Aerospace manufacturing process knowledge retrieval and enhancement generation method based on dynamic graph neural network and large language model cooperation

Abstract

The present application relates to a kind of aerospace manufacturing process knowledge retrieval enhancement generation method based on dynamic graph neural network and large language model cooperation, including the construction containing aerospace manufacturing process information Process document set, based on large language model Construction knowledge graph and associated index matrix;Construction is based on query dependence Dynamic graph neural network model, predict each entity based on the matching score of user query question;Based on matching score and associated index matrix, the further correlation score of process document and user query question is calculated, and the process document highly relevant with user query question is returned;With the user query question of highly relevant process document and user query question input large language model, generate the aerospace manufacturing process knowledge answer matched with user query question.The present application can realize the complex correlation retrieval across document, across knowledge graph, effectively improve the knowledge utilization rate and reasoning accuracy of large language model in aerospace manufacturing field.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and aerospace manufacturing technology, and in particular to an aerospace manufacturing process knowledge retrieval enhancement generation method based on dynamic graph neural network and large language model collaboration. BACKGROUND

[0002] Aerospace products have the characteristics of high single-piece value, high customization degree, long manufacturing process chain, etc., and high requirements for the accuracy, reliability and traceability of process schemes. A large amount of process design and optimization work relies on the experience accumulation of experienced engineers, but these tacit knowledge is often difficult to systematize inheritance and reuse. Large language models (LLM) have strong natural language understanding and generation capabilities, which can make tacit knowledge explicit and build a shareable and trainable knowledge system, thereby alleviating the technical gap caused by talent loss to some extent. Therefore, the large language model research for aerospace manufacturing processes has important engineering value and strategic significance.

[0003] Although the current large language model performs outstandingly in reasoning ability, it still has limitations in processing professional knowledge in specific fields (especially aerospace manufacturing process knowledge), and it is difficult to directly and efficiently utilize domain knowledge outside the pre-training corpus. To make up for this deficiency, retrieval-augmented generation (RAG) has become a widely concerned solution, the core idea of which is to introduce external documents related to the task in the generation process to inject new knowledge into the static large language model. However, existing RAG methods usually retrieve documents independently, making it difficult to capture complex relationships between knowledge fragments across documents, thereby restricting the performance of large language models in multi-hop reasoning and cross-document knowledge integration tasks. Recent research shows that introducing graph structure information into the text indexing and retrieval process can achieve efficient retrieval of entities and their relationships, thereby effectively enhancing the accuracy and relevance of the generation results. However, such methods mostly rely on training on specific graph structures, lacking the ability to migrate to new manufacturing process knowledge graphs. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a retrieval-augmented generation method for aerospace manufacturing process knowledge based on dynamic graph neural network and large language model collaboration, which solves the problems of traditional retrieval-augmented generation RAG methods that are difficult to capture complex relationships between knowledge fragments across documents and lack the ability to migrate to new manufacturing process knowledge graphs.

[0005] To solve the above technical problems, the present application provides the following technical scheme: a retrieval-augmented generation method for aerospace manufacturing process knowledge based on dynamic graph neural network and large language model collaboration, comprising the following steps:

[0006] S1, collect and construct a process document set containing aerospace manufacturing process information, construct a knowledge graph based on a large language model, and establish a correlation index matrix corresponding to the process document set;

[0007] S2, construct a dynamic graph neural network model based on query dependence and optimize its parameters to predict the matching score of each entity in the knowledge graph based on the user query question, to assist in locating the process documents highly related to the user query question;

[0008] S3, further calculate the relevance score of the process document and the user query question based on the matching score and the correlation index matrix, and return the top process documents based on the relevance score in descending order of score as the process documents highly related to the user query question;

[0009] S4, input the process documents highly related to the user query question and the user query question into the large language model to generate the aerospace manufacturing process knowledge answer matched with the user query question.

