Shield intelligent auxiliary type selection system and method based on large language model

Abstract

The invention provides a shield intelligent auxiliary type selection system and method based on a large language model. The model selection system comprises a data input module, a rule knowledge base module, a large model reasoning module and a result generation module. The integrated decision-making system integrating a rule knowledge base, a deep learning model and expert system logic is constructed for the practical problems of complicated geological conditions, multiple rule constraints, high expert dependency and the like in shield construction, and the system combines a structured model selection rule and historical case data, and has the advantages of intelligence, standardization, self-learning, high efficiency and the like. The problems of low efficiency, high subjectivity, insufficient intelligent degree and the like of the existing shield tunneling machine model selection depending on artificial experience and partial standardized guide are solved, and the transformation of shield construction management from artificial experience to intelligent decision can be promoted.

Description

Technical Field

[0001] The present invention relates to the field of integrated application of tunnel construction equipment selection and artificial intelligence technology, and in particular to a shield machine intelligent auxiliary selection system and method based on a large language model. Background Art

[0002] With the acceleration of urbanization and the continuous expansion of infrastructure construction in my country, shield tunneling has become the primary construction method for projects such as underground transportation and municipal pipeline corridors. As core construction equipment, the rational selection of shield machines will directly affect the excavation efficiency, construction safety, and construction costs of the project. Different geological conditions (such as sand and gravel layers, alternating soft and hard layers, and water-rich strata) have significant differences in the adaptability of shield machines. The rational selection of shield types (such as earth pressure balance shields, slurry shields, or hybrid shields) and their key parameters (such as cutterhead configuration, propulsion system parameters, and slurry circulation systems) are key issues in engineering design and construction planning.

[0003] In existing technologies, the selection of shield machines mainly relies on manual experience and some standardized guidelines. Engineering technicians need to manually compare selection rules based on geological survey reports, design drawings, construction requirements, etc. to make judgments and formulate plans. This traditional method has the following major problems:

[0004] (1) Low selection efficiency: Engineers need to consult a large number of geological parameters, equipment manuals and past cases. The overall process is time-consuming and slow to respond, which is not conducive to quick decision-making;

[0005] (2) Highly subjective: The selection process is highly dependent on personal experience. Different technicians may make different judgments on the same working condition, resulting in poor stability and consistency of the selection results.

[0006] (3) Fragmentation of rules: Current engineering practice lacks a structured and unified shield selection rule system. Data is scattered in multiple documents, manuals, or project records, making it difficult to form systematic knowledge accumulation.

[0007] (4) Insufficient intelligence level: Although some companies have tried to introduce expert systems or parameter matching models, they are still unable to meet the selection automation requirements under complex geological conditions with multiple parameters and multiple constraints, and are unable to achieve model self-learning and dynamic optimization.

[0008] With the rapid development of artificial intelligence technology, especially large language models (LLMs), they have demonstrated powerful capabilities in natural language understanding, logical reasoning, and decision-making. Introducing large model technology into the field of shield equipment selection and building an intelligent auxiliary selection system capable of rule understanding, conditional judgment, solution generation, and self-optimization has become a key direction for improving engineering design efficiency and promoting the implementation of intelligent construction. However, to date, there is no mature technical solution that can effectively combine large models, rule knowledge bases, and actual engineering data to achieve automated reasoning and feedback loops for shield equipment selection in the engineering context. Summary of the Invention

[0009] In response to the problems in the background technology, the present invention provides a shield intelligent auxiliary selection system and method based on a large language model. The method makes full use of the reasoning and generation capabilities of the large language model, combines structured selection rules with historical case data, and constructs an intelligent, standardized, and self-learning shield auxiliary selection system. It can solve the problems of low efficiency, strong subjectivity, and insufficient intelligence in the existing shield machine selection that relies on manual experience and partial standardized guidelines, and can promote the transformation of shield construction management from manual experience to intelligent decision-making.

[0010] In order to achieve the above technical objectives, the present invention provides a shield intelligent auxiliary selection system based on a large language model, the selection system includes a data input module, a rule knowledge base module, a large model reasoning module and a result generation module;

[0011] The data input module is used to uniformly collect various original parameters related to shield machine selection in engineering projects, and format and standardize the collected different types of original parameters to ensure the consistency and solvability of different data sources, and provide structured and computable input data for subsequent knowledge rule retrieval and large model reasoning;

[0012] The rule knowledge base module is used to convert the decision logic in traditional expert selection experience, standard specification clauses and historical engineering cases into structured, searchable and inferable rule set data, and input it into the large model reasoning module for combined with the large language model to assist in reasoning;

[0013] The large model reasoning module is the core intelligent decision-making unit of the system, which is used to semantically fuse the structured and computable input data input by the data input module with the rule set data input by the rule knowledge base module, and output the optimal shield selection plan and configuration parameters through the reasoning and generation capabilities of the large language model;

[0014] The result generation module is used to structure and visualize the selection suggestions, reasoning instructions, and parameter configuration information output by the large model reasoning module to form standardized engineering documents and data interface results, so that engineering designers can review, download, archive, or integrate them into the construction management system.

[0015] A further technical solution of the present invention is as follows: the input data accepted by the data input module is divided into four categories based on key influencing factors of shield machine selection, specifically including geological parameters, design parameters, construction constraints and engineering background. The geological parameters include stratum type, permeability coefficient, groundwater pressure, powder content, stratum hardness, stratum particle size and friction coefficient; the design parameters include excavation diameter, tunnel longitudinal slope, minimum curve radius, burial depth, hard soil thickness and linear structure form; the construction constraints include surface settlement control value, propulsion speed requirement, noise / vibration limit level and operation time limit; the engineering background includes project type, regional characteristics, risk level, implementation location and policy restrictions / approval basis; the above data types include enumeration type, floating point number and integer type;

[0016] The process of formatting and standardizing the collected original parameters of different types by the data input module is as follows:

[0017] (1) Enumeration field mapping: For the parameters of the text description field, a preset vocabulary and keyword matching mechanism are used to map them to standard engineering codes to facilitate subsequent rule matching;

[0018] (2) Numerical normalization: For parameters in numerical fields, the input is uniformly processed and normalized in SI units to improve the generalization ability of large models. The details are as follows:

[0019]

[0020] Where x is the original input value, x′ is the normalized value, μ and σ are the historical mean and standard deviation of the parameter respectively;

[0021] (3) Calculate the cross-sectional area of ​​the tunnel excavation to provide an area basis for subsequent thrust calculations. The cross-sectional area of ​​the tunnel excavation is calculated as follows:

[0022]

[0023] Where: A is the cross-sectional area; D is the excavation diameter;

[0024] (4) Calculation of geological comprehensive classification scores: some geological conditions are automatically classified into high / medium / low complexity ranges using a weighted integral model. The calculation method is as follows:

[0025]

[0026] Where: G is the total score; w i is the weight of the i-th parameter; x i is the value of the i-th indicator;

[0027] After the data input module completes the standardization processing, it will generate a unified input vector structure, including numerical tensors and semantic labels; and pass it into the rule knowledge base retrieval module to match the rule entries that meet the conditions, and then pass it into the large model reasoning module to generate semantic prompts and constraints for LLM reasoning.

[0028] A further technical solution of the present invention is as follows: the selection system also includes a feedback and model optimization module, which is used to collect user evaluation and revision feedback on the selection results of the result generation module during actual engineering applications, and convert it into model optimization samples or knowledge base update basis through preset rules and manual review mechanism, thereby realizing a closed-loop mechanism of knowledge accumulation-reasoning optimization-performance enhancement; the feedback and model optimization module integrates an active feedback entry, an automatic scoring system, a sample storage structure and a reward function construction algorithm, and supports the parallel operation of reinforcement learning optimization and expert consensus mechanism;

[0029] The feedback and model optimization module constructs a user feedback reward function, which serves as an important source of scoring for reinforcement learning samples and will be used to design the objective function of the large language model in subsequent policy training. The score and label of the user feedback reward function are passed to the large model inference module to guide the model toward high-quality output. The user feedback reward function is a reinforcement learning reward function constructed using user scores. It acts on the objective function in the policy optimization process and drives the parameters of the reinforcement learning reward model to migrate toward "high-scoring samples."

[0030] The large model inference module introduces a reinforcement learning mechanism based on human feedback and adopts a PPO optimization strategy. This reinforcement learning mechanism takes the form of a reward function. The reward function in the large model inference module is derived from the "user evaluation index system" constructed in the feedback and model optimization module and formalized into a policy optimization function. The reward function is constructed by mapping user scoring labels to structural scoring factors. This function serves as the fitting basis for the target distribution in PPO training and is tailored and optimized in combination with the old strategy ratio to improve the professionalism and practicality of the model output.

