Knowledge-driven underground building design scheme intelligent examination method, system and equipment

Abstract

The invention provides a knowledge-driven underground building design scheme intelligent review method, system and equipment. The method comprises the following steps: identifying a target facility type of a review object, a to-be-reviewed element and a corresponding review material type in a to-be-reviewed underground building design scheme; according to the target facility type, matching a rule structured text entry and an experience structured text entry from the knowledge base of the corresponding facility type; loading a corresponding review material from the application material according to the type of the review material; if the to-be-examined elements relate to the text type examination materials, calling an examination reasoning large model to extract index information corresponding to the to-be-examined elements; if the to-be-examined elements relate to the drawing type examination materials, calling a drawing identification enhanced multi-modal large model to extract index information corresponding to the to-be-examined elements; the extracted index information and the recalled knowledge entries are integrated to construct a cue word template, intelligent review and reasoning are performed on the index information according to the cue word template through a review and reasoning large model, and a review report is output, so that the review efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and underground space facility design review, and particularly relates to a knowledge-driven intelligent review method, system and device for underground building design schemes. Background Art

[0002] The review of the design and update schemes of underground buildings has long relied on empirical judgment and subjective evaluation, and there are problems such as fragmented review logic and insufficient process standardization. Empirical knowledge is difficult to be effectively quantified, and individual judgment differences are likely to affect the consistency and objectivity of decisions. At the same time, the underground space design specification system is huge and dynamically evolving. Existing review systems mostly adopt static threshold determination mechanisms, which can neither integrate specification articles, historical cases and review decision tracks to form a knowledge system, nor can they analyze the implicit logical relationships in empirical knowledge.

[0003] On the other hand, the design of underground buildings contains a large number of drawing application materials. Manual review of paper drawings is cumbersome. Existing systems cannot achieve automatic parsing of the spatial elements of design drawings, lack the ability of cross-modal semantic association analysis between drawing spatial information and design specification texts, cannot achieve automatic recognition and reasoning, and the intelligence level of the review process is insufficient, affecting 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 schemes that overcome the above problems or at least partially solve the above problems.

[0005] In one aspect of the present invention, there is provided a knowledge-driven intelligent review method for underground building design schemes, the method comprising: Identifying the target facility type and the elements to be reviewed of the review object in the underground building design scheme to be reviewed, and determining the review material type corresponding to the elements to be reviewed; Matching the target rule structured text entry from the underground space review rule knowledge base of the corresponding facility type preset according to the target facility type, and matching the target experience structured text entry from the case experience knowledge base of the corresponding facility type preset; Loading the corresponding target review material from the application materials of the current review task according to the review material type corresponding to the elements to be reviewed; If the target review materials include text - type review materials, call the pre - trained large underground space review inference model to extract the first index information corresponding to the elements to be reviewed from the text - type review materials; the underground space review inference model is an intelligent large model fine - tuned based on a preset text corpus of underground space review rules and an instruction alignment corpus. The text corpus of underground space review rules records rule - structured text entries, and the instruction alignment corpus records empirical structured text entries involving review logic descriptions and implicit rule descriptions; If the target review materials include drawing - type review materials, call the pre - trained enhanced multi - modal large model for drawing recognition in the underground space field to extract the second index information corresponding to the elements to be reviewed from the drawing - type review materials; the enhanced multi - modal large model for drawing recognition in the underground space field is a multi - modal large model fine - tuned based on a preset graphic - text alignment corpus. The graphic - text alignment corpus records graphic - text pairs composed of design drawing data and its corresponding text description information; Integrate all the extracted index information, the recalled target rule - structured text entries, and target empirical structured text entries to construct a prompt template, and input the obtained prompt template into the underground space review inference model to review item - by - item whether the index information conforms to each target rule - structured text entry, and infer and output a review report on the index review results in combination with the review logic description and implicit rule description in the target empirical structured text entries.

[0006] Furthermore, the method further includes: If the elements to be reviewed include relevant indicators that require spatial calculation, call the corresponding index calculation tool for the corresponding quantitative indicators according to the elements to be reviewed, so as to calculate and return the corresponding third index information using the index calculation tool.

[0007] Furthermore, the training process of the underground space review inference model includes: Using the preset text corpus of underground space review rules as training data, fine - tune the pre - trained base large model in a dynamic low - rank adaptive fine - tuning manner. During the fine - tuning training process, freeze the weight parameters of the original base large model, inject trainable rank - decomposition matrices into each layer, and assign different rank parameter values to the 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 values according to the complexity of the task; Merge the fine - tuned weights and the base model through a parameter fusion algorithm to generate a knowledge - enhanced large model in the underground space field; Using a pre-set instruction-aligned corpus as training data, a large model enhanced with underground space domain knowledge is fine-tuned using the low-rank fine-tuning method. During the fine-tuning training process, the weight parameters of the large model enhanced with underground space domain knowledge are frozen, and a rank decomposition matrix with a specified rank is injected into the Transformer layer. Through supervised learning optimization training, a large model for underground space review and inference is obtained.

[0008] Further, after obtaining the large model for underground space review and inference, the method further includes: Marking the superiority order of multiple output results of the large model for underground space review and inference under the input of the same prompt word instruction according to a pre-set evaluation criterion to obtain a preference ranking result; Establishing a multi-objective reward model, and performing backpropagation training according to the multi-objective reward model using the preference ranking result. The multi-objective reward model is: ; , wherein, is a given prompt, r is for the N responses generated by the instruction alignment model ; represents a trainable reward model, denotes the rule reward, and the rule-structured text entries in the response r are matched through regular expressions, is the rule-based weight; represents the human preference reward, is the experience correction coefficient; is the objective function of the multi-objective reward model, represents the sigmoid function, the response is superior to , and the training is performed by minimizing the objective function ; Using proximal policy optimization to fine-tune the large model for underground space review and inference, taking the output of the reward model as the reinforcement signal, and iteratively updating the low-rank matrix parameter ΔW of the large model for underground space review and inference to maximize the score of the reward model, thereby obtaining an optimized large model for underground space review and inference.

[0009] Further, the construction steps of the underground space review rule text corpus include: Extracting the key review elements of the underground building design scheme and the quantification indexes of each key review element from the pre-obtained design specifications and guidelines related to underground building design; Establishing the mapping relationship between each quantification index and the corresponding review material type of the quantification index, and / or establishing the connection relationship between each quantification index and the corresponding index calculation tool interface of the quantification index; Standard identification information and chapter content structure of design codes and guidelines related to underground building design are extracted, and the chapter content structure is segmented using a preset multi-level regular template to generate regular structured text entries; According to the different facility types of the review objects, the standard identification information and regular structured text entries corresponding to each design code and guideline related to underground building design are used to construct different categories of underground space review rule text corpora according to the facility types.

