Tunnel engineering carbon emission accounting method based on large model agent

Abstract

The application discloses a tunnel engineering carbon emission accounting method based on a large model agent, takes the large model agent as a dispatching center, classifies layer by layer after receiving non-standard forms, adopts a double-layer funnel architecture of "vector primary screening + large model reasoning" to prevent redundancy and deduplication for open classification, constructs four-dimensional resource coding with self-growth topology and binds with physical position index, normalizes and marks abnormalities, analyzes and generates five-dimensional space coding, triggers differentiated matching routing based on total classification, respectively executes carbon emission factor matching of building material double track and mechanical single track, for abnormal items, the agent calls external API to retrieve and generate traceability evidence cards for review and backfilling, finally executes classification accounting and outputs a detailed table, constructs an OLAP multidimensional data cube, and dynamically generates customized charts by analyzing natural language by the agent. The application realizes intelligent accounting, full-link traceability and customized output of tunnel engineering carbon emission.

Description

Technical Field

[0001] This invention relates to the field of carbon emission accounting and analysis technology for tunnel engineering, and in particular to a method for carbon emission accounting in tunnel engineering based on a large model intelligent agent. Background Technology

[0002] Carbon emission accounting provides data to identify key emission reduction areas and define baselines, serving as the foundation for advancing emission reduction efforts and verifying their effectiveness. Due to its massive scale and extremely high material and energy consumption density, tunnel engineering has become a key area for low-carbon infrastructure construction. Because tunnel construction involves a wide variety of building materials and machinery, and complex procedures, some scholars and teams have developed relevant carbon emission accounting systems to achieve automated calculations. Currently, carbon emission accounting systems for tunnel engineering and related highway and railway engineering projects mainly fall into two categories: those that input data into Excel or cost estimates, automatically match carbon emission factors based on name or number using a built-in carbon emission factor library, and calculate carbon emissions; and those that input data into a digital model, automatically extract relevant parameters, and calculate carbon emissions using a built-in carbon emission factor library.

[0003] However, due to the severe non-standardization of data in actual engineering projects and the emergence of new materials and processes, existing systems struggle to achieve dynamic self-growth of the coding system, making the matching of carbon emission factors a major challenge. Therefore, such accounting software generally requires strict adherence to requirements for the coding, naming, and table structure of Excel form entries, or output files from specific cost estimation software, to ensure the matching of form entries with carbon emission factors. However, organizing these non-standard forms into structured forms is itself a major source of workload in traditional carbon emission accounting. While automatic parameter extraction from digital models can ensure the matching of carbon emission factors, it still only performs calculations on pre-defined entries and requires relatively detailed modeling to ensure accuracy. Furthermore, the fixed output format and rigid interaction of these models make it difficult to support further detailed analysis.

[0004] As tunnel engineering evolves towards green, low-carbon, and refined management, the industry's demand for intelligent accounting methods that adapt to non-standard raw data, dynamically expandable accounting systems, and possess deep interactive analysis capabilities is becoming increasingly urgent.

[0005] CN120975409B discloses an intelligent analysis method for carbon emission data. This method utilizes element extraction and an alias ledger to achieve a unique mapping between name normalization and carbon emission factors. By generating a chain of evidence objects, it ensures the traceability of the accounting process, enabling intelligent calculation and analysis of carbon emission data in bill of exchange projects. However, the successful matching of carbon emission factors is highly dependent on the richness of the alias ledger. Slight changes in aliases or entirely new goods will prevent name normalization and carbon emission factor matching, severely impacting accounting efficiency.

[0006] CN117972157A discloses a rapid carbon emission query system for shield tunnels based on cost and carbon emissions. This system uses carbon emission classification codes and material and equipment list codes to enable rapid querying of carbon emission indicators for shield tunnels. However, it still relies on the traditional list method and does not solve problems such as the difficulty in matching carbon emission factors.

