.

[0186] (1) Explanation of Dependency Graph (DAG):

[0187] The system uses a directed acyclic graph G to represent the dependencies between agents, where: nodes represent each agent; directed edges A→B indicate that agent B depends on the output of agent A; and acyclicity ensures that deadlocks caused by circular dependencies will not occur.

[0188] (2) Agent state mechanism:

[0189] Each Agent has three possible states:

[0190] Waiting: Not all dependency inputs have been received; task cannot begin.

[0191] Ready: All necessary inputs have been received; task execution can begin.

[0192] Completed: Task completed and results output.

[0193] (3) Nodes with no in-degree:

[0194] A node with no in-degree refers to a node with no incoming edges. In a DAG, it represents an Agent that does not depend on the output of other Agents and can be started first.

[0195] (4) Parallel execution mechanism:

[0196] All agents in the "ready" state can execute tasks in parallel, improving system efficiency.

[0197] (5) Message passing and status updates:

[0198] Once the Agent completes its task, it sends its output to downstream Agents that depend on it, updates its own state to "Completed," updates the dependency graph, removes related edges, checks for new nodes with no in-degree, and if so, updates their state to "Ready." Initial dependency graph:

[0199] Maintenance Engineer Agent → Planning and Scheduling Agent, Maintenance Engineer Agent → Safety Expert Agent, Equipment Expert Agent → Maintenance Engineer Agent, Equipment Expert Agent → Operations Expert Agent, Operations Expert Agent → Planning and Scheduling Agent, Safety Expert Agent → Quality Supervision Agent, Planning and Scheduling Agent → Quality Supervision Agent;

[0200] Initialization state:

[0201] All Agents are initialized to "Waiting" state, and nodes with no in-degree are identified:

[0202] H = {Device Expert Agent}

[0203] First round of execution: The device expert agent status is updated to "ready", and the device expert agent executes the task;

[0204] After the task is completed: send the output to the Maintenance Engineer Agent and the Operation Expert Agent, update the status of the Equipment Expert Agent to "Completed", update the dependency graph, and remove the relevant edges of the Equipment Expert Agent;

[0205] Recalculate nodes with no in-degree:

[0206] The new no-in-degree node H = {Maintenance Engineer Agent, Operations Expert Agent} updates the status of these two agents to "Ready" and executes the next round in parallel:

[0207] The maintenance engineer agent and the operations expert agent perform tasks in parallel.

[0208] After both are completed, a new no-in-degree node is calculated: the security expert agent and the scheduling agent continue to execute until completion: the security expert agent and the scheduling agent execute the task;

[0209] Finally, the quality supervision agent executes the task, all agents are in the "complete" status, and the communication process ends.

[0210] The knowledge-sharing module uses a shared semantic space model to achieve knowledge sharing among agents. Inputs: Agent expertise K_i, decision context C; Output: shared knowledge space S. Specific implementation method:

[0211] Step 1: Initialize the shared knowledge space S = {};

[0212] Step 2: For each Agenti, knowledge projection: P_i = projection function (K_i, C), knowledge transformation: Ti = transformation function (P_i), S = S∪Ti;

[0213] Step 3: Perform consistency verification on the knowledge elements in S:

[0214] Detect conflicting knowledge C_k, and for each conflict C_k, apply the conflict resolution rule R;

[0215] Step 4: Optimize the knowledge space S and delete redundant knowledge;

[0216] Step 5: Establish indexing and access mechanisms for shared knowledge

[0217] Return to the shared knowledge space S.

[0218] Decision Generation and Conflict Resolution: The decision generation and conflict resolution mechanism of this invention includes three main processes: multi-agent decision generation, conflict detection and negotiation, and decision fusion and scoring.

[0219] (1) Knowledge projection process

[0220] Knowledge projection is the process of extracting relevant parts of the domain-specific knowledge K_i of each agent based on the current decision context C. The main functions of the projection function are: to filter knowledge elements relevant to the current task, to adjust the importance weights of knowledge according to the context, and to make implicit knowledge explicit for easy sharing.

[0221] For example, when context C is "main steam system valve maintenance", the maintenance engineer agent will extract valve maintenance-related knowledge from its knowledge base, and the safety expert agent will extract steam system safety operating procedures.

[0222] (2) Knowledge transformation process

[0223] Knowledge transformation is the process of converting projected professional knowledge P_i into a general representation in a unified ontology format, enabling knowledge from different agents to interact in a common semantic space.

[0224] The purpose of transformation functions is to convert technical terms into general semantic representations, unify the structured format of knowledge, and establish relationships between knowledge elements.

[0225] (3) Construction of shared knowledge space

[0226] The shared knowledge space S is the collection of knowledge contributed by each agent, forming a unified basis for decision-making. During its construction: each agent contributes knowledge in a specific order.

[0227] (4) Knowledge elements retain their source information.

[0228] Establishing relationships between knowledge points, and detecting and resolving knowledge conflicts: When knowledge contributed by different agents conflicts, the system will:

[0229] Identify conflicting knowledge pairs C_k = (knowledge 1, knowledge 2), analyze the conflict type (factual conflict, opinion conflict, priority conflict, etc.), and apply the corresponding conflict resolution rules R;

[0230] Professional weighting rule: Weighted according to the agent's expertise in a specific field;

[0231] Evidence support rule: Knowledge supported by more evidence takes precedence;

[0232] Conservative principle: More conservative solutions should be adopted for security-related conflicts;

[0233] Manual intervention: Conflicts that cannot be resolved automatically are marked as requiring manual intervention.

[0234] (5) Knowledge Space Optimization

[0235] The optimization process includes:

[0236] Remove redundant knowledge: merge knowledge elements that express the same content; delete irrelevant knowledge: remove knowledge that is not closely related to the current decision; knowledge compression: extract core knowledge points; create indexes: create efficient retrieval structures for knowledge elements.

[0237] Using "main pump shaft seal maintenance" as an example, this demonstrates how a knowledge-sharing mechanism works.

[0238] Decision Context C: Work Order Type: Main Pump Shaft Seal Malfunction Repair; Equipment Location: Primary Circuit System; Priority: High; Time Constraint: Complete within 72 hours;

[0239] Knowledge projection for each agent: Maintenance Engineer Agent: P_Maintenance = {Shaft seal fault diagnosis methods, disassembly and assembly procedures and techniques, sealing surface inspection standards, lubricant specifications}

[0240] Safety Expert Agent: P_Safety = {Primary loop radiation protection requirements, system isolation procedures, safety equipment usage guidelines, emergency response plan}

[0241] Planning and Scheduling Agent: P_Plan = {Historical work hours data, manpower availability, spare parts inventory, and project duration estimation model}

[0242] Knowledge transformation and sharing space construction, S = {shaft seal failure characteristics and cause analysis, maintenance technical route and steps, safety isolation and radiation protection measures, project schedule planning and resource allocation scheme, equipment technical parameters and performance requirements, system operation boundary conditions, quality inspection standards and methods}.

