Implementation method of structural analysis software artificial intelligence question answering system

Abstract

The invention belongs to the technical field of structural analysis software application support, and particularly relates to an implementation method of a structural analysis software artificial intelligence question answering system, which comprises the following steps: converting a structural analysis software related knowledge document into a segmented structured text; based on the structured text, calling a generative artificial intelligence model, generating question and answer pairs, and constructing a text data set; carrying out fine tuning on the basic model based on the text data set by adopting a LoRA fine tuning technology, and training to obtain a LoRA adapter weight; loading the basic model and the LoRA adapter weight to form an expert question and answer model focusing on the related knowledge field of the structural analysis software; constructing a question and answer vector database of related knowledge of the structure analysis software; related knowledge fragments are retrieved from the question and answer vector database for questions input by a structural analysis software user through the RAG technology; questions and related knowledge fragments input by a structural analysis software user are input to the expert question and answer model, and professional answers are generated and output through the expert question and answer model.

Description

Technical Field

[0001] This application belongs to the field of structural analysis software application support technology, specifically relating to an implementation method of an artificial intelligence question-answering system for structural analysis software. Background Technology

[0002] Structural analysis software is a crucial tool for engineering design simulation and analysis, playing a vital role in fields such as aerospace, machinery manufacturing, and civil engineering. However, current structural analysis software generally suffers from problems such as complex functionality and cumbersome operation procedures, posing numerous challenges for users. Specifically, structural analysis software typically includes complex modeling interfaces and numerous parameter setting options, leading users to frequently encounter issues such as unclear operation steps, difficulty in understanding parameter meanings, and a lack of theoretical basis during structural modeling and result interpretation.

[0003] Current technical support for structural analysis software primarily relies on official user manuals, technical white papers, and operation guides. While these documents are comprehensive, they suffer from drawbacks such as fragmented content, inconsistent formatting, and low search efficiency. Users often need to frequently switch between different documents and search for relevant information using keywords, making it difficult to quickly locate the accurate answer to their problem.

[0004] In addition, structural analysis involves fundamental knowledge such as finite element theory, mechanics of materials, and structural dynamics, and contains a large number of technical terms. This specialization further increases the difficulty for users to obtain effective technical support for structural analysis software. For beginners and non-professional users, the learning curve of the software is steep and the threshold for use is high.

[0005] Therefore, there is an urgent need for a method that can understand user problems, quickly extract relevant knowledge, and provide accurate answers to improve the usability of structural analysis software, lower the threshold for using structural analysis, and support efficient engineering design simulation analysis. In view of this, this application is proposed.

[0006] This application is made in view of the aforementioned technical deficiencies. Summary of the Invention

[0007] The purpose of this application is to provide a method for implementing an artificial intelligence question-answering system for structural analysis software, which can quickly respond to user questions and provide effective support for the widespread application of the software.

[0008] The technical solution of this application is:

[0009] A method for implementing an AI question-answering system for structural analysis software includes:

[0010] Step 1: Convert the relevant knowledge documents of the structural analysis software into segmented structured text;

[0011] Step 2: Based on the structured text, invoke the generative artificial intelligence model to generate question-answer pairs and construct a text dataset;

[0012] Step 3: Using LoRA fine-tuning technology, fine-tune the base model based on the text dataset to train and obtain the LoRA adapter weights;

[0013] Step 4: Load the base model and LoRA adapter weights to form an expert question-and-answer model focused on the knowledge domain related to structural analysis software;

[0014] Step 5: Construct a question-and-answer vector database of knowledge related to structural analysis software;

[0015] Step 6: For the questions input by the user of the structural analysis software, use RAG technology to retrieve relevant knowledge fragments from the question-answer vector database;

[0016] Step 7: Input the questions and relevant knowledge fragments entered by the user of the structural analysis software into the expert question-and-answer model, so that the expert question-and-answer model can generate and output professional answers.