[0031] The reinforcement learning reward function is as follows:

[0032] R = α·Accuracy score + β·Term matching rate + γ·Structural completeness

[0033] Among them, α, β, and γ are weighting coefficients;

[0034] The update target of the PPO optimization is:

[0035]

[0036] Where: r t (θ) is the ratio of new and old strategies; is the advantage function; ∈ is the clipping factor, which is generally 0.2.

[0037] A further technical solution of the present invention is as follows: the rule knowledge base module undertakes three major functions in the system: knowledge expression, rule reasoning and model constraint, and is the key interface for realizing the integration of engineering reliability and model intelligence;

[0038] The rule knowledge base module supports the logical expression of "condition → conclusion" and provides rule retrieval capabilities based on attribute matching, conditional relationship modeling capabilities based on graph structures, and semantic parsing capabilities compatible with natural language input. Its underlying architecture is implemented using a hybrid "rule table + graph database + semantically enhanced index" architecture, serving as the retrieval end in the RAG framework and working in collaboration with the large language model.

[0039] The rule knowledge base module supports two rule update methods: manual editing and updating, and model reverse deduction and extraction. The rule knowledge base module also has a confidence evolution mechanism. If a rule is repeatedly fed back as "low applicability", its weight will automatically decrease; otherwise, it will be increased, forming a dynamic trust adjustment capability.

[0040] The rules in the rule knowledge base module are generally divided into the following categories: shield type recommendation rules (classification number R1), cutterhead configuration rules (classification number R2), propulsion system rules (classification number R3), grouting and slurry system rules (classification number R4), and comprehensive adaptability rules (classification number R5). Each rule also contains the following additional metadata: source, applicability level, credibility, and whether it is expert consensus.

[0041] Each rule has the following structure:

[0042]

[0043] Where: R i is the i-th rule; C n is the nth condition; Conclusion is the inference conclusion;

[0044] When multiple rules conflict or compete with each other, the graph edge weight scoring function is used to fuse the confidence and calculate the recommendation score. The specific calculation is as follows:

[0045]

[0046] Where: S j is the composite score of solution j; α i is the rule weight; rij is the score of rule i for solution j.

[0047] The rule knowledge base module uses knowledge graph technology to construct a "parameter-rule-conclusion" triple network, where nodes represent parameters or conclusions and edges represent logical constraints. When multiple parameter combinations trigger different rules, the system performs comprehensive reasoning through graph traversal and confidence fusion algorithms to output the most appropriate conclusion and its confidence score.

[0048] The rule knowledge base module is provided with an interactive mode with the large language model and provides fact constraints and knowledge basis for the large language model reasoning module. The specific process is as follows:

[0049] ①User inputs structured parameters;

[0050] ② Retrieve all rules that meet the conditions in the matching knowledge base;

[0051] ③Build a prompt template example:

[0052] ④ The large model integrates parameters, rules and cases to generate recommendation reasons and configuration plans.

[0053] A further technical solution of the present invention is as follows: the large model inference module performs secondary optimization based on a pre-trained large language model, wherein the large language model includes DeepSeek, Qwen, and Baichuan domestic large models, and the core structure of the large model inference module includes the following:

[0054] Prompt parsing and construction unit: This unit is used to combine input parameters and rule text into natural language prompts as model input. This unit designs a structured prompt template that combines the characteristics of natural language and structured tags to stimulate the contextual reasoning and professional expression capabilities of large models.

[0055] Reasoning generation model: used to load the fine-tuned LLM, perform selection reasoning, solution generation and text output;

[0056] Strategy evaluation and reinforcement learning module: used to evaluate the quality of model output results based on domain feedback and optimize model parameters and weight distribution;

[0057] The processing flow of the large model reasoning module is: structured input → Prompt assembly → large model reasoning → output shield plan + explanation text.

[0058] A further technical solution of the present invention is that the data input module has both parameter interface compatibility and data semantic integrity, and can be connected to the geological data systems of design units, survey units, and construction units, so as to realize one-click import or manual entry of various original parameters related to shield selection in engineering projects;

[0059] The data input module supports multiple input methods. The front-end graphical interface input uses selection boxes, sliders, and data forms to fill in engineering parameters; data file import supports CSV / Excel format to import survey unit report forms; API interface access is connected to the BIM platform and the shield construction platform, and can call geological data services in real time; the interface design adopts the RESTful protocol, and all parameter transmission formats are unified into the JSON structure.

[0060] The data input module is also equipped with an outlier detection and fault tolerance mechanism. If the permeability coefficient input is negative or exceeds a reasonable range, the system will automatically pop up a prompt. If the geological field is empty but designated as a "complex site", the system prompts to fill in detailed parameters or enable the model's self-explanatory function. If the value is missing, the system can make a default inference based on the input field or call the historical case completion mechanism.

[0061] Further technical solutions of the present invention include: the feedback and model optimization module supports user feedback input types including explicit feedback and implicit feedback; all feedback information collected by the feedback and model optimization module is standardized and stored as structured sample data for training the reinforcement learning reward model; the feedback and model optimization module regularly parses and aligns labels of newly added feedback samples every day, and adds them to the model training sample pool; the feedback and model optimization module is equipped with a rule knowledge update mechanism. For new patterns that appear in high-frequency, repeated, and high-scoring feedback, the module automatically analyzes their logical features and generates candidate rules. The candidate rules are presented to experts for review in an artificial language structure. If approved, they are written into the rule knowledge base and marked with the source; the module forms a "rule learning log" to achieve traceable update records;

[0062] After fine-tuning or reward optimization of the large model language, the feedback and model optimization module automatically saves the traceable version and has version freezing, version rollback and version comparison functions; the feedback and model optimization module sets up security mechanisms and expert review guarantees.

[0063] A further technical solution of the present invention is as follows: the result generation module supports multiple output formats, document templates, extensible interfaces, and data interoperability with mainstream CAD / BIM platforms. The output content includes the following five categories of information: recommended shield type, key configuration parameters, reasoning process description, rule and model reference, and structured report data, and is uniformly presented in the form of "structured data + text explanation + specification reference". The result generation module also has a graphical user interface function. The result generation module outputs the following standard formats: PDF engineering report, Excel data table, JSON structured data, and SCR / CAD drawing data.

[0064] When the large model inference module outputs multiple candidate results, the result generation module assigns a confidence score to each recommended candidate result and selects the recommended result with the highest confidence as the final selection result. The confidence score calculation method is as follows:

[0065]

[0066] Where: C is the confidence score; p i is the probability of the model on this judgment; c i Score the effectiveness of the expert assessments.

[0067] The present invention also provides a shield machine intelligent assisted selection method based on a large language model, which uses the above-mentioned shield machine intelligent assisted selection system based on a large language model to perform selection, specifically comprising the following steps:

[0068] S1. Engineering Parameter Collection and Standardization: The data input module receives various raw parameters related to shield selection in the project, including geological conditions, tunnel design parameters, construction constraints, and engineering background. These raw parameter data are converted into structured vector inputs using a standardized algorithm and an engineering semantic dictionary.

[0069] S2. Rule Knowledge Retrieval and Conditional Constraint Reasoning: Using the structured selection rule knowledge base built into the rule knowledge base module, we perform multi-condition matching and logical constraint retrieval on standardized inputs to obtain a set of candidate rules. These rules express engineering experience and standard clauses in an "if-then" format and are modeled through a graph database to form a parameter-condition-conclusion logical graph structure, providing initial constraint prompts and multi-path fusion capabilities for large-scale model reasoning.

[0070] S3. Large-scale model semantic understanding and selection reasoning generation: The combination of conditions retrieved from the knowledge base and the input parameters are combined to form semantic prompts. These prompts are then fed into a large language model fine-tuned for engineering scenarios to perform complex multi-constraint reasoning generation tasks. Based on LoRA fine-tuned parameters, the large language model combines rule semantics, historical cases, and reinforcement learning feedback data to automatically generate the most suitable shield type, key configuration parameters, and complete reasoning explanation text.

[0071] S4. Result Generation and Visualization Output: The result generation module receives the output of the large language model and performs structural reconstruction, format conversion, and multi-format export. The output content includes the recommended type, parameter list, reasoning description, reference rule number, and recommendation confidence score. The result generation module also supports AutoCAD scripts and construction platform data integration interfaces, facilitating direct access and archiving by design units.

[0072] S5. Continuous model optimization driven by user feedback: A user feedback mechanism is provided, supporting explicit scoring and implicit behavior collection. All feedback information will be stored as structured training samples, and combined with reinforcement learning to build a reward function for large-scale model parameter updates. If repeated high-confidence feedback corrections are found, the system automatically generates rule candidates and enters the rule knowledge base review process, building a dynamically evolving "engineering feedback-rule update-model reinforcement" learning loop.

