A spacecraft component intelligent recommendation method based on generative intent subgraph enhancement

Abstract

The application discloses a spacecraft part intelligent recommendation method based on generative intention subgraph enhancement. The application utilizes a large language model to generate six types of design intention queries, i.e., spacecraft type, belonging subsystem, part family, function, interface and applicable environment, from candidate part texts, and scores intention responses of the candidate parts. Based on the intention response scores of the positive matching parts of the design cases, a case intention response vector is aggregated, and accordingly, a design case-candidate part matching graph is decoupled into multiple design intention subgraphs. Graph convolution propagation is performed on each design intention subgraph to learn fine-grained intention embedding, and cross-modal neighborhood retrieval is performed based on the intention embedding similarity to enhance the original matching graph. Finally, bidirectional pairwise contrast learning of the design case-candidate part is adopted, intention embedding and matching embedding are jointly optimized, and a recommendation score is predicted in combination with a quality constraint. The application improves accuracy and interpretability.

Description

Technical Field

[0001] This invention relates to the fields of spacecraft digital design, spacecraft standardized design libraries, and intelligent recommendation technologies. Specifically, it relates to a method for generating design intent based on a large language model, constructing a design intent subgraph based on a graph neural network, and recommending 3D models and parametric design data of candidate components. This invention also relates to technologies such as natural language processing, semantic representation learning, graph convolutional neural networks, cross-modal retrieval, contrastive learning, and computer-aided design, and can be applied to spacecraft cloud-based 3D CAD platforms, spacecraft design knowledge bases, spacecraft standard parts libraries, parametric model libraries, and other spacecraft digital R&D systems. Background Technology

[0002] As space missions evolve towards higher frequency, lower cost, and faster iteration, the spacecraft development process generates a large amount of 3D models of components, parametric design data, interface information, functional descriptions, environmental adaptability data, and historical design cases. How to quickly locate existing component design resources that match current design requirements in new spacecraft design missions is a crucial issue for improving design reuse rates, reducing redundant modeling work, and shortening development cycles.

[0003] Existing methods for retrieving spacecraft components mainly include category catalog retrieval, keyword retrieval, attribute filtering, and similarity-based model retrieval. Category catalog retrieval relies on designers to select spacecraft type, subsystem, and component category step by step, resulting in low retrieval efficiency. Keyword retrieval struggles to handle semantically similar but different engineering terms such as "tank" and "storage tank," or "mounting lug" and "mounting support." Attribute filtering typically requires designers to accurately fill in all attributes in advance, which can easily lead to the omission of reusable design resources when design requirements are incomplete or multiple constraints are coupled.

[0004] The compatibility of spacecraft components typically involves multiple aspects, including spacecraft type, subsystem affiliation, component family, function, interface, environment, and mass. For example, whether a propellant tank is suitable for a deep space probe depends not only on whether it belongs to the propulsion subsystem, but also on whether it has propellant storage and pressure maintenance functions, whether it has corresponding piping interfaces and installation structures, whether it can withstand cryogenic and high-pressure environments, and whether its mass meets design constraints. Traditional single similarity calculation methods struggle to explicitly distinguish these different semantic factors, easily leading to the incorrect recommendation of components with similar attributes but overall incompatibility to designers.

[0005] Existing recommendation methods based on collaborative filtering and graph neural networks can leverage historical matching relationships between cases and components for representation learning. However, positive matching data in spacecraft design is typically sparse. A design case corresponds to only a small number of confirmed usable components, while the large number of components without established matching relationships includes both truly unsuitable components and potentially suitable components that have not yet been verified by designers. Relying solely on sparse case-component matching graphs for graph convolutional propagation makes it difficult to obtain stable and sufficient node representations.

[0006] Furthermore, existing methods typically encode component names, functions, interfaces, applicable environments, and descriptive text into a unified semantic vector, or simply concatenate the text vector with structured attributes. These approaches fail to explicitly distinguish between various design intentions and do not fully leverage the capabilities of large language models in engineering semantic understanding, terminology normalization, and multi-dimensional information summarization.

