Shield tunneling risk assessment method and system based on distributed intelligent agent

Abstract

The invention provides a shield tunneling risk assessment method and system based on a distributed intelligent agent, and belongs to the technical field of shield tunneling. The assessment system comprises a task planning agent, a standard query agent, a parameter prediction agent and a risk assessment agent, and combines semantic understanding and decision-making ability of a large language model with professional knowledge and a quantitative calculation model in the field of shield engineering by constructing a plurality of agents with definite functions and cooperative work. And low-threshold, high-precision and verifiable evaluation of the shield tunneling risk is realized.

Description

Technical Field

[0001] This invention belongs to the field of tunnel boring machine (TBM) technology, specifically relating to a TBM risk assessment method and system based on distributed intelligent agents. Background Technology

[0002] Tunnel boring machines (TBMs) are core construction equipment in urban rail transit, cross-river and cross-sea tunnel projects. Their construction typically takes place in urban built-up areas or complex geological environments, where the surrounding environment is sensitive and operating conditions are highly variable. During TBM tunneling, factors such as ground disturbance and improper control of construction parameters can easily induce engineering accidents such as surface subsidence, collapse, water and sand inrush, and even TBM burial, posing serious risks to project safety and the surrounding environment. Therefore, how to effectively assess and control the risks during TBM tunneling has become an urgent technical problem to be solved in the field of underground space engineering.

[0003] Existing methods for risk assessment in tunnel boring machines (TBMs) primarily rely on expert experience or rule-driven models, such as Bayesian network-based risk assessment methods. These methods typically require manual identification of risk factors and their causal relationships, and depend on expert experience to construct conditional probability tables. This results in significant subjectivity, limited model generalization ability, and difficulty adapting to diverse engineering conditions. On the other hand, some studies employ data-driven quantitative models, such as neural networks, to predict TBM construction parameters or risk indicators. These models can provide more accurate quantitative calculations, but they often require a high level of professional expertise, making them difficult for frontline workers to use in actual engineering projects.

[0004] In recent years, the rapid development of large language model technology has provided a new technical path for the intelligentization of engineering risk assessment. Through natural language interaction, large language models can lower the barrier to entry for system use and, to some extent, assist in construction decision-making. However, existing general-purpose large language models are typically trained on general-domain corpora, and still have significant limitations in vertical engineering fields such as tunnel boring machine (TBM) construction, mainly in the following two aspects: First, the general language model lacks systematic and authoritative professional knowledge support in the field of tunnel boring machine (TBM) construction. When answering questions about TBM construction specifications and parameter control standards, it is easy to make unfounded inferences or even model illusions. In particular, when it comes to specific engineering values, there is a risk of randomly generating or fabricating results.

[0005] Secondly, large language models are essentially text generation models based on probability distributions. They lack the ability to rigorously model engineering physical constraints, mathematical calculation rules, and numerical error propagation processes, making it difficult to independently complete accurate and verifiable quantitative calculations and risk assessments.

[0006] Therefore, existing technologies have not yet formed a technical solution that can effectively combine the natural language understanding and decision-making capabilities of large language models with the quantitative calculation models in the tunnel boring process. It is difficult to achieve low-threshold, intelligent tunnel boring risk assessment and prevention while ensuring the accuracy and verifiability of calculations. Summary of the Invention

[0007] This invention provides a method and system for risk assessment of tunnel boring machines (TBMs) based on distributed intelligent agents. By constructing multiple intelligent agents with clearly defined functions and working collaboratively, the semantic understanding and decision-making capabilities of large language models are combined with professional knowledge and quantitative calculation models in the field of TBM engineering. This enables a low-threshold, high-precision, and verifiable assessment of TBM tunneling risks, thereby effectively solving at least one of the technical problems mentioned in the background art.

[0008] To solve the above-mentioned technical problems, the present invention is implemented as follows: A shield tunneling risk assessment system based on distributed intelligent agents includes: The task planning agent is used to analyze user questions, automatically determine the type of downstream agent to be called based on the question content, and construct the corresponding execution path; The standardized query agent is used to respond to the call requests in the execution plan. It uses retrieval-enhanced generative technology to learn the reserve knowledge related to the user's question from the knowledge vector database, providing knowledge supplementation and reasoning constraints for the reasoning process of other agents. The parameter prediction agent is used to encapsulate various prediction models with different functions and applicable conditions in the form of a tool. Based on the call requirements of the execution plan, it calls the prediction model related to the user problem, uses the construction parameters required by the prediction model provided by the user problem or local data source as the basis for calculation, and outputs parameter prediction results including predicted values ​​and prediction uncertainties. The risk assessment agent is used to calculate the risk coefficient and assess the risk level based on the prediction results of the parameter prediction agent and the control threshold of the risk index. If the risk level exceeds the preset limit, the agent generates the corresponding adjustment plan for the construction parameters and feeds it back to the parameter prediction agent, which then recalculates the adjusted risk index prediction results.

[0009] As a preferred improvement, the task planning agent, the standard query agent, the parameter prediction agent, and the risk assessment agent are each independently configured with a large language model. Through prompt word engineering technology, the role, functional boundaries, workflow, and interaction methods with other agents are defined for each large language model.

[0010] As a preferred improvement, all data within the system is transmitted in JSON format.

[0011] As a preferred improvement, during the internal data transmission process of the system, a numerical data extraction unit is set up on the basis of transmitting the original text. The unit extracts relevant parameters from the text through predefined keywords and forms a structured numerical data file. The numerical data is transmitted to the target intelligent agent together with the original text.

[0012] As a preferred improvement, the task planning agent generates a structured JSON-formatted execution plan based on three fields: "problem analysis," "required downstream agents," and "reason for selection," and packages the user problem and execution plan to send to the corresponding downstream agents.

