Building operation and maintenance corpus construction method based on digital delivery

Abstract

The invention relates to a digital delivery-based building operation and maintenance corpus construction method, which comprises the following four core processes: step 1, constructing a footstone corpus containing structured multi-source information based on building full-life-cycle digital delivery data; 2, generating professional question and answer pairs covering multiple categories in combination with operation and maintenance scenes and expert experience; 3, marking a decision-making thinking chain containing a judgment basis, an analysis process and a processing suggestion for the complex scene; and step 4, establishing a multi-dimensional evaluation system covering basic knowledge, problem solving, professional decision making and innovation ability. By integrating full-life-cycle digital delivery data, hierarchical construction of corpora from foundation to complexity is achieved, the speciality and effectiveness of the corpora are guaranteed through quantitative quality evaluation and a dynamic optimization mechanism, the problems that existing corpora are single in data and lack of deep decision knowledge are solved, and the method has the advantages of being high in practicability and easy to popularize. High-quality data support is provided for special large model training in the field of building operation and maintenance, and intelligent upgrading of operation and maintenance is promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of building informatization and artificial intelligence technology, and particularly relates to a building operation corpus construction method based on digital delivery. BACKGROUND

[0002] With the wide application of BIM technology, Internet of Things and big data in the field of building, the amount of building operation data has increased dramatically, but these data are scattered in different systems, lacking unified semantic standards and knowledge organization framework. Traditional knowledge management methods mainly rely on document libraries and expert experience, which is difficult to effectively support intelligent decision-making. In recent years, large language models have shown great capabilities in many fields, but general large models have limited performance in the field of building operation, mainly because of the lack of high-quality domain professional corpus.

[0003] The existing building operation corpus construction method has many problems. First, the data source is single, mainly relying on text materials, lacking multi-modal data fusion, and unable to fully utilize images, videos, sensor time series data and other data generated in the building operation process; second, the corpus quality is uneven, lacking professional annotation and structured processing, making it difficult for the model to accurately understand building professional terms and complex process flow; third, the corpus construction process lacks systematicness, making it difficult to support the complete knowledge evolution path from basic cognition to professional decision-making; fourth, the evaluation standard is not unified, making it difficult to objectively evaluate the professional ability of the model in the field of building operation; finally, the existing method lacks deep combination with the building digital delivery standard, making it difficult to realize the knowledge continuity from design to operation.

[0004] Therefore, how to provide a building operation corpus construction method based on digital delivery that can cultivate the professional ability of the model in stages and comprehensively improve the professionalism and practicality is a technical problem that needs to be solved by those skilled in the art. SUMMARY

[0005] The present application provides a building operation corpus construction method based on digital delivery to solve the above technical problems.

[0006] To solve the above technical problems, the present application provides a building operation corpus construction method based on digital delivery, comprising the following steps: Step 1: cornerstone corpus construction: based on building full life cycle digital delivery data, collecting and integrating multi-source information required for building operation to form a structured basic data set, the multi-source information at least including digital building asset ledger, specification knowledge base, equipment full life cycle professional knowledge base, project-level knowledge assets and building professional theory knowledge; Step 2: Professional question and answer pair construction: Based on the requirements of building operation and maintenance scenarios and expert experience, generate question and answer pair data covering operation and maintenance tasks, which at least includes fault diagnosis, operation guidance and specification query; Step 3: Decision-making chain construction: For complex scenarios in building operation and maintenance, professional personnel mark the thinking chain data, which at least includes multi-system cooperation, hypothesis verification, alternative scheme comparison and uncertainty handling; The thinking chain data at least includes judgment basis, problem analysis and treatment suggestion; Step 4: Ability evaluation system construction: For the model trained by steps 1 to 3, establish a multi-dimensional evaluation mechanism, which at least includes the evaluation of basic knowledge ability, problem solving ability, professional decision-making ability and innovation ability.

[0007] Preferably, the construction of the cornerstone corpus in step 1 also includes knowledge systematization processing, including: Based on the building operation and maintenance field knowledge graph, organize the scattered knowledge points in the multi-source information into a hierarchical and related knowledge network; Develop automatic association technology based on digital delivery data semantics, establish the topological relationship between building components, equipment and systems; Introduce knowledge integrity checking mechanism and calculate knowledge coverage.