[0243] Conflict Detection and Resolution Examples

[0244] Conflict detected: C_1 = (Maintenance Engineer: "Type A seal should be used", Equipment Expert: "Type B seal should be used"). The conflict is resolved by applying professional weighting and evidence support rules. The equipment expert's recommendation is based on the latest manufacturer's technical bulletin and has a higher weighting. Solution: Use Type B seal, but record the reasons for the decision for manual review.

[0245] The optimized shared knowledge space, S_optimization = {Shaft seal failure cause: "primary seal ring wear", maintenance technical route: "disassembly-inspection-replacement-installation-testing", safety measures: "system pressure reduction and isolation, establishment of radiation control zone", maintenance materials: "Type B seal ring, special lubricant", resource allocation: "4 professional personnel, estimated construction period 48 hours", quality requirements: "leakage rate <0.1L / h, vibration <2mm / s"}.

[0246] The multi-Agent decision generation module takes as input: maintenance work order W, agent set A, and knowledge base K; and outputs: a decision scheme set D, including the following:

[0247] Step 1: Initialize the decision space D = {};

[0248] Step 2: Execute the decision-making process of each agent in parallel:

[0249] For each Agenta∈A, extract the relevant work order information I_a = extraction function (W, a. specialty), load the professional knowledge K_a = loading function (K, a. specialty), and make the decision reasoning: d_a = reasoning function (I_a, K_a), D = D∪d_a;

[0250] Step 3: Initial integration of the decision space D

[0251] Merge similar decisions, remove obviously wrong decisions, and return the decision set D.

[0252] 1. Implementation ideas of multi-agent decision generation algorithm

[0253] Information Extraction and Specialization: Each Agent extracts relevant information from the work order based on its own area of ​​expertise, which ensures the professional relevance of the decision-making process: the maintenance engineer Agent focuses on fault characteristics, the maintenance difficulty, the safety expert Agent focuses on risk factors and safety hazards; the planning and scheduling Agent focuses on time constraints and resource requirements; the information extraction function automatically filters, sorts and enhances the work order information according to the Agent's expertise, enabling the Agent to focus on relevant content.

[0254] 2. Knowledge loading mechanism:

[0255] The knowledge loading function selectively loads relevant professional knowledge from the knowledge base K for each Agent, including: domain rule knowledge (such as safety operating procedures and maintenance process standards); experience case knowledge (such as historical similar maintenance records and results); and equipment-specific knowledge (such as technical specifications and performance parameters). The knowledge loading process dynamically adjusts the scope and level of detail of the knowledge according to the characteristics of the work order to ensure the relevance of the decision.

[0256] 3. Decision-making reasoning process:

[0257] The decision-making reasoning performed by each agent is the core of multi-agent decision generation. The reasoning process typically includes: situation analysis: assessing the current state and identifying key issues; solution generation: proposing feasible solutions based on rules and experience; solution evaluation: assessing the feasibility and security of the solutions; decision formation: determining the recommended solution and the reasons; different agents may use different reasoning methods.

[0258] For example: rule-based reasoning is suitable for domains with clear standards; case-based reasoning is suitable for domains with rich experience; model-based reasoning is suitable for domains with clear physical models; hybrid reasoning combines multiple methods to improve decision quality.

[0259] 4. Decision-making space integration:

[0260] After all agents have completed their respective decisions, the system integrates the initial decision space D: merging similar decisions: identifying and merging decisions with overlapping content; removing obviously wrong decisions: filtering out potentially wrong decisions based on rules or statistics; standardizing decision format: converting the decisions of different agents into a unified format; and establishing relationships between decisions: identifying dependencies or support relationships between decisions.

[0261] Taking the "Main Steam Isolation Valve Failure" repair work order as an example:

[0262] Work order information W, equipment code: MSIV-001, equipment name: main steam isolation valve, fault description: valve does not close tightly, internal leakage exists; task type: planned maintenance; priority: medium; planned time: during the next major overhaul.

[0263] Decision-making process for each agent:

[0264] Maintenance Engineer Agent: Extracts information: {Equipment type: "Isolation valve", Fault: "Internal leakage", Location: "Main steam system"}; Loads knowledge: Valve maintenance procedures, sealing surface repair process, historical case data; Decision reasoning: Analyzes possible causes of internal leakage and evaluates maintenance solutions; Generates decision d1: {Fault diagnosis: "Sealing surface wear or foreign object obstruction", Maintenance solution: "Disassemble and inspect the sealing surface, grind or replace the sealing ring", Tool list: "Special grinding tools, sealing ring spare parts, torque wrench...", Estimated working hours: "24-36 hours"}

[0265] ② Security Expert Agent

[0266] Extracted Information: {System: "Main Steam", Energy Type: "High Temperature and High Pressure"}; Loaded Knowledge: Steam System Safety Regulations, Safety Measures for Thermomechanical Equipment; Decision Reasoning: Analyze safety risks and formulate protective measures; Generate Decision d2: {Risk Assessment: "High Temperature Burn Risk: High, System Overpressure Risk: Medium", Safety Measures: "System Cooling and Pressure Reduction, Double Isolation and Tagging, Safety Monitoring", Protective Equipment: "High Temperature Resistant Gloves, Protective Face Shield, Full-Body Protective Suit", Emergency Plan: "Emergency Isolation Plan, Burn Emergency Treatment Procedure"}

[0267] ③Planning and scheduling Agent

[0268] Extracted information: {Priority: "Medium", Plan: "During Overhaul"}, Loaded knowledge: Overhaul scheduling specifications, Human resource data, Parallel work constraints, Decision reasoning: Optimize time arrangement, Resource allocation to generate decision d3: {Scheduling suggestion: "Execute on the 3rd day of the overhaul, lasting 2 days", Personnel requirements: "2 mechanics, 1 welder, 1 QC, 1 safety supervisor", Material preparation: "2 sets of sealing rings, 1 set of fasteners, 500g of lubricating grease", Related tasks: "Parallel with low-pressure heater maintenance, staggered from main pump maintenance"}.

[0269] ④ Decision-making space integration

[0270] The system integrates and processes all agent decisions: Identifying similarities: similarities in sealing surface treatment solutions between maintenance engineers and equipment experts; conflicts: minor conflicts in the scheduling of quality supervision and planning; removing inapplicable items: a boundary condition proposed by an operations expert that is out of range. The integrated decision set D' comprises: {Technical solutions: {Maintenance route, specific steps, technical requirements...}, Safety measures: {Risk control, protection requirements, emergency plans...}, Resource plans: {Personnel allocation, material preparation, tool list...}, Time arrangements: {Start time, duration, milestones...}, Quality control: {Inspection points, acceptance criteria, record requirements...}, Operational impact: {System status requirements, boundary conditions, recovery plans...}}.