[0073] Further technical solutions of the present invention: In step S3, the fine-tuning of the large language model adopts the following two methods:

[0074] (1) LoRA: Insert the low-rank matrix A, BA, BA, B into the attention layer of the original model and only fine-tune AAA and BBB. The calculation expression is:

[0075] ΔW=A·B, where rank(A), rank(B)< <dim(W)

[0076] Use PEFT framework to load LoRA module;

[0077] (2) SFT: supervised learning is performed using a ternary sample set of “engineering input – output selection result – reasoning explanation”; the loss function is the typical cross entropy loss:

[0078]

[0079] in: is the loss value; P θ Output probability for the model.

[0080] The large-model-based intelligent assisted shield selection system provided by the present invention integrates an engineering rule knowledge base, deep language model reasoning technology and a feedback learning mechanism, aiming to achieve rapid, accurate and intelligent selection of shield equipment types and key parameters under complex geological conditions.

[0081] Through the ternary fusion architecture of "rule knowledge + large model reasoning + expert feedback", the present invention enables the selection of shield equipment to shift from "manual experience decision-making" to an intelligent, standardized and interpretable new paradigm of "data-driven - knowledge-guided - model generation". It has high adaptability, high efficiency and continuous evolution capabilities, and has broad engineering promotion and application prospects in complex geological tunnel construction.

[0082] Aiming at the practical problems of complex geological conditions, multiple constraints on rules, and high dependence on experts in shield construction, the present invention constructs an integrated decision-making system that integrates a rule knowledge base, a deep learning model, and expert system logic. The system combines structured selection rules with historical case data, and is intelligent, standardized, and self-learning. It solves the problems of low efficiency, strong subjectivity, and insufficient intelligence in the existing shield machine selection that relies on manual experience and partially standardized guidelines, and can promote the transformation of shield construction management from manual experience to intelligent decision-making.

[0083] Beneficial effects of the present invention:

[0084] (1) Improve selection efficiency and reduce manual dependence: The present invention uses a large language model to automatically perform shield equipment selection reasoning tasks, replacing the traditional method of relying on manual reference to information and expert judgment, significantly shortening the decision-making time and achieving a leap in decision-making efficiency from "hours" to "minutes", which is suitable for large-scale, fast-response engineering selection scenarios.

[0085] (2) Enhance engineering adaptability and selection accuracy; the system of the present invention integrates the dual mechanisms of structured rule base and deep learning model, which can handle multiple working condition inputs under complex, cross-cutting and nonlinear conditions, output personalized and accurate shield type and parameter recommendations, adapt to typical complex formations such as soft and hard alternation, sand and gravel, and high water pressure, and improve the scientificity and feasibility of selection.

[0086] (3) Make the selection process explainable and standardized. Through rule reference, reasoning path display, and text reasoning generation, the system can output a complete selection logic chain with good explainability. At the same time, the output format supports standard templates and structured archiving, facilitating collaboration among design units, construction units, expert review, and other parties, achieving engineering standardization of the selection process.

[0087] (4) Supporting the model evolution mechanism driven by user feedback; the system of the present invention integrates reinforcement learning and expert consensus mechanism, collects user evaluation information on the selection results, dynamically adjusts the model weights and rule structure, builds a knowledge system that can be continuously evolved, and realizes that the model becomes more accurate with use, and constantly adapts to new projects, new scenarios, and new constraints.

[0088] (5) It is convenient for system integration and multi-platform deployment. The system of the present invention provides a standardized RESTful interface and data output template, which can be seamlessly connected with shield construction platforms, intelligent design systems, BIM platforms, etc. It also supports local deployment and private cloud deployment to meet the scenario-based needs of construction units, design institutes, etc. for data security, engineering isolation, etc.

[0089] (6) Promote the intelligentization of shield construction and the digital inheritance of knowledge; the present invention transforms the selection knowledge scattered in the minds of experts and case experience into structured knowledge graphs and reasonable models, realizes the explicitness, algorithmization and inheritance of engineering knowledge, and helps traditional engineering design enter a new stage of "AI + engineering" integrated development. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 Schematic diagram of the overall architecture of the system of the present invention;

[0091] Figure 2 This is the input parameter structure diagram for shield machine selection in the present invention;

[0092] Figure 3 Schematic diagram of the logical structure of the rule knowledge base in the present invention;

[0093] Figure 4 This is a flowchart of the large model reasoning under the RAG mechanism of the present invention;

[0094] Figure 5 This is a schematic diagram of the inference result generation and output format in the present invention;

[0095] Figure 6 This is a diagram of the interactive interface for multiple scheme comparison and expert revision in the present invention;

[0096] Figure 7 Schematic diagram of user feedback data collection and sample structure in the present invention;

[0097] Figure 8 This is a flow chart of the model enhancement optimization based on the PPO algorithm in the present invention;

[0098] Figure 9 This is a structural diagram of the rule evolution and automatic update mechanism in the present invention;

[0099] Figure 10 Schematic diagram of the integration of the system in the present invention in the shield construction platform DETAILED DESCRIPTION

[0100] The present invention will be further described below with reference to the accompanying drawings and embodiments. Figures 1 to 10 The accompanying drawings are simplified versions of the embodiments and are only used to clearly and concisely illustrate the embodiments of the present invention. The technical solutions shown in the accompanying drawings are specific solutions of the embodiments of the present invention and are not intended to limit the scope of the invention claimed for protection. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0101] The embodiment provides a shield intelligent auxiliary selection system based on a large language model. It addresses the practical problems of complex geological conditions, multiple constraints on rules, and high dependence on experts in shield construction. It builds an integrated decision-making platform that integrates rule knowledge base, deep learning model and expert system logic. The overall system architecture is as follows: Figure 1 As shown, it mainly includes the following modules: data input module, rule knowledge base module, large model reasoning module, result generation module and feedback optimization module.

[0102] The data input module is the starting unit of this system. Its function is to uniformly collect, format and standardize various original parameters related to shield selection in engineering projects, and provide structured and computable input data for subsequent knowledge rule retrieval and large model reasoning. This module has both parameter interface compatibility and data semantic integrity, and can be connected to the geological data systems of design units, survey units and construction units to achieve one-click import or manual entry. Figure 2 The table shows the typical input parameter types and logical structure submitted by users in the data input module. The input parameter types include more than 20 fields in four categories, namely geological parameters, design parameters, construction constraints and engineering background. The details are shown in Table 1:

[0103] Table 1 Input parameter classification and examples

[0104]

[0105]

[0106] The implementation provides a set of input parameter standardization mechanisms to ensure consistency and interpretability of data from different sources. The main processing logic is as follows:

[0107] Enumeration field mapping: For text description fields (such as stratum type "silty clay", hydrological description "soft saturated sand layer"), the system uses a preset vocabulary and keyword matching mechanism to map them to standard engineering codes (such as GG0201, I010201, etc.) to facilitate subsequent rule matching. For example:

[0108] Input: Stratum type = "silty clay" → Mapping: GG0201 = 4 (number represents soft soil)

[0109] Numerical normalization: For numerical fields (such as permeability coefficient and propulsion speed), the input is uniformly processed and normalized in SI units to improve the generalization ability of large models. For example:

[0110]

[0111] Among them, x is the original input value, x′ is the standardized value, μ and σ are the historical mean and standard deviation of the parameter respectively.

[0112] To provide an area basis for subsequent thrust calculations, first calculate the tunnel excavation cross-sectional area:

[0113]

[0114] Where: A is the cross-sectional area; D is the excavation diameter.

[0115] Some geological conditions are automatically classified into high / medium / low complexity ranges using a weighted integral model, calculated as follows:

[0116]

[0117] Where: G is the total score; w i is the weight of the i-th parameter; x i is the index value of item i (such as water content, porosity, etc.).

[0118] To improve system stability, the data input module in this embodiment is equipped with a complete anomaly detection mechanism:

[0119] If the permeability coefficient is negative or exceeds the reasonable range (such as >1×10 -2 m / s), the system will automatically pop up a prompt;

[0120] If the geological field is empty but specified as a "complex site", the system prompts you to fill in the detailed parameters or enable the model self-explanation function;

[0121] If the value is missing, the system can make a default inference based on the entered field or call the historical case completion mechanism.

[0122] The data input module in the embodiment supports multiple input methods, including:

[0123] Front-end graphical interface input: users can fill in project parameters through selection boxes, sliders and data forms;

[0124] Data file import: supports importing survey unit report forms in CSV / Excel format;

[0125] API interface access: connect with BIM platform and shield construction platform to call geological data services in real time;

[0126] The interface design adopts the RESTful protocol, and all parameter transmission formats are unified into JSON structure, for example:

[0127]

[0128] After the data input module completes the normalization process, it generates a unified input vector structure (InputFeature Vector), including numerical tensors and semantic labels, and passes it into:

[0129] Rule knowledge base retrieval module: matches rule entries that meet the conditions;

[0130] Large model reasoning module: generates semantic prompts and constraints for LLM reasoning;

[0131] Feedback and model optimization module: records the correspondence between input and selection recommendations to form training samples.