[0013] As a preferred improvement, the specification query agent segments the text in shield tunneling construction standards, specifications, and technical documents. Using an embedding model, it converts the segmented text fragments into corresponding text vectors, forming a knowledge vector database for the shield tunneling field. When a user asks a question, the embedding model converts the user's question into a question vector, and the question vector is compared with the text vectors in the knowledge vector database. After sorting by similarity, a preset number of highly similar text fragments are retrieved. The retrieved text fragments, as contextual information, are input along with the user's question into the large language model built into the specification query agent to generate a question-and-answer result based on the specification, which is then fed back to the upstream and downstream agents that sent the query request.

[0014] As a preferred improvement, the parameter prediction agent encapsulates various prediction algorithms in the field of tunnel boring machines into callable computing modules, and imports them into the agent framework in binary extended form to form prediction models. Through prompt word engineering, the functions and applicable conditions of each prediction model are predefined, enabling the large language model built into the parameter prediction agent to understand the applicable scenarios and input / output requirements of different prediction models. When a user question is passed to the parameter prediction agent, the large language model built into the parameter prediction agent combines the user question content, agent prompt words, and prompt words of each prediction model to automatically select the prediction model that matches the current question and trigger the call of the corresponding prediction model.

[0015] As a preferred improvement, the prediction model includes a calculation model for stratum identification, surface settlement prediction, soil chamber pressure prediction, and shield attitude prediction; the risk indicators include surface settlement exceeding limits, active instability risk of the excavation face, passive instability risk of the excavation face, shield head horizontal deviation exceeding limits, shield head vertical deviation exceeding limits, shield tail horizontal deviation exceeding limits, shield tail vertical deviation exceeding limits, shield tail gap exceeding limits, roll angle exceeding limits, and pitch angle exceeding limits.

[0016] As a preferred improvement, in the iterative process of parameter prediction and risk assessment, the risk assessment agent regards the multiple risk coefficients corresponding to each candidate parameter combination as a multi-objective optimization problem. In each iteration, it selects the Pareto optimal solution from the candidate parameter combinations and fine-tunes the construction parameters based on the current Pareto optimal solution, continuously iterating and seeking optimization until all risk coefficients are lower than the corresponding threshold.

[0017] An assessment method using the aforementioned distributed agent-based shield tunneling risk assessment system includes the following steps: Step S1: Analyze the user's question, automatically determine the subsequent steps to be executed based on the question content, and construct the corresponding execution path; Step S2: Based on the call request in the execution plan, the retrieval enhancement generation technology is used to learn the reserve knowledge related to the user's question in the knowledge vector database, so as to provide knowledge supplementation and reasoning constraints for the reasoning process in subsequent steps; Step S3: Encapsulate various prediction models with different functions and applicable conditions in the form of tools. Based on the call requirements of the execution plan, call the prediction model related to the user problem. Use the construction parameters required by the prediction model provided in the user problem or local data source as the calculation basis, and output the parameter prediction results including the predicted value and the prediction uncertainty. Step S4: Based on the prediction results of step S3 and the control threshold of the risk indicators, calculate the risk coefficient and assess the risk level. If the risk level exceeds the preset limit, generate the corresponding construction parameter adjustment plan and return to step S3 to re-predict the parameters.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) Through the task planning agent, users only need to ask questions in natural language, and the system can automatically break down complex tasks and plan the execution process. Users do not need to have in-depth algorithm or modeling knowledge, which greatly reduces the threshold for using the professional risk assessment system. Through the standard query agent, the retrieval enhancement generation technology is introduced, which provides authoritative engineering standard basis for the answer of the large language model, effectively suppressing the "illusion" problem that is common in the professional field of general large language models, and ensuring the traceability and accuracy of the conclusions. Through the parameter prediction agent, the semantic understanding and decision-making ability of the large language model is seamlessly integrated with the precise engineering calculation model. Through the "prediction-evaluation-adjustment" closed-loop mechanism constructed by the risk assessment agent, the system can actively iterate and optimize the construction parameters, and develop from passively assessing risks to actively finding safe and efficient combinations of construction parameters, providing strong scientific support for on-site decision-making. (2) By adopting a distributed multi-agent architecture, different functional modules are decoupled. Compared with a single large model, this architecture effectively avoids functional confusion, reduces the computational load of a single module, makes the system easier to maintain, upgrade and expand new functions, and improves the stability and adaptability in complex engineering environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This diagram illustrates the architecture of the shield tunneling risk assessment system based on distributed intelligent agents provided by this invention. Figure 2 This diagram illustrates the execution flow of a task planning agent. Figure 3 A schematic diagram illustrating the execution flow of a canonical query agent; Figure 4 A schematic diagram illustrating the execution flow of a parameter prediction agent; Figure 5 This is a schematic diagram illustrating the execution flow of a risk assessment agent. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figures 1-5 As shown, this embodiment provides a shield tunneling risk assessment system based on distributed intelligent agents, including a task planning intelligent agent, a specification query intelligent agent, a parameter prediction intelligent agent, and a risk assessment intelligent agent. Each of the task planning intelligent agent, specification query intelligent agent, parameter prediction intelligent agent, and risk assessment intelligent agent is independently configured with a large language model. Through sophisticated prompt word engineering technology, the role, functional boundaries, workflow, and interaction methods with other intelligent agents are defined for each large language model.

