Multi-agent collaboration method and system for engineering construction projects

Abstract

This invention provides a multi-agent collaborative method and system for engineering construction projects, relating to the field of railway engineering management. The method includes: training an expert model for railway engineering; processing user-input dialogue information through the expert model to generate prompts; a content-aware agent generating a query statement based on the prompts and returning the query statement to the expert model; the expert model invoking a hybrid retrieval agent to perform a search based on the query statement; a chart generation agent generating a chart based on a chart creation request and returning it to the expert model; a data analysis agent performing data analysis based on a data analysis request and returning it to the expert model; and a language question-answering agent integrating and outputting the search results, charts, and analysis results. This invention effectively coordinates multiple agents to leverage their respective strengths, improving retrieval speed and accuracy, and outputting more comprehensive and understandable search results based on the input natural language.

Description

Technical Field

[0001] The present invention relates to the field of railway engineering management, and in particular to a multi-agent collaboration method and system for engineering construction projects. Background Art

[0002] Railway construction data covers a large amount of document materials throughout the entire construction process, involving multiple disciplines, dozens of file formats, and hundreds of content types, such as geological forecasts, construction logs, progress ledgers, maintenance records, and experimental test reports. These data come from a wide range of sources and are complex in type. How to efficiently retrieve and use this information has become a major technical challenge. Traditional keyword-based retrieval methods have poor robustness, single retrieval results, and poor relevance of retrieval results, and cannot form context-related intelligent results. Although RAG (Retrieval-Augmented Generation) technology has made progress in enhancing generation capabilities, it is difficult to accurately capture the complex relationships between entities, which easily leads to confusion in information associations and is rarely used in complex relationship scenarios. With the continuous expansion of the number and scale of railway engineering projects, an intelligent retrieval technology is urgently needed to meet the growing demand for data retrieval. Summary of the invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art or related technology, and discloses a multi-agent collaboration method and system for engineering construction projects, which effectively collaborates multiple agents to give full play to their respective advantages, improves the retrieval speed and accuracy, and can output more comprehensive and understandable retrieval results based on the input natural language.

[0004] The first aspect of the present invention discloses a multi-agent collaboration method for engineering construction projects, including: training an expert model in the railway engineering field: constructing an expert knowledge system document based on the construction knowledge and specifications related to the engineering construction project; constructing a railway engineering field expert model training data structure based on the expert knowledge system document to obtain a thinking chain knowledge data set; training the railway engineering field expert model based on the thinking chain knowledge data set; wherein the railway engineering field expert model is a large language model, and the large language model is trained based on MixLora; receiving natural language input: the dialogue information input by the user is processed by the railway engineering field expert model to generate prompt instructions; content perception: the content-aware agent generates a query statement based on the prompt instruction, and returns the query statement to the railway engineering field expert model; mixed retrieval: the railway engineering field expert model The domain expert model calls the hybrid retrieval agent to perform retrieval based on the query statement, and the hybrid retrieval agent returns the retrieval results to the railway engineering domain expert model; chart generation: the railway engineering domain expert model generates a chart making request based on the dialogue information and retrieval results, so that the chart generation agent generates a chart based on the chart making request and returns it to the railway engineering domain expert model; data analysis: the railway engineering domain expert model generates a data analysis request based on the dialogue information and retrieval results, so that the data analysis agent performs data analysis based on the data analysis request and returns the analysis results to the railway engineering domain expert model; language question and answer: the railway engineering domain expert model generates prompt instructions based on the retrieval results, charts and analysis results, so that the language question and answer agent integrates and outputs the retrieval results, charts and analysis results.

[0005] According to the multi-agent collaboration method for engineering construction projects disclosed in the present invention, preferably, the content-aware agent is implemented by a large model applied to NL2Sql, a large model applied to NL2Cypher, and an embedding encoding model, which converts natural language queries into query statements at three levels: relation, graph, and vector according to prompt instructions.

[0006] According to the multi-agent collaboration method for engineering construction projects disclosed in the present invention, preferably, the content-aware agents include the following types: work site mileage content-aware agents, process-aware agents, and concrete information-aware agents.