[0010] Furthermore, the steps for constructing the instruction alignment corpus include: Obtain the target text involving the review process and decision-making logic from the pre-acquired historical design scheme review case data, and perform structured expression on the target text to establish a structured review logic description; Obtain the review record text from the historical design scheme review case data, and extract the implicit rule description involving the priority of review element determination from the review record text; Extract the case identification information from the historical design scheme review case data, combine the case identification information, review logic description, and implicit rule description to generate empirical structured text entries, and form all the obtained empirical structured text entries into an instruction alignment corpus.

[0011] Furthermore, the training process of the underground space domain drawing recognition enhanced multi-modal large model includes: Using a preset text-image alignment corpus as training data, fine-tuning and training the general multi-modal large model using the dynamic low-rank adaptive fine-tuning method to obtain the underground space domain drawing recognition enhanced multi-modal large model.

[0012] Furthermore, the general multi-modal large model includes a visual encoder, a feature mapping matrix, and a language model; Using the dynamic low-rank adaptive fine-tuning method to fine-tune and train the general multi-modal large model includes: In the text-image feature alignment stage, freeze the visual encoder of the multi-modal large model, set the rank parameter r = 16, and use the image description generation task dataset in the text-image alignment corpus to only fine-tune and train the feature mapping matrix to achieve text-image semantic alignment; In the joint optimization stage, unfreeze the visual encoder, set the rank parameter r = 32, and use the drawing question-answering task dataset in the text-image alignment corpus to synchronously fine-tune and train the feature matrix and language model parameters to achieve fine-grained feature matching and text-image reasoning; Merge the parameters trained in the text-image feature alignment stage with the parameters trained in the joint optimization stage to obtain the underground space domain drawing recognition enhanced multi-modal large model.

[0013] Furthermore, the steps for constructing the text-image alignment corpus include: Use a multimodal large model to process the design drawing data in the pre-obtained historical renovation design plan of underground buildings to generate text description information corresponding to the drawing data, and establish a semantic correspondence relationship between the design drawing data and the corresponding text description information to generate a corpus of aligned text and images; wherein, the text description information includes design descriptions, technical and economic index information, objective numerical index information, key structural targets and their coordinate frame information, and descriptions of design rationality.

[0014] Further, after the review report is output, the method further includes: Determine whether the review report needs to be manually reviewed according to a preset confidence score model; Collect the correctness feedback of the review report from the reviewers for the review reports that need to be manually reviewed, generate a reinforcement sample set according to the review reports and the corresponding correctness feedback, and feedback the reinforcement sample set to the reinforcement learning layer to update the model parameters through incremental training.

[0015] On the other hand, the present invention also provides an intelligent review system for knowledge-driven underground building design plans, the system includes: A problem classification unit for identifying the target facility type and the elements to be reviewed of the object to be reviewed in the underground building design plan to be reviewed, and determining the type of review materials corresponding to the elements to be reviewed; A knowledge base matching unit for matching target rule structured text entries from a pre-set underground space review rule knowledge base of the corresponding facility type according to the target facility type, and matching target experience structured text entries from a pre-set case experience knowledge base of the corresponding facility type; A review material loading unit for loading corresponding target review materials from the application materials of the current review task according to the type of review materials corresponding to the elements to be reviewed; A text parsing unit for, when the target review materials include text review materials, calling a pre-trained underground space review inference large model to extract the first index information corresponding to the elements to be reviewed from the text review materials; A drawing parsing unit for, when the target review materials include drawing review materials, calling a pre-trained underground space domain drawing recognition enhanced multimodal large model to extract the second index information corresponding to the elements to be reviewed from the drawing review materials; A report generation unit for integrating all the extracted index information, recalled target rule structured text entries, and target experience structured text entries to construct a prompt word template, inputting the obtained prompt word template into the underground space review inference large model to review item by item whether the index information conforms to each target rule structured text entry, and reasoning the index review results in combination with the review logic description and implicit rule description in the target experience structured text entries to output a review report.

[0016] On the other hand, 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.

[0017] On the other hand, the present invention also provides a computer program product, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the knowledge-driven intelligent review method for underground building design schemes as described above are implemented.

[0018] The knowledge-driven intelligent review method, system and device provided by the embodiments of the present invention realize the semantic alignment of multi-modal data such as design specification guidelines for underground space facilities, facility update design schemes and empirical knowledge during the operation and maintenance period, and can integrate the review decisions and implicit logics in the code of practice, historical cases and empirical knowledge to form a knowledge system, realizing the automatic recognition and reasoning of underground building design schemes driven by knowledge, and improving the intelligent level and review efficiency of the review process. Moreover, the present invention can automatically implement the intelligent review of a large number of drawing application materials in the design scheme, realize the automatic analysis of the spatial elements of the design drawings, and improve the cross-modal semantic association analysis ability between the drawing spatial information and the design specification text.

[0019] The above description is only an overview of the technical solution of the present invention. In order to be able to understand the technical means of the present invention more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention are specifically given below. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] By reading the detailed description of the preferred embodiments below, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of illustrating the preferred embodiments and are not considered to be a limitation of the present invention. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings: Figure 1 is a flowchart of the knowledge-driven intelligent review method for underground building design schemes provided by the embodiments of the present invention; Figure 2 is a flowchart of the knowledge-driven intelligent review method for underground building design schemes provided by another embodiment of the present invention; Figure 3 is a structural block diagram of the knowledge-driven intelligent review system for underground building design schemes proposed by the embodiments of the present invention; Figure 4This is a schematic flowchart of the implementation steps of the knowledge-driven intelligent review system for underground building design schemes proposed in the embodiments of the present invention. Specific embodiments

[0021] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the 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. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0022] Those skilled in the art of the present technology can understand that, unless specifically stated otherwise, the singular forms "a", "an", "the" and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present invention means the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups.

[0023] Those skilled in the art of the present technology can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the art to which the present invention belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted in an idealized or overly formal sense unless specifically defined.

[0024] Figure 1 Schematically shows a flowchart of the knowledge-driven intelligent review method for underground building design schemes according to an embodiment of the present invention. Referring to Figure 1 , the knowledge-driven intelligent review method for underground building design schemes according to the embodiments of the present invention specifically includes the following steps: S11. Identify the target facility type and the elements to be reviewed of the review object in the underground building design scheme to be reviewed, and determine the review material type corresponding to the elements to be reviewed. Among them, the facility type is the first-level classification of the review object, such as types like underground shopping malls and civil air defense basements, and the elements to be reviewed are the review objectives included in the review object of this type, that is, the divided review dimensions, such as plane layout, usage function, flow line organization, etc.