[0271] The conflict detection and negotiation module takes as input: a set of decision schemes D and a set of negotiation rules R; and outputs: the negotiated decision scheme D', which includes the following:

[0272] Step 1: Construct the decision conflict diagram G:

[0273] The node represents the decision d_i of each Agent, the edge e(i,j) indicates that there is a conflict between decisions d_i and d_j, and the edge weight w(i,j) indicates the severity of the conflict;

[0274] Step 2: Conflict Clustering:

[0275] Using graph clustering algorithms, identify conflicting clusters C = {c_1, c_2, ..., c_n};

[0276] Step 3: For each conflicting cluster c_i:

[0277] Applying the negotiation algorithm:

[0278] Initialize the negotiation round t=0. When there is a conflict and t<maximum round T: each agent proposes an adjustment plan, evaluates the comprehensive impact of the adjustment plan, selects the optimal adjustment plan, updates the decisions of each agent, and re-detects the conflict, t=t+1.

[0279] Step 4: For conflicts that cannot be resolved through negotiation:

[0280] Mark as a human intervention point and return: the negotiated decision-making solution D'.

[0281] Conflict Detection and Negotiation Algorithm Ideas

[0282] (1) Construction of decision conflict diagram:

[0283] The system first constructs a decision conflict graph G to formally represent the conflict relationships between the decisions of each agent: Nodes: Each node d_i represents an agent's decision scheme; Edges: Edge e(i,j) indicates that there is a conflict between decisions d_i and d_j; Edge weights: Weights w(i,j) quantify the severity of the conflict, usually calculated based on the following factors: the number of parameters involved in the conflict, the degree of impact of the conflict on security, the complexity of conflict resolution, and the magnitude of the impact of the conflict on resources. Conflict detection is based on a series of predefined conflict patterns, such as:

[0284] Parameter value conflict: such as two agents giving different values ​​for the same parameter; constraint conflict: such as mismatch between schedule and resource availability; priority conflict: such as different suggestions on the order of work; methodology conflict: such as different choices of maintenance methods.

[0285] (2) Conflict clustering:

[0286] The system applies a graph clustering algorithm to the conflict graph, organizing related conflicts into conflict clusters: each conflict cluster c_i contains a set of related decision nodes. The clustering process considers the similarity and correlation of the conflicts. The purpose of clustering is to group closely related conflicts for processing, thereby improving negotiation efficiency.

[0287] Commonly used clustering methods include: connected component-based clustering, community detection algorithms, and spectral clustering algorithms.

[0288] (3) Negotiation Algorithm:

[0289] For each conflict cluster, the system initiates an iterative negotiation process; Initialization: Set the negotiation round t=0 and prepare the negotiation environment; Iterative negotiation: When the conflict is unresolved and the maximum round T has not been reached: Adjustment plan generation: Each relevant agent proposes an adjustment plan based on the current state; Plan impact assessment: Evaluate the impact of each adjustment plan on the overall decision; Optimal plan selection: Select the optimal adjustment based on the evaluation results and negotiation rules; Decision update: Update the decisions of relevant agents; Conflict re-check: Check whether conflicts still exist; Round increment: t=t+1.

[0290] (4) Application of negotiation rules

[0291] The rules applied during the negotiation process include: the expert priority rule: conflicts in a specific field are decided by experts in that field; the compromise rule: seeking the middle value of parameters within the feasible range; the security conservatism rule: adopting a more conservative approach for security-related conflicts; and the majority rule: when multiple agents support a similar approach, it is adopted first.

[0292] (5) Marking of artificial intervention points:

[0293] For conflicts that cannot be resolved after the maximum number of rounds T, the system marks them as decision points requiring human intervention, records conflict details, viewpoints of all parties, and negotiation history, and provides a visual representation of the conflict to assist human decision-making.

[0294] Taking "Repair of steam generator feedwater isolation valve" as an example:

[0295] ① Initial Decision Conflict Identification: The system detected the following conflicts: Conflict 1: The maintenance engineer suggested "complete disassembly and repair," while the equipment expert suggested "online sealing surface repair"; Conflict 2: The safety expert requested "complete system isolation and pressure reduction," while the operations expert suggested "maintaining low pressure and performing online repair"; Conflict 3: The planned scheduling allocated a "36-hour timeframe," while the maintenance engineer estimated it would take "48 hours."

[0296] ② Construction of decision conflict diagram: Nodes: d1 (Maintenance Engineer), d2 (Safety Expert), d3 (Planning and Scheduling), d4 (Equipment Expert), d5 (Operation Expert); Edges: e(1,4) weight 0.8, e(2,5) weight 0.9, e(1,3) weight 0.6.

[0297] ③ Conflict clustering: Conflict cluster C1 = {d1, d4} - maintenance method conflict; Conflict cluster C2 = {d2, d5} - system state conflict; Conflict cluster C3 = {d1, d3} - project schedule estimation conflict.

[0298] ④ Example of the negotiation process (taking conflict cluster C1 as an example): Round t=0: The maintenance engineer proposes an adjustment: "Partial disassembly can be used, but all sealing surfaces must be inspected"; The equipment expert proposes an adjustment: "Partial disassembly is acceptable, and the undisassembled parts can be inspected using an endoscope"; Impact assessment: Both adjustment options reduce workload, but the endoscope inspection may reduce inspection quality; Option selection: Adopt a compromise solution of "partial disassembly + endoscope inspection + physical inspection of key parts"; Update decision: Update the decisions of the maintenance engineer and the equipment expert; Conflict detection: The conflict has been significantly reduced but still exists (weight reduced to 0.3). Round t=1: The maintenance engineer proposes an adjustment: "Accept the endoscope inspection, but increase the number of inspection points"; The equipment expert proposes an adjustment: "Accept the increase in inspection points, but optimize the inspection sequence"; Impact assessment: Both options are acceptable and do not increase the total workload; Option selection: Adopt the solution of "optimize the inspection sequence + increase the number of inspection points"; Update decision: Update the decisions of the maintenance engineer and the equipment expert; Conflict detection: The conflict has been resolved.

[0299] ⑤ Negotiation results:

[0300] Conflict clusters C1 and C3 were successfully resolved through negotiation: C1 solution: "Partial disassembly + enhanced endoscopic inspection + physical inspection of key parts"; C3 solution: "The project duration is adjusted to 42 hours, and one additional technician is added"; Conflict cluster C2 was not fully resolved after the maximum number of rounds of negotiation: marked as a point for manual intervention: "The system is in isolation and management needs to make a decision"; negotiation history is recorded: the views and reasons of security experts and operations experts in each round provide decision-making suggestions: possible compromise solutions and their impact are listed.

