Knowledge-driven intelligent review method, system and equipment for underground building design schemes

Abstract

The present invention provides a knowledge-driven intelligent review method, system and equipment for underground building design schemes, the method comprising: identifying the target facility type, the elements to be reviewed and the corresponding review material types of the review object in the underground building design scheme to be reviewed; matching rule-structured text entries and experience-structured text entries from a knowledge base of the corresponding facility type according to the target facility type; loading corresponding review materials from application materials according to the review material type; if text-type review materials are involved, calling a review reasoning big model to extract indicator information corresponding to the elements to be reviewed; if drawing-type review materials are involved, calling a drawing recognition enhanced multimodal big model to extract indicator information corresponding to the elements to be reviewed; integrating the extracted indicator information and recalled knowledge items to construct a prompt word template, performing intelligent review and reasoning on the indicator information according to the prompt word template through the review reasoning big model, and outputting a review report to improve review efficiency.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and underground space facility design review technology, and in particular to a knowledge-driven underground building design scheme intelligent review method, system and equipment. Background Art

[0002] The review of underground building design and renovation plans has long relied on empirical judgment and subjective assessments. This has led to problems such as fragmented review logic and insufficient process standardization. This has made it difficult to effectively quantify empirical knowledge, and individual differences in judgment have easily affected the consistency and objectivity of decision-making. Furthermore, the underground space design specification system is complex and dynamically evolving. Existing review systems often rely on static threshold judgment mechanisms, which are unable to integrate regulatory provisions, historical cases, and review decision-making processes into a knowledge system, nor can they analyze the implicit logical connections within empirical knowledge.

[0003] On the other hand, the design of underground buildings involves a large amount of drawing application materials, and manual review of paper drawings is cumbersome. The existing system cannot realize the automatic analysis of spatial elements of design drawings, lacks the ability to analyze cross-modal semantic associations between drawing spatial information and design specification texts, and cannot realize automatic recognition and reasoning. The review process is not intelligent enough, which affects the review efficiency. Summary of the Invention

[0004] In view of the above problems, the present invention is proposed to provide a knowledge-driven intelligent review method, system and device for underground building design solutions that overcome the above problems or at least partially solve the above problems.

[0005] One aspect of the present invention provides a knowledge-driven intelligent review method for underground building design solutions, the method comprising:

[0006] Identify the target facility type and elements to be reviewed in the underground building design plan to be reviewed, and determine the type of review materials corresponding to the elements to be reviewed;

[0007] According to the target facility type, target rule structured text entries are matched from a preset underground space review rule knowledge base of the corresponding facility type, and target experience structured text entries are matched from a preset case experience knowledge base of the corresponding facility type;

[0008] Load the corresponding target review materials from the application materials of the current review task according to the review material type corresponding to the element to be reviewed;

[0009] If the target review materials include text review materials, the pre-trained underground space review reasoning model is called to extract the first indicator information corresponding to the elements to be reviewed from the text review materials; the underground space review reasoning model is an intelligent large model that has been fine-tuned and trained based on a preset underground space review rule text corpus and an instruction alignment corpus. The underground space review rule text corpus records rule-structured text entries, and the instruction alignment corpus records experience-structured text entries involving review logic descriptions and implicit rule descriptions.

[0010] If the target review materials include drawing review materials, the pre-trained underground space field drawing recognition enhanced multimodal large model is called to extract the second indicator information corresponding to the elements to be reviewed from the drawing review materials; the underground space field drawing recognition enhanced multimodal large model is a multimodal large model that has been fine-tuned and trained based on a preset image-text alignment corpus, which records image-text pairs consisting of design drawing data and corresponding text description information;

[0011] All the extracted indicator information and the recalled target rule structured text entries and target experience structured text entries are integrated to construct a prompt word template, and the obtained prompt word template is input into the underground space review reasoning model to review whether the indicator information complies with each target rule structured text entry one by one, and combine the review logic description and implicit rule description in the target experience structured text entry to reason about the indicator review results and output a review report.

[0012] Furthermore, the method further comprises:

[0013] If the elements to be reviewed include relevant indicators that require spatial calculation, the indicator calculation tool corresponding to the corresponding quantitative indicator will be called according to the elements to be reviewed, so as to use the indicator calculation tool to calculate and return the corresponding third indicator information.

[0014] Furthermore, the training process of the underground space review reasoning model includes:

[0015] Using the preset underground space review rule text corpus as training data, the pre-trained base model is fine-tuned using a dynamic low-rank adaptive fine-tuning method. During the fine-tuning training process, the original base model weight parameters are frozen, and a trainable rank decomposition matrix is ​​injected into each layer. Different rank parameter values ​​are assigned to training data of different specification types according to the different specification types of the rule-structured text entries, so as to dynamically adjust the rank parameter value according to the complexity of the task.

[0016] The parameter fusion algorithm is used to combine the fine-tuning weights and the base model to generate a large model enhanced with the domain knowledge of the underground space;

[0017] The preset instruction alignment corpus is used as training data, and the low-rank fine-tuning method is adopted to fine-tune the underground space domain knowledge enhancement large model. During the fine-tuning training process, the weight parameters of the underground space domain knowledge enhancement large model are frozen, and the rank decomposition matrix with a specified rank is injected into the Transformer layer. The underground space review reasoning large model is obtained through supervised learning optimization training.

[0018] Furthermore, after obtaining the underground space review reasoning large model, the method further includes:

[0019] Multiple output results of the underground space review reasoning model under the same prompt word instruction input are marked in order of merit according to the preset evaluation criteria to obtain the preference ranking result;

[0020] A multi-objective reward model is established. Based on the multi-objective reward model, the preference ranking results are used for back propagation training. The multi-objective reward model is:

[0021] ;

[0022] ,

[0023] in, is a given hint, r is Instructions to align the N responses generated by the model ; represents a trainable reward model, Show rule rewards, by matching regular expressions to the rule structured text entries in the response r, is the rule-based weight; represents human preference for rewards, is the empirical correction coefficient; is the objective function of the multi-objective reward model, Represents the sigmoid function, response Better than , by minimizing the objective function Conduct training;

[0024] The proximal strategy is used to optimize and fine-tune the underground space review reasoning model. The reward model output is used as a reinforcement signal. The low-rank matrix parameter ΔW of the underground space review reasoning model is iteratively updated to maximize the score of the reward model, and the optimized underground space review reasoning model is obtained.

[0025] Furthermore, the steps for constructing the underground space review rules text corpus include:

[0026] Extract key review elements of underground building design schemes and quantitative indicators of each key review element from pre-acquired design specifications and guidelines related to underground building design;

[0027] Establish a mapping relationship between each quantitative indicator and the type of review material corresponding to the corresponding quantitative indicator, and / or establish a connection relationship between each quantitative indicator and the interface of the indicator calculation tool corresponding to the corresponding quantitative indicator;

[0028] Extract the specification identification information and chapter content structure of the design specification guidelines related to underground building design, and use the preset multi-level regular template to segment the chapter content structure to generate regular structured text entries;

[0029] According to the different types of facilities under review, the specification identification information and rule structured text entries corresponding to the relevant design specifications and guidelines for underground building design are used to construct different categories of underground space review rule text corpora according to the facility types.