[0025] S12. Match the target rule structured text entries from the preset knowledge base of underground space review rules for the corresponding facility type, and match the target experience structured text entries from the preset knowledge base of case experiences for the corresponding facility type.

[0026] In this embodiment, the construction method of the underground space review rule knowledge base is as follows: The vector embedding model is used to convert the rule structured text entries in the pre-constructed underground space review rule text corpus of different facility types into vector representations, generating the underground space review rule vector knowledge bases of the corresponding facility types. The underground space review rule text corpus records the rule structured text entries. The construction method of the case experience knowledge base is as follows: The vector embedding model is used to convert the experience structured text entries corresponding to the preset historical design plan review case data into vector representations, and different case experience vector knowledge bases of different facility categories are constructed according to the different facility types of the review objects in the historical design plan review case data. The experience structured text entries include texts involving review logic descriptions and implicit rule descriptions.

[0027] In this embodiment, according to the target facility type of the review object, relevant specification and regulation information and historical review case information are matched from the preset knowledge base. For the input problem (including review questions, review indicators, review processes, etc.), the vector embedding model is used to generate the problem vector embedding for the input problem Based on the problem vector embedding text entry queries are performed on the vector knowledge bases (underground space review rule knowledge base, case experience knowledge base) corresponding to the same facility type, and the vector embeddings of the text entries stored in the knowledge base are compared: , Among them, and are task-specific encoders.

[0028] Cosine similarity calculation is used for semantic matching, and the specific formula is: , Optionally, a re-ranking model can be set to re-rank the results according to relevance. The score threshold is set to 0.5, which is used to set the similarity threshold for text entry screening. Only the text entries with similarity scores greater than 0.5 are recalled, and the parameter is set to retain the top 10 target rule structured text entries with similarity scores as the recall results to ensure that all relevant specifications and regulations for the current review task are recalled; the top 3 target experience structured text entries with similarity scores are retained as references.

[0029] S13. Load the corresponding target review materials from the application materials of the current review task according to the type of review materials 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 reported application materials based on the type of review materials corresponding to the elements to be reviewed.

[0030] S14. If the target review materials include text-based review materials, call the pre-trained large underground space review inference model to extract the first index information corresponding to the elements to be reviewed from the text-based review materials. In this embodiment, the large underground space review inference model is an intelligent large model fine-tuned based on a preset text corpus of underground space review rules and an instruction alignment corpus. The text corpus of underground space review rules records rule-structured text entries, and the instruction alignment corpus records empirical structured text entries involving review logic descriptions and implicit rule descriptions.

[0031] In this embodiment, if the target review materials involve text-based review materials, call the large underground space domain review inference model. According to the problem classification result, that is, the target facility type, call the corresponding prompt word template to parse the text-based review materials such as the design text in the approval materials and extract the corresponding index information. Example of the prompt word template: "This is the design plan text of {project type}, extract the {review index} information from it and output it in JSON format." S15. If the target review materials include drawing-based review materials, call the pre-trained multi-modal large model for enhanced drawing recognition in the underground space domain to extract the second index information corresponding to the elements to be reviewed from the drawing-based review materials. In this embodiment, the multi-modal large model for enhanced drawing recognition in the underground space domain is a multi-modal large model fine-tuned based on a preset graphic-text alignment corpus. The graphic-text alignment corpus records graphic-text pairs composed of design drawing data and its corresponding text description information.

[0032] In this embodiment, if the target review materials involve drawing-based review materials, call the multi-modal large model for enhanced drawing recognition in the underground space domain. According to the problem classification result, that is, the target facility type, call the built-in prompt word template corresponding to the index to be reviewed to identify the corresponding index information in the design drawing. Example of the prompt word template: "This is a {drawing type} design drawing of {project type}, extract the {review index} information from it. Only identify the information clearly marked in the drawing, and output unknown for the unmarked information without self-inference. The result is output in JSON format." S16. Integrate all the extracted index information, the target rule structured text entries recalled, and the target experience structured text entries to construct a prompt template, and input the obtained prompt template into the underground space review inference large model to review item by item whether the index information conforms to each target rule structured text entry, and perform inference on the index review results in combination with the review logic description and implicit rule description in the target experience structured text entry to output a review report.

[0033] The knowledge-driven intelligent review method for underground building design solutions provided by the embodiments of the present invention uses a cross-modal attention mechanism to achieve semantic alignment of multi-modal data such as underground space facility design specification guidelines, facility update design solutions during the operation and maintenance period, and experience knowledge, and can integrate review decisions and implicit logics in code provisions, historical cases, and experience knowledge to form a knowledge system, realizing automatic identification and reasoning of underground building design solutions driven by knowledge, and improving the intelligent 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 solution, realize automatic parsing of spatial elements of design drawings, and improve the cross-modal semantic association analysis ability between drawing spatial information and design specification texts.

[0034] In another optional embodiment of this aspect, as Figure 2 shown, if the elements to be reviewed include relevant indicators that require spatial calculation in addition to text-based review materials or drawing-based review materials, the method further includes step S17; S17. If the elements to be reviewed include relevant indicators that require spatial calculation, call the index calculation tool corresponding to the corresponding quantitative index according to the elements to be reviewed, so as to calculate and return the corresponding third index information by using the index calculation tool.

[0035] In the embodiments of the present invention, a corresponding interaction interface is developed for the index calculation tool based on a lightweight network service architecture. After the interface test is correct, the standard network request instruction of the interface is exported, and the network request instruction is converted into an open interface description specification by using a large language model, which contains the interface definition and function description of the tool, forming a standardized function module that can be called by the intelligent model.

[0036] If the target review material involves calculation-related indicators, call the corresponding calculation tool through a function to return the corresponding index information.