[0008] Preferably, in step 1, the construction method of the specification knowledge base includes: Use natural language processing technology to automatically extract key provisions, technical parameters and applicable conditions in building operation and maintenance industry specifications, and construct a specification knowledge graph; Develop specification dependency chain analysis technology to automatically identify the hierarchical relationship and application priority between different specifications; Through specification application scenario mapping, associate abstract specification clauses with specific building operation and maintenance scenarios; Establish a specification compliance verification mechanism to ensure that the content of the cornerstone corpus conforms to the building industry standards.

[0009] Preferably, step 2 also includes quality optimization of the constructed question and answer pairs, including: Increase the diversity of question expression through synonym replacement and sentence transformation; Generate multi-angle questions for the same operation and maintenance knowledge point; Add common error concept discrimination questions in the field of building operation and maintenance; Construct a question difficulty grading system.

[0010] Preferably, step 2 also includes quantitative evaluation of the quality of the constructed question and answer pairs, and the evaluation formula is:

[0011] wherein, is the equipment coverage; is the specification compliance; is the scenario coverage; is the energy efficiency correlation; , , , are coefficients, and .

[0012] Preferably, step 3 further comprises thought chain structuring guidance, including: Designing a four-step thought framework of “observation-analysis-decision-verification” to guide professionals to systematically express their thinking process for complex scenarios; For each decision point in the thought chain, requiring professionals to clearly state the judgment basis, the operation and maintenance impact factors to be considered, and the alternative solutions, forming a complete and logically coherent thought chain; Through a thought chain integrity checking tool, automatically identifying missing links in the thought chain and prompting professionals to supplement.

[0013] Preferably, step 3 further comprises quantitative evaluation of the quality of thought chain data, and the evaluation formula is:

[0014] wherein, is the information source integrity, including building information model data, real-time sensor data, equipment technical manual, historical operation and maintenance work order, and weather forecast data; is the system coupling degree; is the risk control completeness; is the full life cycle consideration; , , , are coefficients, and .

[0015] Preferably, step 4 further comprises model training optimization, including: Domain pre-training: pre-training a general large model using the cornerstone corpus constructed in step 1, so that the model parameter distribution adapts to the language expression and knowledge system of the building operation and maintenance field; Supervised fine-tuning: adjusting model parameters based on the professional question and answer pair data constructed in step 2, and the optimization objective function is:

[0016] wherein, is the operation and maintenance problem, is the standard answer labeled by professionals, A question-answer pair data set is constructed for the professional field, A probability that the model generates an answer under the parameters, A specification compliance regularization term, if the key operation steps or parameters in the answer conflict with the building operation specification, ; A regularization coefficient; Thought chain reinforcement learning: a proximal policy optimization algorithm is used to optimize the decision-making ability of the model to deal with complex operation scenarios by using the quality evaluation results of the decision-making thought chain data in step 3 as the reward signal.

[0017] Preferably, in step 4, the multi-dimensional evaluation mechanism includes setting a comprehensive score for quantifying the evaluation results, and the formula of the comprehensive score is:

[0018] Wherein, is the basic knowledge ability score, is the problem solving ability score, is the professional decision-making ability score, is the innovation ability score, 、 、 、 are coefficients, respectively, and .

[0019] Preferably, step 4 further includes expert review and dynamic update, including: The model evaluation results are reviewed by experts in the field of building operation to establish an expert scoring mechanism and evaluation feedback process; Based on the evaluation feedback data, continuously optimize the content and structure of the corpus constructed in steps 1 to 3; According to the update of new technologies and new specifications in the field of building operation, dynamically update the evaluation content and judgment standard in the multi-dimensional evaluation mechanism; Construct an open building operation corpus evaluation benchmark to support the technical exchange and iterative optimization of artificial intelligence models in the field of building operation.

[0020] Compared with the prior art, the building operation corpus construction method based on digital delivery provided by the present application has the following advantages: 1. The present application constructs a cornerstone corpus, so that the large model can master the building operation basic knowledge system and reach the level of a professional undergraduate, and can accurately understand building professional terms and basic principles; 2. The present application trains the model to have the ability to handle routine operation tasks through question and answer pairs, which is equivalent to a qualified operation staff, can provide accurate operation guidance and fault handling suggestions, and the practical operation ability is significantly enhanced; 3、The application utilizes the decision-making thinking chain training model to form an expert-level thinking mode and decision-making ability, and can handle complex, multi-system collaborative operation and maintenance problems; 4、The application integrates multi-source data of the whole life cycle of a building, closely combines the corpus with a building digital delivery standard, realizes design-construction-operation and maintenance knowledge penetration, and improves the adaptability of the model to specific projects; 5、The application objectively evaluates the professional ability of the model in the field of building operation and maintenance by establishing a multi-dimensional evaluation method, and provides a clear direction for model iteration and optimization. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The figure shows the flow architecture of the building operation and maintenance corpus construction method based on digital delivery in a specific embodiment of the application. DETAILED DESCRIPTION

