Structural intelligent design and autonomous checking method and system based on search enhancement generated structure

Abstract

The application provides a structure intelligent design and autonomous checking method and system based on retrieval enhancement generation, and belongs to the cross field of artificial intelligence and structure design. The method comprises the following steps: constructing a structure design specification special knowledge base, establishing a standardized vector database by performing text preprocessing, information structuring and vector storage on specification documents; adopting a hybrid retrieval strategy combining vector retrieval and large language model reordering to recall relevant specification clauses, and generating a structured design scheme or checking conclusion based on the retrieval result; and performing multi-level performance evaluation on the generated design scheme. The application effectively solves the problems of low specification clause retrieval efficiency, professional term understanding deviation and insufficient result interpretability, and significantly improves the intelligent level and reliability of railway bridge and tunnel and other infrastructure structure design and checking.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and structural design, and in particular relates to a method and system for intelligent structural design and autonomous verification based on retrieval enhancement generation. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Railway bridges and tunnels are a crucial component of railway engineering, and their design methods are a core element in ensuring project safety and construction efficiency. In recent years, with the continuous expansion of railway construction scale and the increasing complexity of projects, relevant design standards and specifications have become increasingly complex. However, the structural design and verification of railway bridges and tunnels still heavily rely on manual labor. Engineers must consult numerous scattered specification texts and combine their own experience to perform design calculations and manual verification. This traditional method suffers from high subjectivity, low efficiency, and a heavy workload, and is prone to design defects and frequent rework due to human error. Furthermore, the continuous updating of specifications further increases the learning costs and workload for engineers.

[0004] To improve design efficiency, existing research has attempted to apply technologies such as deep learning, computer vision, generative design, and knowledge graphs to areas such as parameter optimization, intelligent span arrangement, seismic design, and health monitoring of bridges and tunnels. However, existing technologies face challenges such as scattered relevant code provisions, complex data types, and implicit design experience. They still focus on local parametric models, making it difficult to integrate multimodal data (such as BIM drawings and monitoring data), and there are still discrepancies in the understanding of professional terminology, resulting in poor data quality. Furthermore, their dynamic updates are inflexible, unable to adapt to the updating and iteration of relevant code provisions, and lack logical reasoning capabilities, making them difficult to apply to practical engineering projects. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a method and system for intelligent structural design and autonomous verification based on retrieval enhancement generation, aiming to solve the problems of low efficiency in structural design specification retrieval, difficulty in matching multi-source information, and insufficient automation and intelligence in the prior art.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for intelligent structural design and autonomous verification based on retrieval enhancement generation; A search-enhanced generation-based intelligent structural design and autonomous verification method includes: Step S1: Construct a dedicated knowledge base for structural design specifications. Through text preprocessing, text information structuring, and vector database creation, unstructured specification documents are transformed into semantically searchable vector indexes. Step S2: Construct a hybrid retrieval and generation system based on retrieval enhancement, adopting a hybrid retrieval strategy that combines vector retrieval with large language model reordering, and generating a structured design scheme based on the retrieval results; Step S3 involves conducting a multi-level performance evaluation of the generated design scheme. The output results of the system are quantitatively evaluated from two dimensions: the performance of proper noun explanation and text retrieval, and the performance of scheme verification and generation, in order to verify the engineering application capability of the system.

[0007] As a further technical solution, the text preprocessing in step S1 includes: extracting content from the PDF standard document recognized by OCR, organizing the chapter hierarchy, converting formulas to LaTeX format, serializing tables to CSV format, and forming a standardized plain text document; The text information structuring includes: identifying the article number and pagination mark through regular expressions, dividing the document into article blocks, and further subdividing them into text blocks, table blocks and formula blocks, constructing a JSON format dictionary containing source, article, page number, type, content and unique identifier, and outputting a JSONL file; The creation of the vector database includes: converting text blocks into vectors using an embedding model and performing L2 normalization, constructing an inner product index using the FAISS library, and forming a vector database.

[0008] As a further technical solution, the hybrid retrieval strategy specifically includes vector retrieval and large language model reordering; Vector retrieval is used to encode user queries into normalized vectors, calculate cosine similarity using FAISS, perform K-nearest neighbor retrieval, and initially recall relevant normative clauses. Large language model re-ranking is used to combine the initial recall results with the user query, input them into the large language model, and use a point-by-point scoring method based on cross-encoder to generate a relevance score for each result. The results with the highest scores are then selected as the final retrieval context.

[0009] As a further technical solution, a structured design scheme is generated based on the search results, including: The highly relevant standard provisions selected after reordering are used as context, and together with the engineering design requirements input by the user and the structured prompt word template containing system instructions and few sample examples, they are input into the large language model. The model generates a structured design scheme containing structural dimensions, material parameters, load values ​​and internal force calculation results through autoregression, and outputs the conclusions, reasoning steps and standard references in JSON format.