[0007] In addition, CN120994930A discloses a system and method for calculating carbon emissions in tunnel sections based on AI smart safety helmets, and CN116894068B discloses a method for calculating and visualizing carbon emissions in highway tunnels based on digital models. These methods rely on expensive hardware sensors or cumbersome detailed modeling and can only include major building materials, making it difficult to fundamentally solve the problems of intelligent matching of carbon emission factors and adaptive cleaning of non-standard data.

[0008] With the breakthroughs in complex semantic parsing and logical reasoning capabilities of Large Language Models (LLMs), and the gradual maturation of LLM agent technology, they have gained the ability to understand unstructured engineering text and autonomously schedule tools, providing a new technical feasibility for breaking down the barriers of traditional rule engines. This is an area that this application needs to focus on improving. Summary of the Invention

[0009] The technical problem this invention aims to solve is to provide a method for calculating carbon emissions in tunnel engineering based on a large-scale intelligent model, achieving intelligent calculation, end-to-end traceability, and customized output of carbon emissions from tunnel engineering.

[0010] To address the aforementioned technical issues, this invention provides a method for calculating carbon emissions in tunnel engineering based on a large-scale intelligent agent. Using the large-scale intelligent agent as the scheduling hub, it receives non-standard forms and categorizes them layer by layer. For open classifications, a two-layer funnel architecture of "vector initial screening + large-scale model inference" is employed to prevent redundancy and deduplication. A four-dimensional resource code with a self-growing topology is constructed and bound to the physical location index. After normalization and anomaly marking, a five-dimensional spatial code is generated. Based on the overall classification, differentiated matching routing is triggered, and carbon emission factor matching is performed for both building material dual-track and mechanical single-track systems. For abnormal entries, the agent calls an external API to retrieve and generate traceability evidence cards, which are then reviewed and reintroduced. Finally, the classification calculation outputs a detailed table, constructing an OLAP multi-dimensional data cube. The agent parses natural language to dynamically generate customized charts, achieving automated processing and intelligent analysis from raw forms to structured results. The specific steps include: S1. The intelligent agent receives the input form, uses a large language model to determine the physical location index of the item name and specification, performs line-by-line matching and classification, constructs a four-dimensional resource category code with a self-growing topology, and binds the four-dimensional resource category code as a structured label to the physical location index in real time. S2. Extract feature parameters from the input form and perform heterogeneous data normalization. Merge the entries and physical location indexes according to the four-dimensional resource category code, and set pending processing marks for abnormal entries with missing parameters. S3. Parse the explanatory text to lock the background database version and generate a five-dimensional engineering space code. Perform differentiated matching routing based on the total classification information of resource categories, complete carbon emission factor matching, and generate a preliminary structured form. S4. For entries marked with the pending flag, trigger the large model agent to perform an external API search based on a preset whitelist. After manual review and verification, the data is fed back and the structured form is updated. S5. Call the deterministic calculation module to perform classification and accounting on the structured form based on the carbon emission factors, and assemble and output a standardized analysis detail table and a carbon emission composition diagram in a preset format by passing through the baseline field and the dynamically derived accounting field. S6. Based on the preset multidimensional coding system, an online analysis and processing multidimensional data cube is constructed. The large model agent parses the user's natural language analysis instructions, and the dynamic scheduling analysis engine performs multidimensional aggregation operations on the multidimensional data cube and generates customized charts.

[0011] In step S1, after receiving the input form, the agent uses the semantic understanding capability of the large language model to determine the column-level physical location index of the form name and specifications based on the table header. Then, it parses the form row by row and constructs a four-dimensional resource category code with a self-growing topology, specifically including: Following the order of general classification (building materials / machinery) - primary classification - secondary classification - tertiary classification, the extracted non-standard item features, namely names and specifications, are compared and classified layer by layer with the standard classification names in the background database. After each layer of classification, the corresponding classification code for that layer is determined according to the preset coding system. The first three layers are closed-loop classifications: the general category has only two subcategories, building materials and machinery, and both the first-level and second-level categories have an "other" subcategory; the large model agent is directly invoked to execute logical routing and distribution based on domain knowledge; The third-level classification does not have an "Other" category. When all third-level classifications cannot be adapted, the "Name + Specification" is used as the new third-level classification name, and its code is set as "End Code + 1" to achieve a self-growing topology of the code.

