Multi-agent based tunnel water and mud inrush disposal scheme recommendation method and system

Abstract

The application relates to the technical field of tunnel water and mud inrush disposal scheme recommendation, in particular to a tunnel water and mud inrush disposal scheme recommendation method and system based on multiple agents, which comprises the following steps: constructing a tunnel water and mud inrush disaster knowledge base based on historical tunnel disaster knowledge data and historical tunnel disaster case data; constructing a tunnel water and mud inrush disaster knowledge graph based on the tunnel water and mud inrush disaster database; acquiring tunnel water and mud inrush disaster site data to be processed, and inputting the data to be processed into a multiple-agent collaborative server; controlling an agent to perform intention recognition on the input data, determining other agent categories and subsequent decision-making processes participating in the task according to the intention recognition result; when the intention recognition is to generate a final disposal scheme, controlling the agent to sequentially dispatch a data preprocessing agent, a retrieval agent, a calculation agent and an inference decision-making agent, and giving the final disposal scheme, so that tunnel water and mud inrush disposal candidate schemes are screened.

Description

Technical Field

[0001] This application relates to the technical field of recommending solutions for handling water and mud inrushes in tunnels, and in particular to a method and system for recommending solutions for handling water and mud inrushes in tunnels based on multi-agent systems. Background Technology

[0002] During tunnel construction, unfavorable geological structures such as water-filled karst caves, underground rivers, and fault fracture zones are frequently encountered. Especially in karst areas with complex groundwater networks, these water-rich structures pose serious hazards of water and mud inrushes during tunnel construction. These disasters severely impact tunnel construction safety, frequently leading to ecological damage, casualties, and significant economic losses, resulting in extremely negative social impacts.

[0003] Current decision-making regarding tunnel water and mud inrush response primarily relies on emergency plans and expert experience. However, water and mud inrushes are closely related to geological, engineering, and construction environments. Furthermore, water inrush disasters are sudden and urgent, making it difficult to adapt decisions based on emergency plans to specific situations. Decisions based on expert experience suffer from poor timeliness and are ill-suited for handling sudden major disasters. While large-scale models possess information understanding and reasoning capabilities, directly using a single large model for tunnel water and mud inrush response decisions still faces issues such as illusions. It struggles to achieve coordinated processing of candidate solution selection, computational verification, and inference optimization, leading to uncontrollable and unreliable results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a method and system for recommending tunnel water and mud inrush disposal solutions based on multi-agent technology. This application is a method for generating water and mud inrush disaster disposal solutions with multi-stage collaborative decision-making and controllable process, which can provide reliable decision support for engineering disaster prevention and control.

[0005] On the one hand, a method for recommending tunnel water and mud inrush disposal solutions based on multi-agent systems is provided, including: Based on historical tunnel disaster knowledge data and historical tunnel disaster case data, a knowledge base for tunnel water inrush and mud inrush disasters is constructed; based on the aforementioned tunnel water inrush and mud inrush disaster database, a knowledge graph for tunnel water inrush and mud inrush disasters is constructed. The system acquires data to be processed from the site of a water inrush and mud inrush disaster in a tunnel, and inputs the data into a multi-agent collaborative server. The multi-agent collaborative server includes: a control agent, a data preprocessing agent, a retrieval agent, a computation agent, and a reasoning and decision-making agent. The controlling agent performs intent recognition on the input data, and determines the categories of other agents participating in the task and the subsequent decision-making process based on the intent recognition results. When the intent recognition indicates that a final disposal plan needs to be generated, the controlling agent sequentially schedules the data preprocessing agent, retrieval agent, computation agent, and reasoning decision agent to provide the final disposal plan.

[0006] On the other hand, a multi-agent-based tunnel inrush water and mud inrush disposal solution recommendation system is provided, including: The knowledge base construction module is configured to: construct a knowledge base for tunnel water inrush and mud inrush disasters based on historical tunnel disaster knowledge data and historical tunnel disaster case data; and construct a knowledge graph for tunnel water inrush and mud inrush disasters based on the aforementioned tunnel water inrush and mud inrush disaster database. The data acquisition module is configured to: acquire data to be processed at the site of a water inrush and mudslide disaster in a tunnel, and input the data to be processed into a multi-agent collaborative server; wherein the multi-agent collaborative server includes: a control agent, a data preprocessing agent, a retrieval agent, a computation agent, and a reasoning and decision-making agent; The output module is configured to: control the agent to perform intent recognition on the input data, determine the categories of other agents participating in the task and the subsequent decision-making process based on the intent recognition results; when the intent recognition indicates that a final disposal plan needs to be generated, the control agent sequentially schedules the data preprocessing agent, retrieval agent, computation agent and reasoning decision agent to provide the final disposal plan.

[0007] The above technical solution has the following advantages or beneficial effects: The original direct generation method based on a single large language model is extended into a multi-stage decision-making process completed collaboratively by multiple functional intelligent agents, realizing the step-by-step processing of candidate solution generation, computational verification and reasoning decision.

[0008] By introducing a control agent to uniformly schedule multiple decision-making processes, task-driven automatic process control is achieved, improving the system's automation and scalability.

[0009] By introducing retrieval constraints and computational feedback mechanisms, the generation process is constrained in multiple stages, transforming the generation of disposal solutions from an uncontrollable language model output into a controllable and reliable generation based on constraint optimization. This enables the collaborative processing of candidate solutions for handling tunnel water and mud inrushes, computational verification, and inference optimization. Attached Figure Description

[0010] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0011] Figure 1This is a flowchart of the method in Example 1. Detailed Implementation

[0012] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0013] Example 1 This embodiment provides a recommended method for handling tunnel water and mud inrush solutions based on multi-agent systems; like Figure 1 As shown, the recommended method for handling tunnel water and mud inrush based on multi-agent systems includes: S101: Construct a knowledge base for tunnel water inrush and mud inrush disasters based on historical tunnel disaster knowledge data and historical tunnel disaster case data; construct a knowledge graph for tunnel water inrush and mud inrush disasters based on the aforementioned tunnel water inrush and mud inrush disaster database; S102: Acquire the data to be processed at the site of the tunnel water and mud inrush disaster, and input the data to be processed into the multi-agent collaborative server; wherein, the multi-agent collaborative server includes: control agent, data preprocessing agent, retrieval agent, computation agent and reasoning and decision-making agent; S103: The control agent performs intent recognition on the input data, determines the categories of other agents participating in the task and the subsequent decision-making process based on the intent recognition results; when the intent recognition indicates that a final disposal plan needs to be generated, the control agent sequentially schedules the data preprocessing agent, retrieval agent, computation agent and reasoning decision agent to provide the final disposal plan.