[0010] As a further technical solution, the performance evaluation of proper noun explanation and text retrieval in step S3 is evaluated by a weighted fusion score of term matching degree and answer fidelity. The term matching degree is calculated using the entropy weight method with F1@K and normalized loss cumulative gain, and the answer fidelity is measured by the METEOR index.

[0011] As a further technical solution, the solution verification and generation performance evaluation in step S3 are evaluated by comprehensively considering three dimensions: problem relevance, solution compliance, and design feasibility. Among them, the relevance of the question is calculated by the cross encoder to determine the relevance score between the query and the generated solution; the compliance of the solution is composed of the weighted score of the accuracy and completeness of the answer; and the design feasibility is determined by verifying the structural bearing capacity, stability and deformation index of the design solution.

[0012] The second aspect of this invention provides a structural intelligent design and autonomous verification system based on retrieval enhancement generation.

[0013] A retrieval-enhanced generation-based intelligent structural design and autonomous verification system includes: The specification knowledge base construction module is configured to: build a dedicated knowledge base for structural design specifications, and transform unstructured specification documents into semantically searchable vector indexes through text preprocessing, text information structuring, and vector database creation; The structured solution generation module is configured to: construct a hybrid retrieval and generation system based on retrieval enhancement, adopt a hybrid retrieval strategy that combines vector retrieval with large language model reordering, and generate structured design solutions based on the retrieval results; The performance evaluation module is configured to perform multi-level performance evaluation on the generated design schemes, and to quantitatively evaluate the system's output results from two dimensions: the performance of proper noun explanation and text retrieval, and the performance of scheme verification and generation, so as to verify the system's engineering application capabilities.

[0014] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the intelligent structural design and autonomous verification method based on retrieval enhancement generation as described in the first aspect of the present invention.

[0015] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the retrieval-enhanced generation-based intelligent structural design and autonomous verification method described in the first aspect of the present invention.

[0016] The fifth aspect of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the retrieval-enhanced generation-based intelligent structural design and autonomous verification method described in the first aspect of the present invention.

[0017] The above one or more technical solutions have the following beneficial effects: (1) This invention constructs a standardized dedicated knowledge base by preprocessing, structuring, and vectorizing structural design specification documents. This knowledge base adopts a modular design, supports rapid import and index reconstruction of newly added or revised specifications, and can adapt to the continuous iteration of design standards, ensuring the timeliness of design basis and overcoming the shortcomings of inflexible dynamic updates and difficulty in adapting to specification changes in existing technologies. A hybrid retrieval strategy combining vector retrieval and large language model re-ranking is adopted, overcoming the shortcomings of traditional vector retrieval which relies solely on embedded vector similarity and cannot understand complex query intent. Through the deep semantic understanding capabilities of the large language model, the preliminary recall results are subjected to secondary screening and ranking, significantly improving the accurate matching ability of professional terms and the accuracy of specification text retrieval, solving the problems of biased understanding of professional terms and poor retrieval quality in existing technologies.

[0018] (2) This invention constructs a structured prompting engineering framework, which guides a large language model to generate a structured JSON object containing conclusions, reasoning steps, and normative references through system instructions, few-sample examples, and explicit output format constraints. This design makes the generation process of design schemes and verification conclusions transparent and traceable, allowing users to clearly understand the normative provisions and reasoning logic on which each conclusion is based, significantly improving the interpretability of the results.

[0019] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a flowchart of the method in the first embodiment.

[0022] Figure 2 This is a system structure diagram of the second embodiment. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] The overall approach proposed in this invention addresses the problems of low efficiency in standard retrieval, misunderstanding of professional terminology, and insufficient interpretability of results in the structural design of infrastructure such as railway bridges and tunnels. This invention constructs a standardized structural design standard knowledge base to achieve structured storage and vector indexing of standard documents; employs a hybrid retrieval strategy to accurately recall relevant clauses; and generates design schemes containing reasoning steps and references based on structured prompts. Finally, a closed-loop mechanism is used to autonomously verify the compliance of the design schemes, forming an interpretable and traceable intelligent design process.

[0027] Example 1 This embodiment discloses a method for intelligent structural design and autonomous verification based on retrieval enhancement generation; like Figure 1 As shown, the intelligent structural design and autonomous verification method based on retrieval enhancement generation includes: Step S1: Construct a dedicated knowledge base for structural design specifications. Through text preprocessing, text information structuring, and vector database creation, unstructured specification documents are transformed into semantically searchable vector indexes.