[0012] Preferably, for open classification of the three-level categories, a two-layer funnel architecture based on "lightweight vector initial screening and deep reasoning of a large language model" is switched to prevent redundancy and deduplication. Specifically, the text embedding model is called to calculate the cosine similarity between the features of the item to be processed and the existing three-level categories in the background database; if the cosine similarity is higher than a first preset threshold, it is determined to be a synonymous heterogeneous item and directly merged into the existing code; if the cosine similarity is in the fuzzy range, the large model agent is triggered to perform logical reasoning in combination with the context; only when the reasoning result is negative or the cosine similarity is lower than a second preset threshold is it determined to be a brand new material, and its "name + specification" is used as a new three-level category name and assigned a new code with incrementing suffix, realizing the self-growing topology of the coding system.

[0013] Once the code is determined, it is immediately used as a structured tag and bound one-to-one with the row-level physical location index of the current entry in the original form in real time. The code is written into the mapping dictionary and stored persistently to provide underlying traceability evidence support for subsequent audits.

[0014] In step S2, when merging entries based on the four-dimensional resource category encoding as the aggregated primary key, the row-level position indices that originally belonged to each independent entry are simultaneously aggregated into an array.

[0015] For entries lacking transportation characteristic parameters, a prompt will appear to upload relevant data; otherwise, the default values ​​will be filled in using industry standards.

[0016] For entries that lack units or key data parameters, the "to be processed" label is automatically applied.

[0017] In step S3, the generated preliminary structured form includes the following core data fields: five-dimensional engineering space code, four-dimensional resource category code, standard classification name, standard unit, standard quantity, transportation distance, carbon emission factor, and transportation carbon emission factor.

[0018] Preferably, a global mapping dictionary mechanism is used when matching transportation carbon emission factors: the transportation mode field of all entries is extracted and hashed to generate a transportation feature set to be matched; the large model agent takes over the feature set, combines semantic understanding to align it with the standard transportation modes in the background database, establishes a mapping index from non-standard transportation descriptions to standard transportation codes, and fills in transportation carbon emission factors in batches accordingly.

[0019] In step S4, the external knowledge search is performed based on a preset authoritative database whitelist mechanism: the large model agent initiates a structured query to the whitelist source, extracts and completes the numerical values ​​while simultaneously capturing metadata, including the data source institution, release time, applicable boundaries and original reference links, generates a structured "source traceability evidence card" and pushes it to the human-machine collaboration interface. Only after review and verification can the dictionary be fed back and the "pending processing" mark be removed, achieving audit-level full-link traceability.

[0020] In step S5, the carbon emissions from building material production and building material transportation are derived for the building material category, while only the carbon emissions from construction machinery are derived for the machinery category, and the calculation content codes are assigned as follows: building material production, building material transportation, and construction machinery.

[0021] The final generated analysis details table includes the following core data fields: five-dimensional engineering spatial code, four-dimensional resource category code, calculation content code, standard name, and carbon emissions (fixed unit).

[0022] In step S6, based on the completed online analytical processing multidimensional data cube, the analysis engine performs dynamic aggregation along multiple coordinate axes based on instruction requirements: specifically, it uses "sub-project - sub-section project - unit project - single project - project" as the engineering space axis, "third-level classification - second-level classification - first-level classification" as the resource category axis, and "calculation content (building material production / building material transportation / construction machinery)" as the carbon emission stage axis, and performs multidimensional roll-up, drill-down, single-dimensional slicing and multi-dimensional block cutting operations.

[0023] In addition to the preset background database, the above steps support custom uploads. However, when performing step S1, a standard encoding system is established for it, and it is automatically organized into the same structure as the preset background database.