[0007] According to the multi-agent collaboration method for engineering construction projects disclosed in the present invention, preferably, the hybrid retrieval agent is implemented by a vector retrieval tool, a graph retrieval tool and a relational database tool, and searches for relevant content according to the instructions of the expert model in the railway engineering field and the query statement; the hybrid retrieval agent includes the following types: a secondary lining construction hybrid retrieval agent and a concrete information hybrid retrieval agent.

[0008] According to the multi-agent collaboration method for engineering construction projects disclosed in the present invention, preferably, the chart generation agent is implemented by a chart generation tool, and charts are generated according to chart generation requests and retrieval results; the chart generation agent includes the following types: surrounding rock measurement chart agent, concrete maintenance record agent.

[0009] According to the multi-agent collaboration method for engineering construction projects disclosed in the present invention, preferably, the data analysis agent is implemented by a semantic understanding large model, including the following types: concrete data analysis agent, construction operation analysis agent.

[0010] According to the multi-agent collaboration method for engineering construction projects disclosed in the present invention, preferably, the language question-answering agent is implemented by a large semantic understanding model, and the retrieval results, charts and analysis results are integrated according to the prompt instructions of the expert model to form a question-answering result.

[0011] According to the multi-agent collaboration method for engineering construction projects disclosed in the present invention, preferably, the step of training a large language model based on MixLora specifically includes:

[0012] MixLora's gating routing uses sparse attention gating:

[0013] The attention coefficient is:

[0014] ;

[0015] in, represents the query matrix, represents the key matrix, represents the value matrix, represents the transposed matrix of K, is the dimension of the bond matrix;

[0016] Through the mask matrix Control the position to be calculated, set the scores that are not within the sparse structure to infinitesimal or zero, and ignore these positions; through the residual connection, the final coefficient attention coefficient is obtained as:

[0017] ;

[0018] The matrix will assign 0 to the positions defined by the sparse structure and a negative maximum value to other positions so that these positions are ignored in the Softmax function;

[0019] A linear transformation is used to generate a gating matrix to control the attention output and give the network the ability to select information flow:

[0020] ;

[0021] in, is the gating matrix, is the activation function, is the weight matrix, is the input matrix, is the bias term;

[0022] To trade off information in the gating matrix:

[0023] ;

[0024] in, is the threshold filter indicator matrix, is the gating matrix after filtering;

[0025] Gated coefficient attention mechanism:

[0026] ;

[0027] The output of the gated coefficient attention mechanism is used as the input of the Lora network layer:

[0028] Lora network layer:

[0029] ;

[0030] in, is the original FFN layer weight matrix, B matrix and A matrix are the Lora layer weight matrices, is the weight coefficient;

[0031] During model training, the rank of the B matrix and the A matrix are set to 8. The coefficient is 0.5.

[0032] The second aspect of the present invention discloses a multi-agent collaboration system for engineering construction projects, including: a memory for storing program instructions; a processor for calling the program instructions stored in the memory to implement a multi-agent collaboration method for engineering construction projects such as any of the above-mentioned technical solutions.

[0033] The beneficial effects of the present invention include at least: based on the collaboration between the expert model and multiple intelligent agents, the retrieval efficiency and accuracy are improved, and the needs of users in different application scenarios can be met, such as on-site construction progress tracking, quality monitoring, construction safety inspection, etc. Users can not only query single questions, but also conduct multi-dimensional analysis of in-depth questions, providing intelligent support for the safety, quality and progress management of railway engineering projects. For example, when a user queries whether a certain construction stage meets the specifications, the expert model will integrate the content retrieved from various databases through the collaborative work of data analysis agents, hybrid retrieval agents and language question-answering agents, and output more accurate and comprehensive conclusions. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of the overall architecture of a multi-agent collaboration method for engineering construction projects according to an embodiment of the present invention is shown.

[0035] Figure 2 A schematic block diagram of a multi-agent collaborative system for an engineering construction project according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0036] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention can also be implemented in other ways different from those described herein, and therefore, the present invention is not limited to the limitations of the specific embodiments disclosed below.