[0028] Existing general-purpose models generally lack specifications or standards related to infrastructure structural design, yet the structural design process must strictly adhere to these specifications. Therefore, a dedicated structural design database is constructed to collect relevant specifications. The original optical character recognition (OCR) format PDF specifications undergo preprocessing, including content extraction, chapter-level organization, and standardization of formulas and tables, ultimately resulting in standardized plain text documents.

[0029] In the text preprocessing process, firstly, the content is extracted and organized by chapter level, and then the standard clauses (including clause number and content) are extracted sequentially for each page. The page number is then recorded separately on a new line after the clause in the format "%%%page number%%%", until all page numbers are recorded. Finally, manual proofreading is performed to ensure the integrity and readability of the standard's structure and content.

[0030] Next, the formulas and tables were standardized. A new line was added after the relevant clause to record the formula converted to LaTeX format and its number (numbering format: clause number—Y), such as: (4.2.1-1); After the relevant clause, present the table content in CSV format, and set a table note on a new line below the table. The table note format is "Note: <Specific Explanation>". For example: "Note: When sufficient evidence is obtained through experiments, other steel materials that meet the requirements of bridge span structure may also be used." During the text information structuring process, regular expressions are used to match the pagination markers ("%%%page number%%%") in the plain text document, splitting the text by page and recording the page number for each page. Within each page, regular expressions are used to identify the article number, and the text from that article number to the next article number is segmented into article blocks, recording the source, page number, and article number for each article. Within each article block, a priority strategy is used for fine-grained segmentation: first, tables in CSV format are extracted and recorded as 'table'; then, formulas are segmented and recorded as 'formula'; finally, the text portion is recorded as a text block and recorded as 'text'. The segmented blocks are combined with the article information, constructing JSON key-value pairs in the format "article number (specification name + page number) * article," generating a dictionary with 'source', 'article', 'page', 'type', 'content', and 'id' as keys and the corresponding text block as values. The final output is a JSONL file of chunks.

[0031] In the process of creating a vector database, the first step is to store metadata. By reading the JSONL file mentioned above, the source, article, page, type, and id are stored as metadata in the metadata JSONL file.

[0032] Read the content text blocks in the JSONL file of the above chunks line by line, and use the locally deployed SentenceTransformer model (bge-small-zh-v1.5) to convert each text block into a 384-dimensional vector representation. To make the inner product available for calculating cosine similarity later, perform L2 normalization on all vectors to make the vector magnitude 1, as shown in the following formula:

[0033] in, For vectors No. One portion, For subscript index variables, Let be the dimension of the vector.

[0034] To achieve efficient semantic retrieval, the 384-dimensional vectors transformed by the deep learning model are extracted from the FAISS library to construct an inner product index (IndexFlatIP). This index is stored in contiguous memory to form a vector matrix, creating a database system for storing and querying high-dimensional vectors after text vectorization. An exhaustive search is used to calculate the inner product between the query vector and each vector in the database. The text most similar to the query vector is quickly found by sorting the inner product values ​​(in descending order). Since all vectors have undergone L2 normalization, the inner product value is exactly equal to the cosine similarity, directly reflecting the semantic relevance between texts and measuring the semantic matching degree.

[0035] in, This is the result after normalizing vector A; This is the result after normalizing vector B.

[0036] Step S2: Construct a hybrid retrieval and generation system based on retrieval enhancement, adopting a hybrid retrieval strategy that combines vector retrieval with large language model reordering, and generating a structured design scheme based on the retrieval results.

[0037] In this embodiment, to improve the efficiency of design specification retrieval and achieve multi-source information matching, a hybrid retrieval strategy is adopted, which includes vector retrieval and large language model reordering.

[0038] Vector retrieval encodes user queries into normalized vectors, calculates cosine similarity using FAISS, performs K-nearest neighbor retrieval, and initially recalls relevant regulatory provisions. Specifically: Inner product retrieval is implemented using spatial partitioning (IVF, HNSW) and product quantization (PQ). A local embedding model (bge-small-zh-v1.5) is loaded using SentenceTransformer to encode the query question as a 384-dimensional vector, and L2 normalization is performed (similar to the text block normalization process described above). Cosine similarity is calculated using matrix multiplication for each vector under the query vector and all IndexFlatIP (inner product index) in the vector database (since the vectors have been normalized, cosine similarity is equivalent to the inner product), yielding all similarities.

[0039] Where q is the query vector. This is the embedding vector of the chunk in the database. The higher the inner product similarity, the more similar the text chunk vector is to the query vector.

[0040]

[0041]

[0042]

[0043] Where q is the query vector. for The query matrix, for The database vector matrix is ​​used to obtain the similarity matrix, where each element represents the cosine similarity between the corresponding vector and the query vector.

[0044] After quickly completing all similarity calculations, a K-nearest neighbor search is performed to retrieve the 30 most similar values ​​(Top-30), and the corresponding index and similarity value are returned. Finally, the corresponding text information is retrieved from the metadata file based on the index.