[0030] Furthermore, the steps for constructing the instruction alignment corpus include:

[0031] Obtain target text involving the review process and decision logic from pre-acquired historical design review case data, and structure the target text to establish a structured review logic description;

[0032] Obtaining review record text from the historical design scheme review case data, and extracting implicit rule descriptions related to review factor priority determination from the review record text;

[0033] Case identification information is extracted from historical design scheme review case data, and the case identification information, review logic description and implicit rule description are combined to generate empirical structured text entries. All the obtained empirical structured text entries are combined into an instruction alignment corpus.

[0034] Furthermore, the training process of the underground space field drawing recognition enhanced multimodal large model includes:

[0035] Using the preset image-text aligned corpus as training data, the general multimodal large model is fine-tuned and trained using a dynamic low-rank adaptive fine-tuning method to obtain an enhanced multimodal large model for drawing recognition in the underground space field.

[0036] Furthermore, the general multimodal large model includes a visual encoder, a feature mapping matrix, and a language model;

[0037] Fine-tune the general multimodal large model using dynamic low-rank adaptive fine-tuning, including:

[0038] In the image-text feature alignment stage, the multimodal large model visual encoder is frozen, the rank parameter r is set to 16, and the image description generation task dataset in the image-text alignment corpus is used to fine-tune the training feature mapping matrix to achieve image-text semantic alignment;

[0039] In the joint optimization phase, the visual encoder is unfrozen, the rank parameter r is set to 32, and the drawing question answering task dataset in the image-text alignment corpus is used to simultaneously fine-tune the training feature matrix and language model parameters to achieve fine-grained feature matching and image-text reasoning.

[0040] The parameters trained in the image-text feature alignment stage are merged with the parameters trained in the joint optimization stage to obtain an enhanced multimodal large model for underground space drawing recognition.

[0041] Furthermore, the steps for constructing the image-text aligned corpus include:

[0042] A multimodal large model is used to process the design drawing data of the pre-acquired historical underground building renewal design plan to generate textual description information corresponding to the drawing data, and a semantic correspondence is established between the design drawing data and the corresponding textual description information to generate a graphic-text aligned corpus; the textual description information includes design description, technical and economic indicator information, objective numerical indicator information, key structural targets and their coordinate frame information, and description of design rationality.

[0043] Furthermore, after outputting the review report, the method further includes:

[0044] Determine whether the review report requires manual review based on the preset confidence scoring model;

[0045] For review reports that require manual review, the reviewers' feedback on the accuracy of the review reports is collected, and a reinforcement sample set is generated based on the review reports and the corresponding accuracy feedback. The reinforcement sample set is fed back to the reinforcement learning layer to update the model parameters through incremental training.

[0046] Another aspect of the present invention provides a knowledge-driven intelligent review system for underground building design solutions, the system comprising:

[0047] A question classification unit is used to identify the target facility type and elements to be reviewed in the underground building design plan to be reviewed, and to determine the type of review materials corresponding to the elements to be reviewed;

[0048] A knowledge base matching unit is used to match target rule structured text entries from a preset underground space review rule knowledge base of the corresponding facility type according to the target facility type, and to match target experience structured text entries from a preset case experience knowledge base of the corresponding facility type;

[0049] An examination material loading unit is used to load corresponding target examination materials from the application materials of the current examination task according to the examination material type corresponding to the element to be examined;

[0050] A text parsing unit, configured to, when the target review materials include text review materials, call a pre-trained underground space review reasoning model to extract first indicator information corresponding to the element to be reviewed from the text review materials;

[0051] A drawing parsing unit is configured to, when the target review materials include drawing review materials, call a pre-trained underground space field drawing recognition enhanced multimodal large model to extract second indicator information corresponding to the elements to be reviewed from the drawing review materials;

[0052] The report generation unit is used to integrate all the extracted indicator information and the recalled target rule structured text entries and target experience structured text entries to construct a prompt word template, and input the obtained prompt word template into the underground space review reasoning model to review whether the indicator information complies with each target rule structured text entry one by one, and combine the review logic description and implicit rule description in the target experience structured text entry to reason about the indicator review results and output a review report.

[0053] Another aspect of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the knowledge-driven intelligent review method for underground building design schemes as described above are implemented.

[0054] Another aspect of the present invention provides a computer program product having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned knowledge-driven intelligent review method for underground building design schemes.

[0055] The knowledge-driven intelligent review method, system, and equipment for underground building design schemes provided by the embodiments of the present invention achieve semantic alignment of multimodal data such as underground space facility design specification guidelines, operation and maintenance period facility update design schemes, and empirical knowledge. It can integrate the review decisions and implicit logic in the specification provisions, historical cases, and empirical knowledge to form a knowledge system, realize automatic recognition and reasoning of underground building design schemes based on knowledge-driven, and improve the intelligence level and review efficiency of the review process. Moreover, the present invention can automatically realize intelligent review of a large number of drawing application materials in the design scheme, realize automated parsing of design drawing space elements, and improve the cross-modal semantic association analysis capability of drawing space information and design specification text.

[0056] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0058] Figure 1 A flow chart of a knowledge-driven intelligent review method for underground building design schemes provided by an embodiment of the present invention;

[0059] Figure 2 A flowchart of a knowledge-driven intelligent review method for underground building design schemes provided in another embodiment of the present invention;

[0060] Figure 3 This is a structural diagram of the knowledge-driven intelligent review system for underground building design schemes proposed in an embodiment of the present invention;

[0061] Figure 4 This is a flowchart of the implementation steps of the knowledge-driven intelligent review system for underground building design solutions proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0063] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0064] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by those skilled in the art in the art to which this invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless specifically defined, should not be interpreted in an idealized or overly formal sense.

[0065] Figure 1The flowchart of the knowledge-driven intelligent review method for underground building design scheme according to one embodiment of the present invention is schematically shown. Figure 1 The knowledge-driven intelligent review method for underground building design schemes according to an embodiment of the present invention specifically includes the following steps:

[0066] S11. Identify the target facility type and elements to be reviewed in the underground building design proposal to be reviewed, and determine the type of review materials corresponding to the elements to be reviewed. The facility type is the first-level classification of the review object, such as underground shopping malls, civil air defense basements, etc., and the elements to be reviewed are the review objectives included in the review object of that type, that is, the review dimensions divided, such as floor plan layout, usage function, and streamlined organization.

[0067] S12. Match target rule structured text entries from a preset underground space review rule knowledge base of the corresponding facility type according to the target facility type, and match target experience structured text entries from a preset case experience knowledge base of the corresponding facility type.

[0068] In this embodiment, the underground space review rule knowledge base is constructed by converting the rule-structured text entries in the pre-constructed underground space review rule text corpus of different facility types into vector representations using a vector embedding model to generate a vector knowledge base of underground space review rule for the corresponding facility type. The underground space review rule text corpus records rule-structured text entries. The case experience knowledge base is constructed by converting the experience structured text entries corresponding to each preset historical design scheme review case data into vector representations using a vector embedding model, and constructing case experience vector knowledge bases for different facility categories based on the different facility types of the review objects in the historical design scheme review case data. The experience structured text entries include texts involving review logic descriptions and implicit rule descriptions.

