Intelligent knowledge retrieval method and system for nuclear fusion device

Abstract

The invention relates to the technical field of knowledge retrieval, and discloses an intelligent knowledge retrieval method and system for a nuclear fusion device, and the method comprises the steps: building a document database; constructing an interactive AI problem representation framework; the interactive AI problem representation framework comprises a sensing component used for converting query information into a structured embedded vector; the brain component is used for embedding each document in the document database into a high-dimensional vector space to obtain a document vector database consisting of document vectors; based on an embedding vector of query information, embedding the query information into the high-dimensional vector space to generate a query vector; and inputting query information of the user into the interactive AI question representation framework to obtain a comprehensive answer finally output by the interactive AI question representation framework. According to the method, the retrieval enhancement generation technology and the large-scale language model are combined, so that the retrieval result is highly related to the user query, and the answer accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of knowledge retrieval technology, specifically to an intelligent knowledge retrieval method and system for nuclear fusion devices. Background Technology

[0002] As fusion energy research continues to advance, scientists face increasingly complex data processing and analysis challenges. These challenges require not only advanced experimental skills and theoretical knowledge but also powerful data processing and analysis capabilities, which traditional methods are no longer sufficient to address. Against this backdrop, artificial intelligence (AI) technology has become an important tool for solving these problems. How to effectively utilize vast amounts of experimental data and literature records to improve the efficiency and accuracy of nuclear fusion researchers in literature retrieval and knowledge acquisition has become an urgent technical problem to be solved.

[0003] In existing fusion energy research, while artificial intelligence (AI) techniques such as retrieval-enhanced generation (RAG) have been applied to data processing and analysis, large-scale language models (LLMs) have not yet been widely used in the fusion field. The lack of large-scale language model support leads to lower efficiency in data processing and knowledge acquisition in fusion research, making it particularly difficult for novice researchers to quickly grasp the relevant content. Furthermore, existing data processing methods have limited efficiency when handling massive amounts of documents and technical reports, making it difficult to quickly extract valuable information, further hindering research progress. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent knowledge retrieval method and system for nuclear fusion devices. By integrating more document data, it expands the retrieval scope, improves the accuracy of generated answers and contextual understanding capabilities, thereby providing real-time and efficient data processing and analysis support. Specific objectives include:

[0005] Large-scale language models can be used to improve the intelligence of data processing and analysis, and enhance the ability to understand and apply complex data.

[0006] It provides efficient document and technical report processing tools, enabling rapid information extraction and knowledge integration, and supporting researchers to complete experimental tasks efficiently.

[0007] To create an easily scalable platform suitable for different fusion devices, promoting the overall progress of fusion science research. To provide novice fusion researchers with convenient access to knowledge, reducing the learning curve and accelerating research.

[0008] This invention will significantly improve the efficiency of data processing and analysis in fusion energy research, enabling researchers to better address complex data challenges and accelerate the realization of research results.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] A smart knowledge retrieval method for nuclear fusion devices, based on retrieval-enhanced generation technology and interactive artificial intelligence technology, achieves knowledge retrieval and generation, specifically including the following steps:

[0011] Step 1: Establish a document database related to data acquisition, data access and management information, and magnetic balance reconstruction information of the nuclear fusion device;

[0012] Step 2: Construct an interactive AI question representation framework; the interactive AI question representation framework includes:

[0013] A perception component is used to convert query information into structured embedding vectors;

[0014] The brain component is used to embed each document in the document database into a high-dimensional vector space to obtain a document vector database composed of document vectors; based on the embedding vector of query information, query information is embedded into the high-dimensional vector space to generate query vectors;

[0015] Step 3: Input the user's query information into the interactive AI question representation framework. Based on the document vector and the query vector corresponding to the query information, the interactive AI question representation framework obtains the k documents that are closest to the query information. The k documents are used as the prompt context and input into multiple large language models. The multiple candidate answers are integrated to obtain the comprehensive answer finally output by the interactive AI question representation framework, and the comprehensive answer is returned to the user.

[0016] Furthermore, the brain component includes a document vector database, a decision-making unit, and a storage unit;

[0017] Document vector databases are used to embed documents from a document database into a high-dimensional vector space, where each document is represented as a document vector, capturing the semantic information in the document;

[0018] The decision-making unit is used to embed the user's query information into a high-dimensional vector space identical to the document vector database, generating a query vector. It then finds the k documents closest to the query vector in the high-dimensional vector space, inputs these k documents as contextual prompts into multiple large language models, integrates the resulting candidate answers to obtain the final comprehensive answer output by the interactive AI question representation framework, and returns the comprehensive answer to the user. The document vector database and the decision-making unit constitute the retrieval enhancement generative model.