[0010] Further, in step S1, the specific process includes the following steps:

[0011] S11, collect a set of process documents related to aerospace manufacturing process information, convert the collected process documents into a processable pure text format, and perform segmentation and sentence processing to obtain a process document set ;

[0012] S12, use a large language model as an information extraction engine to parse each process document in the process document set , identify the entity set and relationship set in the field of aerospace manufacturing, and organize and construct a triple set ; wherein 、 represent the th , the th entity, represents the relationship between the th entity and the th entity;

[0013] S13, use a pre-trained dense vector encoding model to encode each entity and each relationship into entity vectors and relationship vectors, respectively;

[0014] S14, calculate the cosine similarity between entity vectors; if the cosine similarity of two entity vectors exceeds the set threshold, consider the two entities corresponding to them as similar entities; for each pair of similar entities, add additional relationships to obtain an additional set of triples , thereby enriching the connectivity of the knowledge graph;

[0015] S15, final knowledge graph defined as , wherein , represent two entities, represent the relationship between them;

[0016] S16, define a binary matrix, whose row index represents the entity index, and the column index represents the process document index; traverse each process document in the set of process documents , find all entities contained in it and set the elements in the corresponding positions in the binary matrix to 1, forming the association index matrix , which is used to realize the rapid positioning of process documents.

[0017] Further, in step S2, the specific process includes the following steps:

[0018] S21, initialize the user query question, entities and relationships in the knowledge graph using a pre-trained dense vector encoding model to obtain query vectors , entity node feature set , relationship feature set ;

[0019] S22, the dynamic graph neural network model based on query dependency includes a layer Bellman-Ford message passing layer and a final output layer for fusing query information;

[0020] S23, input the query vector , the entity node feature set , the relationship feature set to the dynamic graph neural network model based on query dependency, and after passing through a layer Bellman-Ford message passing layer, obtain the entity vector set ;

[0021] S24, concatenate the query vector after each entity node feature in the entity vector set , and input it into the output layer to obtain the matching score of the entity and the current user query question.

[0022] Furthermore, the Bellman-Ford-style messaging layer primarily updates the characteristics and relationship characteristics of each entity node based on the following formula:

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] in, Indicates the first The first in the layer Each entity node has characteristics, and , Indicates the first The first in the layer The entity and the first The relational characteristics of each entity, and , ⊙ represents element-wise multiplication. Indicates the first The entity to the first The information transmitted by an entity These are attention weights, expressed based on the query vector. , No. The entity relative to the first The importance of other entities in the neighborhood of an entity. This indicates the concatenation of feature dimensions. and These represent the weights of the two linear layers that process the query vector and entity node features, respectively. In knowledge graph Middle and the first A set of indexes of all entities that are related to a given entity. This represents a perceptron consisting of two fully connected layers; the entity node features are output from the last layer of the network. Construct entity vector set .

[0028] Furthermore, the output layer includes a perceptron consisting of a fully connected layer and a sigmoid activation layer, which integrates query information from the query vector into entity node features.

[0029] Furthermore, the parameters of the query-dependent dynamic graph neural network model are optimized by maximizing the likelihood probability of query-related entities to construct the loss function. Defined as:

[0030] ;

[0031] wherein denotes a set of query-related entities, denotes a set of query-unrelated entities, denotes a matching score predicted by a query-dependent dynamic graph neural network model for a query-related entity , denotes a matching score predicted by a query-dependent dynamic graph neural network model for a query-unrelated entity .

[0032] Further, all entities in the knowledge graph that coincide with the entity in the user query question and all entities connected to the entity by a relationship constitute the set of query-related entities ; the remaining entities in the knowledge graph constitute the set of query-unrelated entities .

[0033] Further, in step S3, the specific process includes the following steps:

[0034] S31, after obtaining the matching score of each entity in the knowledge graph based on the user query question , the relevance score of each process document in the process document set to the user query question is calculated based on the association index matrix as follows:

[0035] ;

[0036] S32, sort the data in the relevance score from large to small, take the index number corresponding to the first data, and find the corresponding process document in the process document set based on the index number, which is the process document highly relevant to the user query question.

[0037] Further, step S4 specifically includes: including the process document highly relevant to the user query question into the context input of the large language model, combining the user query question to generate accurate and detailed aerospace manufacturing process knowledge answers to support intelligent question answering and technical guidance in aerospace manufacturing scenarios.

[0038] By the above technical solution, the present application provides an aerospace manufacturing process knowledge retrieval enhancement generation method based on the cooperation of dynamic graph neural network and large language model, which has at least the following beneficial effects:

[0039] ​(1) The application accurately calculates the correlation score by matching the score predicted by the dynamic graph neural network model based on query dependence. This process can dynamically adjust the message passing process according to the specific characteristics and semantic information of the user query problem, and can locate the most relevant information on different knowledge graphs, realizing complex association retrieval and multi-hop reasoning across documents and knowledge graphs.