[0069] In this embodiment, according to the target facility type of the review object, relevant regulatory information and historical review case information are matched from the preset knowledge base. (including review questions, review indicators, review process, etc.), using the vector embedding model to input questions Generate question vector embedding , based on question vector embedding Perform text entry query on the vector knowledge base (underground space review rule knowledge base, case experience knowledge base) corresponding to the same facility type, and compare it with the text entries stored in the knowledge base Vector embedding of For comparison:

[0070] ,

[0071] in, and A task-specific encoder.

[0072] Cosine similarity calculation is used for semantic matching. The specific formula is:

[0073] ,

[0074] Optionally, a re-ranking model can be set to prioritize the results based on relevance. A score threshold of 0.5 is set to set the similarity threshold for text item screening. Only text items with a similarity score greater than 0.5 are recalled. The parameters are set to retain the top 10 target rule structured text items with the highest similarity scores as the recall results, ensuring that all regulations relevant to the current review task are recalled. The top 3 target experience structured text items with the highest similarity scores are retained as references.

[0075] S13. Load the corresponding target review materials from the application materials of the current review task based on the review material types corresponding to the elements to be reviewed. Specifically, the material type requirements can be determined based on the elements to be reviewed of the review object in the current review task, and the corresponding review materials can be loaded from the submitted application materials based on the review material types corresponding to the elements to be reviewed.

[0076] S14. If the target review materials include textual review materials, a pre-trained underground space review reasoning model is invoked to extract first indicator information corresponding to the element to be reviewed from the textual review materials. In this embodiment, the underground space review reasoning model is an intelligent model that has been fine-tuned and trained based on a preset underground space review rule text corpus and a pre-set instruction alignment corpus. The underground space review rule text corpus contains rule-structured text entries, and the instruction alignment corpus contains empirically structured text entries describing review logic and implicit rules.

[0077] In this embodiment, if the target review materials involve text-based review materials, the underground space field review reasoning model is called. Based on the problem classification result, that is, the target facility type, the corresponding prompt word template is called to parse the text-based review materials such as the design text in the approval materials and extract the corresponding indicator information. Prompt word template example:

[0078] "This is the design proposal text for {project type}, from which {review indicator} information is extracted and output in JSON format."

[0079] S15. If the target review materials include drawings, a pre-trained multimodal model for enhanced underground space drawing recognition is invoked to extract second indicator information corresponding to the element to be reviewed from the drawings. In this embodiment, the multimodal model for enhanced underground space drawing recognition is a multimodal model fine-tuned and trained based on a pre-set image-text alignment corpus containing image-text pairs consisting of design drawing data and corresponding textual descriptions.

[0080] In this embodiment, if the target review materials involve drawings, the multimodal large model with enhanced drawing recognition in the underground space field is called. Based on the problem classification result, that is, the target facility type, the built-in prompt word template corresponding to the indicators to be reviewed is called to identify the corresponding indicator information in the design drawing. Prompt word template example:

[0081] "This is a {drawing type} design drawing for {project type}. Extract {review indicator} information from it. Only the information explicitly marked in the drawing will be recognized. Information not marked will be output as unknown. Do not make inferences on your own. The results will be output in JSON format."

[0082] S16. Integrate all the extracted indicator information and the recalled target rule structured text entries and target experience structured text entries to construct a prompt word template, and input the obtained prompt word template into the underground space review reasoning model to review whether the indicator information complies with each target rule structured text entry one by one, and combine the review logic description and implicit rule description in the target experience structured text entry to reason about the indicator review results and output a review report.

[0083] The knowledge-driven intelligent review method for underground building design schemes provided by the embodiment of the present invention uses a cross-modal attention mechanism to achieve semantic alignment of multimodal data such as underground space facility design specifications and guidelines, operation and maintenance period facility update design schemes, and empirical knowledge. It can integrate the review decisions and implicit logic in the specification provisions, historical cases, and empirical knowledge to form a knowledge system, realize automatic recognition and reasoning of underground building design schemes based on knowledge-driven, and improve the intelligence level and review efficiency of the review process. Moreover, the present invention can automatically realize intelligent review of a large number of drawing application materials in the design scheme, realize automated parsing of design drawing spatial elements, and improve the cross-modal semantic association analysis capability of drawing spatial information and design specification text.

[0084] In another optional embodiment of this aspect, Figure 2 As shown, if the elements to be reviewed include related indicators that require space calculation in addition to the text review materials or the drawing review materials, the method further includes step S17;

[0085] S17. If the elements to be reviewed include relevant indicators that require spatial calculation, the indicator calculation tool corresponding to the corresponding quantitative indicator is called according to the elements to be reviewed, so as to calculate and return the corresponding third indicator information using the indicator calculation tool.

[0086] In an embodiment of the present invention, a corresponding interactive interface is developed for the indicator calculation tool based on a lightweight network service architecture. After the interface is tested, the standard network request instructions of the interface are exported, and the network request instructions are converted into an open interface description specification using a large language model, which includes the interface definition and function description of the tool, forming a standardized function module that can be called by the intelligent model.

[0087] If the target review materials involve calculation of related indicators, the corresponding calculation tool is called through the function to return the corresponding indicator information.

[0088] In this example, a large model for review and reasoning in the underground space field is used to design a prompt word template. All extracted indicator information and recalled relevant standards and cases are integrated as context input. Each indicator information is compared to see if it complies with various standards and regulations. The review results are then reasoned and summarized based on relevant cases. A standard review report is output on the front end, and the three most relevant historical cases are pushed to the user for reference. The standard review report is then output. Prompt word template example:

[0089] Generate the following design review report based on the relevant information of the project:

[0090] {Constraint specification} + {Text analysis results} + {Drawing recognition results} + {Spatial calculation results}

[0091] {Example of Audit Report Format}”

[0092] In one embodiment of the present invention, after outputting the review report, the method further includes: determining whether the review report requires manual review based on a preset confidence scoring model; for review reports that require manual review, collecting feedback from reviewers on the accuracy of the review report, generating a reinforcement sample set based on the review report and the corresponding accuracy feedback, and feeding the reinforcement sample set back to the reinforcement learning layer to update the model parameters through incremental training. The specific implementation process is as follows:

[0093] Based on the preset confidence scoring model, it is determined whether the review results require manual review. For situations where manual review is required, manual feedback data is collected and sent to the reinforcement learning layer to achieve incremental training and updates of the system.

[0094] The confidence C is dynamically calculated through the confidence scoring model. The specific formula of the confidence scoring model is:

[0095] C=λ_r×R_rule+λ_c×C_rule+λ_s×S_case

[0096] R_rule is the standard relevance, which is specifically calculated as follows: the average matching rate of the structured text items of the recalled target rule in step S12.

[0097] C_rule is the accuracy of parameter compliance reasoning. It is calculated by extracting each target rule structured text entry for secondary verification and calculating the logical verification pass rate.

[0098] S_case is the historical decision similarity, which is calculated as follows: the average matching rate of the three target experience structured text items recalled in step S12.

[0099] λ_r, λ_c, and λ_s are the weights of the three, which can be optionally set to 0.2, 0.6, and 0.2 respectively.

[0100] When the confidence level C ≥ 90%, it is considered passed; when 80% ≤ C < 90%, secondary reasoning verification is initiated; when C < 80%, manual review is added.