[0037] In this embodiment, the underground space domain review inference large model is called, and a prompt template is designed. All the extracted index information and recalled relevant specifications and cases are integrated as context inputs, and each piece of index information is compared item by item to check whether it conforms to the relevant specification regulations. The review results are inferred and summarized in combination with relevant cases. A review report in the front-end output format is output, and the three most relevant historical cases are pushed for the user's reference, and the review report in the output format is output. Example of the prompt template: "Generate a review report for the following design plan based on the relevant information of the project: {Constraint specification}+{Text parsing result}+{Drawing recognition result}+{Spatial calculation result} {Example of the review report format}" In one embodiment of the present invention, after the review report is output, the method further includes: determining whether the review report needs to be manually reviewed according to a preset confidence scoring model; collecting the correctness feedback of the review report from the review personnel for the review report that needs to be manually reviewed, generating a reinforcement sample set according to the review report and the corresponding correctness 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: Determine whether the review result needs to be manually reviewed according to a preset confidence scoring model, and collect manual feedback data to the reinforcement learning layer for the situation that needs to be manually reviewed to realize system incremental training and update.

[0038] The confidence C is dynamically calculated through the confidence scoring model. The specific formula of the confidence scoring model is: C = λ_r×R_rule + λ_c×C_rule + λ_s×S_case Where R_rule is the standard relevance, and the specific calculation method is: the average matching rate of the structured text entries of the recalled target rules in step S12.

[0039] C_rule is the correct rate of parameter compliance inference, and the specific calculation method is: each structured text entry of the target rule is extracted for secondary verification, and the logical verification passing rate is calculated.

[0040] S_case is the historical decision similarity, and the specific calculation method is: the average matching rate of the three recalled structured text entries of target experiences in step S12.

[0041] λ_r, λ_c, and λ_s are the weights of the three respectively. Optionally, they can be set to 0.2, 0.6, and 0.2.

[0042] When the confidence C≥90%, it is considered passed. When 80%≤C<90%, secondary inference verification is started. When C<80%, manual review is increased.

[0043] An artificial feedback button is set up, and the review personnel provide feedback on the correctness of the review results and record them in the background. A strengthened sample set is generated regularly and fed back to the reinforcement learning layer to update the model parameters through incremental training.

[0044] In a specific example, the output result example is as follows: 1. Review report: Review conclusion: To be corrected - Violation item: The width of the sidewalk is 3.0m (the specification threshold is 4.0 - 6.5m) - Clause basis: Article 4.4.5 of TCECS481 - 2017 - Amendment suggestion: Broaden the width of the sidewalk - Confidence level: 88% (automatically reviewed and passed) 2. Related historical cases: - Case 1: In 2023, the width of the pedestrian passage in a certain shopping mall was insufficient, and the width of the passage was adjusted to 4.5m. (Similarity 0.85) - Case 2: In 2021, the height difference between the first basement floor and the outdoor entrance floor of a certain underground shopping mall exceeded the standard, and the height difference was adjusted to 8m. (Similarity 0.78) - Case 3: In 2022, the evacuation passage of a certain civil air defense project was insufficient (similarity 0.72), and the width of the passage was adjusted to 2.4m.

[0045] Is the reported result accurate: Correct / Incorrect (selected by the artificial feedback button).

[0046] In an embodiment of the present invention, the pre - trained large - model for underground space domain review and reasoning can be called to pre - classify the input review task, that is, the underground building design plan to be reviewed, to judge and identify the facility type of the review object and the review elements required in the current review task, and the large - model outputs the secondary classification result.

[0047] At the same time, the large - model generates dynamic process configuration parameters (in JSON format), including the review material types (text / drawing / calculation) corresponding to the review elements to be reviewed, the selection of the specification knowledge base, the module call sequence, etc.

[0048] The system internally sets up prompt - word templates designed according to the review material basis, key indicators (i.e., review elements to be reviewed), and relevant specification catalogs required for various underground space review objects such as underground shopping malls and civil air defense basements. Then, according to the classification result, the review process is automatically organized, including selecting which unit modules need to be called, whether to call the calculation module, and selecting the corresponding prompt words.

[0049] For example, when the task of "review of the layout plan of an underground shopping mall" is detected, the material loading unit is automatically triggered to load the general layout plan and design description, and the knowledge base matching unit is synchronously activated to retrieve the "knowledge base related to underground shopping malls"; the text parsing unit (prompt word template parameterized injection of "shopping mall / layout plan") and the drawing parsing unit (calling a multi-modal large model and using the configured prompt words to identify key indicators such as the total building area and the width of pedestrian passages) are called in parallel; if the targets to be reviewed in the classification results include calculation requirements such as "proportion of business hall area", the area proportion calculation tool is called through the function call unit; the above call results are automatically incorporated into the context input of the reasoning unit of the report generation unit.

[0050] In order to implement the training of the underground space review inference large model and the underground space field drawing recognition enhanced multi-modal large model, the present invention further includes the step of constructing each corpus.

[0051] First, data collection is carried out. The collection of design specification guidelines is the basis of the entire system. For example, "Regulations on the Development and Utilization of Urban Underground Space", "Design Guidelines for Urban Underground Commercial Spaces", etc. The collection of facility update design plans 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 to be collected, such as underground shopping malls, civil air defense basements, etc.; collect the historical underground building update design plans of these facilities, including project design plan texts, floor plans, elevation views, renderings, etc. The collection of historical case experience knowledge is the core of the entire system. Collect the review results of past design plans, including review opinions, modification suggestions, etc.

[0052] Specifically, the steps for constructing the underground space review rule text corpus include: 1. Extract the key review elements of the underground building design plan and the quantitative indicators of each key review element from the pre-obtained design specification guidelines related to underground building design.

[0053] 2. 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, so as to analyze which materials need to identify which corresponding approval indicators, and establish a system approval process. In order to ensure the comprehensiveness and accuracy of the review process, it is necessary to disassemble the review points in the design specification guidelines, which mainly include the following steps: Step 1, identify the key terms in the design specification guidelines, namely the key review elements, including building scale, use function, streamline organization, plane layout, etc. Step 2, determine the quantitative indicators of these key review elements, such as the total length of the underground commercial street should be less than 1500m, and the height difference between the ground of the first underground floor and the outdoor entrance and exit floor should not be greater than 10m. Step 3, clarify the type of review materials corresponding to each quantitative indicator (such as general plan, elevation, design plan text, etc.), and establish a mapping relationship between the indicator and the review basis. Step 4, for indicators that require spatial calculations such as area proportion, distance between nodes, and dynamic line vector line type, develop corresponding calculation tools based on Python language. For example, extract closed polygons from a CAD floor plan, calculate the area, and then calculate ratio statistics.

[0054] 3. 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. Among them, the specification identification information includes the name title, standard number, release date, implementation date, province, city and district of the design specification guidelines, etc.

[0055] 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 types of facilities.