[0017] Optionally, in the above-mentioned implementation method of the AI ​​question-answering system for structural analysis software, the knowledge documents related to the structural analysis software are from multiple sources and are heterogeneous, including the official user manual, technical white paper, and operation guide of the structural analysis software, as well as professional textbooks and literature on finite element theory, mechanics of materials, and structural dynamics, in formats including txt, pdf, word, and excel.

[0018] Optionally, in the above implementation method of the AI ​​question-answering system for structural analysis software, in step one, the intelligent data extraction tool MinerU is used to extract the text content of relevant knowledge documents of structural analysis software, and the text content is segmented according to chapters and uniformly converted into segmented Markdown format structured text.

[0019] Optionally, in the above implementation method of the structure analysis software artificial intelligence question answering system, in step two, the DeepSeek model API is called to automatically generate question-answer pairs based on structured text, which are used as seed data to expand and construct a text dataset;

[0020] The text dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The text was then standardized, including converting it to Simplified Chinese, removing irrelevant symbols, using a tokenizer for word segmentation and encoding, and standardizing the input and output lengths of question-answer pairs.

[0021] Optionally, in the above-mentioned implementation method of the structural analysis software artificial intelligence question answering system, in step three, the DeepSeek-7B distillation model is used as the basic model, and LoRA fine-tuning technology is used for training.

[0022] Optionally, in the above implementation method of the AI ​​question-answering system for structural analysis software, step four involves merging the basic model and the LoRA adapter weights to form an expert question-answering model focused on the knowledge domain related to structural analysis software.

[0023] Optionally, in the above implementation method of the AI ​​question-answering system for structural analysis software, in step five, a text embedding model is used to convert the relevant knowledge documents of the structural analysis software into question-answer vector representations, and a question-answer vector database is constructed.

[0024] The question-and-answer vector database is updated in real time and supports efficient retrieval;

[0025] Semantic association extension is performed on question-answer pairs in the question-answer vector database to generate a series of related questions with different wording but similar intent.

[0026] Optionally, in the above implementation method of the AI ​​question-answering system for structural analysis software, in step six, the questions input by the user of the structural analysis software are converted into query vectors using the RAG system and a text embedding model. The similarity scores between the query vectors and the question vectors in the question-answering vector database are calculated, and the knowledge fragments corresponding to the top three question vectors with the highest similarity scores are extracted as the most relevant knowledge fragments.

[0027] Optionally, in the above implementation method of the AI ​​question-answering system for structural analysis software, in step seven, the questions input by the user of the structural analysis software and related knowledge fragments are combined into structured prompts, and the structured prompts are input into the expert question-answering model so that the expert question-answering model can generate and output professional answers.

[0028] This application has at least the following beneficial technical effects:

[0029] This paper presents a method for implementing an AI-powered question-answering system for structural analysis software. Based on the DeepSeek-7B distillation model, it employs LoRA fine-tuning technology and trains an expert question-answering model targeting the relevant knowledge domain of the structural analysis software using question-answer pairs generated by the DeepSeek API. Furthermore, it introduces the RAG system to retrieve relevant authoritative knowledge fragments from a question-answer vector database for user-input questions, combining these fragments with the user's question and inputting them into the expert question-answering model to generate accurate, professional, and evidence-based answers. This system can quickly respond to user questions during software use, providing intelligent question-answering services and offering efficient and accurate technical support for the widespread application of the software. Attached Figure Description

[0030] Figure 1 This is a schematic diagram illustrating the implementation method of the artificial intelligence question-answering system for structural analysis software provided in this application embodiment.

[0031] To better illustrate this embodiment, some content in the accompanying drawings may be omitted. They are for illustrative purposes only and should not be construed as limiting the scope of this application. Detailed Implementation

[0032] To make the technical solution and advantages of this application clearer, the technical solution of this application will be described in a clearer and more complete manner below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some embodiments of this application, and are only used to explain this application, not to limit this application. It should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, and other related parts can be referred to the general design.