[0101] A manual feedback button is set up, and the reviewers provide feedback on the accuracy of the review results and record them in the background. Reinforcement sample sets are generated regularly and fed back to the reinforcement learning layer to update the model parameters through incremental training.

[0102] In a specific example, the output is as follows:

[0103] 1. Review Report:

[0104] Review conclusion: To be revised

[0105] - Violation: Sidewalk width 3.0m (regulatory threshold 4.0-6.5m)

[0106] - Based on the clause: TCECS481-2017 Article 4.4.5

[0107] - Correction suggestion: widen the sidewalk width

[0108] - Confidence level: 88% (automatic review passed)

[0109] 2. Related historical cases:

[0110] -Case 1: In 2023, the width of a pedestrian walkway in a shopping mall was insufficient, so the width was adjusted to 4.5m. (Similarity 0.85)

[0111] -Case 2: In 2021, the height difference between the underground first floor and the outdoor entrance of an underground shopping mall exceeded the standard, and the height difference was adjusted to 8m. (Similarity 0.78)

[0112] -Case 3: In 2022, the evacuation passage of a civil air defense project was insufficient (similarity 0.72), and the passage width was adjusted to 2.4m.

[0113] Whether the report result is accurate: True / False (click the manual feedback button).

[0114] In one embodiment of the present invention, the input review task, i.e., the underground building design plan to be reviewed, can be pre-classified by calling a pre-trained underground space field review reasoning model, and the facility type and elements that need to be reviewed of the review object in the current review task can be judged and identified, and the large model outputs the secondary classification results.

[0115] At the same time, dynamic process configuration parameters (JSON format) are generated through the large model, including the review material type (text / drawing / calculation) corresponding to the elements to be reviewed, the selection of the standard knowledge base, the module call sequence, etc.

[0116] The system has built-in prompt word templates designed based on the review materials, key indicators (i.e. elements to be reviewed), and relevant specification catalogs required for review of various underground spaces such as underground shopping malls and civil air defense basements. The system then automatically organizes the review process based on the classification results, including selecting which unit modules need to be called, whether the calculation module needs to be called, and selecting the corresponding prompt words.

[0117] For example, when the "underground shopping mall floor plan review" task is detected, the material loading unit is automatically triggered to load the master plan and design instructions, and the knowledge base matching unit is simultaneously activated to search the "underground shopping mall related knowledge base"; the text parsing unit (the prompt word template is parameterized and injected into "shopping mall / floor plan") and the drawing parsing unit (calling the multimodal large model and using the configured prompt words to identify key indicators such as the total building area and pedestrian walkway width) are called in parallel; if the target to be reviewed in the classification result includes calculation requirements such as "business hall area ratio", the area ratio calculation tool is called through the function call unit; the above call results are automatically incorporated into the context input of the report generation unit's reasoning unit.

[0118] In order to realize the training of the underground space review reasoning large model and the underground space field drawing recognition enhanced multimodal large model, the present invention also includes the step of constructing various corpora.

[0119] First, data collection is carried out. The collection of design specifications and guidelines is the foundation of the entire system. For example, the "Regulations on the Development and Utilization of Urban Underground Space" and the "Design Guidelines for Urban Underground Commercial Space". The collection of design plans for facility updates during the operation and maintenance period is an important part of the entire system. Determine the types of facilities and the scope of design plans that need to be collected, such as underground shopping malls, civil air defense basements, etc.; collect historical underground building renewal design plans for these facilities, including project design plan texts, floor plans, elevations, renderings, etc. The collection of historical case experience knowledge is the core of the entire system, and the collection of past design plan review results, including review opinions, modification suggestions, etc.

[0120] Specifically, the steps for constructing the underground space review rules text corpus include:

[0121] 1. Extract the key review elements of underground building design schemes and the quantitative indicators of each key review element from the pre-acquired design specifications and guidelines for underground building design.

[0122] 2. Establish a mapping relationship between each quantitative indicator and the corresponding review material type, and / or establish a connection between each quantitative indicator and the corresponding indicator calculation tool interface. This will help analyze which materials need to identify which corresponding approval indicators and establish a systematic approval process. To ensure the comprehensiveness and accuracy of the review process, it is necessary to break down the review points in the design specification guidelines. This mainly includes the following steps: Step 1: Identify key clauses in the design specification guidelines, namely key review elements, including building scale, functional use, circulation organization, and floor plan layout. Step 2: Determine the quantitative indicators for these key review elements. For example, the total length of the underground commercial street should be less than 1500 meters, and the height difference between the ground floor of the first underground floor and the outdoor entrance and exit floor should not exceed 10 meters. Step 3: Clarify the corresponding review material type for each quantitative indicator (such as master plan, elevation, design proposal document, etc.) and establish a mapping relationship between the indicator and the review basis. Step 4: For indicators that require spatial calculations, such as area ratio, inter-node distance, and traffic vector line type, develop corresponding calculation tools based on the Python language. For example, extract closed polygons from a CAD plan, calculate the area, and then calculate ratio statistics.

[0123] 3. Extract the specification identification information and chapter content structure of the design specifications and guidelines for underground building design. Use a pre-set multi-level regularization template to segment the chapter content structure to generate regular structured text entries. Specification identification information includes the design specification and guideline title, standard number, release date, implementation date, and the province, city, and district to which the local standard belongs.

[0124] 4. According to the different types of facilities under review, the specification identification information and rule structured text entries corresponding to the relevant design specifications and guidelines for underground building design are used to construct different categories of underground space review rule text corpora according to the facility types.

[0125] In this embodiment, for standard guideline data, most of the data that can be collected on the Internet is stored in the form of PDF. A general multimodal large model is used for text extraction. By designing a prompt word template, key information such as the standard title, standard number, as well as time and space tags such as the release / implementation date and the province, city, and district to which the local standard belongs are extracted.

[0126] Input prompt word:

[0127] "Parse the current page content and extract the following fields: Title: the first complete paragraph, including quotation marks; Standard Number: in accordance with the XX / T XXXXX-XXXX format; Implementation Date: 'Effective from X / X / XXXX'; Location: the 'XX Province / City' keyword from the local standard; Text: exclude headers / footers / watermarks, retaining the chapter structure. Output is JSON, with chapters divided into three levels: 'chapter-section-article'. {Document Image Per Page}"

[0128] Multimodal large model output:

[0129] "{

[0130] "Title": "Design Guidelines for Urban Underground Commercial Spaces",

[0131] "Standard Number": "T / CECS 481-2017",

[0132] "Effective Date": "Effective from December 1, 2017",

[0133] "Territory": "China",

[0134] "Main text": {

[0135] "Chapter": "1 General Provisions",

[0136] "Festival": [

[0137] { "Article": "1.0.1",

[0138] "Content": "To meet the needs of large-scale urban underground commercial space development and construction in my country..."},

[0139] { "Article": "1.0.2",

[0140] "content": "......"}, ]}

[0141] }Use multi-level regular templates to cut the main text and generate regular structured text entries, which are stored as {title}-{standard number}-{chapter number}-{clause number}: {content}.

[0142] All the above rule-structured text entries constitute the underground space review rule text corpus.