[0022] In order to more fully describe the technical solutions of the above-mentioned application, the following specific embodiments are listed to prove the technical effects; it should be emphasized that these embodiments are used to illustrate the application and not to limit the scope of the application.

[0023] The building operation and maintenance corpus construction method based on digital delivery provided by the application, as shown in Figure 1 follows a three-stage ability training path of "basic cognition-operation ability-expert decision-making", and specifically includes the following steps: Step 1: cornerstone corpus construction: based on building whole life cycle digital delivery data, collect and integrate multi-source information required for building operation and maintenance, form a basic data set with structured characteristics, and the multi-source information at least includes digital building asset account, specification knowledge base, equipment whole life cycle professional knowledge base, project level knowledge asset and building class professional theory knowledge. In the project design and construction stage, the data delivery list required for operation and maintenance is defined to ensure that the data has structured and semantic characteristics at the source.

[0024] Step 2: professional question and answer pair construction: combining the needs of building operation and maintenance scene and expert experience, generating question and answer pair data covering operation and maintenance tasks, the question and answer pair data at least contains fault diagnosis, operation guidance and specification query. Specifically, based on typical operation and maintenance work orders and maintenance records, the problems can be generated by field experts, covering daily inspection, routine maintenance and simple fault handling scenes. Design professional problems for different systems such as heating, ventilation, electrical, water supply and drainage, etc., to ensure comprehensive system coverage. Construct multiple types of problems such as fault diagnosis, operation guidance and specification query, and control the difficulty of the problems within the scope that qualified operation and maintenance staff should master. Through expert model interaction verification, ensure the applicability of the problem and answer pair in the actual scene.

[0025] Step 3: Construction of the Decision-Making Thinking Chain (COT): For complex scenarios in building operation and maintenance, professionals annotate the thinking chain data. These complex scenarios must at least include multi-system collaboration, hypothesis testing, comparison of alternative solutions, and uncertainty handling. The thinking chain data must at least include judgment criteria, problem analysis, and handling suggestions. Typical building processes such as air conditioning system commissioning, fire protection system testing, and elevator maintenance can be selected as the basis for COT annotation, which experts then break down into multiple logical steps. For building operation and maintenance monitoring video clips such as equipment malfunctions, personnel operations, and emergencies, experts annotate key observation points, problem analysis, and handling suggestions, forming a detailed thinking chain description. Complex problems are designed for specific scenarios such as logistics, maintenance, security, and energy conservation, and experts annotate the thinking paths and decision-making logic.

[0026] Step 4: Construction of the Competency Assessment System: For the model trained on the corpora from Steps 1 to 3, a multi-dimensional assessment mechanism is established. This mechanism includes assessments of at least basic knowledge competence, problem-solving ability, professional decision-making ability, and innovation ability. An assessment set is constructed specifically to evaluate the model's professional competence in the building operations and maintenance field after the first three stages of corpus training. The aim is to verify whether the model has successfully transitioned from a "junior engineer" to a "mid-level engineer" to a "senior engineer / technical expert."

[0027] This invention integrates digital delivery data throughout the entire building lifecycle to construct a multi-layered corpus system that includes a foundational corpus, professional question-and-answer pairs, and decision-making thought chains, achieving comprehensive coverage from basic data to complex decision-making knowledge.

[0028] In some embodiments, in step 1, the digital building asset ledger, the standard knowledge base, the equipment lifecycle professional knowledge base, the project-level knowledge assets, and the building-related professional theoretical knowledge are respectively: Structured building information, including building component attributes, equipment parameters, system topology relationships and spatial information, is extracted from the as-built BIM model that conforms to IFC standards to form the digital building asset ledger. The system collects building operation and maintenance industry norms and standards, establishes a knowledge base of these norms, and links them to specific components in the BIM model; Integrate technical documents, operation manuals, and digital twin models of equipment provided by construction equipment manufacturers to form a professional knowledge base covering the entire lifecycle of the equipment; Collect design specifications, construction drawings, as-built documentation, and operation and maintenance handover lists from typical building projects to construct the project-level knowledge assets. By selecting fundamental knowledge from architectural college textbooks and professional reference books, a theoretical knowledge system is constructed, and the theoretical knowledge of the architectural profession is acquired.