[0019] The storage unit is used to retain historical data of the interactive AI problem representation framework.

[0020] Furthermore, the interactive AI problem representation framework also includes:

[0021] The action component executes tasks in the network environment based on instructions from the brain component. The corresponding operations include four steps: analyzing the network designer's input, extracting information from the knowledge base, transforming the network problem into a mathematical model, and finally presenting the optimization model.

[0022] Environment components are used to encapsulate application scenarios for interactive AI problem representation frameworks.

[0023] An intelligent knowledge retrieval system for nuclear fusion devices includes:

[0024] The database construction module is used to create a document database related to data acquisition information, data access and management information, and magnetic balance reconstruction information of the nuclear fusion device.

[0025] The framework building module is used to construct an interactive AI question representation framework. The interactive AI question representation framework includes: a perception component, used to convert query information into structured embedding vectors; a brain component, used to embed each document in the document database into a high-dimensional vector space to obtain a document vector database composed of document vectors; and an embedding vector based on the query information to embed the query information into the high-dimensional vector space to generate query vectors.

[0026] The query module inputs the user's query information into an interactive AI question representation framework. Based on the document vector and the query vector corresponding to the query information, the interactive AI question representation framework obtains the k documents that are closest to the query information. The k documents are then used as contextual prompts and input into multiple large language models. The multiple candidate answers are integrated to obtain the final comprehensive answer output by the interactive AI question representation framework, and the comprehensive answer is returned to the user.

[0027] The retrieval system of the present invention corresponds to the retrieval method. It is applicable to the preferred technical solutions of the retrieval method and is also applicable to the retrieval system.

[0028] This invention constructs an Intelligent Augmented Intelligence (IAI) knowledge retrieval system by combining Retrieval Enhanced Generation (RAG) technology with Large Scale Language Models (LLMs). Specifically designed for research on nuclear fusion devices (such as EAST), it can efficiently retrieve experimental literature related to nuclear fusion devices and generate accurate and interpretive professional answers based on historical documents. Compared with existing technologies, the beneficial technical effects of this invention are:

[0029] Improving literature retrieval efficiency: By utilizing Retrieval Enhancement Generation (RAG) technology, the retrieval system can quickly retrieve relevant literature from large-scale document databases, significantly reducing researchers' time and effort in information searching; it can also cite historical documents to provide in-depth explanations of professional issues, improving the professionalism and credibility of the answers.

[0030] Improving retrieval accuracy and relevance: By embedding documents into a high-dimensional vector space and using vector similarity matching, the system ensures that retrieval results are highly relevant to user queries, thus improving the accuracy of responses. This significantly enhances the efficiency of nuclear fusion researchers in quickly locating and retrieving the information they need from massive amounts of literature, reducing the time and effort required for traditional literature retrieval.

[0031] It provides intelligent knowledge support; the retrieval system not only provides search results, but also provides in-depth interpretations based on historical documents, helping researchers to better understand and apply professional knowledge, and enhancing the depth and breadth of research.

[0032] Enhanced system adaptability and versatility: The design is flexible and scalable, and the retrieval system is not only applicable to current nuclear fusion devices, but can also be easily applied to future nuclear fusion devices to meet different research needs and future development requirements. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the interactive AI problem representation framework in an embodiment of the present invention;

[0034] Figure 2 This is a flowchart of the retrieval enhancement generation model in an embodiment of the present invention;

[0035] Figure 3 A comparison chart of BLEU score for different models;

[0036] Figure 4 This is a comparison chart of the question-answering delay times for different models. Detailed Implementation

[0037] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0038] This invention aims to develop a knowledge retrieval method and system for nuclear fusion device research based on a combination of Retrieval Enhanced Generation (RAG) technology and Interactive Artificial Intelligence (IAI) technology. The specific steps of the retrieval method are as follows:

[0039] Step 1: Establish a document database:

[0040] First, a document database is established containing information related to data acquisition, access, management, and magnetic balance reconstruction for nuclear fusion devices. In this embodiment, the document database is 285MB in size and contains documents in formats such as PDF and DOCX. These documents provide rich background knowledge and expert insights, offering an important data foundation for subsequent knowledge retrieval and generation.