[0143] At the same time, the rule-structured text entries that have been segmented are classified according to the facility types of the review objects to construct a text corpus of underground space review rules, such as the underground shopping mall design-related specification corpus and the underground pipeline gallery design-related specification corpus; the structured entries that have been segmented are converted into vector representations using a vector embedding model, and an index is constructed to generate a multi-category underground space review rule vector knowledge base that supports semantic queries, namely the underground space review rule knowledge base.

[0144] Specifically, the steps for constructing the image-text aligned corpus include: using a multimodal large model to process the design drawing data in the pre-acquired historical underground building renewal design plan to generate text description information corresponding to the drawing data, and establishing an effective semantic correspondence between the design drawing data and the corresponding text description information to generate a image-text aligned corpus; wherein the text description information includes design description, technical and economic indicator information, objective numerical indicator information, key structural targets and their coordinate frame information, and description of design rationality.

[0145] In this embodiment, for the floor plans and elevations in the design plan, a universal multimodal large model is used to extract information from the drawings using design prompts. This is combined with human correction to modify information that the large model incorrectly identifies and supplement information that the large model has not identified, ensuring the comprehensiveness and accuracy of the description of the drawing information. Specific content includes:

[0146] (1) Design description: Provide a detailed description of the drawings, including design intent, layout arrangements, etc.

[0147] (2) Technical and economic indicator information: Extract the main technical and economic indicators and other textual data from the master plan to ensure the accuracy and completeness of the data.

[0148] (3) Objective numerical indicator information: numerical extraction and interpretation of objective indicators such as annotations.

[0149] (4) Key structural targets and their coordinate frame information: Identify key structural targets such as entrances and exits and extract their coordinate frames.

[0150] (5) Description of design rationality: Description of design rationality of subjective indicators such as streamline organization.

[0151] We build a structured multimodal prompt word template library, match design drawings with text descriptions, perform pixel-level alignment, and generate an image-text alignment corpus. The image-text alignment corpus specifically includes two types of data: one is the image description generation task dataset obtained from the image description generation task, including image-text pairs of drawings and text descriptions; the other is the drawing question answering task dataset obtained from the drawing-based question answering task, for example:

[0152] “Q: Extract the main technical and economic indicators from the drawings:

[0153] {Review Drawings}

[0154] A:

[0155] 1. Total land area unit: m2 Data: 15794;

[0156] 2. Building area unit: m2 Data: 6485;

[0157] …”

[0158] Specifically, the steps for constructing the instruction alignment corpus include:

[0159] 1. Obtain target text involving the review process and decision logic from pre-acquired historical design review case data, and structure the target text to establish a structured review logic description.

[0160] 2. Obtain the review record text from the historical design scheme review case data, and extract the implicit rule description related to the review factor priority determination from the review record text.

[0161] 3. Extract case identification information from historical design plan review case data, combine the case identification information, review logic description and implicit rule description to generate empirical structured text entries, and form an instruction alignment corpus with all the obtained empirical structured text entries.

[0162] In this example, natural language processing technology is used to establish a structured review logic chain based on empirical knowledge from historical design review cases, systematizing and standardizing the review process and decision-making logic. For example, the sentence "Insufficient evacuation width must be expanded to ≥ 2.4m" is parsed into {Indicator: evacuation width, Current value: less than 2.4m, Specification requirement: ≥ 2.4m (with original specification text), Solution: Widen to 2.4m}. Implicit rules, such as "Prioritize evacuation width > commercial area," are extracted from review records. Cases are combined into structured text {Case name: xx, Time: xx, Location: xx, Details: xxx} to construct a case experience vector knowledge base.

[0163] Specifically, we design a structured prompt word template, construct the instruction alignment corpus, integrate the basic project data (information that can be extracted from the review materials and is consistent with the annotations in the image-text alignment corpus) and the applicable standards in the rule corpus on the input side, and generate a structured review report on the output side according to the thinking chain of "conflict detection → priority determination → solution recommendation", for example:

[0164] “Q: Generate a review report based on the specific information of the following {project type} project:

[0165] {Project basic data}+{Constraint specification}

[0166] A:

[0167] <think>

[0168] 1. Conflict detection: {Indicator A} violates {Specification Clause A} (threshold = {X}, actual value = {Y}); {Indicator B} violates {Specification Clause B} (threshold = {T}, actual value = {Z});

[0169] 2. Priority determination: Based on the "{implicit rule}" rule, it is recommended to prioritize adjusting {Indicator B}.

[0170] 3. Recommended solutions: 1) Adjust {Indicator B} to {Specification B Constraint Value}; 2) Adjust {Indicator A} to {Specification A Constraint Value}

[0171] < / think>

[0172] <answer>

[0173] {Review Conclusion}

[0174] < / answer> ”

[0175] Next, we implement fine-tuning training for each large model based on the underground space review rule text corpus, instruction alignment corpus, and image-text alignment corpus obtained above.

[0176] In the embodiment of the present invention, the training process of the underground space review reasoning model includes:

[0177] 1. Using the preset underground space review rule text corpus as training data, the pre-trained base model is fine-tuned using a dynamic low-rank adaptive fine-tuning method. During the fine-tuning training process, the original base model weight parameters are frozen, and a trainable rank decomposition matrix is ​​injected into each layer. Different rank parameter values ​​are assigned to training data of different standard types according to the different standard types of the rule-structured text entries, so as to dynamically adjust the rank parameter value according to the complexity of the task. Specifically, a pre-trained base model is selected. In this embodiment, the large language model Deepseek-R1 can be selected, and the pre-trained large language model parameters are loaded according to the official documentation of the model. According to the incremental pre-training code in the base model description document, the rule corpus is loaded as training data, and dynamic low-rank adaptive fine-tuning is used to fine-tune the base model by adjusting the training hyperparameters such as the learning rate, the number of training rounds, and the block size. There are two main methods for large model fine-tuning: full parameter fine-tuning and partial parameter fine-tuning. The present invention uses a low-rank fine-tuning method to fine-tune some parameters. By freezing the original pre-trained model weights based on the pre-trained large model and injecting a trainable rank decomposition matrix into each layer, the parameters that need to be trained are greatly reduced, thereby achieving fine-tuning training of the pre-trained model. Furthermore, the present invention designs an improved dynamic low-rank fine-tuning method to sort learning representations of different ranks during the training process. The rank of a matrix is ​​the number of its linearly independent rows or columns. In the context of dynamic low-rank fine-tuning, a low-rank matrix is ​​used to approximate the weight matrix. Rank parameters (r = 8 to 64) can be assigned to different specification types based on different task complexities, such as setting r = 64 for structural safety class, r = 32 for floor plan class, and r = 16 for streamline design class. This allows for dynamic adjustment of the parameter amount according to the complexity of the task and improves training efficiency.

[0178] 2. The parameter fusion algorithm is used to combine the fine-tuning weights and the base model to generate a large model enhanced with underground space domain knowledge. The model is based on the general large language model and incorporates underground space domain knowledge. The parameter fusion algorithm formula is:

[0179] ,

[0180] in, ΔW represents the frozen weight parameters of the base language model, while ΔW represents the low-rank matrix parameters for fine-tuning training. Only the additional ΔW parameters are trained. B and A represent the weight matrices to be trained. The B matrix has dimensions of d × r, corresponding to the combination of the model framework's input dimension d (e.g., 4096) and the rank parameter r. It is used to extract low-rank features related to specific canonical types from the input training data. The A matrix has dimensions of r × k, mapping the low-rank features to the model's output dimension k. This allows the extracted features to be recombined to generate parameter adjustments adapted to the domain knowledge. α is a preset domain adaptation coefficient (empirically set to 0.5), which ensures controllable knowledge injection strength and prevents ΔW from over-covering pre-trained general domain knowledge.