[0024] The technical solution proposed in this invention has at least the following beneficial effects: 1) Achieved intelligent normalization and adaptive expansion of non-standard data: Breaking through the barriers of traditional rule engines, it pioneered a layer-by-layer classification coding method, combined with large model reasoning and a vector double-layer funnel deduplication mechanism, which not only avoids infinite redundancy in the coding library, but also achieves accurate, economical and self-growing coding system. 2) An auditable end-to-end evidence traceability system has been constructed: the classification code is strongly bound to the row-level physical location index of the original form, and a "traceability evidence card" based on a whitelist mechanism has been created for the first time, so that every introduction of external data is traceable and meets the requirements of engineering-level carbon emission audit; 3) It provides highly advanced natural language interaction and multi-dimensional analysis capabilities: It deeply integrates large model intent recognition with the underlying OLAP data cube, and can achieve dynamic slicing and drilling along the three major coordinate axes of engineering space, resource category and carbon emission stage with only natural language commands, which greatly improves the level of management intelligence. Attached Figure Description

[0025] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the encoding system construction and data flow logic provided in a specific embodiment of the present invention. Detailed Implementation

[0026] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The executing entity of a specific embodiment of the present invention is a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, mobile phone, etc., or an electronic device that implements the above functions.

[0027] like Figure 1 and Figure 2 As shown, this embodiment provides a method for carbon emission accounting in tunnel engineering based on a large model intelligent agent. This method uses a large model intelligent agent as the logical scheduling center, and runs through the entire process of data cleaning, factor matching, and multidimensional analysis, including the following steps: S1. The intelligent agent receives the input form 10, uses a large language model to determine the physical location index of the item name and specification, performs line-by-line matching and classification, constructs a four-dimensional resource category code with a self-growing topology, and binds the four-dimensional resource category code as a structured label to the physical location index in real time. S2. Extract feature parameters from input form 10 and perform heterogeneous data normalization. Merge entries and physical location indexes according to the four-dimensional resource category code, and set pending processing marks for abnormal entries with missing parameters. S3. Parse the explanatory text 20 to lock the version of the background database 30 and generate a five-dimensional engineering space code. Perform differentiated matching routing based on the total classification information of resource categories, complete carbon emission factor matching, and generate a preliminary structured form 40. S4. For entries marked with the pending flag, trigger the large model agent to perform an external API search based on a preset whitelist. After manual review and verification, the data is fed back and the structured form 40 is updated. S5. Call the deterministic calculation module, perform classification and accounting on the structured form 40 based on the carbon emission factors, and assemble and output the standardized analysis details table 50 and the carbon emission composition diagram in the preset format by passing through the baseline field and the dynamically derived accounting field. S6. Based on the preset multidimensional coding system, construct an online analytical processing (OLAP) multidimensional data cube. The large model agent parses the user's natural language analysis instructions and dynamically schedules the analysis engine to perform multidimensional aggregation operations on the multidimensional data cube and generate customized charts.

[0028] The specific implementation logic and technical details of the above core steps are as follows: S1. Hierarchical classification and coding, specifically implemented as follows: like Figure 2 As shown, the agent receives a raw input form 10 containing a header, item name, and specifications. First, it utilizes the semantic understanding capability of the Large Language Model (LLM) to determine the column-level physical location index of the item name and specifications based on the header. Then, it performs row-by-row parsing on the input form 10, matching the item features with the standard category names in the background database 30 layer by layer according to the hierarchical order of "general category (building materials / machinery) - primary category - secondary category - tertiary category", and gradually determining the encoding of each layer.

[0029] For the closed-loop classification of the first three layers (general category, first level, and second level), since there is an "other" fallback option, the large model agent is directly invoked to perform logical routing and distribution based on domain knowledge. When performing open-ended three-level classification, to avoid infinite redundancy in coding due to non-standard expressions, the classification mechanism automatically switches to a two-layer funnel architecture of "lightweight vector initial screening + deep reasoning by a large language model": Specifically, firstly, the text embedding model is called to convert the extracted non-standard names and specifications into high-dimensional semantic vectors, and the cosine similarity between them and the existing three-level classifications in the background database 30 is calculated. If the similarity score is higher than the first preset threshold of 95%, it is judged as a synonymous heterogeneous entry and directly merged into the existing code; if the score is in the fuzzy range of 70% to 95%, the large model agent is triggered to perform logical reasoning by combining the form context and knowledge of the tunnel engineering domain; only when the reasoning result of the large model is negative or the initial screening similarity is lower than the second preset threshold is it judged as a new material. For new materials, their "name + specification" is used as the new three-level classification name, and a new code of "current end code + 1" is assigned to them according to the rules to realize the self-growing topology of the coding system.