[0037] According to one embodiment of the present invention, a multi-agent collaboration method for engineering construction projects is disclosed, including: training an expert model in the railway engineering field: constructing an expert knowledge system document based on construction knowledge and specifications related to the engineering construction project; constructing a railway engineering field expert model training data structure based on the expert knowledge system document to obtain a thinking chain knowledge data set; training the railway engineering field expert model based on the thinking chain knowledge data set; wherein the railway engineering field expert model is a large language model, and the large language model is trained based on MixLora; receiving natural language input: the dialogue information input by the user is processed by the railway engineering field expert model to generate prompt instructions; content perception: the content-aware agent generates a query statement based on the prompt instruction, and returns the query statement to the railway engineering field expert model; hybrid retrieval: the railway engineering field expert model The railway engineering domain expert model calls the hybrid retrieval agent to perform retrieval based on the query statement, and the hybrid retrieval agent returns the retrieval results to the railway engineering domain expert model; chart generation: the railway engineering domain expert model generates a chart making request based on the dialogue information and retrieval results, so that the chart generation agent generates a chart based on the chart making request and returns it to the railway engineering domain expert model; data analysis: the railway engineering domain expert model generates a data analysis request based on the dialogue information and retrieval results, so that the data analysis agent performs data analysis based on the data analysis request and returns the analysis results to the railway engineering domain expert model; language question and answer: the railway engineering domain expert model generates prompt instructions based on the retrieval results, charts and analysis results, so that the language question and answer agent integrates and outputs the retrieval results, charts and analysis results.

[0038] like Figure 1As shown, according to the overall architecture diagram of the multi-agent collaboration method for engineering construction projects disclosed in this embodiment, it can be seen that: the user inputs natural language into the expert model, and under the guidance of the expert model, five types of agents, namely content-aware agents, hybrid retrieval agents, chart generation agents, data analysis agents, and language question-answering agents, respectively perform different functions to form a collaborative work system, and finally output the analysis and processing results of multi-agent collaboration to the user. In view of the complex problems in railway engineering, the query, analysis, and generated content provided by each agent are integrated into a visual or structured output, which can meet the needs of users in different application scenarios, such as on-site construction progress tracking, quality monitoring, construction safety inspection, etc. Users can not only query single issues, but also conduct multi-dimensional analysis of in-depth issues, providing intelligent support for the safety, quality and progress management of railway engineering projects.

[0039] According to the above embodiment, preferably, the content-aware agent is implemented by a large model applied to NL2Sql, a large model applied to NL2Cypher, and an embedding coding model, and converts natural language queries into query statements at the three levels of relationship, graph, and vector according to prompt instructions. The content-aware agent includes the following types: work point mileage content-aware agent, process-aware agent, and concrete information-aware agent.

[0040] According to the above embodiment, preferably, the hybrid retrieval agent is implemented by a vector retrieval tool, a graph retrieval tool and a relational database tool, and searches for relevant content according to the instructions of the expert model in the railway engineering field and the query statement; the hybrid retrieval agent includes the following types: secondary lining construction hybrid retrieval agent, concrete information hybrid retrieval agent.

[0041] According to the above embodiment, preferably, the chart generation agent is implemented by a chart generation tool, and the chart is generated according to the chart generation request and the retrieval result; the chart generation agent includes the following types: surrounding rock measurement chart agent, concrete maintenance record agent.

[0042] According to the above embodiment, preferably, the data analysis agent is implemented by a semantic understanding big model, including the following types: concrete data analysis agent, construction operation analysis agent.

[0043] According to the above embodiment, preferably, the language question-answering agent is implemented by a large semantic understanding model, and integrates the retrieval results, charts and analysis results according to the prompt instructions of the expert model to form a question-answering result.

[0044] According to another embodiment of the present invention, the specific implementation of the multi-agent collaboration method for engineering construction projects is disclosed:

[0045] Domain expert knowledge modeling: Summarize the knowledge logic, processing methods, etc. used in the construction scenario, and condense the expert knowledge model. In domain expert knowledge modeling, it is first necessary to comprehensively collect and systematically sort out the knowledge logic, judgment criteria, and processing methods of each link in the construction scenario. This includes detailed domain knowledge, such as operating specifications, compliance judgment basis, quality and safety indicators, etc. By deeply analyzing the internal relationships of these knowledge, a set of structured and operational expert knowledge systems are extracted and converted into the rules and logic required to train the expert model. On this basis, a set of intelligent knowledge models for this scenario is constructed, which can not only support compliance judgments in actual operations, but also provide scientific and reasonable decision-making support.