[0014] Furthermore, S101: Based on historical tunnel disaster knowledge data and historical tunnel disaster case data, a knowledge base for tunnel water inrush and mud inrush disasters is constructed, specifically including: The historical tunnel disaster knowledge data includes: textbooks, books, construction reports, and papers; The historical tunnel disaster case data includes: adverse geological data, stratum lithology data, groundwater level data, topographic data, rock stratum occurrence data, and surrounding rock grade data of the tunnel section where the disaster occurred; Among them, adverse geological data are detected by ground-penetrating radar and seismic wave reflection method to detect hidden faults and karst geological structures, and the distribution of fracture zones or weak interlayers is verified by core drilling. Stratigraphic lithology data rely on in-situ core sampling and geological logging at the working face to record rock types, weathering degree and structural characteristics and to determine rock strength; Groundwater level data is obtained by real-time monitoring of static water level using borehole water level gauges, and piezometers are installed at designated locations to measure pore water pressure. Topographic data, obtained through drone aerial surveys or satellite imagery, provides a basis for disaster analysis; Rock strata attitude data were obtained by directly measuring the strike, dip, and dip angle of the rock strata using a geological compass. The surrounding rock grade data is determined based on rock quality index (RQD), joint density, and groundwater parameters. The grading index is calculated using the RMR or BQ method, and the surrounding rock grade is verified through field rebound hammer tests. A disaster case database provides real-world case constraints for recommending subsequent disaster response plans.

[0015] The beneficial effects of the above technical solution are: by constructing a knowledge base for multi-source tunnel water and mud inrush disasters, a systematic organization of disaster-related information is achieved, covering multi-dimensional data from macro-geological environment to micro-lithological characteristics, providing a real data foundation for subsequent steps.

[0016] Further, step S101: Based on the tunnel water inrush and mud inrush disaster database, construct a knowledge graph of tunnel water inrush and mud inrush disasters, specifically including: Using the first large language model after training, entity-relationship-entity triplet data are extracted from the historical tunnel disaster knowledge data of the tunnel water inrush and mud inrush disaster knowledge base; To ensure consistency alignment between the entities in the triplet data and the entities in the historical tunnel disaster case data of the tunnel water inrush and mud inrush disaster knowledge base: calculate the similarity between the entities in the triplet data and the entities in the historical tunnel disaster case data of the tunnel water inrush and mud inrush disaster knowledge base. If the similarity value is greater than a set threshold, modify the entity name of the triplet data to the entity name of the historical tunnel disaster case data. Based on disaster entities, geological attribute entities, disposal plan entities, and the relationships between entities, a knowledge graph is constructed and then supplemented to obtain a knowledge graph of tunnel water inrush and mud inrush disasters.

[0017] Furthermore, the large language model is used to perform in-depth analysis and structured processing of multi-source heterogeneous data. For unstructured text data and structured case library data, the large model accurately extracts disaster entities and their relationships through semantic recognition, generating a preliminary triplet knowledge framework.

[0018] When performing entity recognition using the first major language model, the text sequence is mapped into a high-dimensional vector through word embeddings and positional encoding, as shown below: ; in, For word embedding matrix, One-hot encoding of the input text. For the position encoding matrix, Represents a high-dimensional vector.

[0019] The first major language model outputs entity labels and relation types through a classification head. The loss function used during its training is the cross-entropy loss function. : ; in, Number of entity types One-hot encoding of the real label. Predicting probabilities for the largest language model.

[0020] The first large language model training set includes: text data with known entity labels and relationship types between entities; the text data is used as the input value of the model, and the entity labels and relationship types between entities are used as the output value of the model to train the first large language model, resulting in the trained first large language model.

[0021] Building upon this foundation, the first major language model further integrates knowledge from multi-source data. By aligning extracted entities with structured databases, it unifies terminology and analyzes and corrects contradictory information, ensuring the consistency of the knowledge system. The first major language model maps entities from different sources to a unified embedding space and measures semantic consistency using cosine similarity. ; in, and This is the entity embedding vector.

[0022] Subsequently, a network of nodes and relationships containing disaster entities, geological attributes, and response plans was constructed using a graph database. At the same time, the reasoning capabilities of the large model were used to fill in sparse areas of the data, and the data was incorporated into the graph after expert review to enhance the completeness of the knowledge.

[0023] The beneficial effects of the above technical solution are as follows: By constructing a multi-source knowledge base for tunnel water and mud inrush disasters, a systematic organization of disaster-related information is achieved, covering multi-dimensional data from macroscopic geological environment to microscopic lithological characteristics, providing a real data foundation for subsequent steps. The first major language model is used to extract and align entity-relation-entity triples from multi-source data. Similarity thresholds are used to unify terminology and fill in sparse knowledge areas. Without requiring extensive manual annotation, a structured and semantically consistent knowledge graph of tunnel water and mud inrush disasters is quickly constructed, providing an accurate and complete knowledge foundation for subsequent retrieval and reasoning.

[0024] Further, in step S102: acquiring the data to be processed at the site of the tunnel water and mud inrush disaster, and inputting the data to be processed into a multi-agent collaborative server; wherein, the multi-agent collaborative server includes: a control agent, a data preprocessing agent, a retrieval agent, a computation agent, and a reasoning and decision-making agent, specifically including: At the tunnel construction site, data related to water and mud inrush disasters is acquired through multi-source information collection methods. This includes adverse geological data, stratigraphic lithology data, groundwater level data, topographic data, rock stratum occurrence data, and surrounding rock grade data. The data to be processed is uniformly accessed and converted according to the preset data interface specifications. After data processing, the data to be processed is encapsulated into input data packets according to a unified data protocol and input to the control interface of the multi-agent collaborative server. This provides the basic input for subsequent intention recognition and task scheduling of the control agents.

[0025] The control agent, data preprocessing agent, retrieval agent, computation agent, and reasoning and decision-making agent were all trained using a large language model.

[0026] The control agent is implemented through a second, trained language model. This control agent formulates decision-making processes based on the input task and schedules other agents to participate in the work according to these processes. The second, trained language model uses an input task with known names of the scheduled agents as its training dataset.

[0027] When the input task is data preprocessing, the control agent schedules the data preprocessing agent to work. When the input task is data retrieval, the control agent schedules the retrieval agent to work. When the input task is data computation, the control agent schedules the computation agent to work. When the input task is decision-making reasoning, the control agent schedules the reasoning and decision-making agents to work. When the input task is to output the final solution for water and mud inrush, the control agent sequentially schedules the preprocessing agent, retrieval agent, computation agent, and reasoning and decision-making agent to work.

[0028] The beneficial effects of the above technical solution are: by unifying and standardizing the access and encapsulation of multi-source heterogeneous data from tunnel water and mud inrush disaster sites, the structure of the data is unified and the flow is efficient, providing a high-quality input foundation for multi-agent collaborative systems, enabling various intelligent agents to efficiently call and process the data.