[0007] Therefore, an intelligent recommendation method for spacecraft design scenarios is needed. This method should retain the matching relationship between design cases and candidate components, use a large language model to explicitly extract multi-dimensional design intentions, decouple complex matching relationships into multiple intention subgraphs, supplement potential neighborhoods through semantic retrieval, and combine engineering constraints to output candidate component design resources with high accuracy and interpretability. Summary of the Invention

[0008] The purpose of this invention is to overcome the problems existing in the current spacecraft component design resource retrieval process, such as inconsistent engineering terminology, entangled multi-dimensional adaptation conditions, sparse positive matching relationships, difficulty in discovering potential adaptable components, and lack of intent interpretation in the recommendation results. The invention provides a spacecraft component intelligent recommendation method based on generative intent subgraph enhancement.

[0009] This invention uses a large language model as a design semantic understanding engine to generate six types of design intent queries from candidate component names, spacecraft types, subsystems, component families, functions, interfaces, applicable environments, and descriptive text, and scores the intent responses of candidate components. Based on confirmed compatible components for design cases, a case intent response vector is formed, and multiple design intent subgraphs are constructed. Graph convolution propagation is performed in each intent subgraph to achieve decoupling learning of fine-grained design intent between design cases and candidate components. Intent embedding is used for cross-modal semantic retrieval, and neighborhood enhancement is performed on the sparse case-component matching graph. The design intent representation and basic matching representation are jointly optimized through pairwise contrastive learning, and Top-N recommendation results are generated based on quality constraints to produce 3D models of candidate components and their parametric design data.

[0010] To achieve the above objectives, this invention provides an intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement, including a design intent generation module based on a large language model, an intent-aware subgraph learning module, a cross-modal retrieval enhancement module, and a recommendation and model optimization module.

[0011] The overall technical framework of this invention is as follows: Figure 1 As shown, its basic processing procedure is as follows.

[0012] First, we acquire spacecraft design case data, candidate component design resource data, and matching sample data between design cases and candidate components. We use design cases as case nodes and the 3D models and parametric design data of candidate components from the standardized spacecraft design library as component nodes. We construct a design case-candidate component matching graph using only confirmed positive matching relationships. Unmatched relationships are not used as connection edges in the graph but are instead used as negative samples during model training.

[0013] Secondly, the DeepSeek 671B large language model is used to perform semantic analysis on the candidate component texts, generating or standardizing the following six types of design intent queries:

[0014] 1. Spacecraft type adaptation intent;

[0015] 2. Consistent intent within the relevant subsystem;

[0016] 3. Component family matching intent;

[0017] 4. The function fulfills the intent;

[0018] 5. Interface compatibility intent;

[0019] 6. Intention to adapt to the environment.

[0020] DeepSeek 671B scores the responsiveness of each candidate component under the six design intents mentioned above, resulting in a six-dimensional component intent response vector. For each spacecraft design case, the mean of the confirmed matching candidate component intent response vectors is aggregated to obtain the six-dimensional intent response vector for that design case.

[0021] Next, based on the response scores of the design cases and candidate components under each design intent, and combined with the preset intent response threshold, the original design case-candidate component matching graph is softly divided into six intent subgraphs. In each intent subgraph, case nodes, component nodes, and their positive matching edges that simultaneously satisfy the corresponding intent response threshold are retained.

[0022] Subsequently, the MiniLM text encoding model was used to encode the design case requirement text and candidate component text respectively, and graph convolution propagation was performed on the original matching graph to obtain the semantic embeddings of the design case and candidate components. The semantic embeddings were then multiplied element-wise with the learnable encodings of different design intentions, and the node semantics were projected onto the corresponding design intention space. Multi-layer graph convolution propagation was then performed in each intention subgraph to obtain six sets of intention-aware embeddings.

[0023] Based on this, the cosine similarity between the design case intent-aware embedding and the unmatched candidate component intent-aware embedding is calculated. Several candidate components with the highest semantic similarity are retrieved from each intent subgraph as semantic cooperative neighbors. The semantic cooperative neighbors are added to the original matching graph in the form of weighted edges to obtain the enhanced matching graph.

[0024] Finally, the intent-aware embedding and basic matching embedding are jointly optimized by using InfoNCE pairwise contrastive losses in both the design case to candidate component and candidate component to design case directions. During the inference phase, the six sets of intent-aware embeddings and basic matching embeddings are concatenated to calculate the recommendation score between the design case and the candidate component. Candidate components whose quality exceeds the upper limit of the design case quality are first eliminated, and then sorted in descending order of recommendation score to output the 3D models of the Top-N candidate components and their parametric design data.

[0025] The invention details of each main module of this method are as follows:

[0026] 1. Design Intent Generation Module Based on Large Language Model

[0027] This module utilizes the DeepSeek 671B large language model to perform unified semantic understanding of the multi-field text of candidate component design resources.