[0022] Data interaction and transmission between agents are conducted via a predefined communication protocol. To ensure the accuracy and consistency of data transmission, all numerical data is structured in JSON format during transmission. Based on the transmitted original text, a numerical data extraction unit is set up to extract relevant parameters from the text using predefined keywords, forming a structured numerical data file. This numerical data, along with the original text, is transmitted to the target agent, effectively avoiding data transmission errors caused by biases in the model's attention mechanism and ensuring the accuracy and reliability of data transfer between agents.

[0023] All agents are developed based on the LangGraph framework, and each agent connects to the DeepSeek-V3.2 model. The differences between agents lie in the functional scope of the prompt words and the tools they connect to. The task planning agent does not need to be connected to other tools; it can complete tasks using its built-in large language model and prompt words. The tool used by the canonical query agent is RAG (Retrieval Augmented Generation), which is used to perform canonical query tasks; The tool accessed by the parameter prediction agent is a machine learning prediction model, which is used to perform prediction tasks with different parameters; The tools accessed by the risk assessment agent are the surface settlement exceeding risk function, the excavation face instability risk function, and the shield attitude exceeding function, which are used to calculate various risk coefficients, as well as the NSGA-Ⅲ genetic optimization algorithm, which is used to optimize construction parameters.

[0024] The task planning agent, as the starting link of the task chain, is used to interact with the user, receive questions input by the user in natural language, parse the user's questions, automatically determine the type of downstream agent to be called based on the question content, and construct the corresponding execution path.

[0025] To ensure the smooth construction of the execution path, the roles and functions of the task planning agents need to be predefined using prompt word engineering techniques before the system goes live. The prompt words also need to embed the functional descriptions and invocation constraints of each downstream agent. The functional descriptions and invocation constraints of the downstream agents are as follows: (1) Standard query agent: Its function is described as "retrieving relevant content in the knowledge vector database based on the user's standard query question".

[0026] (2) Parameter prediction agent: Its function is described as "based on historical tunneling data and geological parameters, predict the current stratum level in front of the excavation face and future soil pressure, shield machine attitude and other state parameters"; its calling constraint is "triggered when the user problem involves a clear prediction requirement".

[0027] (3) Risk assessment agent: Its function is described as "combining real-time tunneling status or relevant parameters obtained by prediction, and using a preset risk calculation function to assess the risks of excessive ground settlement and instability of the excavation face"; its calling constraint is "triggered when the user's problem involves a clear risk judgment requirement, and must be called after the prediction agent outputs a valid result, or triggered when the user directly provides working condition data with potential risks".

[0028] Upon receiving a user's question, the task planning agent's built-in large language model, leveraging its semantic understanding and intent recognition capabilities, extracts key information from the question (such as engineering parameters and task objectives). This extracted key information is then semantically matched with the pre-defined functional descriptions of downstream agents within the task planning agent's prompts. Based on the matching confidence index, the type of downstream agent to be invoked is determined, and the corresponding execution path is constructed. Specifically, this includes the following processes: (1) Intent and entity extraction: Utilize the natural language understanding capability of the large language model built into the task planning agent to parse the text input by the user and extract the operation intent (such as keywords such as state prediction and risk level) and key parameters (such as thrust parameters, geological condition characteristics, and current ring number). (2) Semantic Feature Mapping: The predefined prompts for the task planning sub-agent contain the standard functional descriptions, input requirements, and output formats of each downstream agent. The large language model built into the task planning agent extracts the operational intent and key parameters and performs matching confidence calculations with the functional descriptions of each agent within the prompts. Using text similarity as the metric for this confidence score, the most matching functional description segment is selected through optimization comparison, and its corresponding downstream agent is chosen as the target.

[0029] (3) Limit the output format in the prompt words so that the task planning agent generates a structured JSON execution plan according to the three fields of "problem analysis", "required downstream agent" and "selection reason", and packages the user problem and execution plan and sends them to the corresponding downstream agent (i.e. the downstream agent required by the execution plan).

[0030] The task planning sub-agent can automatically decompose and plan processes for complex engineering problems, avoiding direct user contact with underlying models and algorithm configurations, effectively reducing the system's usage threshold and improving its availability and ease of use in engineering sites.

[0031] The canonical query agent is used to respond to the call request in the execution plan. It uses retrieval-enhanced generation technology to learn the reserve knowledge related to the user's question in the knowledge vector database, providing knowledge supplements and reasoning constraints for the reasoning process of other agents.

[0032] The specification query agent segments the text in shield tunneling construction standards (national standards, local standards, or group standards), specifications, and technical documents. Using an embedding model, it converts the segmented text fragments into corresponding text vectors, forming a knowledge vector database for the shield tunneling field. When a user asks a question, the embedding model converts the user's question into a question vector. The question vector is then compared with the text vectors in the knowledge vector database to calculate vector similarity. A preset number of highly similar text fragments are retrieved based on similarity. These retrieved text fragments, along with the user's question, are input into the large language model built into the specification query agent to generate query results based on the specifications. These results are then fed back to other agents that require query results.

[0033] The aforementioned standard query agent significantly improves the basis and consistency of answers to professional questions in the shield tunneling field by introducing authoritative standard knowledge as the retrieval basis, and reduces the risk of unfounded inferences or model illusions in the shield tunneling construction standard question and answer scenario by using a general large language model.

[0034] The parameter prediction agent is used to encapsulate various prediction models with different functions and applicable conditions in the form of a tool. Based on the call requirements of the execution plan, it calls the prediction model related to the user problem, uses the construction parameters required by the prediction model provided in the user problem or local data source as the calculation basis, and outputs parameter prediction results including predicted values ​​and prediction uncertainties.