[0045] Large language model re-ranking is used to combine the initial recall results with the user query, input them into the large language model, and use a point-by-point scoring method based on cross-encoder to generate a relevance score for each result. The results with the highest scores are then selected as the final retrieval context.

[0046] To improve the accuracy and relevance of search results and to compensate for the shortcomings of vector retrieval (FAISS retrieval) which relies solely on the similarity of embedded vectors and cannot understand complex query intent and context, this embodiment adopts a hybrid retrieval strategy that combines FAISS preliminary retrieval with LLM re-ranking. This strategy utilizes the deep semantic understanding capabilities of a large model to perform secondary filtering and ranking of vector retrieval results.

[0047] The metadata corresponding to the 30 results retrieved by FAISS is sorted according to canonical name, page number, and sequence number to form a structured text. Then, using the Cross-Encoder reordering method based on generative scoring in Pointwise, the query question is... And the above 30 texts Input the local Ollama model (qwen2:1.5b-instruct-q8_0) and calculate the relevance score for each document d:

[0048] Where f is the correlation function (implemented by the large language model, which uses qwen2:1.5b-instruct-q8_0 in this system). The query question is set by the prompt words. ,text Text composed of related instruction templates.

[0049] When calculating the relevance score, the above The text is converted into a token sequence, mapped to a vector, and then rotated to position (RoPE) encoding. For the vector x at position m, a rotation matrix is ​​used. Perform the transformation:

[0050] This matrix acts on each pair of dimensions of the vector, giving each position a unique representation and enabling the model to capture relative positional relationships.

[0051] The above token sequence and position encoding are input into the Transformer model (each Transformer layer includes a multi-head attention mechanism and a feedforward neural network). The multi-head attention mechanism is used to concatenate the outputs of multiple attention heads. For each attention head, the query (Q), key (K), and value (V) are calculated:

[0052]

[0053]

[0054] in, It is the input matrix. , , It is a learnable weight matrix.

[0055] Then, calculate the attention score:

[0056] in, It is the dimension of K, used for scaling to prevent the dot product from becoming too large and causing softmax saturation.

[0057] The attention output described above is fed into a feedforward neural network (FFN) after residual connections and layer normalization. This FFN performs an independent nonlinear transformation on the representation at each position (i.e., the representation of each token). SwiGLU is used in the FFN to combine two linear layers and a gating mechanism.

[0058]

[0059]

[0060] Where β=1, This indicates element-wise multiplication. and Transfer the input x from Dimension mapping to Dimension, then yes Wei, then through Will Dimension Mapping Back dimension.

[0061] After multiple Transformer layers, the model generates each token based on the input conditions, and obtains the entire output after processing the last token.

[0062] Here, input is the prompt word, and output is the JSON string to be output.

[0063] After linear transformation of the output layer, it understands and generates text containing scores from 0 to 10 according to the scoring criteria specified by the prompt words, and generates scoring reasons in an autoregressive manner (that is, it generates a word at a time and adds it to the input, and continues to generate the next word until the maximum length is reached or the end symbol is generated). For the 30 FAISS search results in the input, it obtains a JSON list containing the corresponding index, score and reason.

[0064] Next, the output results are structured. Based on the scores, the 30 FAISS search results are sorted, and the top 10 (Top-10) results are selected. These Top-10 results, the query question, and the constructed prompts (including system instructions, few-shot examples, output format, and contextual text) are input into the local Ollam model (qwen2:1.5b-instruct-q8_0). The model then predicts the next token using autoregression (i.e., token-by-token generation) based on the currently generated tokens and prompts.

[0065]

[0066] in As a prompt word, Model parameters, The probability of generating the output sequence O.

[0067] By specifying the temperature parameter (temperature=0.1), the probability distribution is obtained by applying softmax.

[0068]

[0069] Here, temperature=0.1 is equivalent to a 10-fold amplification. Because the temperature is lower, the model will more confidently select the token with the highest probability.

[0070] The final model outputs a JSON object containing specific fields (conclusion, reasoning_steps, references).

[0071] Step S3 involves conducting a multi-level performance evaluation of the generated design scheme. The output results of the system are quantitatively evaluated from two dimensions: the performance of proper noun explanation and text retrieval, and the performance of scheme verification and generation, in order to verify the engineering application capability of the system.

[0072] To ensure that the constructed specialized intelligent design system for infrastructure structures can accurately and efficiently realize the four levels of functions—definition of proper nouns, intelligent retrieval of clauses, verification of design schemes, and generation of design schemes—and to guarantee its application capability in actual engineering, this study uses the following fusion indicators to conduct multi-level performance evaluation of the system.