[0030] Once the encoding is determined, the four-dimensional resource category encoding is immediately used as a structured label, which is then bound in real time to the row-level physical location index of the current entry. This information is written into the mapping dictionary and stored persistently to provide underlying traceability evidence for subsequent audits.

[0031] S2. Heterogeneous data normalization, specifically implemented as follows: The agent determines the physical location index of the unit, quantity, and transportation characteristic parameters in the input form 10 based on the table header. It then automatically converts the original quantity to a standard quantity based on the corresponding standard unit in the background database 30. Subsequently, it merges the entries using the four-dimensional resource category code as the aggregation key, simultaneously aggregating the row-level location indices that originally belonged to each independent entry into an array. During the normalization process, entries lacking transportation characteristic parameters are filled with reasonable default values ​​according to industry standards; for abnormal entries lacking key data such as unit and quantity, they are automatically marked as "pending processing" and isolated and recorded in the pending processing queue.

[0032] S3. Carbon emission factor matching, specifically implemented as follows: like Figure 2 As shown, the agent receives and parses the unstructured explanatory text 20, extracts the geographical location and construction time features, and thereby locks the corresponding version number in the background database 30; at the same time, it determines the five-dimensional engineering space code containing "project-single project-unit project-sub-project-sub-item project" based on the explanatory text 20. In the matching process, differentiated matching routing is performed by parsing the "General Category" field in the four-dimensional resource category code: for entries with the general category "Building Materials", a dual-track parallel query is triggered, one track matches the corresponding carbon emission factor of building material production in the background database based on the complete four-dimensional resource category code, and the other track matches the corresponding carbon emission factor of building material transportation based on the transportation method extracted in the previous step; for entries with the general category "Machinery", a single-track query is performed, matching the corresponding carbon emission factor of construction machinery only based on the four-dimensional resource category code. To avoid computational redundancy and high-frequency interface calls caused by row-by-row queries, a global mapping dictionary mechanism is introduced in the matching of building material transportation factors. Specifically, the implementation is as follows: First, the transportation mode field in all entries is traversed and extracted. After hash deduplication, a unique set of transportation features to be matched is generated. The large model agent takes over this feature set and, based on semantic understanding, matches it with the standard transportation modes (generally categorized as transportation, distinguishing it from carbon emission factors for building material production and construction machinery) in the background database 30, dynamically generating a mapping index of "original non-standard transportation mode - standard database transportation code". Finally, through an underlying table lookup mechanism, the corresponding transportation carbon emission factors are batch-connected to all entries based on the mapping index and filled in. Ultimately, by integrating the above information, a preliminary structured form 40 containing core fields such as five-dimensional engineering space codes and four-dimensional resource category codes is output.

[0033] S4. Improve structured forms, specifically as follows: For entries marked "Pending Processing," an active completion mechanism is triggered. The large model agent, based on the fields of the missing data and a pre-defined authoritative database whitelist mechanism, calls the corresponding external API to perform a targeted search.

[0034] The intelligent agent initiates a structured query to whitelisted sources. While extracting missing values, such as the unit weight of new materials and special carbon emission factors, it simultaneously captures the metadata of the data, including the data source organization, publication time, applicable boundaries, and original reference links, generating a structured "source traceability evidence card." This card is pushed to the human-machine collaboration interface to prompt the user for review. After manual verification and signature confirmation, the data is fed back into the dictionary, the "pending processing" mark is removed, and the structured form 40 is updated, achieving audit-level end-to-end traceability. If the user does not agree to use the data, they can make custom input or choose to ignore the entry.