[0035] The parameter prediction agent encapsulates various prediction algorithms in the field of tunnel boring machines into callable computing modules, and imports them into the agent framework in binary extended form to form prediction models. Through prompt word engineering, the functions and applicable conditions of each prediction model are predefined, enabling the large language model built into the parameter prediction agent to understand the applicable scenarios and input / output requirements of different prediction models. When a user question is presented to the parameter prediction agent, the large language model built into the agent combines the user question content, agent prompt words, and prompt words from each prediction model to automatically select the prediction model matching the current question and trigger the invocation of the corresponding prediction model.

[0036] The matching principle of the prediction model is the same as that of the intelligent agent. Both involve semantically matching the key information extracted from the user's question with the functional description and applicable conditions of each prediction model, and determining the type of prediction model to be called based on the matching confidence index.

[0037] Each prediction model supports at least two data input methods during the invocation process: when the user question explicitly provides the construction parameters required by the relevant prediction model, the large language model extracts the corresponding input parameters from the user question and sends them to the prediction model; when the user question does not provide the required construction parameters or specifies the use of local data, the prediction model automatically reads the required construction parameters from a preset local data source.

[0038] The prediction model includes computational models for stratum identification, surface settlement prediction, soil chamber pressure prediction, and shield tunnel attitude prediction. After the prediction model is executed, the predicted values ​​and prediction uncertainty information of the corresponding prediction model are output as outputs for subsequent risk assessment.

[0039] The risk assessment agent is used to calculate the risk coefficient and assess the risk level based on the prediction results of the parameter prediction agent and the control threshold of the risk index. If the risk level exceeds the preset limit, an adjustment plan for the corresponding construction parameters is generated and fed back to the parameter prediction agent, which then recalculates the adjusted risk index prediction results.

[0040] The risk assessment agent uses the output of the upstream parameter prediction agent as input, and comprehensively considers the predicted values ​​of risk indicators, prediction uncertainties, and the control values ​​of construction parameters given in the specifications. It quantitatively calculates ten categories of risk indicators, including risks of excessive surface settlement, active instability of the excavation face, passive instability of the excavation face, excessive horizontal deviation of the shield head, excessive vertical deviation of the shield head, excessive horizontal deviation of the shield tail, excessive vertical deviation of the shield tail, excessive shield tail gap, excessive roll angle, and excessive pitch angle, obtaining the corresponding risk coefficients. Specifically: (1) For the risk of settlement exceeding the limit, the core assessment parameter is the maximum surface settlement, which is predicted using the conventional random forest model in this field. At the same time, the inherent ensemble learning characteristics of the random forest model are used to extract uncertainty. The random forest is composed of multiple independent decision trees. When processing the same set of engineering input features, each decision tree generates an independent settlement prediction value according to its own node splitting rules. These discrete prediction results represent the uncertainty of the random forest model.

[0041] Assume that random forest contains The first decision tree, the The settlement prediction value output by each decision tree is This can be calculated. The mean of the prediction results and standard deviation The final prediction result space of this ensemble model can be fitted to a continuous normal distribution, and the distribution density represents the probability of outputting that predicted value. Based on this, settlement control values ​​are set. The risk of settlement exceeding the limit is defined as 10 mm, and is therefore defined as the mathematical probability that the predicted settlement value exceeds the safety control limit. Under the assumption of a normal distribution, this probability of exceeding the limit is equal to the sum of the areas of the probability density curve outside the positive and negative control values. The specific calculation formula is as follows: ; In the formula, This indicates a risk of settlement exceeding the limit; The cumulative distribution function represents the standard normal distribution.

[0042] In the specific implementation of this risk measurement formula, if the extreme case occurs where all decision tree predictions are completely consistent, it will lead to a calculation anomaly where the denominator is zero. Therefore, a very small constant is added to the standard deviation during the underlying calculation to ensure computational stability.

[0043] (2) For the risk of excavation face instability, the core assessment parameters are the predicted values ​​of active and passive earth pressures on the strata and the earth pressure in the tunnel boring machine's chamber. Among them, active and passive earth pressures represent the ultimate resistance range that the natural soil skeleton can withstand. In the specific calculation, the total weight of the overlying soil layer above the tunnel face is first calculated based on the geological elevation data, and a reduction mechanism related to shear strength is introduced to reduce the original cohesion and internal friction tangent values ​​of each soil layer. Combining the above parameters, the active and passive earth pressures at a specific depth of the excavation face are calculated according to the static equilibrium theory of soil mechanics.

[0044] The predicted pressure values ​​for the earth chamber were obtained using a conventional CNN-GRU model. Simultaneously, the model uncertainty was calculated using the Dropout mechanism in the neural network model. Dropout is a regularization technique to prevent overfitting, which increases the model's generalization performance by randomly shutting down some neurons during training. During the inference prediction phase, some neurons were also randomly shut down for multiple forward propagation operations, each outputting independent predicted pressure values ​​for six measuring points. The maximum standard deviation of the predicted values ​​was calculated and normalized using a preset variance control value to obtain the model prediction uncertainty index. .

[0045] Due to the depth of the excavation, the earth pressure at the bottom of the excavation face is usually greater than that at the top, and this difference is even more pronounced in large-diameter tunnels. Simultaneously, the earth pressure within the earth chamber is affected by the density of the internal medium and gravity, exhibiting a significant vertical pressure gradient; the actual earth pressure at the bottom measuring point is typically much greater than that at the top measuring point. Therefore, the predicted values ​​from the top measuring point of the earth chamber and the passive earth pressure above the excavation face are used to assess the passive instability risk of the excavation face, while the predicted values ​​from the bottom measuring point of the earth chamber and the active earth pressure below the excavation face are used to assess the active instability risk. The specific calculation formulas are as follows: ; ; In the formula, and These represent the active and passive instability risks of the excavation face, respectively. This is a risk amplification factor, which can be adjusted according to different projects; and This indicates the predicted pressure at the bottom and top of the soil chamber by the model; and These represent the active earth pressure at the bottom of the excavation face and the passive earth pressure at the top of the excavation face, respectively.

