Spaceflight knowledge extraction method based on generative intelligence

Abstract

The invention relates to an aerospace knowledge extraction method based on generative intelligence, and belongs to the technical field of aerospace reliability. The method comprises the following steps: carrying out cross-modal and automatic knowledge extraction and entity association on a spaceflight information system by utilizing a generative intelligent technology, constructing a spaceflight knowledge graph, and on the basis, realizing information retrieval, capability optimization and intelligent interaction, assisting a user to quickly obtain required knowledge, constructing a thinking system and providing a special policy question and answer function; the problems that the semantic matching precision of professional documents in the aerospace vertical field is low, the fusion analysis effect of multi-modal data is poor, the potential tapping of text data is insufficient, and the utilization degree is low are solved, and an aerospace knowledge extraction technology based on generative intelligence is provided.

Description

Technical Field

[0001] The present invention belongs to the field of aerospace reliability technology and relates to an aerospace knowledge extraction method based on generative intelligence. Background Art

[0002] Aerospace covers highly specialized knowledge and a huge system of technical documents. It builds associations between knowledge in different fields, assists users in semantically matching full-text retrieval of high-quality, multimodal information, recommends relevant content based on historical information, and provides users and designers with information retrieval, knowledge construction, technical support, operation and maintenance, capacity training and other value-added services to achieve cost reduction and efficiency improvement, which has important practical significance and value. At present, there are some problems with existing document parsing. First, most existing technologies rely on rule matching or template matching methods, which often cannot handle long texts or complex documents, resulting in low accuracy and relevance of responses. Existing technologies lack an in-depth understanding of user intent, resulting in responses that may deviate from the actual needs of users.

[0003] Based on this, how to solve the problems of poor document retrieval effect, low potential, low accuracy and relevance of responses has become a technical problem that needs to be solved urgently. Summary of the invention

[0004] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art, propose a method for extracting aerospace knowledge based on generative intelligence, solve the problems of low semantic matching accuracy of professional documents in the aerospace vertical field, poor fusion analysis effect of multimodal data, insufficient potential mining of text data, and low utilization, and provide an aerospace knowledge extraction technology based on generative intelligence.

[0005] The solution of the present invention is:

[0006] A method for extracting aerospace knowledge based on generative intelligence, comprising:

[0007] Use Python to load data of different modalities into memory;

[0008] Preprocess the multimodal data loaded into the memory; obtain the preprocessed data;

[0009] Perform vectorization on the preprocessed data;

[0010] Use the K-means algorithm to cluster the vectorized data, group similar entities and multi-modal data into one category, and realize semantic association between different types of data;

[0011] Use big model technology to extract entities, generate relationships, and build knowledge graphs by grading and classifying;

[0012] Using big models and reasoning engine technology, knowledge can be updated by analyzing entity features and the relationships between entities.

[0013] Receive user request information, including image, text and other modal information;

[0014] Using multimodal vector coding technology to feature encode the user's request information to form an information vector;

[0015] Retrieve entities related to the user information vector from the configured aerospace knowledge graph, and combine keyword retrieval and fuzzy search to realize information retrieval in the knowledge graph and form preliminary retrieval results;

[0016] Based on a large number of preliminary search results, a refined relevance queue is obtained;

[0017] Filter the re-sorted information to obtain the final search results;

[0018] Based on user needs and combined with the retrieved information, a preliminary answer is formed;

[0019] Users provide feedback based on the answer results;

[0020] After receiving user feedback, the algorithm will update the knowledge graph content.

[0021] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, data of different modalities include image data and text data.

[0022] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, the preprocessing content includes image denoising, image enhancement, text segmentation, summary generation, and entity extraction.

[0023] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, the specific method of vectorization processing is:

[0024] Using the embedded Embedding model, data of different modes are uniformly represented in the vector space to achieve vectorized processing of various types of data.

[0025] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, the knowledge graph is divided into two layers. The bottom layer is the original data, which is associated together through vector clustering; the upper layer is the key entity, which is obtained by extracting the key features of the original data from the large model as the label of this part of the data.

[0026] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, the method for obtaining a refined correlation queue is:

[0027] For a large number of preliminary search results, a re-ranking model is used to calculate the relevance of the preliminary search results and the user request information in turn to obtain a refined relevance queue.

[0028] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, the method for screening the reordered information is:

[0029] The reordered information can be filtered by setting a relevance threshold or the number of related information.

[0030] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, the method for forming the initial answer result is:

[0031] Based on user needs and combined with the retrieved information, multimodal large model technology is used to extract key entities and potential knowledge in the original data to form a preliminary version of the answer.

[0032] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, the user's feedback based on the answer result is:

[0033] When the answer meets the user's expectations and the content is correct, the user marks the result, indicating that he or she expects to generate similar answers in the future; when the user's question cannot be answered due to lack of knowledge, the user further describes the question and completes the missing information to obtain the correct answer.

[0034] In the above-mentioned aerospace knowledge extraction method based on generative intelligence, knowledge evolution and capability improvement are achieved by updating the knowledge graph content, supplementing new knowledge and upgrading the knowledge retrieval logic.