[0132] Data input application example in the embodiment:

[0133] In a typical application scenario, enter the following parameters:

[0134] Stratum: soft silt sand layer

[0135] Permeability coefficient: 1×10 -6 m / s

[0136] Groundwater pressure: 0.35MPa

[0137] Tunnel diameter: 9.8m

[0138] Settlement control value: 6mm

[0139] The system normalizes it and generates an input vector, identifies the risk level as "medium", and passes it into the large model inference module. The final output recommends the shield type as "slurry shield", the cutterhead opening rate is 35%, the main drive system thrust is recommended to be 85,000kN, and the propulsion speed is controlled at 70-80mm / min.

[0140] The rule knowledge base module described in the embodiment is the basic component of the system for realizing intelligent decision-making reasoning, which aims to transform the decision logic in traditional expert selection experience, standard specification clauses and historical engineering cases into a structured, searchable and reasonable rule set, and use it as an input condition in combination with the large language model (LLM) for auxiliary reasoning. This module supports the logical expression of "condition → conclusion", and provides rule retrieval capability based on attribute matching, conditional relationship modeling capability based on graph structure, and semantic parsing capability compatible with natural language input. Its underlying structure is implemented with a hybrid architecture of "rule table + graph database + semantic enhanced index", which can be used as the retrieval end in the RAG (Retrieval-Augmented Generation) framework to work in conjunction with the large model. Specifically, Figure 3 and Figure 4 ,in Figure 3 Represents the logical connection between "parameter conditions - rule numbers - recommended conclusions" in the rule knowledge base, including the rule expression structure, the setting method of graph database nodes and edge weights, and the rule reference relationship; Figure 4This paper describes how the system constructs input parameters and rule semantics into prompts under the Retrieval Enhanced Generation (RAG) framework, and calls the fine-tuned large model to execute the generation and reasoning process of shield types and parameters.

[0141] Each rule in the rule knowledge base of the rule knowledge base module adopts the following structured form:

[0142]

[0143] Where: R i is the i-th rule; C n is the nth condition (e.g., "permeability > 10^{-4}m / s");

[0144] Conclusion is an inference (such as "it is recommended to use slurry shield").

[0145] Each rule also contains additional metadata, such as source, applicability level, credibility, whether it is an expert consensus, etc. The rules are generally divided into the following categories:

[0146] Classification number Category Name Main content description R1 Recommended rules for shield types Recommend shield machine type (earth pressure / mud water) based on geological and hydrological parameters R2 Cutter configuration rules Recommend cutterhead diameter, opening ratio, etc. based on soil properties and tunnel diameter R3 Propulsion system rules Estimate thrust range, cylinder layout, upper limit of propulsion speed, etc. R4 Grouting and Slurry System Rules Involving mud discharge speed, flow estimation, mud flow rate and slag discharge capacity R5 Comprehensive Adaptability Rules Hierarchical adaptation rules and exclusion conditions in multi-parameter joint constraint scenarios

[0147] When multiple rules conflict or compete, the system uses the graph edge weight scoring function to fuse the confidence and calculate the recommendation score:

[0148]

[0149] Where: S j is the composite score of solution j; α i is the rule weight; r ij is the score of rule i for solution j.

[0150] The sample rules of the rule knowledge base module are expressed as a JSON structure + table structure:

[0151] Among them, the table structure style (for manual review and design maintenance)

[0152] Rule ID condition Recommended Conclusion Credibility R1-001 <![CDATA[Coefficient of permeability > 10 -4 m / s]]> Slurry shield is recommended high R1-002 <![CDATA[Permeability coefficient < 10 -7 m / s]]> Earth pressure shield is recommended high R1-003 Groundwater pressure>0.3MPa Prioritize the use of slurry shield middle R2-011 Fine particle content>50% Suitable for earth pressure shield high R3-020 Excavation diameter>10m <![CDATA[It is recommended that the propulsion system use a coefficient of 1300 kN / m 2 > high

[0153] JSON structure for model embedding or API calls;

[0154]

[0155] To enhance the comprehensive capabilities of rule-based reasoning in scenarios involving multiple rule conflicts or joint recommendations, this paper further introduces a parameter-rule-conclusion ternary graph modeling approach within the rule knowledge base module. This approach expresses the logical dependencies between multiple rules and quantifies the recommendation strength of each candidate solution through a path scoring function. This mechanism complements the static judgment approach used for single rule matching, empowering the system with stronger combinatorial reasoning and solution integration capabilities. Nodes represent parameters or conclusions, while edges represent logical constraints.

[0156] Node A: "groundwater pressure>0.3MPa";

[0157] Node B: "Slurry shield is recommended";

[0158] Edge A→B: logical dependency, weight 0.9;

[0159] When multiple parameter combinations trigger different rules, the system performs comprehensive reasoning through graph traversal + confidence fusion algorithm to output the most appropriate conclusion and its confidence score.

[0160] After completing graph structure scoring and fusion, to further achieve collaborative reasoning between rules and generative models, this paper adopts the RAG mechanism to construct the matched rules into prompt templates and pass them to the large language model for final recommendation generation. The specific interaction process is as follows:

[0161] ① The user inputs structured parameters (e.g. groundwater pressure = 0.4 MPa);

[0162] ② Retrieve all rules that meet the conditions in the matching knowledge base (such as R1-003);

[0163] ③Build a prompt template (Prompt) example:

[0164] The known geological conditions of the tunnel are as follows:

[0165] -Permeability coefficient is 1e-6m / s;

[0166] - Groundwater pressure is 0.35 MPa;

[0167] -The tunnel excavation diameter is 9.8m.

[0168] Please select the shield machine based on the following rules:

[0169] Rule 1: Groundwater pressure > 0.3 MPa → Slurry shield is recommended;

[0170] Rule 2: Permeability coefficient between 1e-7 and 1e-4 → Earth pressure tunneling or slurry shield can be used;

[0171] Rule 3: Excavation diameter > 10m → Slurry shield is preferred.

[0172] ④ The large model integrates parameters, rules and cases to generate recommendation reasons and configuration plans.

[0173] The rule knowledge base module supports two rule update methods:

[0174] ① Manual editing and updating: Regular maintenance by experts to add new geological scenarios, special cases, etc.

[0175] ② Model reverse deduction and extraction: By analyzing the relationship between the model-generated results and inputs, the underlying logical expressions are extracted, candidate rules are formed and submitted for review.

[0176] The system also has a confidence evolution mechanism: if a rule is continuously fed back as "low applicability", its weight will automatically decrease; otherwise, it will be increased, forming a dynamic and credible adjustment capability.

[0177] In the processing application of the rule knowledge base module in the embodiment, in a real project, the system receives the following input:

[0178] Stratum: sand and gravel mixed with silt;

[0179] Permeability coefficient: 8×10 -5 m / s;

[0180] Groundwater pressure: 0.38MPa;

[0181] Settlement control value: <10mm.

[0182] The knowledge base module searches for matches to the following rules:

[0183] Permeability coefficient > 1e-4 → Slurry shield is recommended;

[0184] Groundwater pressure > 0.3 MPa → Slurry shield is recommended;

[0185] High settlement control requirements → Earth pressure shield is recommended;

[0186] After the graph structure weights are integrated, the weight of the slurry shield is 0.82, and the weight of the earth pressure shield is 0.65. The slurry shield is ultimately recommended, and a complete decision path is output for expert confirmation.

[0187] The rule knowledge base module described in the embodiment undertakes the three major functions of knowledge expression, rule reasoning and model constraint in the system of the present invention, and is the key interface for realizing the integration of engineering reliability and model intelligence. Its structured design and semantic enhancement function provide powerful decision-making support and explainability for subsequent modules. The large model reasoning module is the core intelligent decision-making unit of the invention system, which is responsible for semantically fusing input parameters with knowledge rules, and automatically outputs the optimal shield selection scheme and configuration parameters through the reasoning and generation capabilities of the large language model (LLM), and is accompanied by the reasoning process and reasoning. Its goal is to quickly complete logical reasoning and conclusion generation under multi-source conditions and complex constraints in an expert-like manner, and enable the intelligent decision-making process of tunnel equipment selection. The large model reasoning module successfully automates complex engineering parameter reasoning tasks by integrating rule constraints, language generation and reinforcement learning optimization, demonstrating the deep integration capability of engineering knowledge and general artificial intelligence models.