[0029] Further, S103: The control agent performs intent recognition on the input data, and determines the categories of other agents participating in the task and the subsequent decision-making process based on the intent recognition result, specifically including: The control agent is based on the second major language model, building its task understanding and scheduling capabilities through few-shot examples. During system initialization, various functional agents are standardized, encapsulated, and registered to form an agent resource pool. Each type of agent includes: input interface definition, functional description, invocation toolset, and output format constraints. Specifically: A data preprocessing intelligent agent is used to extract key information; A retrieval agent for knowledge graph queries and case similarity searches; Computational intelligent agents are used for numerical simulation and parameter optimization. A reasoning and decision-making agent used for generating solutions and comparing multiple solutions.

[0030] Controlling intelligent agents involves dynamically identifying and scheduling various intelligent agents by calling the agent registry.

[0031] Upon receiving the input data to be processed, the controlling agent first performs an intent recognition task based on its built-in second-largest language model, performing semantic parsing of the input data, specifically including: (1) Determine whether the data belongs to the field of tunnel water inrush and mud inrush disasters; (2) Preliminary identification of data types (text, images, monitoring data, etc.) and degree of structuring; (3) Identify task objectives (such as solution generation or disposal optimization).

[0032] The intent recognition process can be achieved through a pre-trained and fine-tuned classification model. The training method is as follows: using tunnel disaster data as input and whether it belongs to the domain of water inrush and mud inrush disaster as a label for supervised training, so that the control agent has the ability to filter domains.

[0033] When the identification result does not belong to the target domain, the control agent terminates the subsequent process; when it belongs to the target domain, it enters the subsequent scheduling stage.

[0034] The control agent, based on the intent recognition results, determines the category of agents participating in the task. Specifically: When the input data is mainly unstructured text, the data preprocessing agent is scheduled first. When historical data is needed, schedule the retrieval agent; When parameter calculation and effect evaluation are involved, a computational agent is scheduled; When a final disposal plan needs to be generated, the inference and decision-making agent is scheduled.

[0035] For complex disaster scenarios, multiple intelligent agents can be scheduled to form a collaborative combination.

[0036] Based on this, the controlling agent constructs subsequent decision-making processes according to task type and data characteristics. These processes include two types: linear processes and iterative closed-loop processes. For scenarios with complete data and clearly defined problems, the sequential execution process of "preprocessing - retrieval - calculation - decision" is adopted; For scenarios with high uncertainty, a feedback mechanism is introduced to iterate through multiple rounds of data preprocessing, retrieval, and calculation until the preset conditions are met.

[0037] The beneficial effects of the above technical solution are as follows: by constructing a unified scheduling mechanism with control agents as the core, dynamic selection of multiple types of functional agents can be realized, transforming the original processing process that relied on a single model into a task-driven multi-stage decision-making process; by using intent recognition to perform domain filtering and task parsing on input data, irrelevant data interference can be avoided; by introducing an agent resource pool and a process adaptive construction mechanism, the system can flexibly adjust the path according to different scenarios, significantly improving the automation and adaptability of decision-making.

[0038] Furthermore, the control agent sequentially schedules the data preprocessing agent, retrieval agent, computation agent, and reasoning and decision-making agent to provide the final processing solution, specifically including: S103-1: The control agent schedules the data preprocessing agent to preprocess the data to be processed and obtain key disaster information; S103-2: Control agent schedules and retrieves agent, and retrieves several alternative solutions based on key disaster information in the knowledge graph of tunnel water inrush and mud inrush disasters; S103-3: The control agent schedules the computation agent to perform numerical calculations on each alternative solution, simulate the expected effect of each alternative solution and iteratively optimize to obtain the optimized alternative solution. S103-4: The control agent schedules the reasoning and decision-making agent to obtain the final disposal plan based on the optimized alternatives.

[0039] Further, S103-1: The control agent schedules the data preprocessing agent to preprocess the data to be processed and obtain key disaster information, specifically including: The third language model, after training, is used to extract key disaster information from the data to be processed; the key disaster information includes geological parameters and disaster phenomena.

[0040] It should be understood that all large language models in this application can be implemented using Chat GPT or DeepSeek.

[0041] The data preprocessing agent is built around a large language model, extracting core disaster elements from unstructured text, semi-structured records, and structured monitoring data, and expressing them uniformly according to a preset structured template.

[0042] The key information about the disaster includes: geological parameters and disaster phenomena; Among them, geological parameters include: surrounding rock lithology, rock mass structural characteristics, degree of joint and fissure development, karst development, groundwater type, water pressure and water volume, etc. Among them, disaster phenomena include: location of water inrush, form of water inrush (point, line or area), water inrush volume, accompanying mud and sand, and changes in the working face.

[0043] Specifically, it should be understood that the training set for the trained third language model is tunnel water inrush and mudslide disaster data with known key disaster information; the tunnel water inrush and mudslide disaster data serves as the input value of the third language model, and the known key disaster information serves as the output value of the third language model. This achieves a mapping from unstructured descriptions to standardized parameters.

[0044] Furthermore, S103-1: After the control agent schedules the data preprocessing agent to preprocess the data to be processed and obtains the key disaster information, it also includes: the control agent performs a data integrity check on the key disaster information. If the requirements are not met, the control agent schedules the data preprocessing agent to perform preprocessing again.

[0045] Further, in S103-1: the controlling agent performs a data integrity check on the critical disaster information. If the requirements are not met, the controlling agent schedules the data preprocessing agent to perform preprocessing again, specifically including: Based on a pre-defined data integrity indicator system, key disaster information is evaluated item by item. The indicators include key field coverage and data consistency. Among them, the key field coverage rate is used to measure whether the extracted information covers the predefined set of core elements (such as essential fields for geological parameters and disaster phenomena); the predefined set of core elements is denoted as:

[0046] in, This represents the set of core fields required for handling tunnel water and mud inrush disasters. Indicates the first One core field, This indicates the total number of core fields.

[0047] Key field coverage The calculation formula is:

[0048] in, Indicates the coverage of key fields; Indicates the first The weights of each core field are used to characterize the importance of that field in the recommendation of subsequent disaster response plans; Indicates the first The extraction status of the core field, when the first When all core fields have been extracted and their contents are not empty. ,otherwise When all core fields are of equal importance, all All are set to 1.

[0049] when If the coverage rate is greater than or equal to the preset coverage threshold, it is determined that the disaster key information meets the requirements for subsequent processing in terms of field coverage; otherwise, it is determined that the disaster key information has missing fields, and the control agent reschedules the data preprocessing agent to extract or supplement the data again.

[0050] Data consistency level is used to detect whether there are obvious conflicts between different data sources.

[0051] Data consistency is used to detect whether there are significant conflicts in the information provided by different data sources for the same disaster field. A significant conflict refers to a discrepancy between the values ​​or descriptions provided by different data sources for the same core field that exceeds the allowable error range.

[0052] For example, in the field of surrounding rock grade, if the advanced geological prediction result is grade III surrounding rock, while the tunnel face logging result is grade V surrounding rock, and the two cannot be interpreted through time and location, then it is determined that there is a significant conflict in the field of surrounding rock grade.