[0028] Candidate component text must include at least the component name, applicable spacecraft type, subsystem to which it belongs, component family, function label, interface label, environment label, and component description. Using preset prompts, DeepSeek 671B maps candidate component information to six design intents: spacecraft type, subsystem to which it belongs, component family, function, interface, and applicable environment, and outputs a six-dimensional intent response vector for each candidate component. The intent response vector of a design case is not directly scored using a large language model from its requirement text; instead, it is obtained by aggregating the mean of the intent response vectors of the candidate components that have been confirmed to be a positive match for that design case, to maintain consistency between the case intent and the actual component adaptation relationship. The requirement text in the design case is used to form the semantic embedding of the case text and to describe the case, but it does not directly participate in the generation of the six-dimensional case intent response score.

[0029] 2. Intent-Aware Subgraph Learning Module

[0030] This module divides the original matching graph into six intent subgraphs based on the response scores of case nodes and candidate component nodes under six design intents. For each intent subgraph, only when the response scores of the design case and candidate component under the corresponding design intent both reach the intent response threshold... Only when this condition is met, will the case node, the candidate component node, and the positive matching edge between them be retained in the intent subgraph. The intent response threshold... It can be set to 0.3.

[0031] This module uses MiniLM to obtain the initial text encoding of nodes and unifies the semantic embedding of design cases, semantic embedding of candidate components, design intent encoding, intent-aware embedding, and basic matching embedding to the same embedding dimension. The embedding dimension can be set to 64 dimensions.

[0032] Two-layer graph convolutional propagation is performed on the original design case—the candidate component matching graph—to aggregate the neighborhood semantic information of nodes. Subsequently, the node semantic embedding is element-wise multiplied with the learnable encoding of each design intent to obtain the initial intent-aware embedding of the node in different design intent spaces.

[0033] Two layers of graph convolution propagation are performed in each design intent subgraph to propagate semantic signals such as spacecraft type, subsystem, component family, function, interface and applicable environment in the corresponding graph structure, thereby obtaining a fine-grained design intent representation with clear engineering meaning.

[0034] 3. Cross-modal retrieval enhancement module

[0035] This module uses the design intent subgraph as a semantic bridge, supplementing the potential neighborhood of the design case-candidate component matching graph with similarity relationships in the text semantic space. For each design case, the cosine similarity between the design case and candidate components without a positive matching relationship is calculated in each design intent subgraph, and the highest similarity is selected. Each candidate component serves as a semantic collaborative neighbor for this design case under the corresponding design intent.

[0036] The number of semantic cooperative neighbors It can be set to 5. The semantic cooperative neighbors retrieved from the six design intent subgraphs are summarized, the candidate parts retrieved repeatedly are merged, and semantic similarity is used as the weight of the enhanced edge to add them to the original design case - candidate part matching graph.

[0037] A three-layer graph convolutional propagation is performed on the enhanced matching graph to obtain the basic matching embeddings of design cases and candidate components. This method enables the discovery of candidate components that, while lacking historical matching records, are highly semantically relevant to the current design case, thus alleviating the problem of insufficient recommendation representation caused by the sparsity of spacecraft design matching data.

[0038] 4. Recommendation and Model Optimization Module

[0039] This module employs bidirectional pairwise comparative learning between design cases and candidate components for model optimization. From a basic matching perspective, positively matched design case-candidate component pairs are used as positive samples, while other candidate components in the same training batch, explicitly labeled as unsuitable candidate components, and difficult negative samples are used as negative samples, constructing a comparative loss from design cases to candidate components. Symmetrically, positively matched candidate component-design case pairs are used as positive samples, while other design cases in the same training batch are used as negative samples, constructing a comparative loss from candidate components to design cases.

[0040] The pairwise contrast loss uses InfoNCE loss, with temperature parameters... It can be set to 0.2 to control the smoothness of the Softmax distribution in the contrastive loss. Simultaneously, bidirectional pairwise contrastive losses from the design intent perspective are constructed in each of the six design intent subgraphs.

[0041] The contrast loss from the basic matching perspective is weighted and summed with the contrast losses from the six design intent perspectives. The weighting coefficients of the basic matching contrast loss are as follows: It can be set to 1; the design intent is to compare the weighting coefficients of the loss. It can be set to 10.