[0040] (2) The dynamic graph neural network model based on query dependence constructed by the application can be directly applied to different query scenarios and various aerospace manufacturing process knowledge graphs. The model can consider the semantic features and structural information of the aerospace manufacturing process knowledge graph at the same time, thereby learning the complex association relationship between the query and the entity.

[0041] (3) The application captures the complex relationship between the query, the document and the aerospace manufacturing knowledge graph in a unified and transferable representation space, identifies the relevant entities in the knowledge graph, quickly locates the corresponding process document, and efficiently incorporates these key information into the context of the large language model, thereby generating high-accuracy aerospace manufacturing process knowledge answers, effectively improving the knowledge utilization rate and reasoning accuracy of the large language model in the field of aerospace manufacturing. The method is suitable for process design, optimization and decision support and other scenes. BRIEF DESCRIPTION OF DRAWINGS

[0042] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0043] Figure 1 The flowchart of the aerospace manufacturing process knowledge retrieval enhancement generation method based on the dynamic graph neural network and the large language model cooperation of the application;

[0044] Figure 2 The aerospace manufacturing process knowledge graph structure diagram in the embodiment of the application;

[0045] Figure 3 The first Bellman-Ford type message passing layer diagram of the dynamic graph neural network model based on query dependence constructed in the embodiment of the application;

[0046] Figure 4 The aerospace manufacturing process knowledge retrieval enhancement generation paradigm diagram based on the dynamic graph neural network and the large language model cooperation in the embodiment of the application. DETAILED DESCRIPTION

[0047] ​In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments. Thus, the implementation process of how to apply technical means to solve technical problems and achieve technical effects can be fully understood and implemented.

[0048] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment methods can be completed by programs instructing related hardware, therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer usable program codes.

[0049] Please refer to Figures 1-4 , which shows a specific embodiment of the present embodiment. The present embodiment constructs a knowledge graph and an association index matrix by combining a large language model, locates a process document highly relevant to a user query problem by using a constructed dynamic graph neural network model based on query dependence, and generates an aerospace manufacturing process knowledge answer matching the user query problem again by using a large language model. The present embodiment can realize complex association retrieval and multi-hop reasoning across documents and knowledge graphs, effectively improves the knowledge utilization rate and reasoning accuracy of a large language model in the field of aerospace manufacturing, and is suitable for scenes such as process design, optimization and decision support.

[0050] Please refer to Figure 1 , which shows a specific embodiment of the present embodiment. The present embodiment proposes an aerospace manufacturing process knowledge retrieval enhancement generation method based on the cooperation of a dynamic graph neural network and a large language model, which includes the following steps:

[0051] S1, collect and construct a process document set containing aerospace manufacturing process information, accurately extract entity, relationship and other knowledge elements from the process document based on a large language model, construct a knowledge graph, and establish an association index matrix corresponding to the process document set;

[0052] As a preferred embodiment of step S1, the specific process includes the following steps:

[0053] S11, collect a group of process documents related to aerospace manufacturing process information, convert the collected process documents into a processable pure text format (such as txt format), and perform segmentation and sentence processing to obtain a process document set ;

[0054] S12, use a large language model (such as: OpenBMB / MiniCPM-2.6) as an information extraction engine to analyze the process document set Each process document in the set of process documents, identify a set of entities in the field of aerospace manufacturing (e.g., part name, processing technology, material, equipment model, etc.) and a set of relationships (e.g., "use", "connect", "processed by", etc.), organize a set of triples ; wherein , represent the first , the first entity, represent the relationship between the first entity and the first entity; Figure 2 The structure diagram of the aerospace manufacturing process knowledge graph is shown. For process documents 1-5 in the figure, the processes include welding, milling, heat treatment, surface engineering, and additive manufacturing. The materials include titanium alloy, aluminum alloy, magnesium alloy, high-temperature alloy, carbon fiber composite material, glass fiber composite material, ceramic matrix composite material, metal matrix composite material, and high-strength steel.

[0055] S13, use a pre-trained dense vector encoding model (e.g., Sentence-BERT) to encode each entity and each relationship into entity vectors and relationship vectors, respectively, to form a set of entity vectors and a set of relationship vectors.