[0301] ⑥ Decision-making plan D' after consultation:

[0302] Repair method: Partial disassembly + enhanced endoscopic examination + physical inspection of key components; Work schedule: 42 hours, with one additional technician; Manual intervention points: System isolation status (providing alternative solutions for decision-making) and other coordinated decision items...

[0303] The decision fusion and scoring module takes as input: the negotiated decision scheme D' and scoring rules E; and outputs: the final decision scheme F, which includes the following:

[0304] Step 1: Initialize the fusion decision scheme F = {};

[0305] Step 2: For each dimension m in decision option D':

[0306] Collect the decisions {d_1, m, d_2, m, ..., d_n, m} of each agent regarding dimension m, calculate the decision credibility: C(d_i, m) = f(Agent expertise, strength of decision basis, historical accuracy), and perform weighted fusion.

[0307] F_m=∑(C(d_i,m)×d_i,m) / ∑C(d_i,m), F=F∪F_m;

[0308] Step 3: Evaluate the overall fusion decision F:

[0309] Safety score S_s = safety score function (F), feasibility score S_f = feasibility score function (F), economic score S_e = economic score function (F), comprehensive score S = w_s × S_s + w_f × S_f + w_e × S_e;

[0310] Step 4: If S < threshold T:

[0311] Mark as requiring manual review, and return the final decision option F and its score.

[0312] Decision fusion and scoring methods

[0313] (1) Dimensional decomposition:

[0314] The system first decomposes the decision scheme D' into different dimensions, each representing a key aspect of the decision: maintenance method dimension: specific maintenance technical routes and operating procedures; safety measures dimension: risk control and protection measures; resource allocation dimension: personnel, equipment, and material allocation schemes; time arrangement dimension: project schedule planning and key node times; quality standard dimension: inspection methods and acceptance standards. The division of dimensions ensures the comprehensiveness of the decision and facilitates the integration of the advantages of each agent in different professional fields.

[0315] (2) Calculation of decision credibility:

[0316] For each Agent's decision in each dimension, the system calculates its credibility score C(d_i, m). Factors affecting credibility include: Agent professionalism: the Agent's professional background and authority in a specific dimension; for example, security experts have high professionalism in the security measures dimension; maintenance engineers have high professionalism in the maintenance methods dimension.

[0317] (3) Strength of decision-making basis:

[0318] The sufficiency of supporting evidence and reasons for agent decisions is as follows: strong decision-making basis based on clear procedures, strong decision-making basis based on sufficient data analysis, moderate decision-making basis based on similar historical cases, and weak decision-making basis based on experience judgment.

[0319] (4) Historical accuracy:

[0320] The historical performance of the Agent in similar decisions. The system records the accuracy of each Agent's historical decisions and adjusts the current credibility based on the historical accuracy rate. The credibility calculation function usually adopts a weighted combination form: C(d_i, m) = w1 × professionalism(i, m) + w2 × basis strength(d_i, m) + w3 × historical accuracy rate(i, m).

[0321] Weighted fusion mechanism: The system performs weighted fusion on the decisions of each Agent under each dimension, and the weight is the corresponding credibility score: F_m = ∑(C(d_i, m) × d_i, m) / ∑C(d_i, m). The fusion process varies according to the decision type: Numerical decision: Direct weighted average; such as project duration estimation, resource quantity, etc.;

[0322] Categorical decision: Weighted voting or highest credibility selection; such as maintenance method selection, process route selection, etc.;

[0323] Sequential decision: Weighted fusion after sequence alignment; such as operation step sequence, checklist of inspection points, etc.

[0324] Comprehensive decision scoring: For the fused decision plan F, the system scores it from multiple dimensions:

[0325] (1) Safety score S_s: Evaluate the degree of safety guarantee of the decision

[0326] Integrity of risk identification, effectiveness of protection measures, reliability of emergency plans, sufficiency of personnel protection;

[0327] (2) Feasibility score S_f: Evaluate the actual executability of the decision

[0328] Maturity of the technical route, possibility of resource acquisition, satisfaction of time constraints, accessibility of operating conditions.

[0329] (3) Economic score S_e: Evaluate the resource utilization efficiency of the decision

[0330] Utilization efficiency of human resources, rationality of material use, impact of downtime, overall cost control. The final comprehensive score is the weighted sum of the scores of each dimension S = w_s × S_s + w_f × S_f + w_e × S_e. The weight setting usually reflects the principle of safety first: w_s > w_f > w_e. Decision verification and marking: The system compares the comprehensive score S with the preset threshold T: If S ≥ T, the decision plan passes the verification; if S < T, it is marked as a decision plan that requires manual review. Regardless of the score, the system records the score details for decision reference.

[0331] Take "Reactor Coolant Pump Shaft Seal Repair" as an example:

[0332] Dimension decomposition and credibility calculation:

[0333] (1) Repair method dimension:

[0334] Maintenance engineer decision d1, maintenance: "Completely disassemble and replace all shaft seal assemblies", confidence level C(d1, maintenance) = 0.88; equipment expert decision d2, maintenance: "Disassemble and inspect, then decide on replacement parts based on the degree of wear", confidence level C(d2, maintenance) = 0.82; quality supervision decision d3, maintenance: "Disassemble and replace primary and secondary seals, inspect tertiary seals", confidence level C(d3, maintenance) = 0.65.

[0335] (2) Safety measures dimension:

[0336] Safety expert decision d1, safe: "Comprehensive radiation protection and hydrogen monitoring", credibility C(d1, safe) = 0.92

[0337] The operation expert made the decision d2, which is safe: "System decontamination and depressurization", with a credibility C(d2, safety) = 0.78. The maintenance engineer made the decision d3, which is safe: "Local isolation and identification", with a credibility C(d3, safety) = 0.70.

[0338] (3) Resource allocation dimension:

[0339] Planning and scheduling decision d1, resources: "4 maintenance team members + 1 radiation protection person", confidence level C(d1, resources) = 0.85; Maintenance engineer decision d2, resources: "5 professional technicians", confidence level C(d2, resources) = 0.79

[0340] Safety expert decision d3, resources: "2 safety supervisors", credibility C(d3, resources) = 0.68.

[0341] 2. Weighted fusion calculation: Repair method fusion: F_repair = (0.88×d1, repair + 0.82×d2, repair + 0.65×d3, repair) ÷ (0.88 + 0.82 + 0.65); Specific fusion result: "Disassemble and replace the first and second level seals, check the third level seal, and decide whether to replace it based on the wear condition".