[0073] Step S31: Evaluate the performance of the system in retrieving and generating structural design-related terms.

[0074] To ensure the accuracy and reliability of the system's retrieval of specialized terms and normative clauses related to infrastructure structural design, a weighted average score of the combined indicators of term matching degree and answer fidelity was used to quantitatively evaluate the system's performance in terms of explanation and intelligent retrieval of clauses.

[0075] The term matching degree is evaluated using F1@K and normalized depreciation cumulative gain (NDCG@K) as important indicators to comprehensively assess the accuracy, recall rate and ranking quality of the system's function of retrieving and generating infrastructure structure design-related proper nouns and standard clauses.

[0076] Here, F1@K refers to the harmonic mean of precision@K and recall@K, used to comprehensively measure the system's performance when faced with the first K returned results. That is:

[0077] in, Precision refers to the proportion of relevant results among the first K results returned by the system. Recall rate refers to the number of results that, for a given query, are among the first K results returned by the system and cover all the originally relevant results.

[0078] Normalized Discount Cumulative Gain (NDCG@K) measures the quality of a sorted list (the top K results in the system) using a score between 0 and 1. The closer the score is to 1, the more perfect the sort. It calculates the discount cumulative gain by introducing a discount factor, and finally calculates the normalized discount cumulative gain, which is limited to the range of 0 to 1, where 1 represents a perfect sort and 0 represents the worst, as shown below:

[0079] in, To reduce accumulated gain; The DCG value refers to the ideal loss cumulative gain, which is the value calculated assuming that all relevant results are perfectly sorted from high to low correlation.

[0080] The term matching degree of the system is obtained by weighting the F1@K and normalized depreciation cumulative gain (NDCG@K) results using the entropy weighting method. Specifically: First of all The original data matrix consisting of one evaluation object (sample) and the above two evaluation indicators. The value is standardized. .

[0081] If both of the above evaluation indicators are positive, then:

[0082] Then calculate the first... Entropy value of the item index :

[0083] in, ensure ,like Then define .

[0084] Calculate the next Coefficient of difference of the items :

[0085] in The larger the value, the more information the indicator provides, and the greater its weight should be.

[0086] Final calculation of the first Weight of each indicator :

[0087] Obtain the weights of each indicator And satisfy .

[0088] The weighted average score is calculated based on the weights of each indicator obtained above, and this score is used as the term matching score.

[0089] in, The above calculations yielded The weights; The normalized cumulative gain calculated above is... The weight of the term matching score. The higher the term matching score, the higher the matching degree of the terms generated by the system and the better the performance.

[0090] Answer faithfulness refers to the extent to which the generated answer remains consistent with the source information it is based on (i.e., the system's vector database), without containing any fabricated, distorted, or added content. By employing METEOR (Metric for Evaluation of Translation with Explicit ORdering) as the basic metric to evaluate the system's answer faithfulness, the reliability of its results is ensured, effectively avoiding the illusion problem inherent in large language models.

[0091] First, word alignment is established between the candidate and reference texts. Exact matching prioritizes aligning identical words; among the remaining unaligned words, words with the same root are searched for stem matching; then, among the still unaligned words, synonyms are found using semantic dictionaries such as WordNet for matching.

[0092] Secondly, based on the optimal alignment result determined in the previous step, the most basic precision and recall are calculated.

[0093]

[0094]

[0095] METEOR uses a parameter α to adjust the weighting of recall and precision:

[0096] To evaluate the comprehensiveness of the system's generated results and improve the recall rate weight, the following settings are configured: ,but:

[0097] Then, to evaluate the correctness of the word order, a sequence of consecutively matched words in the candidate and reference translations is defined as a chunk. If the word order is poor, the matched words are scattered throughout the sentence, forming more chunks; conversely, fewer chunks are formed. A fragmentation penalty is calculated based on the number of chunks.

[0098] Where matches is the total number of matched words; chunks is the number of consecutive matched word sequences. This is a parameter that controls the maximum penalty value; set it to 0.5. It is a parameter that controls the shape of the penalty function, set to 3.

[0099] Finally, the weighted harmonic mean score is combined with the fragmentation penalty to obtain the METEOR score, which measures the fidelity of the system-generated answer:

[0100] Combining the two evaluation scores above, the Entropy Weight Method is used to automatically calculate weights based on the degree of variability of the indicators and the dispersion of the data itself, reflecting the distinguishing ability of the indicators. The weighted average score is then taken as the final score.

[0101] in, The first one calculated above The weight of the term matching degree of each indicator; The first one calculated above The weight of the faithfulness of the answers to each indicator; the higher the final score, the better the system's performance in retrieving and generating structural design-related terms.