[0041] Step 2: Design and optimize the knowledge retrieval framework:

[0042] This invention proposes an interactive AI (IAI) problem representation framework, such as... Figure 1 As shown, the framework mainly consists of four components: perception, brain, action, and environment.

[0043] The Perception component is primarily responsible for receiving input data (such as text and speech) and assimilating the information. After receiving the input data, the retrieval system transforms the raw input data into structured embedding vectors. These embedding vectors convert complex input data into a format that interactive AI problem representation frameworks can easily understand.

[0044] The Brain component primarily consists of three parts: a document vector database, a decision-making unit, and a storage unit. The document vector database is used to add new supplementary data and embed all potentially relevant documents into a high-dimensional vector space. Each document is represented as a vector, capturing its key semantic information. The decision-making unit embeds the user's query information into the same vector space as the document vector database, generating a query vector. Then, it finds the k document vectors in the vector space that are closest to the query vector, extracts the documents corresponding to these document vectors, and constructs a prompt context. Finally, the constructed prompt context is input into multiple large language models (LLMs), integrating the candidate answers output by multiple large language models to generate a more comprehensive and accurate answer. The document vector database and the decision-making unit constitute the core Retrieval Augmentation Generation (RAG) model, such as... Figure 2 As shown, this Retrieval Augmentation Generation (RAG) model combines external databases with multiple Large-Scale Language Models (LLMs), significantly improving the performance of language models on specific tasks. This allows the large-scale language models to reference domain-specific information when generating text, thereby improving the accuracy and relevance of the answers. The storage unit primarily retains historical data from the interactive AI question representation framework, enabling it to effectively reference previous memories and strategies during reasoning.

[0045] Action Component: In the interactive AI problem representation framework, the action component performs tasks within the network environment based on instructions from the brain component. Its operation involves four steps: analyzing the network designer's input, extracting information from the knowledge base, transforming the network problem into a mathematical model, and finally presenting the optimized model. The action component ensures that the decisions of the interactive AI problem representation framework can be effectively applied to the actual operation of the network.

[0046] Environment Component: The environment component encapsulates the application scenarios in which the interactive AI problem representation framework operates. It represents the real-world network scenario described by the network designer and forms the foundation of the interactive AI problem representation framework.

[0047] Step 3: Implement knowledge retrieval and generation:

[0048] Using the interactive AI question representation framework described above, the intelligent retrieval method can efficiently achieve knowledge retrieval and generation. The specific process is as follows: Figure 2 As shown:

[0049] Query input: When a user asks a question or makes a query, the system receives and assimilates the query information through the perception component as input data.

[0050] Query vectorization: The retrieval system converts the user's query information into a high-dimensional vector representation for matching in a document vector database.

[0051] Vector database retrieval: In a pre-built document vector database, find the k documents that are most similar to the query vector to ensure the relevance and accuracy of the retrieval results.

[0052] Context building: Extract relevant information from the retrieved documents to build contextual clues, providing necessary background information for large language models.

[0053] Multi-model processing: The constructed prompt context is input into multiple large language models (such as GPT, LLaMA, etc.) for processing to generate multiple candidate answers.

[0054] Results Integration: Integrate candidate answers from multiple large-scale language models to generate a final comprehensive answer, ensuring the completeness and accuracy of the response.

[0055] Return response: Return the generated comprehensive answer to the user, realizing intelligent knowledge retrieval and generation.

[0056] The retrieval system of the present invention corresponds to the retrieval method. It is applicable to the preferred technical solutions of the retrieval method and is also applicable to the retrieval system.

[0057] This invention combines Retrieval Enhanced Generation (RAG) technology with Large-Scale Language Models (LLMs) to construct a large-scale language model specifically for knowledge retrieval and generation for nuclear fusion devices (such as the Advanced Experimental Superconducting Tokamak, EAST). The constructed context is input into multiple large-scale language models (LLMs) to integrate and generate a more comprehensive and accurate answer.

[0058] In a preferred embodiment, the multiple large-scale language models employed may include Chat-GPT and Llama.

[0059] Example 1

[0060] Using common evaluation metrics in natural language processing, including latency, BLEU score, ROUGE score, and METEOR score, the text generation capabilities of the retrieval augmented generation (RAG) model proposed in this invention are compared with those of general models.

[0061] Table 1

[0062]

[0063] As shown in Table 1, although the average response time of the Retrieval Augmented Generation (RAG) model is slightly higher than that of ChatGPT, it exhibits a more stable response speed due to its lower latency standard deviation. Furthermore, the RAG model outperforms ChatGPT on all three scoring metrics, demonstrating better n-gram matching, higher text overlap, and superior semantic and lexical performance. This indicates that the RAG model can provide more accurate and relevant results in generating and understanding natural language, making it a more reliable choice for handling high-quality natural language tasks.