[0035] The beneficial effects of the present invention compared with the prior art are:

[0036] (1) The present invention constructs an aerospace knowledge graph by preprocessing the collected initial multimodal information and extracting entities and generating relationships;

[0037] (2) Based on user requests, the present invention uses hybrid retrieval technology to associate relevant information in the knowledge graph to answer the basic knowledge of content generation; and uses hybrid retrieval technology to achieve on-demand matching and intelligent association of user input;

[0038] (3) The present invention combines user feedback to realize multiple rounds of question and answer requests by users. After the interaction is completed, the knowledge graph and retrieval algorithm are updated to realize knowledge evolution and continuous updating of the fault tree. The large model technology is used to realize multiple rounds of knowledge question and answer and knowledge graph status update. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1The present invention is a flow chart for extracting aerospace knowledge. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with the embodiments.

[0041] The present invention provides a method for extracting aerospace knowledge based on generative intelligence, which belongs to the field of aerospace reliability technology. Generative intelligence technology is used to perform cross-modal and automated knowledge extraction and entity association on the aerospace information system, to construct an aerospace knowledge graph, and on this basis, to realize information retrieval, capability optimization, and intelligent interaction, to assist users in quickly acquiring required knowledge, building a thinking system, and providing a special question-and-answer function. The present invention solves the problems of low semantic matching accuracy of professional documents in the aerospace vertical field, poor fusion analysis effect of multimodal data, insufficient potential mining of text data, and low utilization, and provides an aerospace knowledge extraction technology based on generative intelligence.

[0042] Aerospace knowledge extraction methods based on generative intelligence, such as Figure 1 As shown, the specific steps include:

[0043] (1) Multimodal vector encoding and knowledge graph construction technology aligns different modal data such as images and texts in the vector space to achieve effective association of the same entity and different modal information. The specific steps include:

[0044] Step 1: Information loading: Use Python to load data of different modalities into memory, such as image data and text data;

[0045] Step 2: Data processing: pre-processing the multimodal data loaded into the memory, such as image denoising, image enhancement, text segmentation, summary generation, and entity extraction;

[0046] Step 3: Data vectorization: Use the embedding model to uniformly represent data of different modalities in the vector space and vectorize various types of data.

[0047] Step 4: Vector clustering: Use the K-means algorithm to cluster the vectorized data, group similar entities and multi-modal data into one category, and realize semantic association between different types of data.

[0048] Step 5: Knowledge carrying, using big model technology to extract entities, generate relationships, and construct knowledge graphs by grading and classification. The knowledge graph is divided into two layers. The bottom layer is the original data, which is associated together through vector clustering; the upper layer is the key entity, which is obtained by extracting the key features of the original data from the big model as the label of this part of the data;

[0049] Step 6: Knowledge evolution, using big models and reasoning engine technology to update knowledge by analyzing entity features and the relationships between entities.

[0050] (2) Multimodal information hybrid retrieval technology based on semantic matching. Users upload natural language expressions, document materials, pictures and other information, and use the constructed aerospace knowledge graph combined with multimodal vector coding technology to realize the vectorization of user questions. On this basis, semantic matching, keyword matching, fuzzy search and other technologies are used to realize hybrid retrieval and on-demand association of user questions. The specific steps include the following:

[0051] Step 1: User input, receiving user request information, including images, text and other modal information;

[0052] Step 2: Request vectorization, using multimodal vector coding technology to feature encode the request information uploaded by the user to form an information vector;

[0053] Step 3: Hybrid retrieval: retrieve entities related to the user information vector from the configured aerospace knowledge graph, and combine keyword retrieval and fuzzy search to realize information retrieval in the knowledge graph and form preliminary retrieval results;

[0054] Step 4: Information screening: After obtaining a large number of preliminary search results, a re-ranking model is used to calculate the relevance of the preliminary search results and user requests in turn to obtain a refined relevance queue.

[0055] Step 5: Information acquisition, by setting the relevance threshold or the number of relevant information, the reordered information is screened to obtain the final retrieval results.

[0056] (3) Intelligent question-answering and knowledge evolution technology based on multimodal big models. After retrieving key information from the aerospace knowledge graph based on user input information, the multimodal big model is used to implement intelligent question-answering. When there are questions such as missing retrieval information or logical errors in the answer, the user is supported to re-enter the required relevant knowledge in the form of multiple rounds of dialogue to achieve knowledge evolution and graph update.

[0057] Step 1: Based on user needs and combined with the relevant information retrieved, use multimodal large model technology to extract key entities and potential knowledge in the original data to form a preliminary answer result;

[0058] Step 2: Users provide feedback based on the answer results. When the answer results meet user expectations and the content is correct, users can mark the results to indicate that they expect to generate similar answers in the future. When the user's question cannot be answered due to lack of knowledge, users can further describe the question and complete the missing information to obtain the correct answer.

[0059] Step 3: After receiving user feedback, the algorithm will update the knowledge graph content, add new knowledge and upgrade the knowledge retrieval logic to achieve knowledge evolution and capability improvement.