[0046] In this risk measurement formula, an exponential function is used to amplify the degree of danger of earth chamber pressure approaching the limit. The closer to the active and passive earth pressure, the greater the risk. To unify the magnitude of different risks, if the final excavation face risk value is greater than 1, it is forcibly truncated and set to 1 during the bottom layer calculation.

[0047] (3) Regarding the risk of attitude exceeding limits, the core assessment parameters include six attitude variables: roll angle, pitch angle, shield head horizontal deviation, shield head vertical deviation, shield tail horizontal deviation, and shield tail vertical deviation. The shield tail gap is also considered as an indicator to measure the compression of the tunnel lining structure by the tunnel boring machine. The six attitude components are predicted using the conventional GeoFilter-STL model in this field, while the shield tail gap is calculated from the geometric deviations between the tunnel boring machine's head and tail and the design axis. Specifically, based on the horizontal and vertical deviations of the shield head and tail, the radial displacement of each corresponding point is continuously calculated along the circumference of the shield tail, thereby obtaining the spatial distribution of the shield tail gap. The formula is as follows: ; ; ; In the formula, and These represent the horizontal and vertical relative displacement differences between the front and rear ends of the tunnel boring machine, respectively. Represents the polar coordinate angle along the circumference of the shield tail; This represents the initial standard shield tail clearance; Represents a specific angle on the circumference of the shield tail. Real-time shield tail gap at the location; , , , These represent the horizontal deviation of the shield head, the vertical deviation of the shield head, the horizontal deviation of the shield tail, and the vertical deviation of the shield tail, respectively.

[0048] To quantify model uncertainty, the Dropout mechanism is used to perform multiple independent forward propagations during the prediction process, thereby calculating the mean and standard deviation of the predicted values ​​for each attitude variable. For a single attitude variable, the risk of exceeding limits is the predicted value plus a certain multiple of the standard deviation, divided by the control threshold for that attitude component. The risk of exceeding limits for the shield tail gap is the ratio of the maximum change in the shield tail gap to the maximum allowable gap compression in engineering. The specific quantitative calculation formula is as follows: ; ; In the formula, Indicates the first Risk of exceeding limits for attitude variables These correspond to six attitudes: roll angle, pitch angle, shield head horizontal deviation, shield head vertical deviation, shield tail horizontal deviation, and shield tail vertical deviation, respectively. and These are the mean and standard deviation of the predicted values ​​for the attitude variable, respectively. The uncertainty amplification factor of the representative model; To ensure the safety control values ​​for the corresponding attitude items, the roll angle and pitch angle control values ​​are selected as 5° according to the specifications, and the control values ​​for the four geometric deviations of shield head horizontal deviation, shield head vertical deviation, shield tail horizontal deviation and shield tail vertical deviation are 50mm. This indicates a risk of exceeding the shield tail clearance limit; Minimum shield tail clearance to ensure the safety of the segment structure.

[0049] The risk coefficient is submitted as input to the large language model built into the risk assessment agent for risk discrimination. When any risk coefficient exceeds the preset threshold, the large language model generates multiple sets of construction parameter adjustment schemes based on prompt word rules. At least twenty sets of candidate parameter combinations are preferably generated and the candidate parameter combinations are sent back to the parameter prediction sub-agent for re-prediction.

[0050] During the iterative process of parameter prediction and risk assessment, the risk assessment agent treats the multiple risk coefficients corresponding to each candidate construction parameter combination as a multi-objective optimization problem. In each iteration, it selects the Pareto optimal solution from the candidate construction parameter combinations and fine-tunes the parameters based on the current Pareto optimal solution, continuously iterating and seeking optimization until all risk coefficients are lower than the corresponding thresholds, thereby achieving safe optimization control of shield tunneling parameters.

[0051] The evaluation system provided by this invention consists of four collaborative agents: task planning, specification query, parameter prediction, and risk assessment. Compared to a single-agent architecture, this distributed agent architecture effectively reduces functional confusion and computational load caused by a single agent when there are many tools, improving the overall stability, scalability, and engineering adaptability of the system. Through multi-agent integration, the user-input natural language question first enters the task planning agent for intent analysis and task decomposition. Based on the analysis results, the calling path for downstream agents is determined. Finally, the relevant specification text retrieved by the enhanced generative retrieval technology and the parameter values ​​calculated by the prediction model are returned to the large language model for processing to obtain the final natural language answer.

[0052] This embodiment also provides an assessment method using the above-described shield tunneling risk assessment system based on distributed intelligent agents, comprising the following steps: Step S1: Analyze the user's question, automatically determine the subsequent steps to be executed based on the question content, and construct the corresponding execution path; Step S2: Based on the call request in the execution plan, the retrieval enhancement generation technology is used to learn the reserve knowledge related to the user's question in the knowledge vector database, so as to provide knowledge supplementation and reasoning constraints for the reasoning process in subsequent steps; Step S3: Encapsulate various prediction models with different functions and applicable conditions in the form of tools. Based on the call requirements of the execution plan, call the prediction model related to the user problem. Use the construction parameters required by the prediction model provided in the user problem or local data source as the calculation basis, and output the parameter prediction results including the predicted value and the prediction uncertainty. Step S4: Based on the prediction results of step S3 and the control threshold of the risk indicators, calculate the risk coefficient and assess the risk level. If the risk level exceeds the preset limit, generate the corresponding construction parameter adjustment plan and return to step S3 to re-predict the parameters.