[0056] S14, calculate the cosine similarity between entity vectors; if the cosine similarity of two entity vectors exceeds a set threshold, consider that the two entities corresponding to them are the same entity or strongly semantically related entities, defined as similar entities; for each pair of similar entities, add additional relationships such as "synonymous with" and "alias" to obtain an additional set of triples , thereby enriching the connectivity of the knowledge graph; in this embodiment, the threshold is 0.85;

[0057] S15, the final knowledge graph is defined as , wherein , represent two entities, represent the relationship between them; store the knowledge graph to a graph database to support subsequent retrieval and reasoning of dynamic graph neural networks;

[0058] S16, define a binary matrix whose row index represents the entity index and whose column index represents the process document index; traverse each process document in the set of process documents , find all entities contained therein and set the elements in the corresponding positions in the binary matrix to 1, that is: represents "entity The documents appearing in the document form an association index matrix for realizing fast positioning of process documents; the association index matrix is stored as a high-efficiency queryable data structure (such as a sparse matrix, a hash table).

[0059] S2, a dynamic graph neural network model based on query dependence is constructed, which can dynamically adjust the weights of nodes and edges in the graph structure according to the user input query question, to adapt to the semantic characteristics of different queries; the model is optimized in parameters, to predict the matching score of each entity in the knowledge graph based on the user query question, to assist in positioning the process documents highly related to the user query question;

[0060] As a preferred embodiment of step S2, the specific process includes the following steps:

[0061] S21, the user query question (natural language question / question in retrieval task), entity and relationship in the knowledge graph are respectively initialized by using a pre-trained dense vector encoding model (such as: Sentence-BERT), to obtain query vector , entity node feature set , and relationship feature set ;

[0062] S22, the dynamic graph neural network model based on query dependence includes a layer of Bellman-Ford type message passing layer and a final output layer for fusing query information;

[0063] S23, the query vector , the entity node feature set , and the relationship feature set are input into the dynamic graph neural network model based on query dependence, and after passing through a layer of Bellman-Ford type message passing layer, an entity vector set is obtained.

[0064] S24, each entity node feature in the entity vector set is concatenated with the query vector and sent into the output layer, to obtain the matching score of the entity and the current user query question.

[0065] More specifically, the Bellman-Ford type message passing layer mainly updates each entity node feature and relationship feature based on the following formula:

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] wherein, represents the th entity node feature in the th layer, and , represents the th relation feature between the th entity and the th entity in the th layer, and represents the information transmitted from the th entity to the th entity, is the attention weight, which is expressed based on the query vector , the importance of the th entity relative to other entities in the neighborhood of the th entity, represents the concatenation of feature dimensions, and respectively represent two linear layer weights for processing the query vector and the entity node feature, represents the index set of all entities having a relationship with the th entity in the knowledge graph , represents a perceptron composed of two fully connected layers; the entity node features output by the last layer of the network constitute the entity vector set . Figure 3 The first Bellman-Ford formula message passing layer of the dynamic graph neural network model based on query dependence is shown.

[0071] More specifically, the output layer includes a perceptron composed of a fully connected layer and a Sigmoid activation layer, which integrates query information from the query vector into the entity node feature.

[0072] More specifically, the dynamic graph neural network model based on query dependence optimizes its parameters by maximizing the likelihood probability of the query-related entities, and constructs a loss function defined as:

[0073] ;​

[0074] In the formula This indicates a query for a set of related entities. This indicates a query for a set of irrelevant entities. This indicates that for querying related entities Match scores predicted by a query-dependency-based dynamic graph neural network model. This indicates that the query is for irrelevant entities. The matching score is predicted by a query-dependent dynamic graph neural network model. In this embodiment, the AdamW optimizer is used to update network parameters during the network training process, and the learning rate is set to 5.0e-4.

[0075] More specifically, iterating through user query questions Entity set in Retrieve entity sets from knowledge graphs , All entities Overlapping entities constitute a set of overlapping entities. , Entities in And all of the above Entities that are connected by relationships (referring to the relation set in a knowledge graph) constitute the query-related entity set. The remaining entities, namely The query-irrelevant entity set constitutes the set of entities. .

[0076] In this embodiment, the constructed query-dependent dynamic graph neural network model can be directly applied to different queries and aerospace manufacturing process knowledge graphs. This model can simultaneously consider the semantic features and structural information of the aerospace manufacturing process knowledge graph, thereby learning the complex relationships between queries and entities. All entities in the knowledge graph related to users...