[0053] Specifically, for the first Core fields Perform a consistency check to obtain the field consistency check value. When the first When there are no obvious conflicts in the core fields When the first When there is a significant conflict in a core field .

[0054] When critical disaster information meets the preset integrity threshold, the control agent transmits the critical disaster information as standardized input to subsequent agents. When the requirements are not met, the control agent reschedules the data preprocessing agent to process the data based on the degree of missing information.

[0055] The beneficial effects of the above technical solution are: to achieve standardized processing and key information extraction of multi-source data from tunnel water and mud inrush disaster sites, and to ensure that the input data meets the requirements of subsequent tasks through the quality assessment and feedback mechanism of the control agent.

[0056] Further, in S103-2: the control agent schedules the retrieval agent to search the tunnel water inrush and mudslide disaster knowledge graph based on key disaster information, and obtains several alternative solutions, specifically including: S103-21: Treat key disaster information as key disaster entities; based on geological parameters and disaster phenomena, obtain geological parameter entities and disaster phenomenon entities; S103-22: Based on key disaster entities, a breadth-first search graph traversal algorithm is used to locate the edges directly connected to key disaster entity nodes in the knowledge graph of tunnel water inrush and mud inrush disasters. The nodes connected to the current edge are determined based on the directly connected edges, and the nodes connected to the current edge are used as candidate nodes. S103-23: For each candidate node, calculate the attribute matching degree between the candidate node and the disaster critical entity, retain candidate nodes with attribute matching degree exceeding the set threshold, and delete candidate nodes with attribute matching degree below the set threshold. S103-24: Treat the key disaster entity nodes and the remaining candidate nodes as the current disaster nodes. Search for the top K shortest paths in the knowledge graph based on the current disaster nodes. The shortest path refers to the path between the current disaster node and the successfully handled case node. Extract the handling measures for each path in the top K shortest paths to obtain K alternative handling solutions.

[0057] It should be understood that geological parameter entities refer to structured information used to describe the stratigraphic characteristics and geological environment of the tunnel construction area. Common geological parameters include: surrounding rock grade, stratigraphic lithology, groundwater level, and stratum attitude. For example, "surrounding rock grade V" is a typical geological parameter entity.

[0058] It should be understood that the physical manifestations of a disaster refer to abnormal phenomena observed during tunnel construction, used to describe the specific situation and development trend of water and mud inrush disasters on site. Examples include "mud inrush at the tunnel face," "flowing sand," "surge in water volume," and "turbid water." These disaster phenomena reflect the actual signs that a disaster has begun or is about to occur.

[0059] It should be understood that the "successful handling case node" refers to a node in the tunnel water inrush and mudslide inrush knowledge graph that identifies historical cases where known handling measures have been verified through actual engineering projects and have eliminated the disaster risk. Specifically, a successful handling case node includes elements such as: basic case information, details of the handling measures, and effect evaluation results.

[0060] When the retrieval algorithm starts from a current disaster node to find the shortest path, the successfully handled case node is the target endpoint, which represents an empirical sample that can achieve effective handling results under similar geological and disaster conditions.

[0061] It should be understood that S103-22: Based on the preprocessed key disaster entities (such as surrounding rock level V, groundwater level >10m), associated nodes are located using a breadth-first search-based graph traversal algorithm. The edges directly connected to these nodes are traversed layer by layer according to the breadth-first strategy to expand the range of candidate nodes. During the traversal, the path depth between nodes is recorded to construct an initial candidate subgraph.

[0062] It should be understood that S103-23: For each node in the candidate subgraph, calculate its attribute matching degree with the current disaster scenario. Assume the set of node attributes is... The current set of disaster attributes is The weights of each attribute are The similarity formula is as follows: ; in, For attribute matching functions;

[0063] Retaining similarity exceeding a preset threshold Nodes with attribute matching scores below a set threshold are removed.

[0064] It should be understood that S103-24: After obtaining the key nodes of the knowledge graph, the shortest path from the current disaster node to the success case node is searched in the knowledge graph, the key measures in the path are extracted, and the priority of the scheme is sorted in combination with the path weight.

[0065] Starting from the current disaster node and ending at historical success case nodes, Dijkstra's algorithm is used to search for the top K shortest paths in the knowledge graph. The path weights are determined by the edge type and historical success rate. Key measure nodes are extracted from the top K paths, and their associated parameters are analyzed to serve as the final alternative solutions.

[0066] The beneficial effects of the above technical solution are as follows: by mapping key disaster information to entity nodes in a knowledge graph, using breadth-first search to quickly expand relevant nodes, and using the shortest path algorithm to automatically extract multiple highly relevant treatment paths from the current disaster node to the successful case node, it is possible to quickly and accurately obtain highly matching alternative solutions under the current on-site conditions, providing solid technical support for subsequent assessment and decision-making.

[0067] The retrieval agent is built around the fourth major language model and integrates knowledge graph query tools through a toolcalling mechanism, thereby achieving a collaborative capability of "semantic understanding + structured retrieval". Its configuration includes a graph query interface module, which is encapsulated through a unified interface and has its functions described and parameters defined in the agent's registry.

[0068] Furthermore, S103-2: After the control agent schedules the retrieval agent to retrieve several alternative solutions based on key disaster information in the knowledge graph of tunnel water inrush and mud inrush disasters, the process further includes: the control agent judging the similarity between the alternative solutions and the current situation; if the requirements are not met, the control agent schedules the retrieval agent to retrieve the solutions again.

[0069] Furthermore, the controlling agent determines the similarity between the alternative solution and the current situation. If the similarity does not meet the requirements, the controlling agent schedules the retrieval agent to conduct another retrieval, specifically including: After obtaining alternative solutions, the control agent receives the search results and evaluates the degree of matching between each alternative solution and the current disaster scenario: Specifically, the control agent inputs the current key disaster information, alternative solutions and corresponding historical case information into the big language model, writes the scoring criteria (mainly considering the consistency of disaster situation and the applicability of measures) into the preset Prompt template, and the big language model performs reasoning and judgment based on the situation and outputs the score of each alternative solution.

[0070] When the comprehensive similarity score of the alternative solutions meets the preset threshold, the control agent will pass the current alternative solution as a valid candidate solution to the subsequent calculation and decision-making stage. When the requirements are not met, the control agent dynamically adjusts the retrieval strategy according to the specific reasons for insufficient similarity and reschedules the retrieval agent to perform the retrieval again. The dynamic adjustment of the retrieval strategy includes: relaxing the attribute matching threshold, increasing the graph traversal depth, and increasing the number of candidate nodes.

[0071] The beneficial effects of the above technical solution are as follows: by introducing a retrieval agent and integrating a knowledge graph query tool, efficient retrieval of historical handling experience is achieved; by combining attribute matching degree calculation with shortest path search, the matching accuracy between alternative solutions and the current disaster scenario is improved; and by introducing a similarity evaluation and feedback re-retrieval mechanism under the scheduling of the control agent, the retrieval process has adaptive optimization capabilities.