[0035] S5. Generate the analysis details table, specifically as follows: like Figure 2 As shown, after the data in structured form 40 is complete, the scheduling deterministic calculation module strictly follows the predetermined formula to calculate carbon emissions. Based on the total classification information in the four-dimensional resource category code, a differentiated generation logic is adopted: for the "Building Materials" category, the carbon emission factors of material production and transportation are used to calculate and derive two independent results: "Building Materials Production Carbon Emissions" and "Building Materials Transportation Carbon Emissions"; for the "Machinery" category, only the machinery carbon emission factor is relied upon to calculate and derive the "Construction Machinery Carbon Emissions" result.

[0036] During the assembly and output phase of the detailed table, a strict "transfer and derivation" data flow logic is implemented: the original five-dimensional engineering space code and four-dimensional resource category code in the structured form 40, along with the standard names determined in the previous normalization phase, are used as prerequisite benchmark fields and directly transferred to the result set to ensure that the underlying traceability chain of engineering space and resource attributes remains unbroken. Simultaneously, for each independent result generated by the aforementioned split calculation, the corresponding calculation content code is dynamically derived, precisely identifying whether the result belongs to carbon emissions from building material production, building material transportation, or construction machinery. The final calculated "carbon emissions (fixed unit)" is also output as a derived field. Finally, the aforementioned transferred benchmark fields and dynamically derived calculation fields are horizontally spliced ​​and structurally encapsulated to output a complete standardized analysis detailed table 50, which automatically generates a preset basic carbon emission composition diagram.

[0037] S6. Output customized charts, specifically as follows: Based on the completed analysis details table 50, and relying on the pre-set nine-dimensional coding system (integrating engineering spatial coding, resource category coding, and calculation content coding), an online analytical processing (OLAP) multidimensional data cube is pre-constructed. This data cube contains three core coordinate axes: the engineering spatial axis is "sub-item project - sub-section project - unit project - single project - project", the resource category axis is "third-level classification - second-level classification - first-level classification", and the carbon emission stage axis is "calculation content (building material production / building material transportation / construction machinery)".

[0038] When unstructured natural language analysis commands are issued through the front-end interactive interface, the large model agent performs deep semantic parsing and intent recognition, dynamically converting them into structured query commands for the OLAP data cube, such as Text-to-MDX expressions, Text-to-SQL extended statements, or multidimensional retrieval API parameters. After receiving the commands, the underlying analysis engine, relying on the multidimensional data cube, flexibly performs dynamic aggregation operations along multiple coordinate axes, specifically covering multi-dimensional level roll-up (upward aggregation), drill-down (downward expansion of details), single-dimensional slicing, and multi-dimensional block filtering, ultimately rendering and outputting highly customized visual data charts in real time.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for carbon emission accounting in tunnel engineering based on a large-scale intelligent agent model, characterized in that, include: S1. The intelligent agent receives the input form, uses a large language model to determine the physical location index of the item name and specification, performs line-by-line matching and classification, constructs a four-dimensional resource category code with a self-growing topology, and binds the four-dimensional resource category code as a structured label to the physical location index in real time. S2. Extract feature parameters from the input form and perform heterogeneous data normalization. Merge the entries and physical location indexes according to the four-dimensional resource category code, and set pending processing marks for abnormal entries with missing parameters. S3. Parse the explanatory text to lock the background database version and generate a five-dimensional engineering space code. Perform differentiated matching routing based on the total classification information of resource categories, complete carbon emission factor matching, and generate a preliminary structured form. S4. For entries marked with the pending flag, trigger the large model agent to perform an external API search based on a preset whitelist. After manual review and verification, the data is fed back and the structured form is updated. S5. Call the deterministic calculation module to perform classification and accounting on the structured form based on the carbon emission factors, and assemble and output a standardized analysis detail table and a carbon emission composition diagram in a preset format by passing through the baseline field and the dynamically derived accounting field. S6. Based on the preset multidimensional coding system, an online analysis and processing multidimensional data cube is constructed. The large model agent parses the user's natural language analysis instructions, and the dynamic scheduling analysis engine performs multidimensional aggregation operations on the multidimensional data cube and generates customized charts.