[0188] The large model inference module performs secondary optimization based on pre-trained large language models (such as DeepSeek, Qwen, Baichuan and other domestic large models). The core structure includes the following three components:

[0189] ① Prompt parsing and construction unit: responsible for combining input parameters and rule text into natural language prompts as input to the model;

[0190] ② Reasoning generation model (main model): loads the fine-tuned LLM and performs selection reasoning, solution generation, and text output;

[0191] ③ Strategy evaluation and reinforcement learning module: Evaluate the quality of model output results based on domain feedback and optimize model parameters and weight distribution.

[0192] The overall process is as follows: "Structured input → Prompt assembly → Large model reasoning → Output shield plan + explanatory text".

[0193] To enable large language models to more effectively understand engineering scenarios, the large model inference module has designed a structured prompt template. The typical structure is as follows:

[0194] [Project Background] This project is an urban subway section tunnel with an excavation diameter of 9.5m. The construction stratum is a silty clay interbedded sand layer with a groundwater pressure of 0.38MPa.

[0195] Input parameters

[0196] -Permeability coefficient: 1×10 -6 m / s;

[0197] - Powder content: 52%;

[0198] -Surface settlement control value: 6mm;

[0199] -Curve radius: 300m.

[0200]

Rules Summary

[0201] Rule 1: Permeability coefficient > 10 -4 m / s → Slurry shield is recommended;

[0202] Rule 2: Groundwater pressure > 0.3 MPa → Slurry shield is recommended;

[0203] Rule 3: Settlement control value < 10mm → Earth pressure shield is recommended.

[0204] As a shield design expert, please determine which shield type should be selected based on the above information and explain your reasons.

[0205] This structure combines the characteristics of natural language and structured labels, which can effectively stimulate the contextual reasoning and professional expression capabilities of large models.

[0206] To improve the model's adaptability to engineering selection tasks, the large model inference module fine-tunes the large language model using the following two strategies:

[0207] (1) LoRA (Low-Rank Adaptation)

[0208] Insert the low-rank matrix A, BA, BA, B into the attention layer of the original model and only fine-tune AAA and BBB. The calculation expression is:

[0209] ΔW=A·B, where rank(A), rank(B)<<dim(W)

[0210] Its advantage is that it greatly reduces computing resources and video memory overhead;

[0211] Specifically, the PEFT (Parameter-Efficient Fine-Tuning) framework is used to load the LoRA module.

[0212] (2)SFT(Supervised Fine-Tuning)

[0213] Supervised learning is performed using a ternary sample set of "engineering input - output selection result - reasoning explanation"; the loss function is the typical cross-entropy loss (Cross-Entropy Loss):

[0214]

[0215] in: is the loss value; P θOutput probability for the model.

[0216] The input sample format in the embodiment is as follows:

[0217]

[0218] Based on the input parameters, the system can automatically derive key parameters. For example, the recommended thrust value of the shield propulsion system can be calculated based on the following engineering experience formula:

[0219]

[0220] Where: F is the total thrust of the shield machine (unit: kN); D is the shield excavation diameter (unit: m); k is the thrust coefficient per unit area, and the recommended value is 1100~1300kN / m 2 .

[0221] If D = 9.5m is input, we can get:

[0222]

[0223] This value automatically appears in the selection results as one of the propulsion system design parameters.

[0224] To improve the engineering practicality of the generated results, the large model inference module introduces a reinforcement learning based on human feedback (RLHF) mechanism and adopts the PPO (Proximal Policy Optimization) optimization strategy:

[0225] The rationality, accuracy, and professional expression of the generated answers are evaluated by engineering experts, with a five-level scoring system (0 to 4 points). The reward function is as follows:

[0226] R = α·Accuracy score + β·Term matching rate + γ·Structural completeness

[0227] The module trains the policy model πθ(a|s)\pi_\theta(a|s)πθ(a|s) based on the reward value, increasing the probability of high-quality output. This reward function, used as an immediate reward for policy updates during online training, is derived from the system's automatic scoring and generated content evaluation. The reward function in this module is derived from the "user evaluation index system" constructed in the feedback module and formalized as a policy optimization function. This reward function is constructed by mapping user score labels to structural scoring factors. This function serves as the basis for fitting the target distribution in PPO training and is optimized in combination with the old policy ratio to improve the professionalism and practicality of the model output.

[0228] During the PPO training process, the policy gradient is penalized to avoid drastic changes in the policy;

[0229] Update target:

[0230]

[0231] Where: r t (θ) is the ratio of new and old strategies; is the advantage function; ∈ is the clipping factor, which is generally 0.2.

[0232] The output of the large model inference module includes the following:

[0233] Recommendation results: For example, "Slurry shield is recommended, with a cutterhead opening ratio of 34% and a main thrust of 85,000 kN";

[0234] Explanation of the reasoning process: For example, “the geological permeability is high and the groundwater pressure is high, so the soil in front needs to be stabilized”;

[0235] Reference rule number: such as "Reference rules R1-001, R1-003, R3-020";

[0236] Recommendation level / confidence: such as "strong recommendation (confidence 0.92)".

[0237] All results can be converted into PDF documents and JSON structures, and directly connected to the design review system and construction platform; Figure 5 As shown, Figure 5 This section shows the structured reconstruction process after the model generates the selection results, including the generation path and logic of the recommended type, parameter table, justification, rule references, and output formats (PDF, JSON, Excel).

[0238] An example of implementing the large model reasoning module: For a river-crossing tunnel in a certain city, the input geology is "full-section saturated sand layer", the groundwater pressure is 0.42 MPa, and the settlement is required to be controlled within 10 mm.

[0239] The system automatically identifies it as a "high permeability, water-rich formation" and the matching rules are as follows:

[0240] Permeability coefficient > 1e-4 → Slurry shield is recommended;

[0241] Groundwater pressure > 0.3 → Slurry shield is recommended;

[0242] ·Settlement control → can take earth pressure shield into consideration.

[0243] Large model inference output:

[0244] The recommended tunnel boring machine was a slurry shield, configured with a 35% opening ratio and a total thrust of 88,000 kN. Considering the high permeability of the strata and the stability of the front support, the slurry shield could maintain pressure balance, reducing the risk of gushing and ground subsidence. Experts confirmed the validity of their conclusions, reducing the selection time from the traditional three hours to five minutes, and demonstrating the model's stable and reliable performance.

[0245] The result generation module described in the embodiment is the output terminal module of the system of the present invention, which is responsible for structural processing and visual expression of the selection suggestions, reasoning instructions, parameter configuration and other information output by the large model reasoning module, forming standardized engineering documents and data interface results, so that engineering designers can view, download, archive or integrate them into the construction management system. Figure 6 As shown, Figure 6 The interface shows how users can view multiple selection options, score them, and modify and confirm recommendations. It also supports expert commentary, option ranking, structural sensitivity analysis, and result regeneration. The module supports multiple output formats, document templates, extensible interfaces, and data interoperability with mainstream CAD / BIM platforms, demonstrating excellent system integration capabilities and user-friendly interaction. Serving as a bridge between the system and actual engineering practice, this module achieves a complete closed-loop from intelligent reasoning to engineering documentation, enhancing the visualization, standardization, and engineering feasibility of the selection process, and is a key embodiment of the engineering value of intelligent selection systems.

[0246] The output of the result generation module includes the following five categories of information, which are presented in the form of "structured data + text explanation + standard reference", as shown in the following table:

[0247]

[0248] The result generation module ensures that the selection results have the characteristics of "result clarity + reasoning explainability + specification traceability". This module supports the following four standardized output formats for users to select or batch export:

[0249] (1) PDF engineering report (standard format), with the cover including the project name, generation time, and engineering overview; the main body including the recommended type, parameter list, reasoning explanation, reference rule number, chart illustrations, etc.; the attachments may include CAD diagrams, risk warnings, and expert review opinion areas.

[0250] (2) Excel data table (comparison of multiple schemes), used for comparison and ranking of different shield configuration schemes; each column is a scheme, and each row is a key parameter field; support color marking of the best recommended scheme.

[0251] (3) JSON structured data (API output) is used to integrate into the China Railway 11th Bureau shield construction platform, BIM / CIM platform or its own database; format example:

[0252]

[0253] (4) SCR / CAD drawing data (visual layout). If the user triggers the site layout function while selecting the model, the shield machine layout position, path line and construction boundary map will be automatically generated; the output is in DWG or SCR format, supporting AutoCAD loading or script playback;

[0254] When multiple candidate results are generated, the result generation module generates a confidence score for each recommended selection result. The confidence score calculation method is as follows:

[0255]

[0256] Where: C is the confidence score; p i is the probability of the model on this judgment; c i Score the effectiveness of the expert assessments.