[0072] Further, S103-3: The control agent schedules the computational agent to perform numerical calculations on each candidate solution, simulate the expected effect of each candidate solution, and iteratively optimize to obtain the optimized candidate solution, specifically including: S103-31: The control agent first inputs the initial alternative disposal plan, key disaster information, and boundary conditions related to the current project into the computational agent; S103-32: The computational agent establishes a computational task description of the current disaster response scenario based on the initial alternative response plans, key disaster information, and boundary conditions related to the current project, and determines the key response parameters that need to be numerically calculated. S103-33: Computational intelligent agent, which uses its built-in fifth language model to perform numerical calculations and generate standardized computational inputs; S103-34: The computational agent outputs the expected performance index corresponding to each initial scheme based on the computation results; S103-35: After obtaining the preliminary simulation results, the control agent receives the results output by the computational agent and judges whether the current scheme meets the optimization requirements according to the preset evaluation criteria; when the key indicators of a certain initial scheme do not meet the requirements, the control agent sends an optimization instruction to the computational agent to trigger the parameter iterative optimization process.

[0073] Furthermore, S103-32: The computational agent, based on the initial alternative disposal plans, key disaster information, and boundary conditions relevant to the current project, establishes a computational task description for the current disaster disposal scenario and determines the key disposal parameters requiring numerical calculation, specifically including: The computational agent first identifies the type of the initial alternative treatment plan and determines that the initial alternative treatment plan belongs to one or more of the following measures: grouting and water plugging, drainage and pressure reduction, support and reinforcement, and excavation adjustment.

[0074] Then, based on the identified measure type, the corresponding calculation task template is called from the preset calculation task template library, and the key disaster information and engineering boundary conditions are filled into the calculation task template to form a calculation task description of the current disaster response scenario.

[0075] When determining key treatment parameters, the computational agent extracts parameters that have a major impact on the treatment effect as key treatment parameters based on the type of measures in the initial alternative treatment plans.

[0076] If the initial alternative treatment plan is grouting to plug water, the key treatment parameters to consider are mainly grouting pressure, grouting volume, grout water-cement ratio, and grouting hole spacing. If the initial alternative treatment plan is support and reinforcement measures, the key treatment parameters mainly consider the support type, anchor length, anchor spacing, and steel arch spacing. If the initial alternative disposal plan is excavation adjustment measures, the key disposal parameters should mainly consider the excavation step distance, the method of partial excavation, etc.

[0077] The computational agent uses the fifth language model as the core of task orchestration and integrates numerical calculation models and optimization tools to quantitatively calculate, simulate, and iteratively correct key treatment parameters in alternative solutions. As an intermediate analysis link between steps S103-2 and S103-4, the computational agent verifies the engineering feasibility and optimizes the parameters of the retrieved initial solutions, providing quantitative basis for the final decision.

[0078] Among them, the key disaster information of S103-31 includes the surrounding rock level, lithological characteristics, fracture development, groundwater conditions, water pressure, water volume, water inrush location, and mud inrush degree; The boundary conditions of S103-31 include tunnel depth, excavation method, support conditions, construction status, face stability requirements, and allowable deformation range. The key treatment parameters refer to the engineering parameters in the initial plan that have a decisive impact on the treatment effect, including: grouting pressure, grouting volume, grout diffusion radius, grout water-cement ratio, grouting hole spacing, waterstop wall thickness, and sectional excavation step distance.

[0079] Furthermore, in the numerical calculation stage of S103-33, the computational agent invokes specialized computational tools corresponding to different treatment measures to simulate the expected effects of each alternative scheme and obtain simulation results; specifically, When the initial scheme involves grouting and water plugging, the computational agent calculates the grouting diffusion range, water plugging efficiency, and water pressure attenuation effect based on the formation permeability, water pressure conditions, and grout properties. When the initial scheme involves support reinforcement or excavation adjustment, the computational agent further simulates changes in surrounding rock stability, the probability of face instability, and the stress response of the support.

[0080] The simulation results can be achieved using one or more of the following methods: analytical formulas, finite difference models, and finite element models.

[0081] Furthermore, the expected performance indicators of S103-34 include: the reduction in water inflow, the reduction in water pressure, the improvement in surrounding rock stability, the safety factor of the working face, the consumption of grout materials, the change in construction cycle, and the change in construction risk level.

[0082] Furthermore, the S103-35 evaluation criteria include three categories: effect constraints, construction constraints, and safety constraints. Effect constraints are used to determine whether a solution achieves the expected goals of water blocking, pressure reduction, or stabilization. Construction constraints are used to determine whether a plan meets the on-site equipment capabilities, construction space, and process conditions. Safety constraints are used to determine whether a plan may lead to new instability, roof collapse, secondary water inrush, or secondary disasters.

[0083] Furthermore, the method also includes: S103-36: When the computational agent outputs the results, it not only outputs the optimized parameter values, but also simultaneously outputs a description of the parameter optimization process, a description of key influencing factors, and a summary of the calculation results of each round, so as to enable the control agent and the subsequent reasoning and decision-making agent to perform interpretability analysis.

[0084] The parameter optimization process description reflects how the parameters were adjusted, the key influencing factors description indicates which geological or construction factors dominated the optimization direction, and the summary of calculation results for each round records the changes in simulation effects corresponding to different parameter combinations.

[0085] The beneficial effects of the above technical solution are: by introducing a computational intelligent agent to perform numerical simulation and parameter optimization on the initial treatment plan obtained from the retrieval, the transformation from experience-driven to quantitative analysis-driven is realized, making the treatment plan verifiable and engineering-applicable under the current geological and construction conditions.

[0086] Further, S103-4: The controlling agent schedules the reasoning and decision-making agent to obtain the final disposal plan based on the optimized alternative plans, specifically including: The reasoning and decision-making intelligent agent is a decision-making intelligent agent built around the sixth language model; The input data for the sixth language model includes: raw disaster site data, structured key disaster information, optimized alternative response plans and their corresponding numerical simulation results; The output data of the sixth language model includes: the final handling plan; The sixth language model is trained using a reinforcement learning approach, and its reward function employs a multi-index weighting mechanism, as shown below: ; in, The risk level assessment score of the input strategy is determined by on-site experts based on a combination of factors, using a tiered scoring method (e.g., 1 for high risk, 0.5 for medium risk, and 0 for low risk). The higher the value, the higher the risk. The penalty factor coefficient is used to encourage the model to avoid catastrophic paths.

[0087] ; in, This indicates the current handling plan. Indicates the first Scoring functions for each evaluation dimension This indicates the weight coefficient of the current evaluation dimension. The score for the first evaluation dimension is the standard compliance score; the score for the second evaluation dimension is the expert score; the score for the third evaluation dimension is the construction feasibility score; and the score for the fourth evaluation dimension is the terminology standardization score.