[0053] Example 1 When a user inputs the request "What are the calculation methods and empirical value ranges for tunnel boring machine thrust?", the task planning agent within the risk assessment system first receives the request and performs intent and entity parsing, extracting key entity features such as "tunnel boring machine thrust," "calculation method," and "empirical value." After guiding the user with prompts and calculating the confidence level of the key entity features, the system determines that the request is for definitional and empirical knowledge retrieval, not parameter prediction or risk assessment. Therefore, the task planning agent constructs an execution path of "calling the specification query agent to retrieve the local specification database."

[0054] Subsequently, the specification query agent receives the execution plan and user request, calculates the vector similarity between the key entity features in the user request and the text vectors in the shield tunneling specification vector knowledge base, recalls the top 10 text fragments with the highest similarity, and returns them to the large language model within the specification query agent. The large language model integrates the text fragments based on the recalled text and the user request to form a response in natural language form.

[0055] The specific analysis and processing procedures of the evaluation system are as follows: User issue: What are the calculation methods and empirical ranges for shield tunneling thrust? System response: Task planning agent: Problem Analysis: The user is asking about the calculation method and empirical range of shield thrust. This is a knowledge-based question about the calculation method and empirical values ​​of construction parameters, and does not involve the calculation, prediction or risk assessment of specific ring numbers.

[0056] Required downstream intelligent agent: Standardized query intelligent agent.

[0057] Rationale for selection: The core of the user's question is to seek an explanation of the professional concept of "tunnel boring machine thrust," including its calculation methods (such as theoretical formulas and factors to be considered) and empirical value ranges in actual engineering. This falls under the category of retrieving and answering definitional and empirical knowledge, and does not involve predictive calculations or risk assessments based on specific data. Therefore, it is only necessary to use a standardized query agent to retrieve relevant information from the knowledge vector base to answer the user's question.

[0058] Standardized query agent:

Conclusion

[0059]

in accordance with

[0060] Original text: The empirical calculation method calculates the equipment thrust according to formula (B.7): In the formula: D is the equipment thrust; D is the cutterhead excavation diameter; P is the empirical thrust per unit excavation face. The thrust of an earth pressure balance shield machine is generally 1000~1300 kN / m². "Source": GB / T 34651-2017 Full-face tunnel boring machine earth pressure balance shield machine.

[0061] Original text: The shield thrust F is calculated using the empirical method according to formula (B.7): In the formula: D is the cutterhead excavation diameter; P is the empirical thrust per unit excavation surface. The thrust of a slurry balance shield machine is generally taken as 1000~1400 kN / m². "Source": GB / T 35019-2018 Full-face tunnel boring machine slurry balance shield machine.

[0062] Example 2 When a user inputs the request, "Please predict the settlement value at the current location based on local parameters," the task planning agent within the risk assessment system first receives the request and performs intent and entity parsing, extracting key entity features such as "local parameters," "current location," and "settlement value." After matching and calculating confidence, the system determines the parameter prediction agent as the scheduling target. Since no risk-related keywords are involved, the task planning agent automatically constructs an execution path that "only calls the parameter prediction agent to obtain local parameters for prediction."

[0063] Subsequently, the parameter prediction agent receives the execution plan and user request, and calculates the confidence level by matching it with the functional descriptions of various internal tools (such as surface settlement prediction tools, soil pressure prediction tools, and shield attitude prediction tools). The system reads the monitoring data at the current location from the local database (i.e., the most recently updated data in the parameter database) and inputs it into the surface settlement prediction tool with the highest matching confidence level. The prediction model within this tool automatically extracts the required features from the monitoring data based on preset parameter labels, calculates and outputs the surface settlement prediction results.

[0064] The specific analysis and processing procedures of the evaluation system are as follows: User issue: Based on local parameters, predict the settlement value at the current location.

[0065] System response: Task planning agent: Problem Analysis: The user requests a prediction of the settlement value at the current location. This is a straightforward prediction task and does not involve risk assessment or risk-related keywords.

[0066] Required agent: Parameter prediction agent.

[0067] Reason for selection: Predicting ground subsidence is one of the core functions of a predictive agent. The user's problem only requires predicting the subsidence value and does not involve keywords such as "risk" or "exceeding limits". Therefore, the task can be completed by using a parameter-based predictive agent.

[0068] Parameter prediction agent: I will predict the surface subsidence value at your current location. Let me call up the subsidence prediction tool for analysis.

[0069] According to the prediction results, the predicted surface subsidence at the current location is 2.42 mm.

[0070] This predicted value was calculated using a random forest model based on current tunnel boring machine (TBM) operation data and geological parameters. The settlement value reflects the potential impact of tunnel excavation on the surface.

[0071] It's important to note that this is a predicted value, and actual settlement may be affected by various factors. If you need an assessment of the probability of settlement exceeding the limit, please inquire about a risk assessment.

[0072] Example 3 When a user inputs the request, "Based on local data, please predict the active and passive instability risk of the tunnel boring machine's excavation face at the 469th ring," the task planning agent within the risk assessment system first receives the request and performs intent and entity parsing, extracting key entity features such as "excavation face," "active and passive instability," "risk," and "469th ring." After matching confidence levels, the system initially identifies the risk assessment agent as the primary scheduling target. Simultaneously, the system recognizes preset invocation constraints (i.e., triggered when the user's question involves a clear risk assessment requirement, and must be invoked after the prediction agent outputs valid parameters, or when the user directly provides working condition data). Given that the current request does not carry specific working condition data, the task planning agent automatically constructs an execution path of "first invoking the parameter prediction agent to obtain relevant state parameters, then invoking the risk assessment agent to perform a risk assessment."