[0257] For example, if two candidate results are inferred at the same time, Candidate 1: Earth Pressure Shield and Candidate 2: Slurry Shield, the confidence scores of the two candidate results can be calculated using the above formula, and the selection result with the higher score is output:

[0258] Candidate 1: Earth Pressure Shield → C = 0.86;

[0259] Candidate 2: Slurry shield → C = 0.91 → final output.

[0260] The result generation module in the embodiment can also evaluate and control the results. After the recommended final selection results are output, users can provide feedback and scores based on the accuracy, practicality, interpretability and other dimensions of the results, which can be used for subsequent model optimization. Users can score and evaluate the selection results as shown in the following table:

[0261] Evaluation Dimensions content Rating level accuracy Whether the recommendation meets the actual project needs and geological conditions ★～★★★★★ Practicality Whether the parameter configuration is feasible and meets the capacity of the construction equipment ★～★★★★★ Explainability Whether the reasoning is complete and the citation rules are clear ★～★★★★★ Risk control capabilities Whether the recommendation has reasonable anticipation of abnormal geology or special restrictions ★～★★★★★

[0262] The score will be fed back into the "reinforcement learning reward function" for model optimization.

[0263] The result generation module also has a graphical user interface function, providing the following interactive capabilities:

[0264] Support "one-click report generation" and "custom field output";

[0265] Supports filling in suggestions and modification records in the "Expert Comments Area";

[0266] Supports the generation of “multi-scheme switching comparison” and “key parameter sensitivity analysis” charts;

[0267] Supports importing historical reports for comparison, version control and archiving.

[0268] Examples of application scenarios of the result generation module in the embodiment:

[0269] In a Wuhan subway shield tunneling project, the system automatically imported a geological report (Excel format) provided by the survey unit. After identifying the input parameters, it inferred and recommended the use of a "slurry shield" tunneling system. The parameters were as follows:

[0270]

[0271]

[0272] The system outputs a standard PDF report (4 pages), including a cover page, parameter table, reasoning explanation, rule number, construction precautions, etc. Review experts can mark modification suggestions online in the "Annotation Area".

[0273] The result generation module in this embodiment can be connected to the China Railway 11th Bureau's "Shield Tunneling Intelligent Construction Platform" via a RESTful interface. This allows for the following application scenarios: directly using selection results as input into the construction plan approval process; connecting structured data to the construction BIM system and scheduling system; storing documents on a document cloud platform for archiving and retrieval; and receiving real-time feedback from the platform to support subsequent result optimization and tracking. The interface documentation provides complete field descriptions, authentication mechanisms, sample call code, and other technical information to facilitate system integration and secondary development.

[0274] To ensure the security and compliance of engineering data, the result generation module supports the following configurations: automatic addition of electronic signatures and watermarks to output documents; setting permission control and logging for interface calls; support for synchronization with local server deployment to achieve localized archiving of results; and generation of unique hash values ​​for all result documents for tamper-proof verification.

[0275] The feedback and model optimization module in the embodiment is the core component of the system of the present invention to achieve "continuous learning and dynamic updating". It is responsible for collecting the user's evaluation and revision feedback on the selection results in the actual engineering application process into the system, and converting it into model optimization samples or knowledge base update basis through preset rules and manual review mechanism, thereby realizing a closed-loop mechanism of knowledge accumulation-reasoning optimization-performance enhancement. This module integrates active feedback entry, automatic scoring system, sample storage structure and reward function construction algorithm, supports the parallel operation of reinforcement learning (RLHF) and expert consensus mechanism, and is a key link to ensure the long-term engineering adaptability and professional credibility of the system. Specifically, Figures 7 to 10 As shown, Figure 7 Demonstrates how the system collects users' explicit ratings and implicit behavioral feedback information and converts them into the standard structure of reinforcement learning training samples, including input vectors, raw outputs, rating items, and text suggestion fields. Figure 8 The representation system uses a large language model reinforcement learning process driven by human feedback, including reward function construction, policy update, gradient clipping mechanism and multi-round sample training mechanism. Figure 9 This diagram illustrates the evolutionary process of when consistent revisions occur in high-frequency feedback across multiple projects, the system automatically mines potential rules, generates candidate rule entries, and writes them into the knowledge base after submitting them for expert review. Figure 10 It shows how this system can be integrated into the existing shield intelligent construction platform through the API interface to achieve digital management of the entire process from engineering survey, design selection, construction simulation to equipment scheduling.

[0276] The feedback and model optimization module supports multiple forms of user feedback input, including explicit feedback and implicit feedback.

[0277] After reviewing the selection recommendations, users manually evaluate the results based on the following indicators:

[0278] Feedback describe Rating range Recommendation accuracy Are the recommended shield types and parameters reasonable? 0-5 points Construction feasibility Whether the recommended solution can be actually applied in the project 0-5 points Reasonableness of reasoning Whether the reasoning process is complete and whether the reference rules are appropriate 0-5 points Express professionalism Whether the terminology and language conform to engineering standards 0-5 points

[0279] The feedback and model optimization module automatically collects the following data behaviors: whether the user adopts the solution recommended by the model; whether the user manually modifies the model's recommended content in the document; whether the user re-triggers the re-inference operation under the same input conditions; and the expert revision report version submitted by the user on the platform.

[0280] All feedback information is standardized and stored as structured sample data for training the reinforcement learning reward model. The sample format is as follows:

[0281]

[0282] The module parses and labels newly added feedback samples on a daily basis and adds them to the model training sample pool. The feedback and model optimization module in this embodiment adopts the following two mainstream optimization mechanisms:

[0283] (1) Reinforcement learning optimization (RLHF), using user ratings to construct a reward function as follows:

[0284] R=α·Accuracy+β·Usability+γ·Reasoning

[0285] Where α, β, and γ are weighting coefficients (e.g., 0.5, 0.3, and 0.2).

[0286] The user feedback reward function constructed in this module serves as an important source of scoring for reinforcement learning samples and will be used to design the objective function of the large model in subsequent policy training. The score and label of the reward function will be passed to the large model inference module to guide the model toward high-quality output and drive the model parameters to migrate toward "high-scoring samples."

[0287] (2) Rule knowledge update mechanism: For new patterns that appear in high-frequency, repeated, and excellent scoring feedback, the system automatically analyzes their logical characteristics and generates candidate rules; the candidate rules are presented to experts for review in an artificial language structure. If they pass, they are written into the rule knowledge base and marked with the source (such as "Feedback learning from the 2025-WH05 project"); the system forms a "rule learning log" to achieve traceable update records.

[0288] After each fine-tuning or reward optimization of the large model, the Feedback and Model Optimization module automatically saves a traceable version and has the following capabilities:

[0289] Version freeze: The current best model can be locked for use in the production system to prevent the test model from affecting the project;

[0290] Version rollback: If the new model does not work well, you can restore to the old version with one click;

[0291] Version comparison: Supports A / B testing of different models on the same task. The system records indicators such as accuracy and acceptance to assist in judging the effectiveness of the versions.

[0292] Examples of engineering application and evolution of the feedback and model optimization module in the embodiment:

[0293] The initial model output for a project recommended the use of a slurry shield with a 34% opening ratio and a main thrust of 86,000 kN. However, the site's geology featured highly plastic clay interbedded with soft sand, necessitating extremely stringent settlement control requirements. Experts recommended the use of an earth pressure shield (EPP) instead, reducing the opening ratio to 28% for better settlement control. This feedback was then fed into the feedback and model optimization module, where the expert feedback was rated as follows: Accuracy: 3 / 5; Feasibility: 2 / 5; Reasonableness of Reasoning: 3 / 5.

[0294] The feedback and model optimization module collects the feedback samples and analyzes and modifies the logic;

[0295] The influence of feedback samples decays exponentially over time, and the following function is applied during model iteration:

[0296] w(t)=w0·e -λt

[0297] Where: w(t) is the current sample influence weight; w0 is the initial weight; λ is the decay coefficient; t is the sample time distance (days).

[0298] In the next round of model fine-tuning, the system will include this sample and adjust the recommendation strategy for the same type of parameter combination. When similar conditions are subsequently entered, the recommendation will be changed to:

[0299] “We recommend an earth pressure shield, which is suitable for soft soil areas with high settlement sensitivity. The cutterhead opening ratio should be controlled at 28% to 30%.”

[0300] To prevent model degradation caused by erroneous feedback, the feedback and model optimization module also introduces the following safeguards: all model changes must be confirmed by experts or evaluated on an authoritative test set; each feedback sample must undergo two levels of verification before training: automatic verification + manual spot check; all feedback operations are recorded for auditing and accountability tracking.