[0056] In this embodiment, for standard guideline data, the data that can be collected on the Internet are mostly stored in the form of PDF, and the general multimodal large model is used for text extraction. By designing a prompt word template, key information such as the standard title, standard number, and time and space labels such as the release / implementation date and the province, city, and district to which the local standard belongs are extracted. Input prompt word: "Parse the current page content and extract the following fields: Title: the first complete paragraph, including quotation marks; Standard number: in accordance with the format of XX / T XXXXX-XXXX; Implementation date: 'effective from X / X / XXXX'; Location: the keyword 'XX province / city' in the local standard; Text: exclude headers / footers / watermarks and retain chapter structure. Output is JSON, and chapters are divided into three levels: 'chapter-section-article'. {Document image per page}" Multimodal large model output: “{ "Title": "Design Guidelines for Urban Underground Commercial Spaces", "Standard No.": "T / CECS 481-2017", "Implementation Date": "Implemented as of December 1, 2017", "Jurisdiction": "China", "Text": { "Chapter": "1 General Provisions", "Section": { "Article": "1.0.1", "Content": "To meet the needs of large-scale development and construction of urban underground commercial spaces in China......"}, { "Article": "1.0.2", "Content": "......"}, ]} } Use a multi-level regular template to cut the text of the main body, generate rule-structured text entries, and store them as {Title}-{Standard No.}-{Chapter No.}-{Article No.}: {Content}.

[0057] All the above rule-structured text entries form the text corpus of the underground space review rules.

[0058] Meanwhile, classify the above-mentioned rule-structured text entries that have been segmented according to the facility types of the review objects to construct the text corpus of the underground space review rules, such as the specification corpus for underground shopping mall design and the specification corpus for underground utility tunnel design; use the vector embedding model to convert the above-mentioned segmented structured entries into vector representations, construct an index, and generate a multi-class underground space review rule vector knowledge base that supports semantic queries, namely the underground space review rule knowledge base.

[0059] Specifically, the construction steps of the graphic-text alignment corpus include: using a multi-modal large model to process the design drawing data in the pre-acquired historical updated design schemes of underground buildings 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 the graphic-text alignment corpus; among them, the text description information includes design descriptions, technical and economic index information, objective numerical index information, key structural targets and their coordinate frame information, and descriptions of design rationality.

[0060] In this embodiment, for the plan drawings, elevation drawings, etc. in the design scheme, with the help of a general multi-modal large model, extract the information in the drawings through design prompts, and cooperate with manual correction to modify the information misrecognized by the large model and supplement the information not recognized by the large model to ensure the comprehensiveness and accuracy of the description of the drawing information. The specific content includes: (1)Design description: Provide a detailed description of the drawings, including design intent, layout arrangement, etc.

[0061] (2)Technical and economic indicator information: Extract text data such as the main technical and economic indicators from the general layout plan to ensure the accuracy and integrity of the data.

[0062] (3)Objective numerical indicator information: Extract and interpret the numerical values of objective indicators such as markings.

[0063] (4)Key structure targets and their coordinate frame information: Identify key structure targets such as entrances and exits and extract their coordinate frames.

[0064] (5)Description of design rationality: Describe the rationality of subjective indicators such as streamline organization.

[0065] Construct a structured multi-modal prompt template library, correspond the design drawings with the text descriptions, perform pixel-level alignment, and generate a corpus of aligned graphics and text. The corpus of aligned graphics and text specifically includes two types of data. One is the dataset of image description generation tasks obtained from the image description generation tasks, including the graphics and text pairs of the drawings and text descriptions. The other is the dataset of drawing Q&A tasks obtained from the Q&A tasks based on the drawings. For example: “Q: Extract the main technical and economic indicators from the drawing: {Review the drawing} A: 1. Total land area Unit: ㎡ Data: 15794; 2. Building floor area Unit: ㎡ Data: 6485; ......” Specifically, the steps for constructing the instruction alignment corpus include: 1. Obtain the target text related to the review process and decision-making logic from the pre-acquired historical design scheme review case data, and perform a structured representation of the target text to establish a structured review logic description.

[0066] 2. Obtain the review record text from the historical design scheme review case data, and extract the implicit rule description related to the priority of review element determination from the review record text.

[0067] 3. Extract the case identification information from the historical design scheme 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.

[0068] In this embodiment, for the empirical knowledge of historical design scheme review cases, through natural language processing technology, a structured review logic chain is established to systematize and standardize the review process and decision-making logic. For example, parsing "the evacuation width is insufficient and needs to be expanded to ≥2.4m first" into {indicator: evacuation width, current value: less than 2.4m, code requirement: ≥2.4m (with the original code attached), solution: expand to 2.4m}. Extract implicit rules from the review records, such as "give priority to ensuring that the evacuation width > commercial area", etc. Combine the cases into a structured text {case name: xx, time: xx, location: xx, specific content: xxx} to construct a case experience vector knowledge base.

[0069] Specifically, design a structured prompt template, construct instruction alignment corpus. On the input side, integrate the project basic data (information that can be extracted from the review materials, consistent with the annotations in the graphic-text alignment corpus), and the applicable specification standards in the rule corpus. On the output side, generate a structured review report according to the thinking chain of "conflict detection → priority determination → solution suggestion", for example: "Q: Generate a review report based on the following specific information of a {project type} project: {project basic data}+{constraint specification} A: <think> 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}); 2. Priority determination: According to the rule "{Implicit Rule}", it is recommended to adjust {Indicator B} first; 3. Solution suggestions: 1) Adjust {Indicator B} to {Specification B Constraint Value}; 2) Adjust {Indicator A} to {Specification A Constraint Value} < / think> <answer> {Conclusion of Examination} < / answer> " Based on the above-obtained underground space review rule text corpus, instruction alignment corpus, and graphic-text alignment corpus, fine-tuning training of each large model is realized below.

[0070] In the embodiment of the present invention, the training process of the underground space review inference large model includes: 1. Using the preset text corpus of underground space review rules as training data, the pre-trained base large model is fine-tuned by means of dynamic low-rank adaptive fine-tuning. During the fine-tuning process, the weight parameters of the original base large model are frozen, and trainable rank decomposition matrices are injected into each layer. Different rank parameter values are assigned to the training data of different specification types according to the different specification types to which the rule-structured text entries belong, so as to dynamically adjust the rank parameter values according to the complexity of the task. Specifically, a pre-trained base large 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 model official documentation. According to the incremental pre-training part of the code in the base large model instruction document, the rule corpus is loaded as training data, and dynamic low-rank adaptive fine-tuning is adopted. The base large model is fine-tuned by adjusting training hyperparameters such as learning rate, number of training epochs, and block size. There are mainly two methods for large model fine-tuning, one is full-parameter fine-tuning and the other is partial-parameter fine-tuning. The present invention adopts the low-rank fine-tuning method for partial-parameter fine-tuning. By freezing the weights of the original pre-trained model on the basis of the pre-trained large model and injecting trainable rank decomposition matrices into each layer, the number of parameters to be trained is greatly reduced, so as to fine-tune the pre-trained model. Further, the present invention designs an improved dynamic low-rank fine-tuning method to sort the learned 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, low-rank matrices are used to approximate the weight matrix, and rank parameters (r = 8 - 64) can be assigned to different specification types according to different task complexities. For example, r = 64 is set for structural safety, r = 32 for plane layout, and r = 16 for streamline design, so as to dynamically adjust the number of parameters according to the complexity of the task and improve the training efficiency.