[0033] Furthermore, unless otherwise defined, the technical or scientific terms used in this application description shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The word "comprising" as used in this application description indicates that the concept preceding the word encompasses the concepts listed following the word and their equivalents, without excluding other related concepts.

[0034] To lower the learning and usage barriers of structural analysis software and improve the efficiency of engineering design simulation analysis, this application provides a method for implementing an artificial intelligence question-answering system for structural analysis software, such as... Figure 1 As shown, by combining Retrieval Enhancement Generation (RAG) with Low-Rank Adaptive Fine-Tuning (LoRA), an intelligent question-answering system specifically designed for structural analysis is constructed, which can efficiently respond to questions raised by users during software use.

[0035] The design transforms knowledge documents related to structural analysis software into segmented, structured text. It then uses a generative artificial intelligence model to generate question-and-answer pairs, constructing a text dataset. LoRA technology is used to efficiently fine-tune the parameters of the base model, enabling it to deeply understand domain terminology and internal logic. This results in an expert question-and-answer model for the knowledge domain related to structural analysis software. Additionally, the RAG system is introduced to retrieve relevant authoritative knowledge fragments from the question-and-answer vector database for user-input questions. These knowledge fragments are then provided as context along with the user's question to the LoRA-tuned expert question-and-answer model, which generates the final accurate and professional answer.

[0036] Step 1: Convert the relevant knowledge documents of the structural analysis software into segmented structured text.

[0037] The knowledge documents related to structural analysis software are diverse and heterogeneous, mainly including official user manuals, technical white papers, operation guides, as well as professional textbooks and literature on finite element theory, mechanics of materials, and structural dynamics, in various formats such as txt, pdf, word, and excel.

[0038] The intelligent data extraction tool MinerU can extract the text content of knowledge documents related to structural analysis software, divide the text content into reasonable segments according to chapters, and uniformly convert it into structured text in Markdown format with clear segments and well-defined tags. This forms a standardized, easy-to-search and process formatted database, providing a high-quality data foundation for subsequent model training and vectorization processing.

[0039] Step 2: Based on the structured text, call the generative artificial intelligence model to generate question-answer pairs and build a text dataset.

[0040] It can call model APIs such as DeepSeek to automatically generate a large number of high-quality question-answer pairs based on structured text, which can be used as seed data for further expansion to build a complete text dataset.

[0041] The prompts can be designed to require generative AI models to ask relevant questions and generate standard answers based on the given paragraph content, ensuring that the questions and answers closely align with professional knowledge.

[0042] The text dataset can be further divided into training, validation, and test sets according to a certain ratio, such as 7:2:1, to ensure that no data is missing. The text is then standardized, including converting it to simplified Chinese, removing irrelevant symbols, using a tokenizer for word segmentation and encoding, and standardizing the input and output lengths of question-answer pairs.

[0043] Step 3: Using LoRA fine-tuning technology, fine-tune the base model based on the text dataset to train and obtain the LoRA adapter weights.

[0044] The DeepSeek-7B distillation model can be used as the base model, and LoRA fine-tuning technology can be used for efficient training.

[0045] For an original weight matrix The LoRA method updates it to To limit the update amount and improve training efficiency, a low-rank decomposition method is used to represent the weight update matrix. , where the matrix ,matrix ,rank During training, the original weight matrix Frozen, will not be updated, all trainable parameters are concentrated in the matrix , middle.

[0046] Original weight matrix With the weight update matrix Sharing the same input vector Their outputs are added together on the coordinates, and the forward propagation process matrix is... It can be represented as:

[0047] .

[0048] Regarding parameter initialization, for the matrix Initialize using random Gaussian and set the matrix Set it to a zero matrix to ensure that the weight update matrix is ​​used at the start of training. The initial output is zero, and the original weight matrix is ​​the same as the original output. Consistent.