2. The method for calculating carbon emissions from tunnel engineering based on a large-scale intelligent agent as described in claim 1, characterized in that, The construction of a four-dimensional resource category code with a self-growing topology in S1 specifically includes: The classification is carried out hierarchically in the order of general classification, primary classification, secondary classification, and tertiary classification; For the closed-loop classification of the first three layers, the large model agent is invoked to perform logical routing and distribution based on domain knowledge; For open classification of three-level categories, a two-layer funnel architecture of "lightweight vector initial screening + deep reasoning of large language model" is adopted to prevent redundancy and deduplication: the text embedding model is called to calculate the cosine similarity between the features and the existing three-level categories in the background database; if the cosine similarity is higher than a first preset threshold, it is merged into the existing code; if the cosine similarity is in the fuzzy range, the large model agent is triggered to perform logical reasoning in combination with the context; only when the reasoning result is negative or the cosine similarity is lower than a second preset threshold is it determined to be a new material, and a new code is assigned to it according to the rules to realize the self-growing topology of the coding system.

3. The method for carbon emission accounting in tunnel engineering based on a large-scale intelligent agent as described in claim 1, characterized in that, The differential matching routing based on the overall classification of resource categories in S3 specifically includes: If the overall category is building materials, then a dual-track parallel query is triggered: one track matches the carbon emission factor of building materials production based on the complete four-dimensional resource category code, and the other track matches the carbon emission factor of building materials transportation based on the extracted transportation method. If the overall category is machinery, then a single-track query is performed: the carbon emission factor of construction machinery is matched only based on the complete four-dimensional resource category code; Specifically, a global mapping dictionary mechanism is used when matching the carbon emission factors of building material transportation: the transportation mode field of all entries is extracted and hashed to generate a transportation feature set to be matched, which is then aligned with the standard transportation modes in the background database by the large model agent to generate a mapping index, and the carbon emission factors of building material transportation are batch-filled into the relevant entries based on the mapping index.

4. The method for calculating carbon emissions from tunnel engineering based on a large-scale intelligent agent as described in claim 1, characterized in that, In step S4, triggering the large model agent to perform an external API retrieval based on a preset whitelist specifically includes: The large model agent initiates a structured query to the whitelisted sources, extracting missing values ​​while simultaneously capturing metadata, which includes at least the data source organization, publication time, and original reference link. Based on the numerical data and metadata, a structured traceability evidence card is generated and pushed to the human-machine collaboration interface; after manual signature confirmation, the data is fed back into the dictionary to remove the pending markers.

5. The method for calculating carbon emissions from tunnel engineering based on a large-scale intelligent agent as described in claim 1, characterized in that, The S5 process assembles and outputs a standardized analysis detail table by transmitting a baseline field and dynamically derived accounting fields, specifically including: The five-dimensional engineering space code, the four-dimensional resource category code, and the standard name determined in the previous normalization stage are used as the pre-reference fields and directly passed to the result set. For each independent result generated by the classification and accounting breakdown, a corresponding calculation content code is dynamically derived. The calculation content code is used to accurately identify whether the result belongs to carbon emissions from building material production, carbon emissions from building material transportation, or carbon emissions from construction machinery. The aforementioned baseline fields, the derived calculation content codes, and the calculated carbon emissions are horizontally spliced ​​and structurally encapsulated to output the standardized analysis details table.

6. The method for carbon emission accounting in tunnel engineering based on a large-scale intelligent agent according to claim 1, characterized in that, The dynamic scheduling analysis engine in S6 performs multidimensional aggregation operations on the multidimensional data cube, specifically including: Construct the online analysis and processing multidimensional data cube containing three core coordinate axes: engineering space axis, resource category axis, and carbon emission stage axis; The large model agent performs intent recognition, dynamically converting the unstructured natural language analysis instructions into structured multidimensional query instructions for the multidimensional data cube; After receiving the structured multidimensional query command, the analysis engine performs multidimensional level roll-up, drill-down, single-dimensional slicing, and multidimensional block operations along multiple coordinate axes to output customized charts.

Images