[0088] No. The scoring functions for each evaluation dimension, and the corresponding scoring methods are as follows: if If it fully complies with the design specifications, then ;if If it basically meets the design specifications, then ;if If it does not comply with design specifications, then ; if If it is completely consistent with expert knowledge, then ;if If it is only partially consistent with expert knowledge, then ;if If there is a conflict with expert knowledge, then ; if If it is completely consistent with the construction constraints, then ;if If only part of the construction constraints are consistent, then ;if If it is inconsistent with the construction constraints, then ; if If the terminology is standardized, the structure is clear, and the steps are complete, then... ;if If the terminology is ambiguous or the steps are incomplete, then ;if If the terminology is seriously incorrect or the structure is confusing, then .

[0089] Understandably, to enable the large-scale model to grasp the professional knowledge and disaster response logic in the field of tunnel water and mud inrush, supervised fine-tuning and reinforcement learning are applied to enhance its deep thinking ability and adapt it to engineering decision-making scenarios. Annotated real historical disaster case data (input: disaster situation + alternative solutions; label: final response plan) are used to update the parameters of the large-scale language model, and the cross-entropy loss function is adopted. : ; in, For input, The final solution text annotated by experts.

[0090] Based on this, professional standards for tunnel construction are introduced as an expert rule base to train the reward model. Through reinforcement learning, the compliance of the generated scheme is optimized, and the model's thinking and reasoning abilities are effectively improved.

[0091] To effectively guide the generation of tunnel inrush water and mud disposal solutions using large language models, constructing a reward model capable of quantifying solution quality is a crucial step in the reinforcement learning phase. This reward model evaluates the compliance, effectiveness, and feasibility of the solutions generated by the model, and uses the evaluation results as feedback signals to optimize the model strategy, enabling it to continuously approach expert levels through multiple iterations.

[0092] It should be understood that the rating dimensions include: The compliance score S1 is calculated by analyzing structured industry design standards, such as the "Railway Tunnel Design Standard". Compliance constraint rules are extracted from these standards. If the model output solution meets these structured standard clauses, a higher score is obtained; otherwise, a lower score is obtained. The expert knowledge fit score S2 is based on the matching degree between the evaluation scheme constructed in the knowledge graph and the expert experience. If the model scheme is consistent with this, the score is high. The construction feasibility score S3 determines whether the plan meets the actual constraints of the construction site, such as physical limitations, construction technology, and resource allocation. For example, whether the equipment can be deployed in the existing working face space and whether the required materials are on site. The terminology standardization score is S4, which includes: technical terminology standardization, logical structure clarity, and step completeness.

[0093] In addition, disaster risk modeling with a high penalty factor is introduced. Tunnel water inrush and mud inrush are serious disaster events. Incorrect handling strategies can easily trigger secondary disasters, causing serious casualties or structural damage.

[0094] Therefore, a high penalty factor is specifically introduced in the reward model design to quantify and punish the catastrophic risk caused by strategy misjudgment.

[0095] The beneficial effects of the above technical solution are as follows: by inputting disaster information and alternative disposal plans into the sixth language model that has been supervised fine-tuned and optimized by reinforcement learning, and by using a multi-index weighted reward function combined with a high penalty factor, the trained model has the ability to generate disposal plans that meet industry standards, expert experience and on-site operability, thus ensuring the quality of the generated plans.

[0096] Furthermore, the method also includes: guiding the trained sixth language model using a thought chain approach when the reasoning and decision-making agent performs a task, specifically including: S103-51: Using the sixth language model, semantic parsing is performed on the data to be processed to extract the entities to be processed and align the entities to be processed with the entity nodes in the knowledge graph. S103-52: Using the sixth language model, extract several disaster factors from the data to be processed; rank all disaster factors according to the pre-defined risk level of each disaster factor, and select the top N disaster factors as key triggers; S103-53: Input the key triggers and several alternative solutions into the sixth language model to analyze the expected effect of each solution; S103-54: Compare the expected effects of different treatment plans using the sixth language model, output the final treatment plan, and verify the reliability of the final treatment plan.

[0097] It should be understood that disaster factors refer to the key conditions that directly lead to or affect the occurrence of water and mud inrush events in tunnels, including geological environment, construction conditions, or groundwater characteristics; adverse geology: such as faults, fracture zones, karst cavities, and other hidden geological structures, which can provide channels or weak annular surfaces for water and mud inrush; advanced geological forecasting: if the forecast results are incomplete or misjudged, potential risk areas cannot be identified in time, which can easily lead to the tunnel face entering high-risk areas; stratum lithology: different lithologies have significant differences in permeability and deformation resistance, which will directly affect the outflow trend of water and mud flow; excavation and support: if the support design or construction is insufficient, the surrounding rock is prone to instability or collapse under the action of water, which can lead to mud inrush or chain collapse.

[0098] The beneficial effects of the above technical solution are as follows: by introducing chain-like thinking guidance into the sixth language model, the model can accurately align the entities to be processed with the nodes in the knowledge graph, automatically extract key disaster factors and track relevant risk factors, and then perform step-by-step deduction and effect comparison of each alternative solution. Finally, the generated results are verified by combining the expert rule base, realizing the whole process of deep reasoning from semantic parsing to causal analysis, solution deduction to result verification, and generating reliable disposal solutions under on-site conditions.

[0099] Based on the concept of a thought chain, this application proposes a large-scale model for deep reasoning and solution generation. After problem analysis and alternative solutions are input into the large model, a "thought chain" mechanism drives the model to perform deep reasoning. The proposed thought chain mechanism introduces a highly domain-specific reasoning path into the scenario of handling water and mud inrush disasters in tunnels. This mechanism relies on the structured expression of disaster semantics, combined with the identification of key triggers, the deduction of treatment effects, and result verification based on expert rules, to construct a deep logical reasoning process that conforms to the characteristics of tunnel water and mud inrush engineering.

[0100] Based on the large-scale model of thought chain deep reasoning and solution generation, including: (1) Disaster scenario localization and input standardization: After receiving the description input related to tunnel water inrush and mud inrush disaster, the large model first uses semantic parsing and entity linking technology to align the non-standardized text information with the concept nodes in the disaster knowledge graph to ensure that the model's understanding of the problem has semantic consistency, which serves as the starting point for subsequent thought chains.

[0101] (2) Key factor extraction and priority ranking: Based on the standardized expression, the graph neural network is used to perform propagation calculation on the graph to obtain the causal relationship in the knowledge graph, identify the key factors in the disaster formation process, and prioritize the risk factors.

[0102] (3) Potential effect deduction and multi-stage solution construction: For each top-ranked factor, a prompt word template is constructed with "key factor + candidate measure" as input. An example of the prompt word template is shown below to guide the large model to generate targeted evaluation analysis. After completing the effect deduction of a single measure, the model constructs a multi-stage disposal strategy through the logical combination of disposal order.