[0071] 2. The fine-tuned weights and the base model are merged through a parameter fusion algorithm to generate a large model enhanced with underground space domain knowledge. The model injects underground space domain knowledge on the basis of the general large language model. The parameter fusion algorithm formula is: ,

[0072] where, Denote the frozen base large language model weight parameters as \(W\), and \(\Delta W\) as the low-rank matrix parameters for fine-tuning training. Only the additional \(\Delta W\) parameters are trained. \(B\) and \(A\) represent the weight matrices to be trained. The dimension of matrix \(B\) is \(d\times r\), corresponding to the combination of the input dimension \(d\) (such as 4096) of the model framework and the rank parameter \(r\), which is used to extract low-rank features related to specific specification types from the input training data. The dimension of matrix \(A\) is \(r\times k\), 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 the parameter adjustment amount adapted to the domain knowledge. \(\alpha\) is a preset domain adaptation coefficient (empirical value 0.5), which realizes controllable knowledge injection intensity and prevents \(\Delta W\) from overwriting the pre-trained general domain knowledge.

[0073] 3. Use the preset instruction-aligned corpus as the training data, and adopt the low-rank fine-tuning method to fine-tune the large model enhanced with underground space domain knowledge. During the fine-tuning training process, freeze the weight parameters of the large model enhanced with underground space domain knowledge, inject a rank decomposition matrix with a specified rank value in the Transformer layer, and optimize the training through supervised learning to obtain the underground space review and inference large model. Specifically, by loading the instruction-aligned corpus, adopting fixed-value low-rank fine-tuning, freezing the base model parameters, only injecting a low-rank matrix with a rank of 32 in the Transformer layer, and optimizing the instruction-response matching through supervised learning. Fine-tune the large model enhanced with underground space domain knowledge to obtain the instruction-aligned model, that is, the initial underground space review and inference large model.

[0074] Further, after obtaining the initial underground space review and inference large model, the method further includes: 4. Label the multiple output results of the underground space review and inference large model under the input of the same prompt instruction according to the preset evaluation criteria. For example, label the superiority and inferiority order according to the three-level standard of "specification citation accuracy > calculation logic integrity > expression clarity" to obtain the preference ranking result; 5. Establish a multi-objective reward model, and use the preference ranking result to perform backpropagation training according to the multi-objective reward model. The multi-objective reward model is: ; , where, is a given prompt, \(r\) is for the \(N\) responses generated by the instruction-aligned model ; represents the trainable reward model, represents the rule reward, which matches the rule-structured text entries in the response \(r\) through regular expressions, that is, matches the corresponding rule-structured text entry content from the underground space review rule text corpus through regular expressions. For the rule-based weight, it is set to 0.5 - 0.7. Each valid clause match is recorded points; represents the human preference reward, and the scoring result is obtained through human preference ranking, is the experience correction coefficient, set to 0.3 - 0.5; denotes the sigmoid function, and according to the scoring result, it responds superior to , and is trained by minimizing the objective function of the multi-objective reward model .

[0075] 6. Use proximal policy optimization to fine-tune the large underground space review and reasoning model. Take the output of the reward model as the reinforcement signal, and iteratively update the low-rank matrix parameter ΔW of the large underground space review and reasoning model to maximize the score of the reward model, and obtain the optimized large underground space review and reasoning model.

[0076] The present invention uses a low-rank fine-tuning method to perform instruction alignment fine-tuning training on the large underground space domain knowledge enhancement model fine-tuned 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.

[0077] In the embodiment of the present invention, the training process of the large underground space domain drawing recognition enhanced multi-modal model includes: using a preset text-image alignment corpus as training data, and performing fine-tuning training on the general multi-modal model in a dynamic low-rank adaptive fine-tuning manner to obtain the large underground space domain drawing recognition enhanced multi-modal model, which can extract information from the design drawings in the underground space domain according to the prompt words.

[0078] In this embodiment, first, a pre-trained base model is selected. In this embodiment, the multi-modal model Qwen2.5-VL is selected, and the pre-trained large language model parameters are loaded according to the model official documentation. The general multi-modal 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 feature vectors; 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 outputs. Further, by running the fine-tuning training part of the code in the base multi-modal model file, the underground space domain text-image alignment corpus is loaded for fine-tuning. The general multi-modal model is fine-tuned and trained in a dynamic low-rank adaptive fine-tuning manner, specifically including: In the text-image feature alignment stage, freeze the visual encoder of the multi-modal model, set the rank parameter r = 16, and only fine-tune and train the feature mapping matrix using the image description generation task dataset in the text-image alignment corpus to achieve text-image semantic alignment; In the joint optimization stage, thaw the visual encoder, set the rank parameter r = 32, and use the drawing Q&A task dataset in the graph-text alignment corpus to synchronously fine-tune and train the feature matrix and language model parameters to achieve fine-grained feature matching and graph-text reasoning; According to the code for parameter merging in the base large model specification document, merge the parameters trained in the graph-text feature alignment stage with the parameters trained in the joint optimization stage to obtain an enhanced multi-modal large model for underground space domain drawing recognition.

[0079] For the method embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the embodiments of the present invention are not limited by the described action sequence, because according to the embodiments of the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0080] Another embodiment of the present invention also provides a knowledge-driven intelligent review system for underground building design schemes, and the system includes functional modules for implementing the knowledge-driven intelligent review method for underground building design schemes as described in any one of the above. Figure 3 Schematically shows a structural block diagram of a knowledge-driven intelligent review system for underground building design schemes according to another embodiment of the present invention. Refer to Figure 3 , the knowledge-driven intelligent review system of this embodiment specifically includes a problem 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 generation unit 306, where: The problem 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 scheme to be reviewed, and determine the type of review materials corresponding to the elements to be reviewed.