[0029] In some embodiments, the data collected specifically includes: IFC data: Extracted from BIM models of 3 typical large public building projects, containing 12850 building components and 7420 equipment objects; Industry standards: Collected 86 national and local standards, including 23 mandatory standards and 63 recommended standards; Technical documents: Integrated technical manuals of 12 major equipment manufacturers, covering HVAC, electrical, water supply and drainage systems, etc.; Teaching materials: Selected 15 core textbooks of Building Environment and Energy Application Engineering, extracted 3200 key knowledge points.

[0030] Step 1 also includes systematizing building operation and maintenance knowledge, including: Based on the building operation and maintenance knowledge graph, the scattered knowledge points in the multi-source information are organized into a hierarchical and correlated knowledge network; Develop automatic association technology based on digital delivery data semantics, establish topological relationship between building components, equipment and systems; Introduce knowledge integrity checking mechanism, system automatically identify knowledge fault and logical loophole; Calculate knowledge coverage, ensure the integrity of equipment system knowledge base, building material knowledge base, operation and maintenance process knowledge base, spatial information knowledge base and energy consumption benchmark knowledge base.

[0031] Typical relationship example: (Chiller) - [: connection] -> (Cooling tower) (Chiller) - [: belongs to] -> (HVAC system) (Cooling tower) - [: needs maintenance] -> (Quarterly cleaning) (Quarterly cleaning) - [: according to] -> (GB50243-2016).

[0032] In some embodiments, the construction method of the specification knowledge base includes: Use natural language processing (NLP) technology to automatically extract key clauses, technical parameters and application conditions in building operation and maintenance industry standards, and construct specification knowledge graph; Develop specification dependency chain analysis technology, automatically identify hierarchical relationship and application priority between different specifications; Through specification application scenario mapping, abstract specification clauses are associated with specific building operation and maintenance scenarios; Establish specification compliance verification mechanism to ensure that the content of the cornerstone corpus complies with the building industry standards.

[0033] In some embodiments, step 2 includes: Step 21: Expert-led question and answer pair generation flow: A three-way mapping mechanism of problem scenario knowledge is established to ensure that each question is closely related to specific operation and maintenance scenarios and professional knowledge. Organizational domain experts brainstorm to generate questions from multiple angles, and experts write standard answers. Answers include operation steps, safety precautions, and relevant regulatory basis. Through expert consensus evaluation, controversial questions are reviewed in multiple rounds to ensure the authority of the answers.

[0034] Step 22: Quality optimization of constructed question and answer pairs Increase the diversity of question expressions through synonym replacement and sentence transformation to improve the model's understanding of different expressions.

[0035] For example, the original question: "How to check the refrigerant charge of the air conditioning system?" Enhanced variants can include "The air conditioning system has a refrigeration effect, suspecting that the refrigerant is insufficient, how to detect?" "According to GB50243-2016, what is the method for checking the refrigerant charge of the air conditioning system?" "How to determine whether the air conditioning system needs to supplement the refrigerant? What are the detection indicators?" "What will happen if the refrigerant charge of the air conditioning system is insufficient? How to handle it?" etc.

[0036] Generate multi-angle questions for the same operation and maintenance knowledge point to enhance the model's knowledge transfer ability.

[0037] Add common error concept analysis questions in the field of building operation and maintenance to improve the model's error identification and correction ability.

[0038] Construct a question difficulty grading system to support progressive cultivation of model capabilities.

[0039] In some embodiments, step 2 also includes quantitative evaluation of the quality of the constructed question and answer pairs, and the evaluation formula is:

[0040] Wherein,

[0041]

[0042]

[0043]

[0044] , , , are coefficients, and , usually, .

[0045] In some embodiments, step 3 includes: Step 31: Structured Guidance of Thought Chain Design a four-step thinking framework of "observation-analysis-decision-verification" to guide professionals to systematically express their thinking process for complex scenarios; For each decision point in the thought chain, require the professional to clearly state the judgment basis, the operation and maintenance factors to be considered, and the alternative solutions, forming a complete and logical thought chain; Through the thought chain integrity checking tool, automatically identify the missing links in the thought chain and prompt the professional to supplement.