[0103] Example of a prompt word template: The tunnel currently faces the following hazards: water inrush and mud inrush: {key factors}; Candidate measures: {candidate measures}; Based on this risk factor and measures, please conduct the following analysis: (31) Analyze the applicability and limitations of this measure under this risk condition; (32) Analyze the expected effects of implementing this measure, and if there are any shortcomings, please propose more suitable measures.

[0104] (4) Expert rule inversion verification: After forming a complete disposal plan, the model will analyze the reasoning results one by one in a structured manner, retrieve the most relevant rules in the expert rule base and historical disposal experience with keywords, construct prompt word template with "reasoning results + expert rules" as input, and guide the model to identify potential conflicts by constructing comparison verification prompts to verify the reliability of the results.

[0105] The reasoning and decision-making agent is configured through a model + rules + reasoning strategy, specifically including: (1) The large language model serves as the core reasoning engine; (2) The expert rule base serves as an external constraint tool; (3) The reasoning strategy of the thinking chain serves as a control mechanism for the reasoning process; (4) The reward model is used as a tool for evaluating the quality of the scheme.

[0106] Example 2 This embodiment provides a multi-agent-based tunnel water and mud inrush disposal solution recommendation system, including: The knowledge base construction module is configured to: construct a knowledge base for tunnel water inrush and mud inrush disasters based on historical tunnel disaster knowledge data and historical tunnel disaster case data; and construct a knowledge graph for tunnel water inrush and mud inrush disasters based on the aforementioned tunnel water inrush and mud inrush disaster database. The data acquisition module is configured to: acquire data to be processed at the site of a water inrush and mudslide disaster in a tunnel, and input the data to be processed into a multi-agent collaborative server; wherein the multi-agent collaborative server includes: a control agent, a data preprocessing agent, a retrieval agent, a computation agent, and a reasoning and decision-making agent; The output module is configured to: control the agent to perform intent recognition on the input data, determine the categories of other agents participating in the task and the subsequent decision-making process based on the intent recognition results; when the intent recognition indicates that a final disposal plan needs to be generated, the control agent sequentially schedules the data preprocessing agent, retrieval agent, computation agent and reasoning decision agent to provide the final disposal plan.

[0107] It should be noted that the knowledge base construction module, data acquisition module, and output module mentioned above correspond to steps S101 to S103 in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should be noted that these modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0108] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for recommending solutions for tunnel water and mud inrush incidents based on multi-agent systems, characterized in that, include: Based on historical tunnel disaster knowledge data and historical tunnel disaster case data, a knowledge base for tunnel water inrush and mud inrush disasters will be constructed. Based on the aforementioned knowledge base on tunnel water inrush and mud inrush disasters, a knowledge graph of tunnel water inrush and mud inrush disasters is constructed. The system acquires data to be processed from the site of a water inrush and mud inrush disaster in a tunnel, and inputs the data into a multi-agent collaborative server. The multi-agent collaborative server includes: a control agent, a data preprocessing agent, a retrieval agent, a computation agent, and a reasoning and decision-making agent. The controlling agent performs intent recognition on the input data, and determines the categories of other agents participating in the task and the subsequent decision-making process based on the intent recognition results. When the intent recognition indicates that a final solution needs to be generated, the controlling agent sequentially schedules the data preprocessing agent, retrieval agent, computation agent, and inference decision-making agent to provide the final solution, which specifically includes: The control agent schedules a data preprocessing agent to preprocess the data to be processed, obtaining critical disaster information. The control agent then performs a data integrity check on this critical information. If the check fails to meet the requirements, the control agent schedules the data preprocessing agent to perform preprocessing again, specifically including: Based on a pre-defined data integrity indicator system, key disaster information is evaluated item by item. The indicators include key field coverage and data consistency. Among them, the key field coverage rate is used to measure whether the extracted information covers a predefined set of core elements, which is denoted as: in, This represents the set of core fields required for handling tunnel water and mud inrush disasters. Indicates the first One core field, Indicates the total number of core fields; Key field coverage The calculation formula is: in, Indicates the coverage of key fields; Indicates the first The weights of each core field are used to characterize the importance of that field in the recommendation of subsequent disaster response plans; Indicates the first The extraction status of the core field, when the... When all core fields have been extracted and their contents are not empty. ,otherwise When all core fields are of equal importance, all All are set to 1; when When the coverage rate is greater than or equal to the preset coverage threshold, it is determined that the disaster key information meets the requirements for subsequent processing in terms of field coverage; otherwise, it is determined that the disaster key information has missing fields, and the control agent reschedules the data preprocessing agent to extract or supplement the data again. The data consistency level is used to detect whether there are obvious conflicts between different data sources. When the requirements are not met, the control agent reschedules the data preprocessing agent to process the data according to the degree of missing information. The control agent schedules the retrieval agent to search the tunnel water inrush and mudslide disaster knowledge graph based on key disaster information, obtaining several alternative solutions. The control agent then judges the similarity between the alternative solutions and the current situation. If the alternative solutions do not meet the requirements, the control agent schedules the retrieval agent to search again, specifically including: After obtaining alternative solutions, the control agent receives the search results and evaluates the degree of matching between each alternative solution and the current disaster scenario; When the overall similarity score of the alternative solutions meets the preset threshold, the control agent passes the current alternative solution as a valid candidate solution to the subsequent calculation and decision-making stage; when the requirements are not met, the control agent dynamically adjusts the retrieval strategy according to the specific reasons for insufficient similarity and reschedules the retrieval agent to perform the retrieval again. The control agent schedules the computational agent to perform numerical calculations on each alternative solution, simulate the expected effect of each alternative solution and iteratively optimize it to obtain the optimized alternative solution. This includes: the control agent first inputs the initial alternative solution, key disaster information and boundary conditions related to the current project into the computational agent. The computational agent establishes a computational task description for the current disaster response scenario based on initial alternative response plans, key disaster information, and boundary conditions relevant to the current project. It also identifies key response parameters requiring numerical calculation. When determining these parameters, the agent extracts those with the primary impact on the response effectiveness based on the types of measures in the initial alternative response plans. The agent uses a fifth language model as the core of its task orchestration and integrates numerical calculation models and optimization tools to quantitatively calculate, simulate, and iteratively correct the key response parameters in the alternative response plans. As an intermediate analysis step, the agent verifies the engineering feasibility of the retrieved initial plans and optimizes their parameters, providing quantitative evidence for the final decision. During the numerical calculation phase, the computational agent invokes specialized computational tools corresponding to different treatment measures to simulate the expected effects of each alternative scheme and obtain simulation results. Specifically: when the initial scheme involves grouting and water plugging, the computational agent calculates the grouting diffusion range, water plugging efficiency, and water pressure attenuation effect based on the formation permeability, water pressure conditions, and grout properties; when the initial scheme involves support reinforcement or excavation adjustment, the computational agent further simulates changes in surrounding rock stability, the probability of face instability, and the support stress response; the simulation results can be achieved using one or more of the following methods: analytical formulas, finite difference models, and finite element models. After obtaining the initial simulation results, the control agent receives the results output by the computational agent and judges whether the current scheme meets the optimization requirements according to the preset evaluation criteria. When the key indicators of an initial scheme do not meet the requirements, the control agent sends an optimization command to the computational agent to trigger the parameter iterative optimization process. The evaluation criteria include three categories: effect constraints, construction constraints, and safety constraints. Effect constraints are used to determine whether the scheme achieves the expected goals of water blocking, pressure reduction, or stabilization. Construction constraints are used to determine whether the scheme meets the on-site equipment capabilities, construction space, and procedural conditions. Safety constraints are used to determine whether the scheme may cause new instability, roof collapse, secondary water inrush, or secondary disasters. The computational agent outputs the expected effect indicators for each initial scheme based on the calculation results. The expected effect indicators include: the reduction in water inflow, the reduction in water pressure, the improvement in surrounding rock stability, the safety factor of the working face, the consumption of grout materials, the change in construction cycle, and the change in construction risk level. The control agent schedules the reasoning and decision-making agents to obtain the final solution based on the optimized alternatives. The sixth language model is trained using a reinforcement learning approach, and its reward function employs a multi-index weighting mechanism, as shown below: ； in, The risk level assessment score of the input strategy is determined by on-site experts based on a variety of influencing factors. A graded scoring method is used, with 1 for high risk, 0.5 for medium risk, and 0 for low risk. The higher the value, the higher the risk. The penalty factor coefficient is used to incentivize the model to avoid catastrophic paths; ； in, This indicates the current handling plan. Indicates the first Scoring functions for each evaluation dimension This indicates the weight coefficient of the current evaluation dimension. The score for the first evaluation dimension is the standard compliance score; the score for the second evaluation dimension is the expert score; the score for the third evaluation dimension is the construction feasibility score; and the score for the fourth evaluation dimension is the terminology standardization score.