[0073] Subsequently, the parameter prediction agent receives the execution plan and user request, and performs confidence calculations by matching them with the functional descriptions of various internal tools (such as surface settlement prediction tools, earth chamber pressure prediction tools, and shield attitude prediction tools). The system reads the corresponding monitoring data of the 469th ring of the shield tunnel from the local database and inputs it into the earth chamber pressure prediction tool with the highest matching confidence. The prediction model within this tool automatically extracts the required features from the monitoring data based on preset parameter labels, calculates and outputs the earth chamber pressure prediction results. Further, guided by its built-in prompts, the parameter prediction agent determines that additional active and passive earth pressure data are needed for "excavation face instability assessment," and thus autonomously triggers and calls the active and passive earth pressure calculation tools. After the calculation is completed, the large language model within the parameter prediction agent integrates the above multi-dimensional parameters and outputs the prediction results in natural language or structured form.

[0074] Finally, the system transmits the acquired earth chamber pressure and active / passive earth pressure data, along with the original user request, to the risk assessment agent. This agent, also based on matching confidence calculations, accurately invokes its internal active / passive instability risk assessment tools to calculate the excavation face instability risk value based on the input parameters. The system determines that since this risk value is within a low safety threshold range, it does not trigger the internal parameter optimization tool according to preset logic. Finally, the large language model within the risk assessment agent performs semantic reconstruction and logical organization of the numerical assessment results, outputting an accurate and fluent natural language risk assessment report to the user.

[0075] The specific analysis and processing procedures of the evaluation system are as follows: User issue: Based on local data, please predict the active and passive instability risks of the tunnel boring machine's excavation face on the 469th ring.

[0076] System response: Task planning agent: Problem Analysis: The user requests an assessment of the active and passive instability risks at the excavation face of the 469th ring tunnel boring machine. This is a clearly defined risk assessment task involving the stability analysis of the excavation face.

[0077] Required downstream agents: parameter prediction agent and risk assessment agent.

[0078] Reasons for selection: 1. First, the parameter prediction agent needs to be invoked to calculate key parameters such as the active and passive earth pressures at that location, the predicted earth pressure distribution in the ballast chamber, and the type of the preceding strata, based on the ring number 469. These are the basic data for assessing the risk of excavation face instability. 2. Then, the risk assessment agent must be invoked, utilizing the output of the parameter prediction agent (especially earth pressure, stratum type, etc.) to specifically assess the risk of excavation face instability. Since the problem explicitly points to "active and passive instability risk," this falls within the core functional scope of the risk assessment agent.

[0079] Parameter prediction agent: Risk assessment, including the assessment of active and passive instability risks at the excavation face, falls within the responsibilities of the risk assessment agent. I can provide you with relevant prediction data for Ring 469, which can serve as the basis for the risk assessment: Let me calculate the relevant parameters for ring 469 for you: Based on the calculation results, the predicted data for ring 469 is as follows: Formation pressure parameters: Passive earth pressure: 19.55 bar; Active earth pressure: 0.22 bar; Earth chamber pressure prediction: Bottom resistance (P_lower_point): 1.36 bar; Mid-section pressure: 1.61 bar; Top pressure (P_upper_point): 1.84 bar; Predicted pressure confidence interval: 0.037 bar; Uncertainty coefficient for pressure prediction: 0.38; If you require a risk assessment, please consult a risk assessment agent. I can then provide you with calculation data for other ring numbers or explain the engineering implications of these predictions.

[0080] Risk assessment agent: Based on the prediction results you provided, I will use the `champer_pressure_risk` tool (which encapsulates the excavation face instability risk function) to assess the active and passive instability risks of the 469th ring tunnel boring machine's excavation face. According to the assessment results of the `champer_pressure_risk` tool, the active and passive instability risks of the 469th ring tunnel boring machine's excavation face are as follows: Risk assessment results: 1. Risk of Active Instability Risk value: 0.0054 (approximately 0.54%); Risk level: Very low risk.

[0081] Note: Active instability risk refers to the risk of excavation face collapse caused by the soil chamber pressure falling below the active earth pressure. The current risk value is only 0.54%, indicating that the active instability risk is very low.

[0082] 2. Passive instability risk Risk value: 5.40×10-39 (almost zero); Risk level: Very low risk.

[0083] Note: Passive instability risk refers to the risk that the pressure in the soil chamber exceeds the passive earth pressure, leading to ground heave or gushing. The current risk value is almost zero, indicating that the passive instability risk is extremely low.

[0084] Risk Analysis Based on the provided forecast data: Pressure Comparison Analysis: Passive earth pressure: 19.55 bar; Active earth pressure: 0.22 bar; Earth chamber pressure range: 1.36-1.84 bar; The pressure in the soil chamber is much lower than the passive earth pressure and much higher than the active earth pressure. Safety margin analysis: Compared to passive earth pressure: the earth chamber pressure has a sufficient safety margin (approximately 17.7 bar). Compared to active earth pressure: the earth chamber pressure has a sufficient safety margin (approximately 1.1 bar). Impact of uncertainty: Uncertainty coefficient for pressure prediction: 0.375; Although there is some uncertainty, the current pressure setting has a large enough safety margin to effectively cope with prediction errors. in conclusion: The excavation face of the tunnel boring machine on the 469th ring is in a very safe condition. Active instability risk: extremely low (0.54%). Passive instability risk: extremely low (almost zero); The pressure setting of the earth chamber is reasonable and safe under the current geological conditions; It is recommended to continue tunneling according to the current parameters, while maintaining real-time monitoring of pressure parameters.

[0085] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention without departing from the spirit of the present invention, and all of these modifications are within the protection scope of the present invention.