[0046] Taking the timing of the second lining construction as an example, whether the timing of the second lining construction is in compliance with regulations requires judging the safety step distance, the consistency of the second lining construction time, the time sequence, and the comparison of the surrounding rock deformation measurement time. Here is a brief description of some knowledge content. The surrounding rock deformation should be stable, the horizontal convergence is less than 0.2mm / d, the arch sinking speed is less than 0.15mm / d, and the distance between the arch and the heading face in the safety step distance should not be greater than 90m for the surrounding rock of grades I, II, and III, and should not be greater than 35 for the surrounding rock of grades IV, V, and VI. The recording time of the construction log should correspond to the surrounding rock measurement time and the concrete production time. According to the scene content, a knowledge model for solving the scene problem is constructed. These knowledge models are the main steps and ideas for solving the scene problem.

[0047] Dataset preparation: The present invention divides the model knowledge learning content into two parts. One is the basic concept relationship in the field of railway engineering. The general large model of this part of data is already available; the other is the railway engineering application scenario knowledge. The specific step is to convert the knowledge model into a model training data set, and through training, the model can learn knowledge. The present invention constructs a multi-round dialogue instruction data set. The instructions include instruction, input, output, and history, which correspond to instructions, input, output, and historical dialogues respectively. The knowledge models corresponding to different scenarios are prepared into different example data.

[0048] The sample data for judging the timing of secondary lining construction are as follows:

[0049] {

[0050] "instruction": "You are a very rigorous expert in the field of railway engineering. The following is the process you need to master in the scenario of determining the timing of secondary lining construction."

[0051] "input": "This is the data analysis result of ** tunnel entrance main tunnel DK*70+285, which determines whether ** tunnel entrance main tunnel DK*70+285 meets the timing of second lining construction. The section (surrounding rock level) measuring points (lines) in the surrounding rock monitoring data file of ** tunnel entrance main tunnel DK*70+285 are DK*70+285 (Ⅳ) GD00 (monitoring type is arch settlement) and DK*70+285 (Ⅳ) SL01-SL02 (monitoring type is peripheral convergence), and the buried mileage is DK*70+285. The measured deformation of the surrounding rock is 0.3mm / d, which is required to be less than 0.2mm / d, which does not meet the timing of construction."

[0052] "output": "Through data analysis, it is known that the surrounding rock convergence and arch sinking speed of the ** tunnel entrance main tunnel DK*70+285 do not meet the requirements of the specification, and the secondary lining of the ** tunnel entrance main tunnel DK*70+285 has not been constructed."

[0053] “history”: [ [

[0056] "** Tunnel entrance main tunnel DK*70+285, is the timing of secondary lining construction in compliance with the specifications?"

[0057] "The work point to be queried is the main tunnel, the location is ** tunnel DK*70+285, the scene is the timing of the second lining construction, and the content to be queried has corresponding mileage information query data including mileage, work point, construction log time, surrounding rock type, and surrounding rock measurement records. Construction timing requirements: Based on the keyword "second lining construction timing", determine the standard timing requirements for this construction stage that need to be retrieved, extract the safe step distance, construction condition information, production ledger information, construction second lining front end mileage, face end mileage, and invert front end mileage."

[0058] ], [

[0060] "You need to query the relevant information of ** tunnel entrance main tunnel DK*70+285. You need to instruct the content-aware model to generate SQL, Cypher and other query instructions."

[0061] "You need to query all the surrounding rock measurement data of the ** tunnel entrance main tunnel DK*70+285, query all the data in the construction log, and query all the data in the production ledger. Query all the 2-neighbor relationship node information of the ** tunnel entrance main tunnel DK*70+285. Query the production specification information."

[0062] ], [

[0064] “This is the query statement for the ** tunnel entrance main tunnel DK*70+285 information query: SELECT * FROM surrounding rock measurement WHERE location = DK*70+285;

[0065] SELECT * FROM construction_log WHERE location = DK*70+285;

[0066] SELECT * FROM production ledger WHERE location = DK*70+285;

[0067] MATCH (n {name: "DK*70+285"})-[:RELATION*1..2]-(neighbor)

[0068] RETURN neighbor;

[0069] Production specification: [0.2..,0.5..,…,0.9..] (1024-dimensional vector)

[0070] You instruct the hybrid search agent to determine which databases to use for hybrid searches."