[0102] Step S32: Verify the system solution and evaluate the performance of the generated solution. To ensure the compliance and feasibility of the structural design schemes and parameters generated by the system for actual infrastructure projects, this study uses three evaluation indicators—problem relevance, scheme compliance, and design feasibility—to calculate their weighted average score and comprehensively evaluate the system's scheme verification and scheme generation performance.

[0103] To assess the relevance of the problem, a cross-encoder is used to obtain a relevance score between the user query and the generated solution document, which is then used as an indicator to evaluate the problem relevance of the system.

[0104] For a given query (length is) (each token) and generation scheme document (length is) (each token) First, the query and document are concatenated into a sequence and separated by a special marker:

[0105] in, These are special markers used to aggregate sequence information; Used to separate queries and documents; Indicates query The first in One token; Indicates the generation of solution documents The first in A token.

[0106] The initial representation of each token is the sum of three parts:

[0107] in, Map the token to word vectors; Inject location information; Used to distinguish between queries (segment 0) and documents (segment 1), helping the model identify the structure of input pairs.

[0108] Finally, the input matrix is ​​obtained. ,in For the hidden layer dimension.

[0109] Input the above-obtained matrix Input by The encoder structure is based on Transformer and consists of stacked layers. Each layer contains multi-head self-attention and a feedforward network (FFN), along with residual connections and layer normalization.

[0110] For the Layer input ,in Multi-head attention calculation is performed for the sequence length:

[0111] Each of them for ; This is the output projection matrix for multi-head attention. , and It is the first Trainable projection matrices for each attention head are used to map the input sequence to the Query, Key, and Value spaces, respectively.

[0112] Attention function combined with normalized exponential function By scaling factor Using scaled dot product:

[0113] Let the query be... The query vector for each token is Document No. The key vector of each token is When calculating attention weights:

[0114] Combining residual connectivity with layer normalization:

[0115]

[0116]

[0117] in Typically, this involves two linear transformations plus an activation function:

[0118] go through The final hidden state is obtained after the layer. .

[0119] use The last hidden state of the labeled layer is used as the aggregate vector for the entire input pair:

[0120] Finally, aggregate the vector Input a linear layer, mapped to scalar fractions:

[0121] in, , .

[0122] To assess the compliance of the solutions, a weighted average score was calculated using a combination of answer correctness and answer integrity metrics to determine the compliance score of the system-generated solutions.

[0123] Answer Correctness is calculated based on the ground truth and the answer, with a score ranging from 0 to 1. A higher score indicates that the generated answer is closer to the true answer, thus demonstrating higher accuracy. Answer correctness encompasses the similarity between the generated answer and the ground truth. The semantic similarity between the generated answer and the ground truth. Two aspects.

[0124] The F1 score is used to measure the similarity between the generated answer and the ground truth. .

[0125]

[0126] in, For accuracy ; For recall rate .but:

[0127] In this context, TP stands for True Positive, meaning a fact or statement exists in both the ground truth and the generated answer; FP stands for False Positive, meaning a fact or statement exists in the generated answer but not in the ground truth; and FN stands for False Negative, meaning a fact or statement exists in the ground truth but not in the generated answer.

[0128] Vector similarity is used to measure the semantic similarity between the generated answer and the ground truth. .

[0129]

[0130] Using a weighted average, the two are combined as the final result of the answer's correctness:

[0131] Since the system must ensure that the generated answers are based on factual similarity to the normative provisions, it takes... .

[0132] Answer integrity is based on ground truth and is calculated by covering key words. and information density To measure whether the generated answer is complete.

[0133] Extract keywords from the ground truth using a Large Language Model (LLM) and calculate keyword coverage. :

[0134] in, This indicates the number of keywords contained in the generated answer; This indicates the number of ground truth keywords.

[0135] The ground truth and the answer are segmented into words and sentences, and their information density is calculated. :

[0136] in, Segment the ground truth; Divide the answer into clauses; Segment the ground truth; Segment the answer into words; add recall to the density using crossover and union ratio.

[0137] Combining the two yields the answer integrity:

[0138] The feasibility of the system is obtained by using an entropy weighting method to calculate a weighted average of the accuracy and completeness of the answers.

[0139] in, The weighting of the accuracy of the answers calculated above; This is the weight for the completeness of the answer calculated above.

[0140] To assess the feasibility of the design, the load-bearing capacity, stability, and deformation of the designed structure are comprehensively considered to verify the feasibility of the design scheme.

[0141] Taking a simply supported T-shaped reinforced concrete bridge as an example, the loads borne by the structure are first calculated by comprehensively considering both permanent and variable actions, and then combined according to the specifications to obtain the bearing capacity combination:

[0142] in, The structural importance coefficients are set to 1.1, 1.0, and 0.9 for levels one, two, and three, respectively. This is the partial factor for the dead load; take 1.2 if the load is unfavorable. The partial factor for vehicle load is 1.4. This is the live load partial factor; The combination value coefficient for live load; For the first The effect of a standard value of constant load; For the first The effect of a standard live load value; This represents the number of types of permanent loads. This represents the number of live load types.