[0042] During model inference, the intent-aware embeddings of design cases and candidate components under six design intent categories are concatenated with the basic matching embeddings, and the recommendation score is calculated through the inner product. Quality constraints are treated separately as hard engineering constraints; candidate components whose quality exceeds the upper limit of design case quality are not included in the final recommendation ranking.

[0043] The beneficial effects of this invention are reflected in the following aspects:

[0044] First, this invention is the first to utilize a large language model to perform unified semantic understanding of spacecraft component names, functions, interfaces, environments, and descriptive texts. It can identify synonyms and near-synonyms in engineering terms, reducing the problem of missed detections caused by relying solely on keyword matching.

[0045] Second, this invention explicitly divides the semantics of spacecraft component adaptation into six categories of design intent: spacecraft type, subsystem, component family, function, interface, and applicable environment. This can avoid multiple design constraints from becoming entangled in a single embedding space and improve the fineness and interpretability of the recommendation results.

[0046] Third, this invention uses only confirmed and adapted positive matching relationships to construct the graph structure, and uses unfit samples and difficult negative samples for comparative learning, which can reduce the interference caused by erroneous connection edges to graph convolution propagation.

[0047] Fourth, this invention supplements the sparse case-component matching graph with potential semantic neighbors through cross-modal semantic retrieval, which can discover candidate component design resources that have not yet been used in practice but have high reuse value.

[0048] Fifth, this invention uses quality restrictions as a hard constraint in the recommendation stage for filtering, which can prevent semantically similar but quality-unsatisfactory components from entering the final recommendation results.

[0049] Sixth, the objects recommended by this invention are the three-dimensional models of components and their parametric design data in the spacecraft standardized design library, which can help designers quickly complete model retrieval, component selection and reuse of design results, and reduce repetitive modeling work.

[0050] Seventh, this invention can be deployed in an independent 3D cloud CAD platform, as well as in a local spacecraft design system, a private cloud design platform, or other spacecraft digital R&D systems, and has good application adaptability.

[0051] In summary, this invention provides an intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement. It enables intent-based extraction of engineering semantics for spacecraft components, fine-grained decoupling of design cases and candidate component fit relationships, and semantic neighborhood enhancement of sparse matching graphs within a unified end-to-end learning framework. Furthermore, it combines quality constraints to recommend 3D models and parametric design data of candidate components. This provides an effective, robust, and well-interpretable solution for intelligent component retrieval, selection, and design result reuse in the digital design process of spacecraft. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall framework of the intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement of the present invention.

[0053] Figure 2 This is a schematic diagram illustrating the process of generating intent response scores and constructing intent subgraphs in this invention. Detailed Implementation

[0054] This invention discloses an intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement. A preferred embodiment of this invention is given below with reference to the accompanying drawings. The specific implementation steps of this invention are as follows:

[0055] Step 1: Data preprocessing and partitioning of training and test sets:

[0056] First, spacecraft design case data, candidate component design resource data, and matching data between design cases and candidate components are acquired. This embodiment uses three data files, cases.csv, items.csv, and pairs.csv, for training and validation.

[0057] The files are as follows: cases.csv contains 500 spacecraft design cases, each including case number, spacecraft type, target subsystem, target part family, required functions, required interfaces, required environment, quality limit, and design requirement text; items.csv contains 1000 candidate component design resources, each including component number, component name, applicable spacecraft type, subsystem, part family, function label, interface label, environment label, quality, and description text; and pairs.csv contains 10,000 sets of design cases—candidate component matching samples.

[0058] Design cases are used as case nodes, and candidate component 3D models and their parametric design data from the spacecraft standardized design library are used as component nodes. For each set of design case-candidate component matching samples, when the label=1, it means that the candidate component meets the requirements of the corresponding design case and is regarded as a positive matching sample; when the label=0, it means that the candidate component does not meet all the requirements of the corresponding design case and is regarded as a negative matching sample.

[0059] Design case studies are constructed based solely on positive matching samples—candidate component matching diagrams. Negative matching samples are not used as connection edges in the graph, but are used for constructing negative samples in the subsequent model training process.

[0060] Simultaneously, the name, spacecraft type, subsystem, part family, functional label, interface label, environment label, and descriptive text of each candidate component are concatenated to form the contextual text of the corresponding candidate component. This constitutes a candidate component text set:

[0061] ,

[0062] Use the query_text in cases.csv as the corresponding design case. Requirement text .