[0071] "We will use the relational database, graph database, and vector database for retrieval, and pass the query content to these three databases respectively."

[0072] ], [

[0074] “This is the surrounding rock measurement data table retrieved from the relational database:

[0075] Measurement time: 2024-05-24

[0076] Measured value: 3319.4562

[0077] Single deformation: 0.3

[0078] Cumulative deformation: 5.4

[0079] (The actual data is a data table, only one copy of the data is retained here for illustration)

[0080] It is necessary to generate instructions and pass them to the chart generation agent to generate BI charts."

[0081] “Pass the surrounding rock measurement data information to the chart generation agent to generate BI charts” ] [

[0084] "This is the data retrieved from the graph database, relational database, and vector database. You need to generate analysis instructions and pass them to the data analysis intelligent body. **The section (surrounding rock level) measuring points (lines) in the surrounding rock monitoring data file of the DK*70+285 surrounding rock at the entrance of the tunnel are DK*70+285 (Ⅳ) GD00 (monitoring type is arch subsidence) and DK*70+285 (Ⅳ) SL01-SL02 (monitoring type is peripheral convergence), the buried mileage is DK*70+285, and the measured deformation of the surrounding rock is 0.3mm / d."

[0085] "The timing of the second lining construction needs to determine the safe step distance, the consistency of the second lining construction time, the time sequence, and the comparison of the surrounding rock deformation measurement time. The surrounding rock deformation should be stable, the horizontal convergence should be less than 0.2mm / d, the arch sinking speed should be less than 0.15mm / d, and the distance between the arch and the tunnel face in the safe step distance should not be greater than 90m for the surrounding rock of grades I, II, and III, and should not be greater than 35 for the surrounding rock of grades IV, V, and VI. The recording time of the construction log should correspond to the surrounding rock measurement time and concrete production time.

[0086] This is the data. In the DK*70+285 surrounding rock monitoring data file of the entrance of the tunnel, the cross-sectional (surrounding rock level) measuring points (lines) are DK*70+285 (Ⅳ) GD00 (monitoring type is arch subsidence) and DK*70+285 (Ⅳ) SL01-SL02 (monitoring type is peripheral convergence), the buried mileage is DK*70+285, and the measured deformation of the surrounding rock is 0.3mm / d”

[0087] ]]}

[0088] Expert model training: The expert model of the present invention uses the LLM (Large Language Models) model. The MOE expert model in the LLM model is suitable for processing larger knowledge domains. Therefore, the present invention adopts the ordinary LLM model, but improves the training method, adopting MixLora, a mixed low-rank training method that is more suitable for processing small-scale knowledge domains, and is more suitable for railway engineering scenarios.

[0089] The present invention designs the FFN feedforward network layer of LLM as 8 Lora-processed FFN network layers. Specifically: a gated route is designed for the output after Self-Attention, that is, the input of FFN, and two optimal FFN layers are selected for the output after routing. The FFN layers are processed by Lora, and finally the outputs of the two FFNs are obtained by Concat. The function of the gated route is to select the Lora network layer that is most suitable for solving the input problem. Because different scenarios in the field of railway engineering require different data for analysis and processing, the present invention uses a threshold filtering indicator matrix to solve this problem. The LLM model is relatively common, and only the principles of the algorithm optimization content are introduced:

[0090] Gated routing: Because there is a certain degree of overlap in the knowledge of various scenarios in the field of railway engineering, the gated routing adopts a sparse attention gating design.

[0091] Step 1, attention coefficient for

[0092] ;

[0093] in, A query matrix is ​​used to ask questions or requests; is the key matrix, which is equivalent to the "feature label" or "index" of each word; The value matrix represents the information content that each word will eventually convey.

[0094] Step 2, through the mask matrix Control the position to be calculated, set the scores that are not within the sparse structure to infinitesimal or zero, and ignore these positions. Through the residual connection, the final coefficient attention coefficient is

[0095] ;

[0096] The matrix will be assigned 0 at the positions defined by the sparse structure and a negative maximum value at other positions so that These locations are ignored.

[0097] In step 3, the gating mechanism uses an independent linear transformation to generate a gating matrix to control the attention output and give the network the ability to select information flow.