[0046] Step 32: Construction of Multi-Scenario Professional Problem COT Design complex problems for specific scenarios such as logistics, maintenance, security, and energy saving. The problem should involve multi-system coordination and include long-term impact assessment, cost-benefit analysis, and risk trade-off. Label expert-level thinking paths and decision logic, including hypothesis verification, alternative solution comparison, and uncertainty handling. Construct decision models under the influence of dynamic factors such as seasonality and time period to improve the model's situational adaptability.

[0047] Quantitative evaluation of the quality of thought chain data, the evaluation formula is:

[0048] Among them,

[0049] Information sources include BIM models, real-time sensor data, equipment manuals, historical work orders, weather forecasts, etc., reflecting the ability to make decisions based on multi-source data fusion.

[0050]

[0051] For example, "adjusting the air conditioning system causes the power load to rise, and needs to coordinate the power system" is a cross-system interaction, reflecting the complex correlation of building systems.

[0052]

[0053] Forced models consider risks in decision-making.

[0054]

[0055] Senior operation and maintenance decisions must consider the long-term value and cost of assets, not just immediate problems.

[0056] 、 、 、 is the coefficient, and , usually, .

[0057] In some embodiments, step 4 includes: Step 41: Model training optimization: Domain pre-training: Pre-train the general large model using the cornerstone corpus built in step 1 (including BIM data, specifications, equipment manuals, etc.), so that the model parameter distribution adapts to the language expression and knowledge system in the building operation and maintenance field.

[0058]

[0059] Supervised fine-tuning: Adjust the model parameters based on the professional question and answer pair data set built in step 2, and the optimization objective function is:

[0060] Where, is the operation and maintenance problem; is the standard answer marked by professionals; is the professional question and answer pair data set; is the probability of the model generating an answer under the parameters; is the specification compliance regularization term, which punishes the model for generating answers that violate building specifications, and its value is calculated through specification knowledge graph matching. If the key operation steps or parameters in the answer conflict with the building operation and maintenance specifications, then ; is the regularization coefficient, which controls the importance of specification compliance.

[0061] Chain of thought reinforcement learning (COT-RL): Use the proximal policy optimization (PPO) algorithm to optimize the model's decision-making ability in complex operation and maintenance scenarios, using the quality evaluation results of the decision-making chain data in step 3 as the reward signal:

[0062] Where: is the ratio of new and old strategies.

[0063] is the advantage function, calculated using the COT data quality index as the reward signal.

[0064] is the policy entropy, which encourages exploration.

[0065] is the energy efficiency optimization reward term, which gives an additional positive reward when the model proposes suggestions to reduce energy consumption (such as adjusting operating parameters or start-stop strategies) during the reasoning process.

[0066] is the hyperparameter.

[0067] Step 42: Multi-Dimensional Capability Assessment Design Foundational Knowledge Assessment: Test the model's grasp of building operation fundamentals, including terminology understanding, code reference, principle application, etc.

[0068] Problem-Solving Capability Assessment: Test the model's ability to solve typical operation problems, evaluate the accuracy and safety of its operation guidance.

[0069] Expert Decision-Making Capability Assessment: Test the model's professional decision-making ability in complex situations, including multi-objective trade-offs and long-term impact assessment.

[0070] Innovation Capability Assessment: Test the model's ability to analyze and solve new problems, evaluate its knowledge transfer and innovative thinking.

[0071] Step 43: Quantitative Assessment Results The comprehensive score formula representing the model's professional ability is:

[0072] Where,

[0073] Covering equipment principles, material properties, code provisions, etc.

[0074]

[0075] The expert review score (100 full marks) for the i th problem, the operation criticality weight of the problem (such as fire safety problems involving life safety =1.5, general lighting problems =1.0).

[0076]

[0077] Where, is the score of the j th complex decision-making problem, is the scenario complexity coefficient of the problem (comprehensively evaluated according to the number of systems involved, risk level, economic impact, etc., range 1.0-2.0).

[0078]

[0079] Where, ( Innovation Quotient , innovation quotient) is scored by experts according to the originality of the scheme, technical feasibility, economic benefit, range 0.8-1.0.

[0080] 、 、 、 are coefficients, and , which can be dynamically adjusted according to the evaluation target, for example, when evaluating a middle-level engineer a higher value is taken.