[0301] The feedback and model optimization module in the embodiment establishes a "dialogue bridge" between the model and engineers; transforms expert experience into continuously updated "digital rules"; drives the model to automatically adapt to the selection requirements of different regions and projects; and improves the system's robustness and engineering adaptability in complex geological and non-standard scenarios.

[0302] The large language model-based intelligent shield machine selection system provided in the embodiments integrates an engineering rule knowledge base, deep language model reasoning technology, and a feedback learning mechanism to achieve rapid, accurate, and intelligent selection of shield equipment types and key parameters under complex geological conditions. The system operates through a closed-loop process of "input-inference-output-optimization" and provides a large language model-based intelligent shield machine selection system, specifically including the following steps:

[0303] S1. Engineering parameter collection and standardization: The system first receives the input parameters required for the project through the data input module, which mainly include: geological conditions (such as stratum type, permeability coefficient, groundwater pressure), tunnel design parameters (such as excavation diameter, longitudinal slope, minimum curve radius) and construction constraints (such as surface settlement control value, advancement speed limit), etc.; these data are converted into structured vector input through standardization algorithms (such as Z-score normalization) and engineering semantic dictionary conversion.

[0304] S2. Rule knowledge retrieval and conditional constraint reasoning: The system calls the built-in structured selection rule knowledge base, performs multi-condition matching and logical constraint retrieval on the standardized input, and obtains a series of candidate rule sets. These rules express engineering experience and standard terms in the form of "IF-THEN", such as "Permeability coefficient>1×10 -4 m / s→Recommend slurry shield", and form a parameter-condition-conclusion logical graph structure through graph database modeling, providing initial constraint prompts and multi-path fusion capabilities for large model reasoning.

[0305] S3. Large-scale model semantic understanding and selection reasoning generation: The condition combination and input parameters after knowledge base retrieval are jointly constructed into semantic prompts (Prompt), which are input into a large language model (such as DeepSeek / Qwen) that has been fine-tuned for engineering scenarios. The system performs complex multi-constraint reasoning generation tasks. Based on LoRA fine-tuning parameters, the model combines rule semantics, historical cases and reinforcement learning feedback data to automatically generate the most suitable shield type (such as slurry / earth pressure shield), key configuration parameters (such as cutterhead opening ratio, main thrust) and complete reasoning explanation text.

[0306] S4. Result Generation and Visualization: The result generation module receives the large model output and performs structural reconstruction, format conversion, and multi-format export (supporting PDF reports, Excel spreadsheets, JSON interfaces, etc.). The output includes the recommended type, parameter list, reasoning explanation, reference rule number, and recommendation confidence score. The system also supports AutoCAD scripts and construction platform data integration interfaces, facilitating direct access and archiving by design teams.

[0307] S5. Continuous model optimization driven by user feedback: The system provides a user feedback mechanism that supports explicit scoring (accuracy, feasibility, and explainability) and implicit behavior collection (whether suggestions were adopted, whether manual modifications were made). All feedback information will be stored as structured training samples, and combined with reinforcement learning (PPO) to build a reward function to update large model parameters. If repeated high-confidence feedback corrections are found, the system automatically generates rule candidates and enters the rule knowledge base review process, establishing a dynamically evolving "engineering feedback-rule update-model reinforcement" learning loop.

[0308] The system and method of the present invention, through the ternary fusion architecture of "rule knowledge + large model reasoning + expert feedback", enables the selection of shield equipment to shift from "manual experience decision-making" to an intelligent, standardized and interpretable new paradigm of "data-driven - knowledge-guided - model generation". It has high adaptability, high efficiency and continuous evolution capabilities, and has broad engineering promotion and application prospects in complex geological tunnel construction.

Claims

1. A shield machine intelligent auxiliary selection system based on a large language model, characterized by: The selection system includes a data input module, a rule knowledge base module, a large model reasoning module and a result generation module; The data input module is used to uniformly collect various original parameters related to shield machine selection in engineering projects, and format and standardize the collected different types of original parameters to ensure the consistency and solvability of different data sources, and provide structured and computable input data for subsequent knowledge rule retrieval and large model reasoning; The rule knowledge base module is used to convert the decision logic in traditional expert selection experience, standard specification clauses and historical engineering cases into structured, searchable and inferable rule set data, and input it into the large model reasoning module for combined with the large language model to assist in reasoning; The large model reasoning module is the core intelligent decision-making unit of the system, which is used to semantically fuse the structured and computable input data input by the data input module with the rule set data input by the rule knowledge base module, and output the optimal shield selection plan and configuration parameters through the reasoning and generation capabilities of the large language model; The result generation module is used to structure and visualize the selection suggestions, reasoning instructions, and parameter configuration information output by the large model reasoning module to form standardized engineering documents and data interface results, so that engineering designers can review, download, archive, or integrate them into the construction management system.

2. The shield machine intelligent auxiliary selection system based on a large language model according to claim 1 is characterized by: The input data accepted by the data input module are divided into four categories based on the key influencing factors of shield selection, specifically including geological parameters, design parameters, construction constraints and engineering background. The geological parameters include stratum type, permeability coefficient, groundwater pressure, powder content, stratum hardness, stratum particle size and friction coefficient; the design parameters include excavation diameter, tunnel longitudinal slope, minimum curve radius, burial depth, hard soil thickness and linear structure form; the construction constraints include surface settlement control value, advancement speed requirement, noise / vibration limit level and operation time limit; the engineering background includes project type, regional characteristics, risk level, implementation location and policy restrictions / approval basis; the above data types include enumeration type, floating point number and integer type; The process of formatting and standardizing the collected original parameters of different types by the data input module is as follows: (1) Enumeration field mapping: For the parameters of the text description field, a preset vocabulary and keyword matching mechanism are used to map them to standard engineering codes to facilitate subsequent rule matching; (2) Numerical normalization: For parameters in numerical fields, the input is uniformly processed and normalized in SI units to improve the generalization ability of large models. The details are as follows: Where x is the original input value, x′ is the normalized value, μ and σ are the historical mean and standard deviation of the parameter respectively; (3) Calculate the cross-sectional area of ​​the tunnel excavation to provide an area basis for subsequent thrust calculations. The cross-sectional area of ​​the tunnel excavation is calculated as follows: Where: A is the cross-sectional area; D is the excavation diameter; (4) Calculation of geological comprehensive classification scores: some geological conditions are automatically classified into high / medium / low complexity ranges using a weighted integral model. The calculation method is as follows: Where: G is the total score; w i is the weight of the i-th parameter; x i is the value of the i-th indicator; After the data input module completes the standardization processing, it will generate a unified input vector structure, including numerical tensors and semantic labels; and pass it into the rule knowledge base retrieval module to match the rule entries that meet the conditions, and then pass it into the large model reasoning module to generate semantic prompts and constraints for LLM reasoning.

3. The shield machine intelligent auxiliary selection system based on a large language model according to claim 1 or 2, characterized in that: The selection system also includes a feedback and model optimization module, which is used to collect user evaluation and revision feedback on the selection results of the result generation module during actual engineering applications, and convert it into model optimization samples or knowledge base update basis through preset rules and manual review mechanisms, thereby realizing a closed-loop mechanism of knowledge accumulation-reasoning optimization-performance enhancement; the feedback and model optimization module integrates an active feedback entry, an automatic scoring system, a sample storage structure and a reward function construction algorithm, supporting the parallel operation of reinforcement learning optimization and expert consensus mechanism; The feedback and model optimization module constructs a user feedback reward function, which serves as an important scoring source for reinforcement learning samples and will be used to design the objective function in subsequent strategy training of the large language model. The rating and label of the user feedback reward function are passed to the large model inference module to guide the model toward high-quality output. The user feedback reward function is a reinforcement learning reward function constructed using user scores. It acts on the objective function in the strategy optimization process and drives the parameters of the reinforcement learning reward model to migrate towards "high-scoring samples"; The large model reasoning module introduces a reinforcement learning mechanism based on human feedback and adopts a PPO optimization strategy; The reinforcement learning mechanism uses a reward function. The reward function in the large model inference module is derived from the "user evaluation index system" constructed in the feedback and model optimization module and formalized into a policy optimization function. The reward function is constructed by mapping user rating labels to structural scoring factors. This function serves as the fitting basis for the target distribution in PPO training and is tailored and optimized in combination with the old policy ratio to improve the professionalism and practicality of the model output. The reinforcement learning reward function is as follows: R = α·Accuracy score + β·Term matching rate + γ·Structural completeness Among them, α, β, and γ are weighting coefficients; The update target of the PPO optimization is: Where: r t (θ) is the ratio of new and old strategies; is the advantage function; ∈ is the clipping factor, which is generally 0.

2.