[0049] To further stabilize training, the low-rank update item was updated. Apply scaling factor ,in, It is a constant. When using adaptive optimizers such as Adam, This is approximately equivalent to the learning rate in full parameter fine-tuning, and the scaling rate should be set appropriately. In practice, it is usually set to... Afterwards, it may no longer follow the order. By making adjustments, this method can significantly reduce the need for hyperparameter retuning.

[0050] In practical reasoning, the merged weights can be explicitly calculated and stored. To maintain inference efficiency. When it is necessary to switch to different downstream tasks, this can be achieved by subtracting the current... Quickly recover the original weight matrix By adding adaptation weights for new tasks This yields a new weight matrix. This operation has extremely low memory overhead, and compared with traditional fine-tuning models, the LoRA method does not introduce any additional latency in inference.

[0051] During training, specific parameters for LoRA fine-tuning, such as rank, application module, learning rate, batch size, and training epochs, are set. After training, lightweight LoRA adapter weights are obtained, typically only tens of MB in size.

[0052] Step 4: Load the base model and LoRA adapter weights to form an expert question-and-answer model focused on knowledge related to structural analysis software.

[0053] By merging the base model and LoRA adapter weights, an expert question-and-answer model can be formed that focuses on the knowledge domain related to structural analysis software. This model can deeply understand the software usage and professional terminology in the knowledge domain related to structural analysis software.

[0054] Step 5: Construct a question-and-answer vector database of knowledge related to structural analysis software.

[0055] The text embedding model is used to convert knowledge documents related to structural analysis software into question-answer vector representations, thus constructing a question-answer vector database.

[0056] The question-answer vector database can be updated in real time and supports efficient retrieval. To enhance the accuracy and recall of vector retrieval, the question-answer pairs in the question-answer vector database are semantically extended to generate a series of related questions with different wording but similar intent.

[0057] Step Six: For the questions input by the user of the structural analysis software, use RAG technology to retrieve relevant knowledge fragments from the question-answer vector database.

[0058] For questions input by users of the structural analysis software, the RAG system uses a text embedding model to convert the questions into query vectors. The similarity scores between the query vectors and the question vectors in the question-answer vector database are calculated. The knowledge fragments corresponding to the answer vectors of the top three question vectors with the highest similarity scores are extracted as the most relevant authoritative knowledge fragments, which are then used as extended content for subsequent prompts.

[0059] Step 7: Input the questions and relevant knowledge fragments entered by the user of the structural analysis software into the expert question-and-answer model, so that the expert question-and-answer model can generate and output professional answers.

[0060] The structural analysis software combines the user-input questions with relevant knowledge fragments into a structured prompt. The specific format is as follows:

[0061] As an expert in the field of structural analysis software, please answer the questions strictly according to the text provided below:

[0062] [User Issues]

[0063] [User-input questions for structural analysis software]

[0064] [Knowledge Fragment]

[0065] [Related Knowledge Fragment 1]

[0066] [Related Knowledge Fragment 2]

[0067] [Related Knowledge Fragment 3]

[0068]

Answer

[0069] The structured prompts are input into the expert question-and-answer model, which generates and outputs professional answers as the final accurate, professional, and evidence-based responses, which are then returned to the user.

[0070] The implementation method of the AI ​​question-answering system for structural analysis software disclosed in the above embodiments is based on the DeepSeek-7B distillation model and employs LoRA fine-tuning technology. It trains an expert question-answering model for the relevant knowledge domain of the structural analysis software using question-answer pairs generated by the DeepSeek API. Furthermore, it introduces the RAG system to retrieve relevant authoritative knowledge fragments from the question-answering vector database for user-input questions, combines these fragments with the user's question, and inputs them into the expert question-answering model to generate a final accurate, professional, and evidence-based answer. This system can quickly respond to user questions during software use, providing efficient and accurate technical support and intelligent question-answering services, effectively supporting the widespread application of the software.