[0081] In some embodiments, the expert decision scoring rules are as follows:

[0082] wherein Accuracy = 0.4 x Technical Correctness + 0.3 x Specification Compliance + 0.3 x Data Traceability Technical Correctness refers to whether the scheme is feasible in engineering; Specification Compliance refers to whether mandatory provisions are met; and Data Traceability refers to whether real data sources such as BIM and sensors are referenced.

[0083] Safety = 0.6 x Risk Identification + 0.4 x Emergency Plan Risk Identification refers to whether operational risks, equipment risks, and environmental risks are comprehensively identified; and Emergency Plan refers to whether effective emergency disposal and personnel evacuation schemes are included.

[0084] Economy = 0.3 x Short-term Cost + 0.4 x Long-term Benefit + 0.3 x Resource Utilization Rate Long-term Benefit refers to a life cycle cost (LCC) analysis considering equipment residual life, maintenance cost, and energy saving.

[0085]

[0086] For example, adjusting the impact of air conditioning systems on power load, the impact of elevator maintenance on fire safety systems, etc.

[0087] Foresight = 0.4 x Trend Prediction + 0.3 x Sustainability + 0.3 x Scalability Sustainability refers to whether energy saving and emission reduction, and green operation and maintenance are considered; and Scalability refers to whether the scheme is easy for future system upgrade or function expansion.

[0088] 、 、 、 、 The sum of = 0.4, = 0.25, = 0.2, = 0.1, = 0.05.

[0089] Step 44: capability stage determination criteria

[0090] Step 45: expert review and dynamic update The evaluation results of the model are reviewed by experts in the field of building operation and maintenance, and an expert scoring mechanism and evaluation feedback process are established; Based on the evaluation feedback data, the content and structure of the corpus constructed in steps 1 to 3 are continuously optimized; According to the update of new technologies and new specifications in the field of building operation and maintenance, the evaluation content and determination criteria in the multi-dimensional evaluation mechanism are dynamically updated to maintain the timeliness of the evaluation system; A public building operation and maintenance corpus benchmark is constructed to support the technical exchange and iterative optimization of artificial intelligence models in the field of building operation and maintenance.

[0091] In summary, the building operation and maintenance corpus construction method based on digital delivery provided by the present application integrates building full life cycle digital delivery data, constructs a multi-level corpus system including cornerstone corpus, professional question and answer pairs, and decision thinking chain, and realizes comprehensive coverage from basic data to complex decision knowledge. At the same time, a perfect quality evaluation and model training optimization mechanism is established to ensure the professionalism, accuracy and effectiveness of the corpus.

[0092] Obviously, those skilled in the art can make various modifications and variations to the application without departing from the spirit and scope of the application. Therefore, if these modifications and variations of the application fall within the scope of the claims of the application and their equivalents, the application also intends to include these modifications and variations.

Claims

1. A method for constructing a building operation corpus based on digital delivery, characterized in that, Comprising the following steps: Step 1: cornerstone corpus construction: based on building full life cycle digital delivery data, collect and integrate multi-source information required for building operation and maintenance, form a structured basic data set, the multi-source information at least includes digital building asset ledger, specification knowledge base, equipment full life cycle professional knowledge base, project level knowledge assets and building professional theory knowledge; Step 2: professional question and answer pair construction: combining building operation and maintenance scene demand and expert experience, generating question and answer pair data covering operation and maintenance tasks, the question and answer pair data at least contains fault diagnosis, operation guidance and specification query content; Step 3: decision thinking chain construction: for complex scenarios in building operation and maintenance, professional personnel mark thinking chain data, the complex scenarios at least include multi-system cooperation, hypothesis verification, alternative scheme comparison and uncertainty handling; the thinking chain data at least includes judgment basis, problem analysis and treatment suggestion; Step 4: ability evaluation system construction: for the model trained by steps 1 to 3, establish a multi-dimensional evaluation mechanism, the multi-dimensional evaluation mechanism at least includes the evaluation of basic knowledge ability, problem solving ability, professional decision making ability and innovation ability.

2. The digital delivery based building operation corpus construction method of claim 1, wherein, The cornerstone corpus construction of step 1 also includes knowledge systematization processing, including: Based on building operation and maintenance field knowledge graph, organize the scattered knowledge points in the multi-source information into a hierarchical and related knowledge network; Develop automatic association technology based on digital delivery data semantics, establish the topological relationship between building components, equipment and system; Introduce knowledge integrity checking mechanism and calculate knowledge coverage.