4. The shield machine intelligent auxiliary selection system based on a large language model according to claim 1 or 2 is characterized by: The rule knowledge base module undertakes three major functions in the system: knowledge expression, rule reasoning and model constraint. It is the key interface for realizing the integration of engineering reliability and model intelligence. The rule knowledge base module supports the logical expression of "condition → conclusion" and provides rule retrieval capabilities based on attribute matching, conditional relationship modeling capabilities based on graph structures, and semantic parsing capabilities compatible with natural language input. Its underlying architecture is implemented using a hybrid "rule table + graph database + semantically enhanced index" architecture, serving as the retrieval end in the RAG framework and working in collaboration with the large language model. The rule knowledge base module supports two rule update methods: manual editing and updating, and model reverse deduction and extraction. It also features a confidence evolution mechanism. If a rule receives continuous feedback of "low applicability," its weight will automatically decrease; otherwise, it will be increased, forming a dynamic confidence adjustment capability. The rules in the rule knowledge base module are generally divided into the following categories: shield type recommendation rules (classification number R1), cutterhead configuration rules (classification number R2), propulsion system rules (classification number R3), grouting and slurry system rules (classification number R4), and comprehensive adaptability rules (classification number R5). Each rule also contains the following additional metadata: source, applicability level, credibility, and whether it is expert consensus. Each rule has the following structure: Where: R i is the i-th rule; C n is the nth condition; Conclusion is the inference conclusion; When multiple rules conflict or compete with each other, the graph edge weight scoring function is used to fuse the confidence and calculate the recommendation score. The specific calculation is as follows: Where: S j is the composite score of solution j; α i is the rule weight; r ij is the score of rule i for solution j. The rule knowledge base module uses knowledge graph technology to construct a "parameter-rule-conclusion" triple network, where nodes represent parameters or conclusions and edges represent logical constraints. When multiple parameter combinations trigger different rules, the system performs comprehensive reasoning using a graph traversal + confidence fusion algorithm to output the optimal conclusion and its confidence score. The rule knowledge base module is provided with an interactive mode with the large language model and provides fact constraints and knowledge basis for the large language model reasoning module. The specific process is as follows: ①User inputs structured parameters; ② Retrieve all rules that meet the conditions in the matching knowledge base; ③Build a prompt template example: ④ The large model integrates parameters, rules and cases to generate recommendation reasons and configuration plans.

5. The shield machine intelligent auxiliary selection system based on a large language model according to claim 1 or 2, characterized in that: The large model inference module performs secondary optimization based on the pre-trained large language model, which includes the domestic large models of DeepSeek, Qwen, and Baichuan. The core structure of the large model inference module includes the following: Prompt parsing and construction unit: This unit is used to combine input parameters and rule text into natural language prompts as model input. This unit designs a structured prompt template that combines the characteristics of natural language and structured tags to stimulate the contextual reasoning and professional expression capabilities of large models. Reasoning generation model: used to load the fine-tuned LLM, perform selection reasoning, solution generation and text output; Strategy evaluation and reinforcement learning module: used to evaluate the quality of model output results based on domain feedback and optimize model parameters and weight distribution; The processing flow of the large model reasoning module is: structured input → Prompt assembly → large model reasoning → output shield plan + explanation text.

6. The shield machine intelligent auxiliary selection system based on a large language model according to claim 2 is characterized by: The data input module has both parameter interface compatibility and data semantic integrity, and can connect to the geological data systems of design units, survey units, and construction units, enabling one-click import or manual entry of various original parameters related to shield selection in engineering projects; The data input module supports multiple input methods, and the front-end graphical interface inputs project parameters through selection boxes, sliders and data forms; Data file import supports CSV / Excel format to import survey unit report form; The API interface is connected to the BIM platform and the shield construction platform, and can call geological data services in real time; the interface design adopts the RESTful protocol, and the transmission format of all parameters is unified into the JSON structure. The data input module also features outlier detection and fault tolerance mechanisms. If a permeability coefficient input is negative or exceeds a reasonable range, the system automatically prompts a prompt. If a geological field is empty but designated as a "complex site," the system prompts you to complete detailed parameters or enable the model's self-explanatory function. If a value is missing, the system can make a default inference based on the entered field or invoke a historical case completion mechanism.

7. The shield machine intelligent auxiliary selection system based on a large language model according to claim 3 is characterized by: The feedback and model optimization module supports user feedback input types including explicit feedback and implicit feedback. All feedback information collected by the feedback and model optimization module is standardized and stored as structured sample data for training the reinforcement learning reward model. The feedback and model optimization module parses and labels newly added feedback samples on a daily basis and adds them to the model training sample pool. The feedback and model optimization module has a rule knowledge update mechanism. For new patterns that appear in high-frequency, repeated, and high-scoring feedback, the module automatically analyzes their logical characteristics and generates candidate rules. The candidate rules are presented to experts for review in an artificial language structure. If approved, they are written into the rule knowledge base and marked with the source. The module forms a "rule learning log" to achieve traceable update records. After fine-tuning or reward optimization of the large model language, the feedback and model optimization module automatically saves the traceable version and has version freezing, version rollback and version comparison functions; the feedback and model optimization module sets up security mechanisms and expert review guarantees.

8. The shield machine intelligent auxiliary selection system based on a large language model according to claim 3 is characterized by: The result generation module supports multiple output formats, document templates, extensible interfaces, and data interoperability with mainstream CAD / BIM platforms. Its output includes the following five categories of information: recommended shield type, key configuration parameters, reasoning process description, rule and model references, and structured report data, all presented in a unified format of "structured data + textual explanation + specification reference." The module also features a graphical user interface and outputs the following standard formats: PDF engineering reports, Excel data tables, JSON structured data, and SCR / CAD drawing data. When the large model inference module outputs multiple candidate results, the result generation module assigns a confidence score to each recommended candidate result and selects the recommended result with the highest confidence as the final selection result. The confidence score calculation method is as follows: Where: C is the confidence score; p i is the probability of the model on this judgment; c i Score the effectiveness of the expert assessments.

9. A shield machine intelligent auxiliary selection method based on a large language model, characterized by The shield machine intelligent auxiliary selection system based on the large language model described in claims 1 to 8 is used for selection, specifically comprising the following steps: S1. Engineering Parameter Collection and Standardization: The data input module receives various raw parameters related to shield selection in the project, including geological conditions, tunnel design parameters, construction constraints, and engineering background. These raw parameter data are converted into structured vector inputs using a standardized algorithm and an engineering semantic dictionary. S2. Rule Knowledge Retrieval and Conditional Constraint Reasoning: Using the structured selection rule knowledge base built into the rule knowledge base module, we perform multi-condition matching and logical constraint retrieval on standardized inputs to obtain a set of candidate rules. These rules express engineering experience and standard clauses in an "if-then" format, and are modeled through a graph database to form a parameter-condition-conclusion logical graph structure, providing initial constraint prompts and multi-path fusion capabilities for large-scale model reasoning. S3. Large-scale model semantic understanding and selection reasoning generation: The combination of conditions retrieved from the knowledge base and the input parameters are combined to form semantic prompts. These prompts are then fed into a large language model fine-tuned for engineering scenarios to perform complex multi-constraint reasoning generation tasks. Based on LoRA fine-tuned parameters, the large language model combines rule semantics, historical cases, and reinforcement learning feedback data to automatically generate the most suitable shield type, key configuration parameters, and complete reasoning explanation text. S4. Result Generation and Visualization Output: The result generation module receives the output of the large language model and performs structural reconstruction, format conversion, and multi-format export. The output content includes the recommended type, parameter list, reasoning description, reference rule number, and recommendation confidence score. The result generation module also supports AutoCAD scripts and construction platform data integration interfaces, facilitating direct access and archiving by design units. S5. Continuous model optimization driven by user feedback: A user feedback mechanism is provided, supporting explicit scoring and implicit behavior collection. All feedback information will be stored as structured training samples, and combined with reinforcement learning to build a reward function to update large model parameters. If repeated high-confidence feedback corrections are found, the system automatically generates rule candidates and enters the rule knowledge base review process, establishing a dynamically evolving "engineering feedback-rule update-model reinforcement" learning loop.

10. The shield machine intelligent auxiliary selection method based on a large language model according to claim 9 is characterized in that: In the S3 step, fine-tuning of the large language model uses the following two methods: (1) LoRA: Insert the low-rank matrix A, BA, BA, B into the attention layer of the original model and only fine-tune AAA and BBB. The calculation expression is: ΔW=A·B, where rank(A), rank(B)<<dim(W) Use PEFT framework to load LoRA module; (2) SFT: supervised learning is performed using a ternary sample set of "engineering input - output selection result - reasoning explanation"; the loss function is the typical cross entropy loss: in: is the loss value; P θ Output probability for the model.

Images