[0077] S3. Based on the matching score and the associated index matrix, further calculate the relevance score between the process documents and the user's query, and return the previous documents in descending order of relevance score. Each process document is a process document that is highly relevant to the user's query question;

[0078] As a preferred embodiment of step S3, the specific process includes the following steps:

[0079] S31, after obtaining the knowledge graph Each entity in the table is matched based on a score for the user's query. Then, based on the association index matrix Calculation process document collection The relevance score of each process document to the user's query question. As follows:

[0080] ;

[0081] S32, the relevance score is calculated based on the matching score predicted by the dynamic graph neural network model based on query dependency The data in the data are sorted in descending order of value, and the index numbers corresponding to the first The index numbers are used to find the corresponding process documents in the process document set, which are the process documents highly related to the user query question .

[0082] In this embodiment, the relevance score is calculated based on the matching score predicted by the dynamic graph neural network model based on query dependency, which can dynamically adjust the message passing process according to the user query question, locate the most relevant information on different knowledge graphs, and realize complex association retrieval and multi-hop reasoning across documents and knowledge graphs.

[0083] S4, input the filtered process documents highly related to the user query question and the user query question into the large language model to generate highly matched, accurate and professional aerospace manufacturing process knowledge answers;

[0084] As a preferred embodiment of step S4, the specific process includes the following steps:

[0085] The process documents highly related to the user query question are included in the context input of the large language model LMM, combined with the user query question , to generate highly matched, accurate and professional aerospace manufacturing process knowledge answers to support intelligent question answering and technical guidance in aerospace manufacturing scenarios, that is:

[0086] ;

[0087] Figure 4 This embodiment shows the aerospace manufacturing process knowledge retrieval enhancement generation paradigm based on dynamic graph neural network. In this embodiment, the dynamic graph neural network model based on query dependency receives the aerospace manufacturing process question proposed by the user, constructs the aerospace manufacturing process knowledge graph, sorts the aerospace manufacturing process documents and locates the high correlation process documents, and then sends the aerospace manufacturing process question proposed by the user and the located high correlation process documents into the large language model LMM to obtain the aerospace manufacturing process reply.

[0088] In the embodiment, by capturing the complex relationship between the query, the document and the aerospace manufacturing knowledge graph in a unified and migratable representation space, identifying related entities in the knowledge graph, quickly locating the corresponding process document, and incorporating it into the context of the large language model, a high-accuracy aerospace manufacturing process knowledge answer is generated, effectively improving the knowledge utilization rate and reasoning accuracy of the large language model in the aerospace manufacturing field, and is suitable for process design, optimization and decision support scenarios.

[0089] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0090] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a list of executable instructions for implementing logic functions, which can be specifically embodied in any computer readable medium for use by an instruction execution system, device or apparatus, such as a computer-based system, a system including a processor or other system that can fetch and execute instructions from the instruction execution system, device or apparatus, or in conjunction with these instructions.

[0091] The above embodiments have been described in detail, and the principles and embodiments of the present application have been described herein by applying specific examples; the above embodiment descriptions are only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific embodiments and application scope will be changed; in view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. An aerospace manufacturing process knowledge retrieval and enhancement generation method based on dynamic graph neural network and large language model collaboration, characterized in that, Comprise the following steps: S1, collect and construct a process document set containing aerospace manufacturing process information, construct a knowledge graph based on a large language model, and establish a correlation index matrix corresponding to the process document set; The specific process comprises the following steps: S11, collect a set of process documents related to aerospace manufacturing process information, convert the collected process documents into a processable pure text format, and perform segmentation and sentence processing to obtain a process document set ; S12, using a large language model as an information extraction engine, parsing each process document in the process document set to identify a set of entities in the aerospace manufacturing field , a set of relationships , organize a set of triples , represent the first , entity, represent the relationship between the first entity and the second entity S13, using a pre-trained dense vector encoding model to encode each entity and each relationship into entity vectors and relationship vectors respectively; S14、calculate the cosine similarity between entity vectors; if the cosine similarity between two entity vectors exceeds a set threshold, consider the two entities corresponding to the two entity vectors as similar entities; for each pair of similar entities, add an additional relationship to obtain an additional set of triples , thereby enriching the connectivity of the knowledge graph; S15, final knowledge graph defined as wherein , denotes two entities, denotes the relationship between them; S16, define a binary matrix, whose row index represents entity index and column index represents process document index; traverse each process document in the process document set, find all entities contained in the process document and set the element in the corresponding position in the binary matrix to 1, form the association index matrix for realizing fast positioning of the process document;​ S2, construct a dynamic graph neural network model based on query dependence and optimize its parameters to predict the matching score of each entity in the knowledge graph based on the user query question, to assist in locating the process documents highly related to the user query question; S3, further calculate the relevance score of the process document and the user query question based on the matching score and the association index matrix, and return the top process documents as the process documents highly relevant to the user query question in the order of the score from high to low based on the relevance score. S4, input the process documents highly related to the user query question and the user query question into the large language model to generate an aerospace manufacturing process knowledge answer matched with the user query question.