[0181] 3. Using the preset instruction alignment corpus as training data, the underground space domain knowledge enhancement model is fine-tuned using a low-rank fine-tuning method. During the fine-tuning training process, the weight parameters of the underground space domain knowledge enhancement model are frozen, and a rank decomposition matrix with a specified rank is injected into the Transformer layer. The underground space review and reasoning model is obtained through supervised learning optimization training. Specifically, by loading the instruction alignment corpus, a fixed-value low-rank fine-tuning method is used, the base model parameters are frozen, and only a low-rank matrix with a rank of 32 is injected into the Transformer layer. The instruction response matching is optimized through supervised learning. Fine-tuning the underground space domain knowledge enhancement model to obtain the instruction alignment model, that is, the initial underground space review and reasoning model.

[0182] Furthermore, after obtaining the initial underground space review reasoning model, the method further includes:

[0183] 4. Multiple output results of the underground space review reasoning model under the same prompt word instruction input are marked in order of merit according to the preset evaluation criteria. For example, the order of merit is marked according to the three-level criteria of "accuracy of specification reference > calculation logic completeness > expression clarity" to obtain the preference ranking result;

[0184] 5. Establish a multi-objective reward model and perform backpropagation training based on the preference ranking results. The multi-objective reward model is:

[0185] ;

[0186] ,

[0187] in, is a given hint, r is Instructions to align the N responses generated by the model ; represents a trainable reward model, Represents a rule reward, matching the rule structured text entry in the response r through a regular expression, that is, matching the corresponding rule structured text entry content from the underground space review rule text corpus through a regular expression, The weight is based on the rule, set to 0.5-0.7. Each valid clause is recorded point; Represents human preference rewards, and ranks the scoring results by human preference. is the empirical correction coefficient, set to 0.3-0.5; The sigmoid function is shown. According to the scoring results, the response Better than , by minimizing the objective function of the multi-objective reward model Conduct training.

[0188] 6. Use the proximal strategy to optimize and fine-tune the underground space review reasoning model, use the reward model output as the reinforcement signal, iteratively update the low-rank matrix parameter ΔW of the underground space review reasoning model to maximize the score of the reward model, and obtain the optimized underground space review reasoning model.

[0189] The present invention uses a low-rank fine-tuning method to perform instruction alignment fine-tuning training on the underground space domain knowledge enhancement large model obtained by fine-tuning in the knowledge injection stage, then uses human feedback to train the reward model, and then guides the large language model through reinforcement learning to ensure that the model can output professional answers with reasoning logic according to the instructions.

[0190] In an embodiment of the present invention, the training process of the enhanced multimodal large model for underground space field drawing recognition includes: using a preset image-text aligned corpus as training data, and adopting a dynamic low-rank adaptive fine-tuning method to fine-tune the general multimodal large model to obtain an enhanced multimodal large model for underground space field drawing recognition. This large model can extract information from underground space field design drawings based on prompt words.

[0191] In this embodiment, first, a pre-trained base large model is selected. In this embodiment, the multimodal large model Qwen2.5-VL is selected, and the pre-trained large language model parameters are loaded according to the official documentation of the model. The general multimodal large model includes a visual encoder, a feature mapping matrix, and a language model. Among them, the visual encoder is used to convert image information into a feature vector; the feature mapping matrix is ​​used to project image features into the text semantic space; the language model is used to process text instructions and generate output. Furthermore, by running the fine-tuning training part code in the base multimodal large model file, the underground space field image-text alignment corpus is loaded for fine-tuning. The general multimodal large model is fine-tuned and trained using a dynamic low-rank adaptive fine-tuning method, which specifically includes:

[0192] In the image-text feature alignment stage, the multimodal large model visual encoder is frozen, the rank parameter r is set to 16, and the image description generation task dataset in the image-text alignment corpus is used to fine-tune the training feature mapping matrix to achieve image-text semantic alignment;

[0193] In the joint optimization phase, the visual encoder is unfrozen, the rank parameter r is set to 32, and the drawing question answering task dataset in the image-text alignment corpus is used to simultaneously fine-tune the training feature matrix and language model parameters to achieve fine-grained feature matching and image-text reasoning.

[0194] According to the parameter merging code in the base model description document, the parameters trained in the image-text feature alignment phase are merged with the parameters trained in the joint optimization phase to obtain an enhanced multimodal large model for underground space drawing recognition.

[0195] For simplicity of description, the method embodiments are described as a series of actions. However, those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because certain steps can be performed in other orders or simultaneously according to the embodiments of the present invention. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0196] Another embodiment of the present invention further provides a knowledge-driven intelligent review system for underground building design solutions, which includes a functional module for implementing the knowledge-driven intelligent review method for underground building design solutions as described in any of the above items. Figure 3 The following schematically shows a structural block diagram of a knowledge-driven intelligent review system for underground building design solutions according to another embodiment of the present invention. Figure 3 The knowledge-driven intelligent review system for underground building design solutions of this embodiment specifically includes a question classification unit 301, a knowledge base matching unit 302, a review material loading unit 303, a text parsing unit 304, a drawing parsing unit 305, and a report generating unit 306, wherein:

[0197] The question classification unit 301 is used to identify the target facility type and the elements to be reviewed of the review object in the underground building design plan to be reviewed, and determine the review material type corresponding to the elements to be reviewed.

[0198] In this embodiment, the problem classification unit 301 can perform pre-classification on the input review task, i.e., the underground building design plan to be reviewed, by calling the pre-trained underground space field review reasoning big model, judge and identify the facility type and elements that need to be reviewed of the review object in the current review task, and the big model outputs the secondary classification results.

[0199] Furthermore, the problem classification unit 301 can also generate dynamic process configuration parameters (JSON format) from the large model, including the type of review materials (text / drawings / calculations) corresponding to the review elements, the specification knowledge base selection, and the module call sequence. Specifically, the system includes built-in prompt word templates designed based on the review materials, key indicators (i.e., review elements), and relevant specification catalogs required for review of various underground spaces such as underground shopping malls and civil air defense basements. Based on the classification results, the system automatically organizes the review process, including selecting which modules to call, whether to call the calculation module, and selecting the corresponding prompt words.

[0200] The knowledge base matching unit 302 is used to match the target rule structured text entries from the preset underground space review rule knowledge base of the corresponding facility type according to the target facility type to obtain relevant specification fragments, and to match the target experience structured text entries from the preset case experience knowledge base of the corresponding facility type to obtain relevant case experience fragments.

[0201] The review material loading unit 303 is used to load corresponding target review materials from the application materials of the current review task according to the review material type corresponding to the element to be reviewed.

[0202] The text parsing unit 304 is used to call the pre-trained underground space review reasoning model to extract the first indicator information corresponding to the element to be reviewed from the text review materials when the target review materials include text review materials.