2. The method for recommending tunnel water and mud inrush disposal solutions based on multi-agent technology as described in claim 1, characterized in that, Based on the aforementioned knowledge base on tunnel water inrush and mud inrush disasters, a knowledge graph of tunnel water inrush and mud inrush disasters is constructed, specifically including: Using the first large language model after training, entity-relationship-entity triplet data are extracted from the historical tunnel disaster knowledge data of the tunnel water inrush and mud inrush disaster knowledge base; To ensure consistency alignment between the entities in the triplet data and the entities in the historical tunnel disaster case data of the tunnel water inrush and mud inrush disaster knowledge base: calculate the similarity between the entities in the triplet data and the entities in the historical tunnel disaster case data of the tunnel water inrush and mud inrush disaster knowledge base. If the similarity value is greater than a set threshold, modify the entity name of the triplet data to the entity name of the historical tunnel disaster case data. Based on disaster entities, geological attribute entities, disposal plan entities, and the relationships between entities, a knowledge graph is constructed and then supplemented to obtain a knowledge graph of tunnel water inrush and mud inrush disasters.

3. The method for recommending tunnel water and mud inrush disposal solutions based on multi-agent technology as described in claim 1, characterized in that, The control agent schedules and retrieves agents to search the tunnel water inrush and mudslide disaster knowledge graph based on key disaster information, resulting in several alternative solutions, including: Key disaster information is used as key disaster entities; based on geological parameters and disaster phenomena, geological parameter entities and disaster phenomenon entities are obtained; Based on key disaster entities, a breadth-first search graph traversal algorithm is used to locate the edges directly connected to the nodes of key disaster entities in the knowledge graph of tunnel water inrush and mud inrush disasters. The nodes connected to the current edge are determined based on the directly connected edges, and the nodes connected to the current edge are selected as candidate nodes. For each candidate node, calculate the attribute matching degree between the candidate node and the disaster critical entity, retain candidate nodes with attribute matching degree exceeding the set threshold, and delete candidate nodes with attribute matching degree below the set threshold. The critical disaster entity nodes and the remaining candidate nodes are all regarded as the current disaster node. Based on the current disaster node, the top K shortest paths are searched in the knowledge graph. The shortest path refers to the path between the current disaster node and the successfully handled case node. The handling measures of each path in the top K shortest paths are extracted to obtain K alternative handling solutions.

4. The method for recommending tunnel water and mud inrush disposal solutions based on multi-agent technology as described in claim 1, characterized in that, The control agent schedules the reasoning and decision-making agents to arrive at the final solution based on the optimized alternatives, specifically including: The reasoning and decision-making intelligent agent is a decision-making intelligent agent built around the sixth language model; The input data for the sixth language model includes: raw disaster site data, structured key disaster information, optimized alternative response plans and their corresponding numerical simulation results; The output data of the sixth language model includes: the final disposal plan.

5. The method for recommending tunnel water and mud inrush disposal schemes based on multi-agent technology as described in claim 4, characterized in that, The method further includes: guiding the trained sixth language model using a thought chain approach when the reasoning and decision-making agent performs tasks, specifically including: Using the sixth language model, semantic parsing is performed on the data to be processed, the entities to be processed are extracted, and the entities to be processed are aligned with the entity nodes in the knowledge graph. Using the sixth language model, several disaster factors are extracted from the data to be processed; all disaster factors are ranked according to the pre-defined risk level of each disaster factor, and the top N disaster factors are taken as key triggers. Input the key triggers and several alternative solutions into the sixth language model to analyze the expected effect of each solution. The expected effects of different treatment plans are compared using the sixth language model, the final treatment plan is output, and the reliability of the final treatment plan is verified.

6. A multi-agent-based tunnel water and mud inrush disposal solution recommendation system, characterized in that, A method for recommending tunnel water and mud inrush response schemes based on multi-agent technology as described in any one of claims 1-5, comprising: The knowledge base construction module is configured to: construct a knowledge base for tunnel water inrush and mud inrush disasters based on historical tunnel disaster knowledge data and historical tunnel disaster case data; and construct a knowledge graph for tunnel water inrush and mud inrush disasters based on the aforementioned knowledge base. The data acquisition module is configured to: acquire data to be processed at the site of a water inrush and mudslide disaster in a tunnel, and input the data to be processed into a multi-agent collaborative server; wherein the multi-agent collaborative server includes: a control agent, a data preprocessing agent, a retrieval agent, a computation agent, and a reasoning and decision-making agent; The output module is configured to: control the agent to perform intent recognition on the input data, determine the categories of other agents participating in the task and the subsequent decision-making process based on the intent recognition results; when the intent recognition indicates that a final disposal plan needs to be generated, the control agent sequentially schedules the data preprocessing agent, retrieval agent, computation agent and reasoning decision agent to provide the final disposal plan.

Images