3. The digital delivery based building operation corpus construction method of claim 1, wherein, In step 1, the construction method of the specification knowledge base includes: Using natural language processing technology to automatically extract key clauses, technical parameters and applicable conditions in building operation and maintenance industry specifications, and constructing specification knowledge graph; Develop specification dependency chain analysis technology to automatically identify the hierarchical relationship and application priority between different specifications; Through specification application scene mapping, associate abstract specification clauses with specific building operation and maintenance scenarios; Establish specification compliance verification mechanism to make the content of the cornerstone corpus comply with the building industry standards.

4. The digital delivery based building operation corpus construction method of claim 1, wherein, Step 2 also includes quality optimization of the constructed question and answer pairs, including: Increase the diversity of question expression through synonym replacement and sentence transformation; Generate multi-angle questions for the same operation and maintenance knowledge point; Add common error concept discrimination questions in the field of building operation and maintenance; Construct a question difficulty grading system.

5. The digital delivery based building operation corpus construction method of claim 4, wherein, Step 2 also includes quantitative evaluation of the quality of the constructed question and answer pairs, the evaluation formula is: wherein, is a device coverage; is a specification compliance; is a scenario coverage; is an energy efficiency correlation; , , , are coefficients, respectively, and .

6. The digital delivery based building operation corpus construction method of claim 1, wherein, Step 3 also includes thinking chain structuring guidance, including: Design a four-step thinking framework of "observation-analysis-decision-verification" to guide professionals to systematically express their thinking process for complex scenarios; For each decision point in the thinking chain, require professionals to clearly state the judgment basis, operation and maintenance factors to be considered and alternative solutions to form a logically coherent complete thinking chain; Through the thinking chain integrity checking tool, automatically identify the missing links in the thinking chain and prompt the professional to supplement.

7. The digital delivery based building operation corpus construction method of claim 1, wherein, Step 3 also includes quantitative evaluation of the quality of the thinking chain data, the evaluation formula is: wherein, for information source completeness, the information source including building information modeling data, real-time sensor data, equipment technical manuals, historical operations and maintenance work orders, and weather forecast data; for system coupling degree; for risk control completeness; for full life cycle consideration; , , , for coefficient, and .

8. The digital delivery based building operation corpus construction method of claim 1, wherein, Step 4 also includes model training optimization, including: Domain pre-training: Pre-training the general large model using the cornerstone corpus constructed in Step 1, adapting the model parameter distribution to the language expression and knowledge system of the building operation and maintenance field; Supervised fine-tuning: Adjusting the model parameters based on the professional question and answer pair data constructed in Step 2, with the optimization objective function being: wherein, is the operation and maintenance problem, is the standard answer annotated by professionals, is the professional Q&A pair dataset, is the probability of the model generating the answer under the parameters, is the specification compliance regular term, if the key operation steps or parameters in the answer conflict with the building operation and maintenance specification, ; is the regularization coefficient; Thought chain reinforcement learning: Using the proximal policy optimization algorithm, taking the quality evaluation results of the decision-making thought chain data in Step 3 as the reward signal, and optimizing the model's decision-making ability to handle complex operation and maintenance scenarios.

9. The digital delivery based building operation corpus construction method of claim 1, wherein, In Step 4, the multi-dimensional evaluation mechanism includes setting a comprehensive score for quantifying the evaluation results, and the formula for the comprehensive score is: wherein, is a score for the knowledge ability, is a score for the problem solving ability, is a score for the professional decision making ability, is a score for the innovation ability, , , , are coefficients, respectively, and .

10. The digital delivery based building operation corpus construction method of claim 9, wherein, Step 4 also includes expert review and dynamic update, including: Professional review of model evaluation results by experts in the field of building operation and maintenance, establishing an expert scoring mechanism and evaluation feedback process; Based on the evaluation feedback data, continuously optimize the content and structure of the corpus constructed in Steps 1 to 3; According to the update of new technologies and new specifications in the field of building operation and maintenance, dynamically update the evaluation content and judgment standards in the multi-dimensional evaluation mechanism; Constructing an open building operation and maintenance corpus evaluation benchmark to support technical exchanges and iterative optimization of artificial intelligence models in the field of building operation and maintenance.

Images