[0342] Safety measures integration: F_safety = (0.92×d1, safety + 0.78×d2, safety + 0.70×d3, safety) ÷ (0.92 + 0.78 + 0.70), specific integration results: "System decontamination and depressurization, comprehensive radiation protection, continuous hydrogen monitoring, and regional isolation markings".

[0343] Resource allocation integration, F_resource = (0.85×d1, resource + 0.79×d2, resource + 0.68×d3, resource) ÷ (0.85 + 0.79 + 0.68).

[0344] The specific integration results are: "4 maintenance technicians, 1 radiation protection technician, and 1 safety supervisor."

[0345] Overall decision score:

[0346] Safety score: assesses the safety protection integrity of the maintenance plan, the reliability of the emergency plan, and the comprehensiveness of risk control. Calculation score: S_s = 0.90. Feasibility score: assesses the feasibility of the technical solution, the rationality of resource allocation, and the availability of the time schedule.

[0347] Calculated score: S_f = 0.86, economic efficiency score: assesses the efficiency of human resource utilization, assesses the rationality of material use, and assesses the impact of downtime.

[0348] Score calculated: S_e = 0.82.

[0349] Overall Score: Weighting: Safety (w_s = 0.5), Feasibility (w_f = 0.3), Economy (w_e = 0.2) (This weighting is based on nuclear power industry safety regulations and risk assessment standards, reflecting the "safety first" principle in decision-making). Final Score: S = 0.5 x 0.90 + 0.3 x 0.86 + 0.2 x 0.82 = 0.87.

[0350] Decision Verification: Scoring threshold T = 0.75, comprehensive score S = 0.87 > T, the decision scheme passes verification, and the final decision scheme F: Repair method: Disassemble and replace the primary and secondary seals, inspect the tertiary seal, and decide whether to replace it based on the wear condition; Safety measures: System decontamination and depressurization, comprehensive radiation protection, continuous hydrogen monitoring, and area isolation markings; Resource allocation: 4 maintenance technicians, 1 radiation protection technician, and 1 safety supervisor; Time arrangement: Estimated duration 48 hours, critical path includes system isolation, disassembly, replacement, and testing; Quality standards: Static leakage rate <0.1L / h, dynamic operating parameters meet technical specifications; Comprehensive score: 0.87 (Safety: 0.90, Feasibility: 0.86, Economy: 0.82).

[0351] The human-machine hybrid decision-making framework takes as input: system decision F, human expert E, and interaction interface I, and outputs: final execution plan P, including the following:

[0352] Step 1: The system generates an initial decision plan F, which includes: maintenance plan suggestions, safety measure suggestions, resource allocation suggestions, and risk analysis report.

[0353] Step 2: Based on the decision risk level R:

[0354] If R < low-risk threshold T_l: the system decision is automatically approved; otherwise, if R < medium-risk threshold T_m: the decision is pushed to the supervisory experts for quick review; otherwise, the decision is pushed to the expert team for detailed review.

[0355] Step 3: Experts review the decision through interactive interface I:

[0356] Visualize decision-making options, provide the basis for decision-making and reasoning process, and allow experts to modify specific decision items.

[0357] Step 4: The system records expert modifications:

[0358] Capture modified items δ = {δ_1, δ_2, ..., δ_n}, perform modification pattern analysis and determine the cause, and update relevant Agent knowledge and rules.

[0359] Step 5: Generate the final execution plan P:

[0360] Integrate system decisions F and expert modifications δ, automatically generate execution documents, distribute them to relevant execution systems, and return the final execution plan P.

[0361] Implementation ideas of human-machine hybrid decision-making framework

[0362] (1) Initial system decision generation:

[0363] The system first generates an initial decision plan F based on a multi-agent collaborative decision-making process, which includes the following key components: maintenance plan suggestions: specific maintenance steps, methods and technical routes, fault diagnosis results, maintenance technical routes, operation step sequences, and technical parameter requirements.

[0364] (2) Safety measures recommendations: Protective measures to ensure safety during the maintenance process

[0365] Risk identification list, safety protection requirements, isolation plan, and emergency response plan.

[0366] (3) Resource allocation recommendations: personnel, equipment, and material allocation plans

[0367] Staffing plan, professional skill requirements, tool and equipment list, spare parts and material requirements.

[0368] (4) Risk Analysis Report: Potential Risk Points and Their Impact Assessment

[0369] Technical risk analysis, safety risk points, schedule risk factors, and quality control challenges.

[0370] (5) Decision-making risk assessment and classification:

[0371] The system conducts a risk assessment on decision-making plan F, generates a risk level R, and adopts different processing strategies according to the risk level: Low-risk decision (R < T_l): Characteristics: Standard process, sufficient experience, limited impact. Processing: The system decision automatically passes without manual intervention.

[0372] Examples: Routine maintenance, replacement of standard components, etc.

[0373] Medium-risk decision (T_l ≤ R < T_m): Characteristics: Some non-routine operations, limited uncertainty, certain system impact; Processing: Push it to relevant supervision experts for rapid review; Key points for review: Key decision points, parameter settings, risk control measures; Examples: Partial maintenance of important equipment, system parameter adjustment, etc.; High-risk decision (R ≥ T_m): Characteristics: Non-routine maintenance, high uncertainty, significant system impact; Processing: Push it to the expert team for detailed review and evaluation; Review method: Interactive in-depth review meeting; Examples: Major overhaul of important safety equipment, implementation of innovative maintenance plans, etc.; Expert interaction review mechanism: The system provides a decision review environment for experts through the interaction interface I to ensure that experts can effectively understand and adjust; System decision: Visual display of decisions: Decision tree view: Displays decision logic and branches; Parameter correlation diagram: Displays the dependency relationships between parameters; Risk hot spot map: Intuitively shows the risk distribution; Resource allocation diagram: Displays the allocation of personnel, equipment, and time; Transparency of the reasoning process: Knowledge application chain: Displays the knowledge path used in the decision; Sources of decision basis: Cited specifications, standards, and cases; Uncertainty marking: Identifies the uncertain parts in the decision.