[0060] The present invention also specifically designs the following contents:

[0061] Multimodal vector encoding and knowledge graph construction technology

[0062] Step 1: Information loading: Read information of various modalities such as documents and images through Langchain’s loader, and pre-process the information. For example, separate the text, images and other data in the document and store them in different categories.

[0063] The Langchain loader used in this step can support reading data in multiple modes, such as pdf, word, html, png, etc., and preprocess the original data on this basis, such as extracting text, images, and removing irrelevant characters.

[0064] Step 2: Data processing: Process the obtained text, image and other modal data, including document segmentation, summary generation, etc.; perform data enhancement such as denoising, rotation, and scaling on the image to form processed data.

[0065] In this step, Langchain is used to automate text processing, and OpenCV is used to automate image processing.

[0066] Step 3: Data vectorization: Use the embedding model, take the processed data as input, and use the vector as output to obtain data features.

[0067] The embedding models used in this step include text vectorization (Text Embedding) model and image vectorization model (Image Embedding). The text vectorization model can convert text data (words, sentences, documents) into vectors, which can be implemented using models such as Word2Vec and BERT; image vectorization can convert image data into vectors (png, jgp, etc.), which is often implemented using models such as CNN and Transformer.

[0068] Step 4: Vector clustering: Use a clustering algorithm to cluster the multiple vectors obtained after vectorizing the data, with each cluster corresponding to several entity information.

[0069] In this step, the K-means algorithm is used to cluster data in the vector space, assuming that data in the same category represent the same entity information.

[0070] Step 5: Knowledge carrying: Based on the data categories formed in the previous step, associate the original data, extract entity features on this basis, and build a knowledge graph.

[0071] The knowledge graph constructed in this step is divided into two layers. The first layer is the original data, which is clustered and associated by the vectors in the previous step; the second layer is high-dimensional features, which uses a large model to extract key entities and associations in the data and associate them with the first layer.

[0072] Step 6: Knowledge evolution: Based on the knowledge graph that has been built, use the reasoning engine and big model technology to verify the correctness of the relationship between entities, and build new associations based on the original relationships to achieve knowledge evolution and logic optimization.

[0073] The present invention pre-processes the collected initial multimodal information, constructs an aerospace knowledge graph through entity extraction and relationship generation; then based on user requests, the hybrid retrieval technology is used to associate relevant information in the knowledge graph to answer the basic knowledge of content generation; then combined with user feedback, multiple rounds of question and answer are implemented for user requests. After the interaction is completed, the knowledge graph and retrieval algorithm are updated to realize knowledge evolution.

[0074] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for extracting aerospace knowledge based on generative intelligence, characterized by: include: Use Python to load data of different modes into memory; Preprocess the multimodal data loaded into memory; Obtain preprocessed data; Perform vectorization on the preprocessed data; Use the K-means algorithm to cluster the vectorized data, group similar entities and multi-modal data into one category, and realize semantic association between different types of data. Use big model technology to extract entities, generate relationships, and build knowledge graphs by grading and classifying; Using big models and reasoning engine technology, knowledge can be updated by analyzing entity features and the relationships between entities. Receive user request information, including image, text and other modal information; Using multimodal vector coding technology to feature encode the user's request information to form an information vector; Retrieve entities related to the user information vector from the configured aerospace knowledge graph, and combine keyword retrieval and fuzzy search to realize information retrieval in the knowledge graph and form preliminary retrieval results; Based on a large number of preliminary search results, a refined relevance queue is obtained; Filter the re-sorted information to obtain the final search results; Based on user needs and combined with the retrieved information, a preliminary version of the answer is formed; Users provide feedback based on the answer results; After receiving user feedback, the algorithm will update the knowledge graph content.

2. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: Data of different modalities include image data and text data.

3. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: Preprocessing includes image denoising, image enhancement, text segmentation, summary generation, and entity extraction.

4. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: The specific method of vectorization processing is: Using the embedded Embedding model, data of different modes are uniformly represented in the vector space to achieve vectorized processing of various types of data.

5. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: The knowledge graph is divided into two layers. The bottom layer is the original data, which is associated together through vector clustering; the upper layer is the key entity, which is obtained by extracting the key features of the original data from the large model as the label of this part of the data.

6. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: The method to obtain a refined relevance queue is: For a large number of preliminary search results, a re-ranking model is used to calculate the relevance of the preliminary search results and the user request information in turn to obtain a refined relevance queue.

7. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: The method for filtering the reordered information is: The reordered information can be filtered by setting a relevance threshold or the number of related information.

8. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: The method for forming the initial answer results is: Based on user needs and combined with the retrieved information, multimodal large model technology is used to extract key entities and potential knowledge in the original data to form a preliminary version of the answer.

9. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: The user's feedback based on the answer results is as follows: When the answer meets the user's expectations and the content is correct, the user marks the result, indicating that he or she expects to generate similar answers in the future; when the user's question cannot be answered due to lack of knowledge, the user further describes the question and completes the missing information to obtain the correct answer.

10. The aerospace knowledge extraction method based on generative intelligence according to claim 1 is characterized by: By updating the content of the knowledge graph, adding new knowledge and upgrading the knowledge retrieval logic, knowledge evolution and capability improvement can be achieved.

Images