2. The aerospace manufacturing process knowledge retrieval and enhanced generation method based on dynamic graph neural network and large language model collaboration according to claim 1, characterized in that: The specific process of step S2 comprises the following steps: S21, the user query question, the entity and the relationship in the knowledge graph are respectively initialized by using a pre-trained dense vector coding model to obtain query vectors , entity node feature set , relationship feature set ; S22, the dynamic graph neural network model based on query dependence comprises a Bellman-Ford type message passing layer and a last output layer for fusing query information; S23、the query vector , the entity node feature set , the relationship feature set is input into a dynamic graph neural network model based on query dependency, and after passing through a layer of Bellman-Ford type message passing layer, a set of entity vectors is obtained ; S24, the entity vector set each entity node feature in the entity vector set the post-splicing query vector , and sent to the output layer to obtain the matching score of the entity and the current user query question.

3. The aerospace manufacturing process knowledge retrieval and enhanced generation method based on dynamic graph neural network and large language model collaboration according to claim 2, characterized in that: The Bellman-Ford formula message passing layer mainly updates the features of each entity node and the relationship based on the following formula: ； ； ； ； wherein, represents the layer in the entity node feature, and , represents the layer in the entity and the entity relationship feature, and , ⊙ represents element-wise multiplication, represents the entity to the entity transmitted information, is an attention weight, which is expressed based on the query vector , the entity relative to the entity neighborhood of other entities, represents the concatenation of feature dimensions, and respectively represent two linear layer weights for processing query vectors and entity node features, represents the index set of all entities related to the entity in the knowledge graph , represents a perceptron composed of two fully connected layers; the entity node feature output by the last layer of the network constitutes an entity vector set .

4. The aerospace manufacturing process knowledge retrieval and enhanced generation method based on dynamic graph neural network and large language model collaboration according to claim 2, characterized in that: The output layer includes a perceptron composed of a fully connected layer and a Sigmoid activation layer, which integrates the query information from the query vector into the entity node features.

5. The aerospace manufacturing process knowledge retrieval and reinforcement generation method based on dynamic graph neural network and large language model cooperation according to claim 1, characterized in that: The query-dependent dynamic graph neural network model, parameters of which are optimized by maximizing the likelihood probability of the query-related entities, constructs a loss function is defined as: ； wherein denotes a set of query dependent entities, denotes a set of query independent entities, denotes a matching score predicted by the query dependent dynamic graph neural network model for a query dependent entity denotes a matching score predicted by the query dependent dynamic graph neural network model for a query independent entity .​​ 6. The aerospace manufacturing process knowledge retrieval and enhanced generation method based on dynamic graph neural network and large language model collaboration according to claim 5, characterized in that: all entities in the knowledge graph that coincide with the entities in the user query question and all entities connected to this entity by a relationship, form the query relevant entity set ; the remaining entities in the knowledge graph form the query irrelevant entity set .

7. The aerospace manufacturing process knowledge retrieval and reinforcement generation method based on dynamic graph neural network and large language model cooperation according to claim 1, characterized in that: The specific process of step S3 comprises the following steps: S31、in the knowledge graph, each entity based on the user query question matching score After, based on the association index matrix , calculate the relevance score of each process document in the process document set and the user query question As follows: ； S32, the relevance score The middle data is sorted in descending order of value, and the index number corresponding to the first data is taken, and the corresponding process document in the process document set is found based on the index number, that is, the process document highly related to the user query question.

8. The aerospace manufacturing process knowledge retrieval and reinforcement generation method based on dynamic graph neural network and large language model cooperation according to claim 1, characterized in that: Step S4 specifically includes: including the process documents highly related to the user query question into the context input of the large language model, combining the user query question to generate accurate and detailed aerospace manufacturing process knowledge answers to support intelligent question answering and technical guidance in aerospace manufacturing scenarios.

Images