[0374] Agent professional opinions: Displays the original opinions of each Agent; Interactive modification mechanism: Parameter adjustment: Modify specific decision parameters; Option change: Select an alternative technical route; Constraint addition: Add additional limiting conditions; Resource adjustment: Modify the resource allocation plan; Expert modification records and analysis: The system records all modifications made by experts to the decision-making plan and conducts in-depth analysis for continuous learning: Modification item capture: Records each specific modification δ i; Parameter values ​​before and after modification; Decision branches after modification; Explanation of reasons for modification; Modification pattern analysis: Identifying recurring modification patterns; Common adjustments for specific types of decisions; Modification preferences of specific experts; Modification characteristics in different scenarios; Classification of modification reasons: Knowledge gaps: The system lacks knowledge in specific domains; Overly strict rules: The system applies rules that are too strict; Experience-based judgment: Based on expert experience that is difficult to formalize; External constraints: Additional constraints not considered by the system; Knowledge updates and system learning: Based on the analysis of expert modifications, the system updates the relevant knowledge base and decision-making mechanism: Agent knowledge update: Update the content of the professional knowledge base; Adjust the weight of knowledge items; Add missing knowledge points; Decision rule optimization: Correct inaccurate decision rules. Then; adjust the threshold for rule application; add new decision paths; expand the case library: store the current decision as a new case; associate expert modifications and final effects; establish case retrieval features; generate the final execution plan: the system integrates the initial decision F and expert modifications δ to generate the final execution plan P; plan integration: apply all expert modifications to the original plan; re-verify the consistency of the modified plan; update related decision branches; generate execution documents: detailed work instructions; technical specifications; quality inspection standards; safe operating procedures; task decomposition and allocation: decompose the plan into specific work packages; clarify the responsible persons for each work package; define deliverables and acceptance standards; execution tracking mechanism: deviation handling mechanism for execution status monitoring points; feedback collection channels.

[0375] Taking "reactor coolant pump maintenance decision-making" as an example:

[0376] ① Initial system decision generation:

[0377] The system generates an initial decision plan F: Repair plan: Disassemble the pump shaft seal assembly; inspect the wear of the sealing surface; replace the primary and secondary sealing rings; clean the sealing water system; reassemble and test; Safety measures: system isolation and drainage; radiation protection zone delineation; dose monitoring and control; implementation of pollution prevention measures; Resource allocation: 4 maintenance technicians; 2 radiation protection personnel; 1 quality inspection personnel; Risk analysis of special tools and equipment list: bearing damage risk (medium); radiation exposure risk (high); project delay risk (medium); start-up leakage risk (medium); Decision risk assessment and classification: The system assessment yields a risk level R = 0.78, exceeding the medium risk threshold T_m = 0.7, classifying it as a high-risk decision. Reasons: involves important safety equipment; there are high-risk points for radiation protection; includes unconventional maintenance steps; The system decides to push the plan to the expert team for detailed review.

[0378] ②Expert Interactive Review Process

[0379] Experts reviewed the decision-making plan through an interactive interface: viewing the decision tree: focusing on the decision branch of shaft seal inspection and replacement; analyzing risk hotspots: focusing on radiation protection measures and startup leakage risks; querying the reasoning basis: checking the shaft seal replacement standards and basis; experts found the following problems: the shaft seal inspection standards were too conservative; radiation protection measures need to be strengthened; the sealing water system treatment plan was not detailed enough; the project period estimate was too short.

[0380] ③ Expert revision record:

[0381] The experts made the following modifications: δ1: Adjusted the shaft seal inspection standard, replacing only seal rings worn more than 0.3mm; δ2: Added local shielding measures and personal dosing alarms; δ3: Added detailed cleaning steps for the sealing water system; δ4: Adjusted the project duration from 3 days to 4 days, increasing testing time; δ5: Added an experienced technician to participate in critical operations.

[0382] ④ Modification analysis and knowledge update:

[0383] System analysis experts modified the mode to address the following issues: knowledge gap: inaccurate application of shaft seal wear standards; overly strict rules: overly conservative standard replacement; experience-based judgment: increased testing time based on experience-based judgment; resource optimization: adding experienced technicians for key operations.

[0384] ⑤ Corresponding knowledge updates:

[0385] Update shaft seal wear assessment standards; add detailed knowledge on radiation protection; adjust the time estimation model for similar work; create a new "sealing water system cleaning" case.

[0386] ⑥ Final execution plan generation:

[0387] Repair plan: Disassemble the pump shaft seal assembly; inspect the wear of the sealing surface (replace if >0.3mm); selectively replace severely worn sealing rings; perform detailed cleaning of the sealing water system (including filters and heat exchangers); reassemble and extend the test (72 hours).

[0388] Safety measures: system isolation and drainage; enhanced radiation protection (increased local shielding); personal dose alarm and real-time monitoring; pollution prevention measures and waste sorting and disposal.

[0389] Resource allocation: 4 maintenance technicians (1 of whom must have more than 5 years of experience); 2 radiation protection personnel; 1 quality inspection personnel; updated list of special tools and equipment.

[0390] Execution plan: Total duration 4 days, daily work objectives and acceptance criteria, key milestones and decision points, and contingency backup plan.

[0391] The key point of this invention is:

[0392] Multi-professional agent collaborative network construction method: Using six core professional agents (maintenance engineer agent, safety expert agent, planning and scheduling agent, equipment expert agent, operation expert agent, and quality supervision agent) as nodes, the method upgrades single-professional decision-making to multi-professional parallel collaboration through a "task decomposition-suitability assessment-dynamic allocation" algorithm based on work order information; it achieves intelligent sub-task matching, load balancing, and precise knowledge adaptation, significantly improving collaborative efficiency and professional coverage.

[0393] The DAG-based agent communication protocol uses a directed acyclic graph to manage agent dependencies, sets a "wait-ready-complete" three-state mechanism and a parallel execution strategy to ensure the correct order of information flow, avoid deadlock, and can trigger parallel computation in batches according to nodes with no in-degree, thus compressing the overall decision-making link latency.

[0394] Shared Semantic Space Knowledge Fusion Mechanism: A four-step method of "knowledge projection - unified transformation - conflict governance - index optimization" is proposed to map the proprietary knowledge of each agent to a unified semantic space; knowledge conflicts are automatically resolved through professional weights, evidence support and conservative security principles to ensure the consistency, traceability and high callability of the decision-making basis.

[0395] Multi-Agent Decision Generation and Conflict Negotiation Algorithm: Constructs a process of "preliminary reasoning - conflict graph clustering - iterative negotiation - manual intervention labeling". Graph clustering is used to locate highly coupled conflict clusters, and automatic negotiation is carried out according to rules such as professional priority, compromise solution, safety conservatism, and majority decision-making. Visual intervention points are generated for intractable conflicts to ensure the reliability and safety redundancy of decision-making in high-risk scenarios.

[0396] Credibility-weighted decision fusion and multi-dimensional scoring model: The credibility of each agent solution is calculated and weighted according to the three dimensions of security, feasibility and economy. Then, the solution that needs to be manually reviewed is automatically screened by threshold comparison, so as to achieve comprehensive optimization of "security first and economic coordination".