[0203] The drawing parsing unit 305 is used to call the pre-trained underground space field drawing recognition enhanced multimodal large model to extract the second indicator information corresponding to the elements to be reviewed from the drawing review materials when the target review materials include drawing review materials.

[0204] The report generation unit 306 is used to integrate all the extracted indicator information and the recalled target rule structured text entries and target experience structured text entries to construct a prompt word template, and input the obtained prompt word template into the underground space review reasoning model to review whether the indicator information complies with each target rule structured text entry one by one, and combine the review logic description and implicit rule description in the target experience structured text entry to reason about the indicator review results and output a review report.

[0205] In an embodiment of the present invention, the system also includes a function calling unit, which is used to call the indicator calculation tool corresponding to the corresponding quantitative indicator according to the element to be reviewed when the element to be reviewed includes relevant indicators that require spatial calculation, so as to use the indicator calculation tool to calculate and return the corresponding third indicator information.

[0206] In an embodiment of the present invention, the system also includes a feedback verification unit, which is used to determine whether the review report requires manual review based on a preset confidence scoring model after outputting the review report; for review reports that require manual review, the reviewer's feedback on the accuracy of the review report is collected, and a reinforcement sample set is generated based on the review report and the corresponding accuracy feedback, and the reinforcement sample set is fed back to the reinforcement learning layer to update the model parameters through incremental training.

[0207] Figure 4 The following is a schematic diagram showing the implementation steps of the knowledge-driven intelligent review system for underground building design schemes proposed in an embodiment of the present invention. Figure 4 As shown, for an underground building design proposal to be reviewed, the question classification unit identifies the target facility type, review elements, and corresponding review material types. The knowledge base matching unit matches rule-structured text entries and empirical structured text entries from the knowledge base of the corresponding facility type based on the target facility type. The review material loading unit loads the corresponding review materials from the application materials based on the review material type. For review materials that do not involve calculations, if the review materials involve text, the text parsing unit invokes the review reasoning macromodel to extract the indicator information corresponding to the review elements. If the review materials involve drawings, the drawing parsing unit invokes the drawing recognition enhanced multimodal macromodel to extract the indicator information corresponding to the review elements. If the review elements include relevant indicators requiring spatial calculation, the function calling unit invokes the indicator calculation tool corresponding to the corresponding quantitative indicator based on the review elements. The indicator calculation tool calculates and returns the corresponding third indicator information. Finally, the report generation unit integrates the extracted indicator information and recalled knowledge items to construct a prompt word template. The review reasoning macromodel performs intelligent review and reasoning on the indicator information based on the prompt word template and outputs a review report. After the review report is output, the feedback verification unit determines whether the review report requires manual review based on a preset confidence scoring model. For review reports that require manual review, feedback on the reviewer's accuracy is collected. A reinforcement sample set is generated based on the review report and the corresponding accuracy feedback. The reinforcement sample set is fed back to the reinforcement learning layer to update the model parameters through incremental training. When the confidence score model dynamically calculates a confidence level C ≥ 90%, it is considered verified and no manual review is performed. When 80% ≤ C < 90%, secondary reasoning verification is initiated. When C < 80%, manual review is added.

[0208] As for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0209] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0210] In addition, another embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor; when the computer program is executed by the processor, the steps of the knowledge-driven intelligent review method for underground building design schemes as described above are implemented.

[0211] In addition, another embodiment of the present invention further provides a computer program product, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the above-mentioned knowledge-driven intelligent review method for underground building design schemes.

[0212] Those skilled in the art will appreciate that although some embodiments herein include certain features included in other embodiments but not other features, the combination of features from different embodiments is intended to be within the scope of the present invention and to form different embodiments. For example, any of the claimed embodiments may be used in any combination.

[0213] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A knowledge-driven intelligent review method for underground building design schemes, characterized in that: The method comprises: Identify the target facility type and elements to be reviewed in the underground building design plan to be reviewed, and determine the type of review materials corresponding to the elements to be reviewed; According to the target facility type, target rule structured text entries are matched from a preset underground space review rule knowledge base of the corresponding facility type, and target experience structured text entries are matched from a preset case experience knowledge base of the corresponding facility type; Load the corresponding target review materials from the application materials of the current review task according to the review material type corresponding to the element to be reviewed; If the target review materials include text review materials, the pre-trained underground space review reasoning model is called to extract the first indicator information corresponding to the element to be reviewed from the text review materials; If the target review materials include drawing review materials, the pre-trained underground space drawing recognition enhanced multimodal large model is called to extract the second indicator information corresponding to the elements to be reviewed from the drawing review materials; Integrate all extracted indicator information and recalled target rule structured text entries and target experience structured text entries to construct a prompt word template, input the obtained prompt word template into the underground space review reasoning model to review whether the indicator information complies with each target rule structured text entry one by one, and combine the review logic description and implicit rule description in the target experience structured text entry to reason the indicator review results and output a review report; The training process of the underground space review reasoning model includes: Using the preset underground space review rule text corpus as training data, the pre-trained base model is fine-tuned using a dynamic low-rank adaptive fine-tuning method. During the fine-tuning training process, the original base model weight parameters are frozen, and a trainable rank decomposition matrix is ​​injected into each layer. Different rank parameter values ​​are assigned to training data of different specification types according to the different specification types to which the rule-structured text entries belong. Specifically, the learning representations of different ranks are sorted during the training process, and the weight matrix is ​​approximated using a low-rank matrix. Different rank parameter values ​​are assigned to different specification types according to different task complexities, so as to dynamically adjust the rank parameter value according to the complexity of the task. The parameter fusion algorithm is used to combine the fine-tuning weights and the base model to generate a large model enhanced with underground space domain knowledge. The parameter fusion algorithm formula is: ， in, Represents the frozen base language model weight parameters, To fine-tune the low-rank matrix parameters for training, B and A represent the weight matrices to be trained. The dimension of the B matrix is ​​d×r, corresponding to the combination of the input dimension d and the rank parameter r of the model framework, and is used to extract low-rank features related to specific specification types from the input training data; the dimension of the A matrix is ​​r×k, which is used to map the low-rank features to the output dimension k of the model, so as to recombine the extracted features to generate parameter adjustments that adapt to the domain knowledge. α is the preset domain adaptation coefficient. The preset instruction alignment corpus is used as training data, and the low-rank fine-tuning method is adopted to fine-tune the underground space domain knowledge enhancement large model. During the fine-tuning training process, the weight parameters of the underground space domain knowledge enhancement large model are frozen, and the rank decomposition matrix with a specified rank is injected into the Transformer layer. The underground space review reasoning large model is obtained through supervised learning optimization training.

2. The method according to claim 1, characterized in that The method further comprises: If the elements to be reviewed include relevant indicators that require spatial calculation, the indicator calculation tool corresponding to the corresponding quantitative indicator will be called according to the elements to be reviewed, so as to use the indicator calculation tool to calculate and return the corresponding third indicator information.