Claims

1. A shield tunneling risk assessment system based on distributed intelligent agents, characterized in that, include: The task planning agent is used to analyze user questions, automatically determine the type of downstream agent to be called based on the question content, and construct the corresponding execution path; The standardized query agent is used to respond to the call requests in the execution plan. It uses retrieval-enhanced generative technology to learn the reserve knowledge related to the user's question from the knowledge vector database, providing knowledge supplementation and reasoning constraints for the reasoning process of other agents. The parameter prediction agent is used to encapsulate various prediction models with different functions and applicable conditions in the form of a tool. Based on the call requirements of the execution plan, it calls the prediction model related to the user problem, uses the construction parameters required by the prediction model provided by the user problem or local data source as the basis for calculation, and outputs parameter prediction results including predicted values ​​and prediction uncertainties. The risk assessment agent is used to calculate the risk coefficient and assess the risk level based on the prediction results of the parameter prediction agent and the control threshold of the risk index. If the risk level exceeds the preset limit, the agent generates the corresponding adjustment plan for the construction parameters and feeds it back to the parameter prediction agent, which then recalculates the adjusted risk index prediction results.

2. The shield tunneling risk assessment system based on distributed intelligent agents according to claim 1, characterized in that, The task planning agent, the standard query agent, the parameter prediction agent, and the risk assessment agent are each independently configured with a large language model. Through prompt word engineering technology, the role, functional boundaries, workflow, and interaction methods with other agents are defined for each large language model.

3. The shield tunneling risk assessment system based on distributed intelligent agents according to claim 1, characterized in that, All data within the system is transmitted in JSON format.

4. The shield tunneling risk assessment system based on distributed intelligent agents according to claim 3, characterized in that, During the internal data transmission process, a numerical data extraction unit is set up on the basis of transmitting the original text. It extracts relevant parameters from the text through predefined keywords and forms a structured numerical data file. The numerical data is transmitted to the target intelligent agent along with the original text.

5. The shield tunneling risk assessment system based on distributed intelligent agents according to claim 3, characterized in that, The task planning agent generates a structured JSON execution plan based on three fields: "problem analysis", "required downstream agents", and "reason for selection", and packages the user problem and execution plan and sends them to the corresponding downstream agents.

6. The shield tunneling risk assessment system based on distributed intelligent agents according to claim 1, characterized in that, The specification query agent segments the text in shield tunneling construction standards, specifications, and technical documents. Using an embedding model, it converts the segmented text fragments into corresponding text vectors, forming a knowledge vector database for the shield tunneling field. When a user asks a question, the embedding model converts the user's question into a question vector. The question vector is then compared with the text vectors in the knowledge vector database to calculate vector similarity. A preset number of highly similar text fragments are retrieved based on the similarity scores. These retrieved text fragments, along with the user's question, are input into the large language model built into the specification query agent to generate a question-and-answer result based on the specification. This result is then fed back to the upstream and downstream agents that sent the query request.

7. The shield tunneling risk assessment system based on distributed intelligent agents according to claim 1, characterized in that, The parameter prediction agent encapsulates various prediction algorithms in the field of tunnel boring into callable computing modules, and imports them into the agent framework in binary extended form to form prediction models. Through prompt word engineering, the functions and applicable conditions of each prediction model are predefined, enabling the large language model built into the parameter prediction agent to understand the applicable scenarios and input / output requirements of different prediction models. When a user question is passed to the parameter prediction agent, the large language model built into the parameter prediction agent combines the user question content, agent prompt words, and prompt words of each prediction model to automatically select the prediction model that matches the current question and trigger the call of the corresponding prediction model.

8. The shield tunneling risk assessment system based on distributed intelligent agents according to claim 1, characterized in that, The prediction model includes calculation models for stratum identification, surface settlement prediction, soil chamber pressure prediction, and shield attitude prediction; the risk indicators include surface settlement exceeding limits, active instability risk of excavation face, passive instability risk of excavation face, shield head horizontal deviation exceeding limits, shield head vertical deviation exceeding limits, shield tail horizontal deviation exceeding limits, shield tail vertical deviation exceeding limits, shield tail gap exceeding limits, roll angle exceeding limits, and pitch angle exceeding limits.

9. The shield tunneling risk assessment system based on distributed intelligent agents according to claim 1, characterized in that, During the iterative process of parameter prediction and risk assessment, the risk assessment agent treats the multiple risk coefficients corresponding to each candidate parameter combination as a multi-objective optimization problem. In each iteration, it selects the Pareto optimal solution from the candidate parameter combinations and fine-tunes the construction parameters based on the current Pareto optimal solution, continuously iterating and seeking optimization until all risk coefficients are lower than the corresponding thresholds.

10. An assessment method for a shield tunneling risk assessment system based on a distributed intelligent agent as described in any one of claims 1-9, characterized in that, Includes the following steps: Step S1: Analyze the user's question, automatically determine the subsequent steps to be executed based on the question content, and construct the corresponding execution path; Step S2: Based on the call request in the execution plan, the retrieval enhancement generation technology is used to learn the reserve knowledge related to the user's question in the knowledge vector database, so as to provide knowledge supplementation and reasoning constraints for the reasoning process in subsequent steps; Step S3: Encapsulate various prediction models with different functions and applicable conditions in the form of tools. Based on the call requirements of the execution plan, call the prediction model related to the user problem. Use the construction parameters required by the prediction model provided in the user problem or local data source as the calculation basis, and output the parameter prediction results including the predicted value and the prediction uncertainty. Step S4: Based on the prediction results of step S3 and the control threshold of the risk indicators, calculate the risk coefficient and assess the risk level. If the risk level exceeds the preset limit, generate the corresponding construction parameter adjustment plan and return to step S3 to re-predict the parameters.

Images