[0081] In this embodiment, the problem classification unit 301 can perform pre-classification on the input review task, that is, the underground building design scheme to be reviewed, by calling a pre-trained review inference large model for the underground space domain, judge and identify the facility type of the review object and the elements that need to be reviewed in the current review task, and output a secondary classification result by the large model.

[0082] Meanwhile, the problem classification unit 301 can also generate dynamic process configuration parameters (in JSON format) through a large model, including the types of review materials corresponding to the elements to be reviewed (text / drawings / calculation), the selection of the specification knowledge base, the module call sequence, etc. Specifically, the system has built-in prompt templates designed based on the review material basis, key indicators (i.e., elements to be reviewed), and relevant specification catalogs required for various types of underground space review objects such as underground shopping malls and civil air defense basements. Then, according to the classification results, the review process is automatically organized, including selecting which unit modules need to be called, whether to call the calculation module, and selecting the corresponding prompt words.

[0083] 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 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.

[0084] The review material loading unit 303 is used to load the corresponding target review materials from the application materials of the current review task according to the type of review materials corresponding to the elements to be reviewed.

[0085] The text parsing unit 304 is used to, when the target review materials include text review materials, call the pre-trained underground space review inference large model to extract the first index information corresponding to the elements to be reviewed from the text review materials.

[0086] The drawing parsing unit 305 is used to, when the target review materials include drawing review materials, call the pre-trained drawing recognition enhanced multi-modal large model in the underground space field to extract the second index information corresponding to the elements to be reviewed from the drawing review materials.

[0087] The report generation unit 306 is used to integrate all the extracted index information, the recalled target rule structured text entries, and the target experience structured text entries to construct a prompt template, and input the obtained prompt template into the underground space review inference large model to review whether each index information meets each target rule structured text entry item by item, and infer and output a review report on the index review results in combination with the review logic description and implicit rule description in the target experience structured text entries.

[0088] In the embodiment of the present invention, the system further includes a function call unit, which is used to, when the elements to be reviewed include relevant indicators that require spatial calculation, call the index calculation tool corresponding to the corresponding quantitative indicators according to the elements to be reviewed, so as to calculate and return the corresponding third index information by using the index calculation tool.

[0089] In an embodiment of the present invention, the system further includes a feedback verification unit, which is used to determine whether the review report needs manual review according to a preset confidence scoring model after the review report is output; collect the correctness feedback of the review report from the review personnel for the review report that needs manual review, generate a reinforcement sample set according to the review report and the corresponding correctness feedback, and feedback the reinforcement sample set to the reinforcement learning layer to update the model parameters through incremental training.

[0090] Figure 4 Schematically shows 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. As Figure 4 shown, for the underground building design solution to be reviewed, the target facility type of the review object, the review elements to be reviewed, and their corresponding review material types in the underground building design solution to be reviewed are identified through the problem classification unit. The rule-structured text entries and experience-structured text entries are matched from the knowledge base of the corresponding facility type according to the target facility type through the knowledge base matching unit. The corresponding review materials are loaded from the application materials according to the review material types through the review material loading unit. For the review materials that do not involve the calculation part, if they involve text-based review materials, the index information corresponding to the review elements is extracted by calling the review inference large model through the text parsing unit. If they involve drawing-based review materials, the index information corresponding to the review elements is extracted by calling the drawing recognition enhanced multi-modal large model through the drawing parsing unit. When the review elements include relevant indicators that require spatial calculation, the corresponding index calculation tool for the corresponding quantitative indicators is called according to the review elements through the function call unit to calculate and return the corresponding third index information using the index calculation tool. Finally, the report generation unit integrates the extracted index information and the recalled knowledge entries to construct a prompt word template, and the review inference large model performs intelligent review and reasoning on the index information according to the prompt word template and outputs a review report. After the review report is output, it is determined whether the review report needs manual review according to a preset confidence scoring model through the feedback verification unit; the correctness feedback of the review report from the review personnel is collected for the review report that needs manual review, a reinforcement sample set is generated according to the review report and the corresponding correctness feedback, and the reinforcement sample set is feedback to the reinforcement learning layer to update the model parameters through incremental training. When the confidence C dynamically calculated by the confidence scoring model is ≥ 90%, it is considered to pass the verification and no manual review is performed. When 80% ≤ C < 90%, secondary reasoning verification is started. When C < 80%, manual review is increased.

[0091] 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.

[0092] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0093] In addition, another embodiment of the present invention further 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, it implements the steps of the knowledge-driven intelligent review method for underground building design solutions as described above.

[0094] In addition, another embodiment of the present invention further provides a computer program product, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the knowledge-driven intelligent review method for underground building design solutions as described above.

[0095] Those skilled in the art can understand that although some of the embodiments herein include certain features included in other embodiments rather than other features, the combination of the features of different embodiments means that it is within the scope of the present invention and forms different embodiments. For example, any one of the claimed embodiments can be used in any combination.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or equivalently replace some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate 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 includes: Identifying the target facility type and the elements to be reviewed of the review object in the underground building design plan to be reviewed, and determining the review material type corresponding to the elements to be reviewed; Matching the target rule structured text entry from the underground space review rule knowledge base of the preset corresponding facility type according to the target facility type, and matching the target experience structured text entry from the case experience knowledge base of the preset corresponding facility type; Loading the corresponding target review material from the application materials of the current review task according to the review material type corresponding to the elements to be reviewed; If the target review material includes text review materials, calling the pre-trained underground space review inference large model to extract the first index information corresponding to the elements to be reviewed from the text review materials; If the target review material includes drawing review materials, calling the pre-trained drawing recognition enhanced multi-modal large model in the underground space field to extract the second index information corresponding to the elements to be reviewed from the drawing review materials; Integrating all the extracted index information, the recalled target rule structured text entry, and the target experience structured text entry to construct a prompt template, and inputting the obtained prompt template into the underground space review inference large model to review item by item whether the index information conforms to each target rule structured text entry, and reasoning the index review results in combination with the review logic description and implicit rule description in the target experience structured text entry to output a review report.

2. The method according to claim 1, wherein The method further includes: If the elements to be reviewed include relevant indicators that require spatial calculation, calling the index calculation tool corresponding to the corresponding quantitative index according to the elements to be reviewed to calculate and return the corresponding third index information by using the index calculation tool.