[0397] A human-machine hybrid closed-loop learning framework: Comprehensive system solutions are categorized by risk level and pushed to experts; expert modifications (δ) are recorded and used to update the knowledge base, rules, and case database. Continuous learning and iteration improve the agent's inference accuracy and rule adaptability, forming a self-evolving closed loop of "system-expert-system." Improved efficiency of the entire workflow: In traditional manual preparation scenarios exceeding 2 hours, the above collaborative mechanism compresses the diagnosis, assessment, planning, and step generation processes to within 15 minutes, while maintaining or improving compliance with nuclear safety regulations.

Claims

1. A nuclear power plant maintenance decision-making system based on multi-agent collaboration, characterized in that: The system comprises an input layer, a multi-agent collaboration layer, a knowledge support layer, a decision processing layer, an output and interaction layer, and a feedback learning layer. The input layer receives initial work order information, including basic equipment information and fault descriptions. The multi-agent collaboration layer assigns subtasks to six predefined professional agents based on the work order information using task decomposition and suitability algorithms, and completes information interaction using the DAG communication protocol. The knowledge support layer includes a nuclear power professional knowledge base, a historical maintenance database, and a regulations database. The decision processing layer performs conflict detection, negotiation, and decision fusion on the agent outputs. The output and interaction layer provides decision result display and a human-computer interaction interface. The feedback learning layer updates knowledge and rules based on execution results and expert modifications to achieve a comprehensive improvement in the efficiency and safety of nuclear power plant maintenance decisions.

2. The nuclear power plant maintenance decision-making system based on multi-agent collaboration as described in claim 1, characterized in that: It includes a task decomposition and allocation module, an agent communication protocol module, a knowledge sharing module, a multi-agent decision generation module, a conflict detection and negotiation module, a decision fusion and scoring module, and a human-machine hybrid decision-making module; The task decomposition and allocation module takes as input: basic maintenance work order information W, and outputs: a set of subtasks T and an allocation result A. The Agent communication protocol module is based on a directed acyclic graph to construct a task dependency graph G, and uses asynchronous message passing to realize state synchronization and information interaction between Agents. Input: A set of agents, G communication dependency graph; Output: S set of agent states after communication is completed. The knowledge sharing module uses a shared semantic space model to realize knowledge sharing among agents. The inputs are: agent expertise K_i and decision context C. Output: Shared knowledge space S; The multi-Agent decision generation module takes as input: maintenance work order W, agent set A, and knowledge base K; Output: Set of decision options D; The conflict detection and negotiation module takes as input: decision scheme set D, negotiation rule set R; Output: The negotiated decision plan D'; The decision fusion and scoring module takes as input: the negotiated decision scheme D' and the scoring rule E; Output: Final decision plan F; The human-machine hybrid decision-making framework takes the following inputs: system decision F, human expert E, and interaction interface I, and outputs: final execution plan P.

3. The nuclear power plant maintenance decision-making system based on multi-agent collaboration as described in claim 2, characterized in that: The task decomposition and allocation module processes tasks as follows: Step 1: Parse work order W and extract key information set I = {equipment code, fault description, task type...}; Step 2: Initialize the subtask set T = {}; Step 3: Generate subtasks based on the preset task decomposition template M; For each template m∈M, if m.condition(I)==True, T=T∪m.generatingsubtask(I); Step 4: For each subtask t∈T: Calculate the suitability score for each Agent: S(t, a) = f(t is the knowledge requirement, a is the professional field, and a is the historical performance). Select the Agent with the highest score to assign tasks and return the subtask set T and the assignment result A.

4. A nuclear power plant maintenance decision-making system based on multi-agent collaboration as described in claim 2, characterized in that, The Agent communication protocol module processes the following steps: Step 1: Initialize all Agents to "Waiting" state; Step 2: Find the set H of Agents in graph G that have no in-degree nodes; Step 3: Set the status of all Agents in H to "Ready"; Step 4: When there are Agents in the "ready" state: Execute all Agent tasks in the "ready" state in parallel; For each Agent a that completes the task: send a message to all Agents that depend on a, update the status of a to "complete", update the dependency graph G, recalculate the set of no-in-degree nodes H, and set the Agents in H with the status of "waiting" to "ready". Step 5: When all Agents are in the "Completed" state, the communication process ends; Return the Agent state set S.

5. A nuclear power plant maintenance decision-making system based on multi-agent collaboration as described in claim 2, characterized in that, The specific implementation method of the knowledge sharing module is as follows: Step 1: Initialize the shared knowledge space S = {}; Step 2: For each Agenti, knowledge projection: P_i = projection function (K_i, C), knowledge transformation: Ti = transformation function (P_i), S = S∪Ti; Step 3: Perform consistency verification on the knowledge elements in S: Detect conflicting knowledge C_k, and for each conflict C_k, apply the conflict resolution rule R; Step 4: Optimize the knowledge space S and delete redundant knowledge; Step 5: Establish indexing and access mechanisms for shared knowledge Return to the shared knowledge space S.

6. A nuclear power plant maintenance decision-making system based on multi-agent collaboration as described in claim 2, characterized in that, The multi-Agent decision generation module is implemented as follows: Step 1: Initialize the decision space D = {}; Step 2: Execute the decision-making process of each agent in parallel; For each Agenta∈A, extract the relevant work order information I_a = extraction function (W, a. specialty), load the professional knowledge K_a = loading function (K, a. specialty), and make the decision reasoning: d_a = reasoning function (I_a, K_a), D = D∪d_a; Step 3: Initial integration of the decision space D Merge similar decisions, remove obviously wrong decisions, and return the decision set D.

7. A nuclear power plant maintenance decision-making system based on multi-agent collaboration as described in claim 2, characterized in that, The conflict detection and negotiation module is implemented as follows: Step 1: Construct the decision conflict diagram G: The node represents the decision d_i of each Agent, the edge e(i,j) indicates that there is a conflict between decisions d_i and d_j, and the edge weight w(i,j) indicates the severity of the conflict; Step 2: Conflict Clustering: Using graph clustering algorithms, identify conflicting clusters C = {c_1, c_2, ..., c_n}; Step 3: For each conflicting cluster c_i: Initialize the negotiation round t = 0. When there is a conflict and t < maximum round T: each agent proposes an adjustment plan, evaluates the comprehensive impact of the adjustment plan, selects the optimal adjustment plan, updates the decisions of each agent, and re-detects the conflict, t = t + 1; Step 4: For conflicts that cannot be resolved through negotiation: Mark as a human intervention point and return: the negotiated decision-making solution D'.