3. The method according to claim 1, characterized in that After obtaining the underground space review reasoning model, the method further includes: Multiple output results of the underground space review reasoning model under the same prompt word instruction input are marked in order of merit according to the preset evaluation criteria to obtain the preference ranking result; A multi-objective reward model is established. Based on the multi-objective reward model, the preference ranking results are used for back propagation training. The multi-objective reward model is: ； ， in, is a given hint, r is Instructions to align the N responses generated by the model ; represents a trainable reward model, Indicates rule rewards, is the rule-based weight; represents human preference for rewards, is the empirical correction coefficient; is the objective function of the multi-objective reward model, Represents the sigmoid function, response Better than , by minimizing the objective function Conduct training; The proximal strategy is used to optimize and fine-tune the underground space review reasoning model. The reward model output is used as a reinforcement signal. The low-rank matrix parameter ΔW of the underground space review reasoning model is iteratively updated to maximize the score of the reward model, and the optimized underground space review reasoning model is obtained.

4. The method according to claim 1, wherein The steps for constructing the underground space review rules text corpus include: Extract key review elements of underground building design schemes and quantitative indicators of each key review element from pre-acquired design specifications and guidelines related to underground building design; Establish a mapping relationship between each quantitative indicator and the type of review material corresponding to the corresponding quantitative indicator, and / or establish a connection relationship between each quantitative indicator and the interface of the indicator calculation tool corresponding to the corresponding quantitative indicator; Extract the specification identification information and chapter content structure of the design specification guidelines related to underground building design, and use the preset multi-level regular template to segment the chapter content structure to generate regular structured text entries; According to the different types of facilities under review, the specification identification information and rule structured text entries corresponding to the relevant design specifications and guidelines for underground building design are used to construct different categories of underground space review rule text corpora according to the facility types.

5. The method according to claim 1, wherein The steps for constructing the instruction alignment corpus include: Obtain target text involving the review process and decision logic from pre-acquired historical design review case data, and structure the target text to establish a structured review logic description; Obtaining review record text from the historical design scheme review case data, and extracting implicit rule descriptions related to review factor priority determination from the review record text; Case identification information is extracted from historical design scheme review case data, and the case identification information, review logic description and implicit rule description are combined to generate empirical structured text entries. All the obtained empirical structured text entries are combined into an instruction alignment corpus.

6. The method according to claim 1, characterized in that The training process of the underground space field drawing recognition enhanced multimodal large model includes: Using the preset image-text aligned corpus as training data, the general multimodal large model is fine-tuned and trained using a dynamic low-rank adaptive fine-tuning method to obtain an enhanced multimodal large model for drawing recognition in the underground space field.

7. The method according to claim 6, characterized in that The general multimodal large model includes a visual encoder, a feature map matrix, and a language model; Fine-tune the general multimodal large model using dynamic low-rank adaptive fine-tuning, including: In the image-text feature alignment stage, the multimodal large model visual encoder is frozen, the rank parameter r is set to 16, and the image description generation task dataset in the image-text alignment corpus is used to fine-tune the training feature mapping matrix to achieve image-text semantic alignment; In the joint optimization phase, the visual encoder is unfrozen, the rank parameter r is set to 32, and the drawing question answering task dataset in the image-text alignment corpus is used to simultaneously fine-tune the training feature matrix and language model parameters to achieve fine-grained feature matching and image-text reasoning. The parameters trained in the image-text feature alignment stage are merged with the parameters trained in the joint optimization stage to obtain an enhanced multimodal large model for underground space drawing recognition.

8. The method according to claim 6, characterized in that The steps to construct the image-text alignment corpus include: A multimodal large model is used to process the design drawing data of the pre-acquired historical underground building renewal design plan to generate textual description information corresponding to the drawing data, and a semantic correspondence is established between the design drawing data and the corresponding textual description information to generate a graphic-text aligned corpus; the textual description information includes design description, technical and economic indicator information, objective numerical indicator information, key structural targets and their coordinate frame information, and description of design rationality.

9. The method according to claim 1, characterized in that After outputting the review report, the method further includes: Determine whether the review report requires manual review based on the preset confidence scoring model; For review reports that require manual review, the reviewers' feedback on the accuracy of the review reports is collected, and a reinforcement sample set is generated based on the review reports and the corresponding accuracy feedback. The reinforcement sample set is fed back to the reinforcement learning layer to update the model parameters through incremental training.

10. A knowledge-driven intelligent review system for underground building design schemes, characterized by: The system comprises: A question classification unit is used to identify the target facility type and elements to be reviewed in the underground building design plan to be reviewed, and to determine the type of review materials corresponding to the elements to be reviewed; A knowledge base matching unit is used to match target rule structured text entries from a preset underground space review rule knowledge base of the corresponding facility type according to the target facility type, and to match target experience structured text entries from a preset case experience knowledge base of the corresponding facility type; An examination material loading unit is used to load corresponding target examination materials from the application materials of the current examination task according to the examination material type corresponding to the element to be examined; A text parsing unit, configured to, when the target review materials include text review materials, call a pre-trained underground space review reasoning model to extract first indicator information corresponding to the element to be reviewed from the text review materials; A drawing parsing unit is configured to, when the target review materials include drawing review materials, call a pre-trained underground space field drawing recognition enhanced multimodal large model to extract second indicator information corresponding to the elements to be reviewed from the drawing review materials; A report generation unit is used to integrate all extracted indicator information and recalled target rule structured text entries and target experience structured text entries to construct a prompt word template, input the obtained prompt word template into the underground space review reasoning model to review whether the indicator information complies with each target rule structured text entry one by one, and combine the review logic description and implicit rule description in the target experience structured text entry to reason the indicator review result and output a review report; The training process of the underground space review reasoning model includes: Using the preset underground space review rule text corpus as training data, the pre-trained base model is fine-tuned using a dynamic low-rank adaptive fine-tuning method. During the fine-tuning training process, the original base model weight parameters are frozen, and a trainable rank decomposition matrix is ​​injected into each layer. Different rank parameter values ​​are assigned to training data of different specification types according to the different specification types to which the rule-structured text entries belong. Specifically, the learning representations of different ranks are sorted during the training process, and the weight matrix is ​​approximated using a low-rank matrix. Different rank parameter values ​​are assigned to different specification types according to different task complexities, so as to dynamically adjust the rank parameter value according to the complexity of the task. The parameter fusion algorithm is used to combine the fine-tuning weights and the base model to generate a large model enhanced with underground space domain knowledge. The parameter fusion algorithm formula is: ， in, Represents the frozen base language model weight parameters, To fine-tune the low-rank matrix parameters for training, B and A represent the weight matrices to be trained. The dimension of the B matrix is ​​d×r, corresponding to the combination of the input dimension d and the rank parameter r of the model framework, and is used to extract low-rank features related to specific specification types from the input training data; the dimension of the A matrix is ​​r×k, which is used to map the low-rank features to the output dimension k of the model, so as to recombine the extracted features to generate parameter adjustments that adapt to the domain knowledge. α is the preset domain adaptation coefficient. The preset instruction alignment corpus is used as training data, and the low-rank fine-tuning method is adopted to fine-tune the underground space domain knowledge enhancement large model. During the fine-tuning training process, the weight parameters of the underground space domain knowledge enhancement large model are frozen, and the rank decomposition matrix with a specified rank is injected into the Transformer layer. The underground space review reasoning large model is obtained through supervised learning optimization training.

11. A computer device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 9 are implemented.

12. A computer program product, characterized in that The computer program product stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images