[0098] ;

[0099] in, is the gating matrix, is the activation function, is the weight matrix, is the input matrix, is the bias term.

[0100] Step 4, Make information choices. is the threshold filter indicator matrix, is the gating matrix after filtering.

[0101] ;

[0102] Step 5, finally get the gated coefficient attention mechanism as the input of the Lora network layer.

[0103] ;

[0104] Lora network layer: The formula of Lora low-rank adaptive algorithm is:

[0105] ;

[0106] in is the original FFN layer weight matrix, The matrix is ​​the Lora layer weight matrix, is the weight coefficient. During model training, set the Lora network layer weight matrix The rank size is 8, The coefficient is 0.5, and three rounds of training are set.

[0107] Multi-agent collaboration based on expert models: Multi-agents include five core agent groups: railway engineering content perception, railway engineering hybrid retrieval, railway engineering diagram generation, railway engineering data analysis, and railway engineering language question and answer. The expert model will collaborate with these agents to handle analysis problems. The dialogue information input by the user first passes through the expert model to form prompt instructions. The content-aware agent generates query statements based on the prompt instructions and query content input by the expert model. The query statements are returned to the expert model, and the expert model calls the hybrid retrieval agent for retrieval. The hybrid retrieval agent will return the generated results to the expert model. Based on the dialogue information and retrieval results, the expert model generates prompt instructions for the relevant query results and passes them to the diagram generation agent and the data analysis agent. Finally, the results of these two agents are input to the expert model. The expert model will generate prompt instructions and pass them to the question and answer agent to finally get the answer to the question. The implementation methods of each agent group are as follows:

[0108] 1) Railway Engineering Content-Aware Agent Group:

[0109] The main body of the railway engineering content-aware agent is mainly composed of a large model applied to NL2Sql and NL2Cypher and an embedding coding model, including an input module, a model coding tool, and an output module. The input module accepts instructions from the expert model and user input, and the model coding tool includes three model tools: NL2Sql, NL2Cypher, and embedding, which can convert content into query statements at the three levels of relationship, graph, and vector respectively.

[0110] The railway engineering content-aware agent group includes scene agents such as work-site mileage content-aware agents, process-aware agents, and concrete information-aware agents. These agents can perceive the input content and convert it into query statements. Taking the work-site mileage content-aware agent as an example, this agent can convert the mileage information involved in the question into relevant query statements. The expert model will form instructions based on the input, and the content-aware agent will convert the query into relevant query statements based on the prompt word instructions.

[0111] Example: "**Tunnel entrance main tunnel DK*70+285, what is the timing of the second lining construction, and whether it complies with the specifications" (the following intelligent agents all use this problem as an example). The domain expert model will form a prompt word: "According to the problem, the following two points must be determined: Work point positioning: Use "DK*69+405" to locate the specific construction location, and call the corresponding mileage information query data including mileage, work point, construction log time, surrounding rock type, and surrounding rock measurement records. Construction timing requirements: Based on the keyword "second lining construction timing", determine the need to retrieve the standard timing requirements for this construction stage, extract the safe step distance, construction condition information, production ledger information, construction second lining front mileage, face end mileage, and invert front mileage." The content-aware intelligent experience will convert this prompt word and query content into relevant query statements.

[0112] Some of the query statements in the example are:

[0113] SELECT * FROM surrounding rock measurement WHERE location = DK*70+285;

[0114] SELECT * FROM construction_log WHERE location = DK*70+285;

[0115] SELECT * FROM production ledger WHERE location = DK*70+285;

[0116] MATCH (n {name: "DK*70+285"})-[:RELATION*1..2]-(neighbor)

[0117] RETURN neighbor;

[0118] Production specification: [0.2..,0.5..,…,0.9..] (1024-dimensional vector)

[0119] 2) Railway Engineering Hybrid Retrieval Agent Group:

[0120] The hybrid retrieval agent is implemented by the retrieval tool, which accepts the input processed by the expert model and performs retrieval. The retrieval tool includes an input processing module, a retrieval module, and an output module. Input processing module: receives the processing results from the expert model and the content-aware agent. Retrieval module: implements different types of retrieval methods through a variety of tools, including vector retrieval tools, graph retrieval tools, and relational database tools. Output module: generates retrieval results based on the return results of the tool.