3. The method according to claim 1, wherein The training process of the underground space review inference large model includes: Using the preset underground space review rule text corpus as the training data, adopting the dynamic low-rank adaptive fine-tuning method to fine-tune the pre-trained base large model. During the fine-tuning training process, freeze the weight parameters of the original base large model, inject a trainable rank decomposition matrix into each layer, and assign different rank parameter values to the 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 values according to the complexity of the task; Combining the fine-tuned weights and the base model through a parameter fusion algorithm to generate a knowledge-enhanced large model in the underground space field; Using the preset instruction alignment corpus as the training data, adopting the low-rank fine-tuning method to fine-tune the knowledge-enhanced large model in the underground space field. During the fine-tuning training process, freeze the weight parameters of the knowledge-enhanced large model in the underground space field, inject a rank decomposition matrix with a specified rank into the Transformer layer, and optimize the training through supervised learning to obtain the underground space review inference large model.

4. The method according to claim 3, wherein After obtaining the underground space review inference large model, the method further includes: Marking the multiple output results of the underground space review inference large model under the input of the same prompt instruction in the order of superiority and inferiority according to the preset evaluation criteria to obtain a preference ranking result; Establishing a multi-objective reward model, and performing backpropagation training according to the multi-objective reward model by using the preference ranking result. The multi-objective reward model is: ； ， Among them, is a given hint, and r is for N responses generated by the instruction alignment model ; represents a trainable reward model, represents a rule-based reward, is a rule-based weight; represents a human preference reward, is an empirical correction coefficient; is the objective function of the multi-objective reward model, represents the sigmoid function, and the response is better than , and is trained by minimizing the objective function ; Use proximal policy optimization to fine-tune the large underground space review inference model, take the output of the reward model as the reinforcement signal, and iteratively update the low-rank matrix parameter ΔW of the large underground space review inference model to maximize the score of the reward model, thus obtaining the optimized large underground space review inference model.

5. The method according to claim 1, characterized in that, The construction steps of the text corpus of underground space review rules include: Extract the key review elements of the underground building design scheme and the quantification indicators of each key review element from the pre-acquired design specifications and guidelines related to underground building design; Establish the mapping relationship between each quantification indicator and the corresponding review material type, and / or establish the connection relationship between each quantification indicator and the corresponding index calculation tool interface; Extract the specification identification information and the chapter content structure of the design specifications and guidelines related to underground building design, and use a preset multi-level regular template to segment the chapter content structure to generate rule-structured text entries; According to the different facility types of the review objects, construct different categories of text corpora of underground space review rules based on the specification identification information and rule-structured text entries corresponding to each design specification and guideline related to underground building design according to the facility type.

6. The method according to claim 1, characterized in that The construction steps of the instruction alignment corpus include: Obtain the target text involving the review process and decision-making logic from the pre-acquired historical design scheme review case data, and perform a structured representation on the target text to establish a structured review logic description; Obtain the review record text from the historical design scheme review case data, and extract the implicit rule description involving the determination priority of review elements from the review record text; Extract the case identification information from the historical design scheme 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.

7. The method according to claim 1, characterized in that, The training process of the enhanced multi-modal large model for drawing recognition in the underground space field includes: Use the preset graph-text alignment corpus as the training data, and fine-tune and train the general multi-modal large model in a dynamic low-rank adaptive fine-tuning manner to obtain the enhanced multi-modal large model for drawing recognition in the underground space field.

8. The method according to claim 7, wherein The general multi-modal large model includes a visual encoder, a feature mapping matrix and a language model; Using the dynamic low-rank adaptive fine-tuning method to fine-tune and train the general multi-modal large model includes: In the graph-text feature alignment stage, freeze the visual encoder of the multi-modal large model, set the rank parameter r = 16, and use the image description generation task dataset in the graph-text alignment corpus to only fine-tune and train the feature mapping matrix to achieve graph-text semantic alignment; In the joint optimization stage, unfreeze the visual encoder, set the rank parameter r = 32, and use the drawing Q&A task dataset in the graph-text alignment corpus to synchronously fine-tune and train the feature matrix and language model parameters to achieve fine-grained feature matching and graph-text reasoning; Merge the parameters trained in the graph-text feature alignment stage with the parameters trained in the joint optimization stage to obtain the enhanced multi-modal large model for drawing recognition in the underground space field.

9. The method according to claim 7, wherein The construction steps of the graph-text alignment corpus include: Use a multi-modal large model to process the design drawing data in the pre-obtained historical renovation design plan of underground buildings to generate text description information corresponding to the drawing data, and establish a semantic correspondence relationship between the design drawing data and the corresponding text description information to generate a corpus of aligned graphics and text; wherein, the text description information includes design descriptions, technical and economic index information, objective numerical index information, key structural targets and their coordinate frame information, and descriptions of design rationality.

10. The method according to claim 1, characterized in that After the review report is output, the method further includes: Determine whether the review report needs to be manually reviewed according to a preset confidence score model; Collect the correctness feedback of the review personnel on the review report for the review report that needs to be manually reviewed, generate a reinforcement sample set according to the review report and the corresponding correctness feedback, and feedback the reinforcement sample set to the reinforcement learning layer to update the model parameters through incremental training.

11. A knowledge-driven intelligent review system for underground building design solutions, characterized in that, The system includes: A problem classification unit for identifying the target facility type and the elements to be reviewed of the object to be reviewed in the underground building design plan to be reviewed, and determining the type of review materials corresponding to the elements to be reviewed; A knowledge base matching unit for matching target rule structured text entries from a preset knowledge base of underground space review rules for the corresponding facility type according to the target facility type, and matching target experience structured text entries from a preset case experience knowledge base for the corresponding facility type; A review material loading unit for loading corresponding target review materials from the application materials of the current review task according to the type of review materials corresponding to the elements to be reviewed; A text parsing unit for, when the target review materials include text-based review materials, calling a pre-trained underground space review inference large model to extract the first index information corresponding to the elements to be reviewed from the text-based review materials; A drawing parsing unit for, when the target review materials include drawing-based review materials, calling a pre-trained drawing recognition enhanced multi-modal large model in the underground space field to extract the second index information corresponding to the elements to be reviewed from the drawing-based review materials; A report generation unit for integrating all the extracted index information, recalled target rule structured text entries, and target experience structured text entries to construct a prompt template, inputting the obtained prompt template into the underground space review inference large model to review item by item whether the index information conforms to each target rule structured text entry, and reasoning the index review results in combination with the review logic description and implicit rule description in the target experience structured text entries to output a review report.

12. A computer device, characterized in that, 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 method according to any one of claims 1-10 are implemented.

13. A computer program product, characterized in that, A computer program is stored on the computer program product, and when the computer program is executed by the processor, the steps of the method according to any one of claims 1-10 are implemented.

Images