[0071] The technical solution of this application has been described in conjunction with the preferred embodiments shown in the accompanying drawings. Those skilled in the art should understand that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A method for implementing an artificial intelligence question-answering system for structural analysis software, characterized in that, include: Step 1: Convert the relevant knowledge documents of the structural analysis software into segmented structured text; Step 2: Based on the structured text, invoke the generative artificial intelligence model to generate question-answer pairs and construct a text dataset; Step 3: Using LoRA fine-tuning technology, fine-tune the base model based on the text dataset to train and obtain the LoRA adapter weights; Step 4: Load the base model and LoRA adapter weights to form an expert question-and-answer model focused on the knowledge domain related to structural analysis software; Step 5: Construct a question-and-answer vector database of knowledge related to structural analysis software; Step 6: For the questions input by the user of the structural analysis software, use RAG technology to retrieve relevant knowledge fragments from the question-answer vector database; Step 7: Input the questions and relevant knowledge fragments entered by the user of the structural analysis software into the expert question-and-answer model, so that the expert question-and-answer model can generate and output professional answers.

2. The implementation method of the artificial intelligence question-answering system for structural analysis software according to claim 1, characterized in that, The knowledge documents related to structural analysis software are from multiple sources and are heterogeneous, including official user manuals, technical white papers, and operation guides for structural analysis software, as well as professional textbooks and literature on finite element theory, mechanics of materials, and structural dynamics, in formats including txt, pdf, word, and excel.

3. The implementation method of the artificial intelligence question-answering system for structural analysis software according to claim 2, characterized in that, In step one, the intelligent data extraction tool MinerU is used to extract the text content of relevant knowledge documents of the structural analysis software. The text content is then segmented according to chapters and uniformly converted into segmented Markdown format structured text.

4. The implementation method of the artificial intelligence question-answering system for structural analysis software according to claim 3, characterized in that, In step two, the DeepSeek model API is called to automatically generate question-answer pairs based on structured text, which are then used as seed data to expand and construct a text dataset. The text dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The text was then standardized, including converting it to Simplified Chinese, removing irrelevant symbols, using a tokenizer for word segmentation and encoding, and standardizing the input and output lengths of question-answer pairs.

5. The implementation method of the artificial intelligence question-answering system for structural analysis software according to claim 4, characterized in that, In step three, the DeepSeek-7B distillation model is used as the base model, and LoRA fine-tuning technology is used for training.

6. The implementation method of the artificial intelligence question-answering system for structural analysis software according to claim 5, characterized in that, In step four, the base model and LoRA adapter weights are merged to form an expert question-and-answer model focused on knowledge related to structural analysis software.

7. The implementation method of the artificial intelligence question-answering system for structural analysis software according to claim 6, characterized in that, In step five, a text embedding model is used to convert the knowledge documents related to the structural analysis software into question-answer vector representations, thereby constructing a question-answer vector database. The question-and-answer vector database is updated in real time and supports efficient retrieval; Semantic association extension is performed on question-answer pairs in the question-answer vector database to generate a series of related questions with different wording but similar intent.

8. The implementation method of the artificial intelligence question-answering system for structural analysis software according to claim 7, characterized in that, In step six, for the questions input by the user of the structural analysis software, the RAG system uses a text embedding model to convert the questions into query vectors, calculates the similarity score between the query vectors and the question vectors in the question-answer vector database, and extracts the knowledge fragments corresponding to the answer vectors of the top three question vectors with the highest similarity scores as the most relevant knowledge fragments.

9. The implementation method of the artificial intelligence question-answering system for structural analysis software according to claim 8, characterized in that, In step seven, the questions input by the user of the structural analysis software and related knowledge fragments are combined into structured prompts. These structured prompts are then input into the expert question-and-answer model, which generates and outputs professional answers.

Images