[0121] The railway engineering hybrid retrieval agent group includes the secondary lining construction hybrid retrieval agent, the concrete information hybrid retrieval agent, etc. The data storage carriers are different. The expert model determines the query content and inputs the output of the content-aware agent into the hybrid retrieval agent. The agent will search for relevant content according to the query statement and instructions. In the example, after the hybrid retrieval agent gets the above retrieval instructions, it can retrieve mileage content, surrounding rock content, construction date and other data in the graph database. These data are in Json format. It can search for data descriptions such as tunnel construction safety step requirements and surrounding rock construction requirements in the vector database. These data are also in Json format. It can search for surrounding rock measurement records in the relational database. These data are in table form. After the retrieval is completed, the expert model obtains these results feedback and inputs the data information into the chart generation agent or data analysis agent according to the problem understanding.

[0122] 3) Railway Engineering Graph Generation Agent Group:

[0123] The railway engineering diagram generation agent group is implemented by the diagram generation tool. It will accept the results of the expert model and the hybrid search to generate the diagram, including the input processing module, the diagram generation module, and the output module. Input processing module: receives the processing results from the expert model and the hybrid search agent. Diagram generation module: responsible for tabulation. Output module: generates the diagram according to the return results of the tool.

[0124] The railway engineering chart generation agent group includes surrounding rock measurement agents, concrete maintenance record agents, etc. The agent can be composed of chart generation components. The expert model will indicate which data needs to generate charts based on the query content, and then the agent group generates the chart. After the expert model obtains the output results of the hybrid retrieval agent, it will input its results into the chart generation agent and the data analysis agent. The selection is based on the judgment of the expert model. In the example, the expert model will instruct the agent to input the surrounding rock measurement data into the chart generation agent to generate a chart.

[0125] 4) Railway Engineering Data Analysis Agent Group:

[0126] The railway engineering data analysis agent group is composed of a large semantic understanding model, which can analyze whether the scenario is reasonable based on the prompt information and data of the expert model. It includes input processing module, data analysis module and output module. Input processing module: receives the processing results from the expert model and the hybrid retrieval agent. Data analysis module: responsible for data proofreading. Output module: returns the analysis results.

[0127] The railway engineering data analysis agent group includes concrete data analysis agents, construction operation analysis agents, etc. The expert model will input the query rules and corresponding data generation instructions to the agent, and the agent will understand and analyze. Taking the example as an example, the query rules include judging the safe step distance, the consistency of the second lining construction time, the time sequence, and the comparison of the surrounding rock deformation measurement time. These contents are the query results obtained from the vector library by the hybrid retrieval agent. The expert model will form prompt instructions for these rule information and the corresponding data information and input them to the data analysis agent.

[0128] 5) Railway Engineering Language Question Answering Agent Group:

[0129] The railway engineering language question-answering agent group is composed of a large semantic understanding model, which can analyze whether the scenario is reasonable based on the prompt information and data of the expert model. It includes input processing module, semantic understanding module, and output module. Input processing module: receives the processing results from the expert model and data analysis agent. Semantic understanding module: responsible for content understanding. Output module: returns the analysis results. The expert model will combine the analysis results into prompt words, input them into the question-answering agent, and the question-answering agent will generate a reply.

[0130] like Figure 2 As shown, according to another embodiment of the present invention, a multi-agent collaboration system 200 for engineering construction projects is also disclosed, including: a memory 201, for storing program instructions; a processor 202, for calling the program instructions stored in the memory to implement the multi-agent collaboration method for engineering construction projects as in the above-mentioned embodiment.

[0131] All or part of the steps in the various methods of the above embodiments can be completed by controlling the relevant hardware through a program, and the program can be stored in a readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other readable medium that can be used to carry or store data.