[0143] Calculate the internal forces at the control section:

[0144]

[0145] in, To calculate the span, take the sum of the net span and the support length; For uniformly distributed loads; For concentrated loads.

[0146] Perform flexural capacity calculation of the normal section (here) This is a first-class T-beam, classified by width. (Calculation of rectangular beams)

[0147]

[0148] in, The height of the concrete compression zone must meet the following requirements. To prevent over-reinforcement, The relative height of the pressure zone is the boundary. This refers to the tensile strength of ordinary steel bars; This is the design value for the axial compressive strength of concrete. This represents the cross-sectional area of ​​the longitudinal reinforcement in the tension zone; The effective width of the T-section flange; This is the effective height of the cross-section. It must also meet the following requirements: To prevent muscle loss.

[0149] Calculate the shear capacity of the inclined section:

[0150]

[0151]

[0152] The combined shear capacity of the concrete and stirrups is:

[0153] in, This is the design value for shear force; This refers to the standard value of the compressive strength of a concrete cube. The influence coefficient for opposite-sign bending moments is set to 1.0. The prestress influence coefficient is taken as 1.0; The influence coefficient of the compressed flange; ; .

[0154] Perform crack width verification:

[0155] in, This is the surface shape factor for reinforcing bars; for ribbed reinforcing bars, it is taken as 1.0. This represents the long-term effect coefficient. The component's stress property coefficient; The stress in the reinforcing steel at the crack section is calculated based on the combination of short-term effects. The diameter of the longitudinal tensile reinforcement; This refers to the longitudinal tensile reinforcement ratio.

[0156] Perform deflection calculation:

[0157] in, The elastic modulus of concrete; To convert the moment of inertia of the cross section; This is the coefficient of non-uniform strain of the longitudinal tensile reinforcement between cracks; The reinforcement factor is the pressure flange.

[0158] Finally, verification is performed. The clear spacing between reinforcing bars should not be less than the diameter of the reinforcing bars and not less than 30mm; the thickness of the protective layer is determined according to the environmental category; the basic anchorage length of tensile reinforcing bars is... The values ​​must meet the specifications; the maximum spacing of stirrups should not exceed [the specified value]. And it should not exceed 400mm. It must also meet the following requirements:

[0159]

[0160]

[0161]

[0162] If all the above calculations meet the requirements, the design feasibility is considered to be satisfied.

[0163] Analogous to this simply supported T-shaped reinforced concrete bridge, corresponding verification formulas are used for different structures to verify the results.

[0164] Finally, taking into account the evaluation results from three aspects—relevance to the problem, compliance of the solution, and feasibility of the design—a comprehensive evaluation of the system solution verification and solution generation performance was completed.

[0165] The entropy weight method is used again to automatically calculate weights based on the degree of variability of the indicators and the dispersion of the data itself, reflecting the distinguishing ability of the indicators. The weighted average score is then taken as the final score.

[0166] in, The weights for the problem relevance calculated above; The weights for the compliance of the schemes calculated above are used; the higher the final score and the more feasible the scheme, the better the performance of the system's scheme verification and scheme generation.

[0167] The system has undergone four levels of small-sample verification, including proper noun explanation, intelligent text retrieval, design scheme verification, and design scheme generation, and its functions meet the requirements.

[0168] Example 2 This embodiment discloses a structural intelligent design and autonomous verification system based on retrieval enhancement generation; like Figure 2 As shown, the structure intelligent design and autonomous verification system based on retrieval enhancement generation includes: The specification knowledge base construction module is configured to: build a dedicated knowledge base for structural design specifications, and transform unstructured specification documents into semantically searchable vector indexes through text preprocessing, text information structuring, and vector database creation; The structured solution generation module is configured to: construct a hybrid retrieval and generation system based on retrieval enhancement, adopt a hybrid retrieval strategy that combines vector retrieval with large language model reordering, and generate structured design solutions based on the retrieval results; The performance evaluation module is configured to perform multi-level performance evaluation on the generated design schemes, and to quantitatively evaluate the system's output results from two dimensions: the performance of proper noun explanation and text retrieval, and the performance of scheme verification and generation, so as to verify the system's engineering application capabilities.

[0169] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0170] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the retrieval-enhanced generation-based intelligent structural design and autonomous verification method as described in Example 1.

[0171] Example 4 The purpose of this embodiment is to provide an electronic device.

[0172] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the retrieval-enhanced generation-based intelligent structural design and autonomous verification method as described in Embodiment 1.