[0064] Example 2

[0065] In this embodiment, BLEU scores and latency metrics are used to roughly evaluate the degree of match between the outputs of the Retrieval Augmentation Generation (RAG) and ChatGPT models and the reference answer, as well as their response times. It is important to note that since this evaluation is not performed within the specific context of a machine translation task, the reference answer is not the only correct answer. This may result in generally low BLEU scores, which is understandable considering the broader application scenarios. Figure 3 As shown, the Retrieval Augmentation (RAG) model has a relatively high BLEU score, indicating that the responses generated by the RAG model are closer to the reference answer. From a response time perspective, as... Figure 4 As shown, the Retrieval Enhancement Generation (RAG) model consistently exhibits a shorter response time than the ChatGPT model.

[0066] Example 3

[0067] Comparison of answers to questions about nuclear fusion devices (such as the EAST device):

[0068] Table 2: Test Question "What is P-EFIT magnetic balance reconstruction?"

[0069]

[0070] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0071] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An intelligent knowledge retrieval method for a nuclear fusion device, characterized in that: Based on the retrieval enhancement generation technology and interactive artificial intelligence technology, knowledge retrieval and generation are realized, which specifically includes the following steps: Step 1: Establish a document database related to data acquisition information, data access and management information, and magnetic balance reconstruction information of the nuclear fusion device; Step 2: Construct an interactive AI problem representation framework; the interactive AI problem representation framework includes: The perception component is used to convert query information into a structured embedding vector; The brain component is used to embed each document in the document database into a high-dimensional vector space to obtain a document vector database composed of document vectors; based on the embedding vector of the query information, the query information is embedded into the high-dimensional vector space to generate a query vector; Step three: input the user's query information into the interactive AI question representation framework. The interactive AI question representation framework obtains the k documents closest to the query information based on the document vector and the query vector corresponding to the query information, inputs the k documents as prompt contexts into multiple large language models, integrates the multiple candidate answers obtained, obtains the comprehensive answer finally output by the interactive AI question representation framework, and returns the comprehensive answer to the user.

2. The intelligent knowledge retrieval method for a nuclear fusion device according to claim 1, characterized in that: The brain component includes a document vector database, a decision-making unit, and a storage unit; The document vector database is used to embed documents in the document database into a high-dimensional vector space. Each document is represented as a document vector to capture the semantic information in the document. The decision-making unit is used to embed the user's query information into the same high-dimensional vector space as the document vector database to generate a query vector; find the k documents closest to the query vector in the high-dimensional vector space, input the k documents into multiple large language models as prompt contexts, integrate the multiple candidate answers obtained, obtain the comprehensive answer finally output by the interactive AI question representation framework, and return the comprehensive answer to the user; the document vector database and the decision-making unit constitute a retrieval enhancement generation model; The storage unit is used to retain historical data of the interactive AI problem representation framework.

3. The intelligent knowledge retrieval method for a nuclear fusion device according to claim 1, characterized in that: The interactive AI problem representation framework also includes: The action component performs tasks in the network environment according to the instructions from the brain component. The corresponding operation includes four steps: analyzing the input of the network designer, extracting information from the knowledge base, converting the network problem into a mathematical model, and finally presenting the optimization model; Environment component, used to encapsulate the application scenarios of the interactive AI problem representation framework.

4. An intelligent knowledge retrieval system for a nuclear fusion device, characterized in that: include: A database construction module is used to establish a document database related to data acquisition information, data access and management information, and magnetic balance reconstruction information of the nuclear fusion device; A framework building module is used to build an interactive AI problem representation framework; the interactive AI problem representation framework includes: a perception component, which is used to convert query information into a structured embedding vector; a brain component, which is used to embed each document in the document database into a high-dimensional vector space to obtain a document vector database composed of document vectors; based on the embedding vector of the query information, the query information is embedded into the high-dimensional vector space to generate a query vector; The query module inputs the user's query information into the interactive AI question representation framework. The interactive AI question representation framework obtains k documents closest to the query information based on the document vector and the query vector corresponding to the query information, inputs the k documents as prompt contexts into multiple large language models, integrates the multiple candidate answers obtained, obtains the comprehensive answer that is finally output by the interactive AI question representation framework, and returns the comprehensive answer to the user.

Images