8. A nuclear power plant maintenance decision-making system based on multi-agent collaboration as described in claim 2, characterized in that, The decision fusion and scoring module is implemented as follows: Step 1: Initialize the fusion decision scheme F = {}; Step 2: For each dimension m in decision option D': Collect the decisions {d_1, m, d_2, m, ..., d_n, m} of each agent regarding dimension m, calculate the decision credibility: C(d_i, m) = f(Agent expertise, strength of decision basis, historical accuracy), and perform weighted fusion. F_m=∑(C(d_i,m)×d_i,m) / ∑C(d_i,m), F=F∪F_m; Step 3: Evaluate the overall fusion decision F: Safety score S_s = safety score function (F), feasibility score S_f = feasibility score function (F), economic score S_e = economic score function (F), comprehensive score S = w_s × S_s + w_f × S_f + w_e × S_e; Step 4: If S < threshold T: Mark as requiring manual review, and return the final decision option F and its score.

9. A nuclear power plant maintenance decision-making system based on multi-agent collaboration as described in claim 2, characterized in that, The human-machine hybrid decision-making framework includes: Step 1: The system generates an initial decision plan F, which includes: maintenance plan suggestions, safety measure suggestions, resource allocation suggestions, and a risk analysis report; Step 2: Based on the decision risk level R: If R < low-risk threshold T_l: the system decision is automatically approved; otherwise, if R < medium-risk threshold T_m: the decision is pushed to the supervisory experts for quick review; otherwise, the decision is pushed to the expert team for detailed review. Step 3: Experts review the decision through interactive interface I: Visualize decision-making options, provide the basis for decision-making and reasoning process, and allow experts to modify specific decision items; Step 4: The system records expert modifications: Capture modified items δ = {δ_1, δ_2, ..., δ_n}, analyze modification patterns and reasons, and update relevant Agent knowledge and rules; Step 5: Generate the final execution plan P: Integrate system decisions F and expert modifications δ, automatically generate execution documents, distribute them to relevant execution systems, and return the final execution plan P.

10. A nuclear power plant maintenance decision-making method based on multi-agent collaboration, comprising the following steps: Task decomposition and allocation process: Step 1: Parse work order W and extract key information set I = {equipment code, fault description, task type...}; Step 2: Initialize the subtask set T = {}; Step 3: Generate subtasks based on the preset task decomposition template M; For each template m∈M, if m.condition(I)==True, T=T∪m.generatingsubtask(I); Step 4: For each subtask t∈T: Calculate the suitability score for each Agent: S(t, a) = f, where t is the knowledge requirement, a is the professional domain, and a is the historical performance. Select the Agent with the highest score to assign tasks, and return the subtask set T and the assignment result A. The Agent communication protocol and its processing steps are as follows: Step 1: Initialize all Agents to "Waiting" state; Step 2: Find the set H of Agents in graph G that have no in-degree nodes; Step 3: Set the status of all Agents in H to "Ready"; Step 4: When there are Agents in the "ready" state: Execute all Agent tasks in the "ready" state in parallel; For each Agent a that completes the task: send a message to all Agents that depend on a, update the status of a to "complete", update the dependency graph G, recalculate the set of no-in-degree nodes H, and set the Agents in H with the status of "waiting" to "ready". Step 5: When all Agents are in the "Completed" state, the communication process ends; Return the Agent state set S; Knowledge sharing can be implemented as follows: Step 1: Initialize the shared knowledge space S = {}; Step 2: For each Agenti, knowledge projection: P_i = projection function (K_i, C), knowledge transformation: Ti = transformation function (P_i), S = S∪Ti; Step 3: Perform consistency verification on the knowledge elements in S: Detect conflicting knowledge C_k, and for each conflict C_k, apply the conflict resolution rule R; Step 4: Optimize the knowledge space S and delete redundant knowledge; Step 5: Establish indexing and access mechanisms for shared knowledge Return to the shared knowledge space S; Multi-agent decision generation, the implementation process is as follows: Step 1: Initialize the decision space D = {}; Step 2: Execute the decision-making process of each agent in parallel; For each Agenta∈A, extract the relevant work order information I_a = extraction function (W, a. specialty), load the professional knowledge K_a = loading function (K, a. specialty), and make the decision reasoning: d_a = reasoning function (I_a, K_a), D = D∪d_a; Step 3: Initial integration of the decision space D Merge similar decisions, remove obviously wrong decisions, and return the decision set D; The conflict detection and negotiation process is as follows: Step 1: Construct the decision conflict diagram G: The node represents the decision d_i of each Agent, the edge e(i,j) indicates that there is a conflict between decisions d_i and d_j, and the edge weight w(i,j) indicates the severity of the conflict; Step 2: Conflict Clustering: Using graph clustering algorithms, identify conflicting clusters C = {c_1, c_2, ..., c_n}; Step 3: For each conflicting cluster c_i: Initialize the negotiation round t = 0. When there is a conflict and t < maximum round T: each agent proposes an adjustment plan, evaluates the comprehensive impact of the adjustment plan, selects the optimal adjustment plan, updates the decisions of each agent, and re-detects the conflict, t = t + 1; Step 4: For conflicts that cannot be resolved through negotiation: Mark the point as a human intervention point and return: the negotiated decision-making solution D'; The decision fusion and scoring process is as follows: Step 1: Initialize the fusion decision scheme F = {}; Step 2: For each dimension m in decision option D': Collect the decisions {d_1, m, d_2, m, ..., d_n, m} of each agent regarding dimension m, calculate the decision credibility: C(d_i, m) = f(Agent expertise, strength of decision basis, historical accuracy), and perform weighted fusion. F_m=∑(C(d_i,m)×d_i,m) / ∑C(d_i,m), F=F∪F_m; Step 3: Evaluate the overall fusion decision F: Safety score S_s = safety score function (F), feasibility score S_f = feasibility score function (F), economic score S_e = economic score function (F), comprehensive score S = w_s × S_s + w_f × S_f + w_e × S_e; Step 4: If S < threshold T: Mark as requiring manual review, and return the final decision option F and its score; The human-machine hybrid decision-making framework includes: Step 1: The system generates an initial decision plan F, which includes: maintenance plan suggestions, safety measure suggestions, resource allocation suggestions, and a risk analysis report; Step 2: Based on the decision risk level R: If R < low-risk threshold T_l: the system decision is automatically approved; otherwise, if R < medium-risk threshold T_m: the decision is pushed to the supervisory experts for quick review; otherwise, the decision is pushed to the expert team for detailed review. Step 3: Experts review the decision through interactive interface I: Visualize decision-making options, provide the basis for decision-making and reasoning process, and allow experts to modify specific decision items; Step 4: The system records expert modifications: Capture modified items δ = {δ_1, δ_2, ..., δ_n}, analyze modification patterns and reasons, and update relevant Agent knowledge and rules; Step 5: Generate the final execution plan P: Integrate system decisions F and expert modifications δ, automatically generate execution documents, distribute them to relevant execution systems, and return the final execution plan P.

Images