[0173] Example 5 Embodiment 5 of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the retrieval-enhanced generation-based intelligent structural design and autonomous verification method as described in Embodiment 1.

[0174] The steps and methods involved in the apparatuses of Embodiments 2, 3, 4, and 5 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0175] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0176] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for intelligent structural design and autonomous verification based on retrieval-enhanced generation, characterized in that, include: Step S1: Construct a dedicated knowledge base for structural design specifications. Through text preprocessing, text information structuring, and vector database creation, unstructured specification documents are transformed into semantically searchable vector indexes. Step S2: Construct a hybrid retrieval and generation system based on retrieval enhancement, adopting a hybrid retrieval strategy that combines vector retrieval with large language model reordering, and generating a structured design scheme based on the retrieval results; Step S3 involves conducting a multi-level performance evaluation of the generated design scheme. The output results of the system are quantitatively evaluated from two dimensions: the performance of proper noun explanation and text retrieval, and the performance of scheme verification and generation, in order to verify the engineering application capability of the system.

2. The structural intelligent design and autonomous verification method based on retrieval enhancement generation as described in claim 1, characterized in that, The text preprocessing in step S1 includes: extracting content from the PDF standard document recognized by OCR, organizing the chapter hierarchy, converting formulas to LaTeX format, serializing tables to CSV format, and forming a standardized plain text document. The text information structuring includes: identifying the article number and pagination mark through regular expressions, dividing the document into article blocks, and further subdividing them into text blocks, table blocks and formula blocks, constructing a JSON format dictionary containing source, article, page number, type, content and unique identifier, and outputting a JSONL file; The creation of the vector database includes: converting text blocks into vectors using an embedding model and performing L2 normalization, constructing an inner product index using the FAISS library, and forming a vector database.

3. The structural intelligent design and autonomous verification method based on retrieval enhancement generation as described in claim 1, characterized in that, The hybrid retrieval strategy specifically includes vector retrieval and large language model reordering; Vector retrieval is used to encode user queries into normalized vectors, calculate cosine similarity using FAISS, perform K-nearest neighbor retrieval, and initially recall relevant normative clauses. Large language model re-ranking is used to combine the initial recall results with the user query, input them into the large language model, and use a point-by-point scoring method based on cross-encoder to generate a relevance score for each result. The results with the highest scores are then selected as the final retrieval context.

4. The structural intelligent design and autonomous verification method based on retrieval enhancement generation as described in claim 1, characterized in that, A structured design scheme is generated based on the search results, including: The highly relevant standard provisions selected after reordering are used as context, and together with the engineering design requirements input by the user and the structured prompt word template containing system instructions and few sample examples, they are input into the large language model. The model generates a structured design scheme containing structural dimensions, material parameters, load values ​​and internal force calculation results through autoregression, and outputs the conclusions, reasoning steps and standard references in JSON format.

5. The structural intelligent design and autonomous verification method based on retrieval enhancement generation as described in claim 1, characterized in that, The performance evaluation of proper noun explanation and text retrieval in step S3 is evaluated using a weighted fusion score of term matching degree and answer fidelity. The term matching degree is calculated using the entropy weight method with F1@K and normalized loss cumulative gain, and the answer fidelity is measured by the METEOR index.

6. The structural intelligent design and autonomous verification method based on retrieval enhancement generation as described in claim 1, characterized in that, The solution verification and generation performance evaluation in step S3 are evaluated from three dimensions: problem relevance, solution compliance, and design feasibility. Among them, the relevance of the question is calculated by the cross encoder to determine the relevance score between the query and the generated solution; the compliance of the solution is composed of the weighted score of the accuracy and completeness of the answer; and the design feasibility is determined by verifying the structural bearing capacity, stability and deformation index of the design solution.

7. A structural intelligent design and autonomous verification system based on retrieval-enhanced generation, characterized in that, include: The specification knowledge base construction module is configured to: build a dedicated knowledge base for structural design specifications, and transform unstructured specification documents into semantically searchable vector indexes through text preprocessing, text information structuring, and vector database creation; The structured solution generation module is configured to: construct a hybrid retrieval and generation system based on retrieval enhancement, adopt a hybrid retrieval strategy that combines vector retrieval with large language model reordering, and generate structured design solutions based on the retrieval results; The performance evaluation module is configured to perform multi-level performance evaluation on the generated design schemes, and to quantitatively evaluate the system's output results from two dimensions: the performance of proper noun explanation and text retrieval, and the performance of scheme verification and generation, so as to verify the system's engineering application capabilities.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the intelligent structural design and autonomous verification method based on retrieval enhancement generation as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the structural intelligent design and autonomous verification method based on retrieval enhancement generation as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps in the retrieval-enhanced generation-based intelligent structural design and autonomous verification method as described in any one of claims 1-6.

Images