[0132] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-agent collaboration method for engineering construction projects, characterized in that: include: Train the expert model in the railway engineering field: Build expert knowledge system documents based on the construction knowledge and specifications related to the engineering construction project; Constructing a railway engineering field expert model training data structure according to the expert knowledge system document to obtain a thinking chain knowledge data set; The railway engineering domain expert model is trained according to the thought chain knowledge data set; wherein the railway engineering domain expert model is a large language model, and the large language model is trained based on MixLora; the steps of training the large language model specifically include: The gating routing of MixLora adopts sparse attention gating: The attention coefficient is: ； in, represents the query matrix, represents the key matrix, represents the value matrix, represents the transposed matrix of K, is the dimension of the bond matrix; The positions to be calculated are controlled by the mask matrix, and the scores that are not within the sparse structure are set to infinitesimal or zero, thereby ignoring these positions; through the residual connection, the final attention coefficient is obtained: ； Mask Matrix The sparse structure defines the position where the value is 0, and the other positions are assigned a negative maximum value so that these positions are ignored in the Softmax function; A linear transformation is used to generate a gating matrix to control the attention output and give the network the ability to select information flow: ； in, is the gating matrix, is the activation function, is the weight matrix, is the input matrix, is the bias term; To select information from the gating matrix: ； in, is the threshold filter indicator matrix, is the gating matrix after filtering; Gated coefficient attention mechanism: ； The output of the gated coefficient attention mechanism is used as the input of the Lora network layer: Lora network layer: ； in, is the original FFN layer weight matrix, B matrix and A matrix are the Lora layer weight matrices, is the weight coefficient; During model training, the rank of the B matrix and the A matrix are set to 8. The coefficient is 0.5; Receiving natural language input: the dialogue information input by the user is processed by the railway engineering domain expert model to generate prompt instructions; Content-aware: the content-aware agent generates a query statement according to the prompt instruction, and returns the query statement to the railway engineering domain expert model; Hybrid retrieval: the railway engineering domain expert model calls the hybrid retrieval agent to perform retrieval according to the query statement, and the hybrid retrieval agent returns the retrieval result to the railway engineering domain expert model; Chart generation: the railway engineering domain expert model generates a chart making request according to the dialogue information and the search results, so that the chart generating agent generates a chart according to the chart making request and returns it to the railway engineering domain expert model; Data analysis: the railway engineering domain expert model generates a data analysis request according to the dialogue information and the search results, so that the data analysis agent performs data analysis according to the data analysis request and returns the analysis result to the railway engineering domain expert model; Language question answering: The railway engineering domain expert model generates prompt instructions based on the retrieval results, the charts and the analysis results, so that the language question answering agent integrates and outputs the retrieval results, the charts and the analysis results.

2. The multi-agent collaboration method for engineering construction projects according to claim 1 is characterized in that: The content-aware agent is implemented by a large model applied to NL2Sql, a large model applied to NL2Cypher, and an embedding encoding model, and converts natural language queries into query statements at three levels: relation, graph, and vector according to the prompt instructions.

3. The multi-agent collaboration method for engineering construction projects according to claim 2 is characterized in that: The content-aware intelligent agents include the following types: work site mileage content-aware intelligent agents, process-aware intelligent agents, and concrete information-aware intelligent agents.

4. The multi-agent collaboration method for engineering construction projects according to claim 1, characterized in that: The hybrid retrieval agent is implemented by a vector retrieval tool, a graph retrieval tool and a relational database tool, and searches for relevant content according to the instructions of the railway engineering field expert model and the query statement; the hybrid retrieval agent includes the following types: a secondary lining construction hybrid retrieval agent and a concrete information hybrid retrieval agent.

5. The multi-agent collaboration method for engineering construction projects according to claim 1, characterized in that: The chart generation agent is implemented by a chart generation tool, and generates a chart according to the chart generation request and the search result; the chart generation agent includes the following types: surrounding rock measurement chart agent, concrete maintenance record agent.

6. The multi-agent collaboration method for engineering construction projects according to claim 1, characterized in that: The data analysis agent is implemented by a large semantic understanding model, and includes the following types: concrete data analysis agent and construction operation analysis agent.

7. The multi-agent collaboration method for engineering construction projects according to claim 1, characterized in that: The language question-answering agent is implemented by a large semantic understanding model, and integrates the retrieval results, the charts and the analysis results according to the prompt instructions of the expert model to form a question-answering result.

8. A multi-agent collaborative system for engineering construction projects, characterized in that: include: A memory for storing program instructions; A processor, configured to call the program instructions stored in the memory to implement the multi-agent collaboration method for engineering construction projects as described in any one of claims 1 to 7.

Images