[0063] For each spacecraft design case, its positive and negative matching samples are randomly divided into training and test sets in an 8:2 ratio, and each design case is ensured to retain at least one positive matching sample in the test set.

[0064] Step 2: Generating Intent Queries Based on Large Language Models:

[0065] Set of context text for all candidate components A design intent prompt and related keywords are simultaneously input into the large language model. The large language model generates... A design intent query, denoted as:

[0066] ,

[0067] in, This represents the first result obtained by the large language model from the contextual text of candidate components. Each design intent query. In this embodiment, the number of design intents is set to... The six design intent queries are spacecraft type, subsystem, component family, function, interface, and applicable environment. These six design intents reflect the compatibility between candidate components and spacecraft design tasks under different engineering semantic dimensions.

[0068] Step 3: Evaluation of Intent Response Scores for Candidate Components and Design Cases:

[0069] Based on the generation of six design intent queries, a large language model is further utilized to analyze each candidate component. In each design intent query The response level is scored accordingly.

[0070] The large language model outputs an intent response score normalized to the interval between 0 and 1 for each candidate component, expressed as:

[0071] ,

[0072] in, Indicates candidate components Design Intent The responsiveness reflects the degree of correlation between candidate components and their corresponding design intent semantic space. Candidate components... In all The response scores under each design intent query constitute the intent response vector:

[0073] ,

[0074] For each spacecraft design case The set of candidate components that positively match the design case in the training set. The design case is obtained by averaging the intent response scores within the data. In the Response score under each design intent:

[0075] ,

[0076] This leads to design case studies. Intent response vector:

[0077] ,

[0078] in, Design Cases The number of corresponding positive matching candidate parts.

[0079] The six-dimensional intent response score of the design case is obtained by aggregating the intent response scores of its positively matching candidate components, and the design case requirement text. It is used for subsequent semantic embedding encoding and is not directly used to generate the six-dimensional case intent response scores mentioned above.

[0080] Step 4: Construction of the Intent Subgraph:

[0081] Based on the intent response scores of design cases and candidate components and Combined with preset intent response thresholds Matching diagrams of candidate components with original design examples Soft division into A design intent sub-graph. In this embodiment, the intent response threshold is preferably set to... For each design intent Assign nodes and edges to the following rules: Sub-image of design intent middle:

[0082]

[0083] That is, only when design cases and candidate components In the The response scores under each design intent are not lower than the threshold. Only then are the two incorporated into the corresponding design intent sub-graph. A design case study – matching edges of candidate components. A matching edge is only preserved if both of its two endpoints exist in the same design intent subgraph and the matching edge is a positive matching edge in the original matching graph G.

[0084] Using the threshold control strategy described above, the original design case—the candidate component matching graph G—was softly partitioned into six semantically coherent design intent subgraphs:

[0085] ,

[0086] Each design intent sub-graph corresponds to a design intent semantic space for spacecraft type, subsystem, component family, function, interface, or applicable environment.

[0087] like Figure 2 As shown, firstly, the large language model scores the response level of each candidate component under six design intents. Then, the mean of the intention response scores of the candidate components that positively match the design case is aggregated to obtain the intention response vector of the corresponding design case. Finally, the case nodes, candidate component nodes and their positive matching edges are filtered according to the preset intention response threshold to construct the design intent subgraph corresponding to different design intents.

[0088] Step 5: Semantic Embedding Pre-training:

[0089] To obtain stable initial semantic representations, semantic embedding pre-training is first performed on the original design case—the candidate component matching graph G. Specifically, the pre-trained lightweight BERT model MiniLM is used for each candidate component. Context text Encode the initial text embeddings of the candidate components. MiniLM was used for each design case. Requirement text Encode the initial text embedding of the design case. Then, multi-layer symmetric normalized graph convolution propagation is performed on the design case—the candidate component matching graph G. The specific formula is as follows:

[0090] ,

[0091] ,

[0092] in, and These represent design cases. and candidate components In the Semantic embedding of layers For graph convolutional layer index; Design Cases The set of positively matched candidate parts Indicates the relationship with candidate components A collection of design examples with positive matching. Symmetrical normalization term. This is used to balance the influence of different node degrees on the neighborhood aggregation results and prevent nodes with high degrees from dominating the aggregation process.

[0093] The obtained The semantic embeddings of the layers are averaged to obtain the final semantic embeddings of the design cases and candidate components:

[0094] ,

[0095] ,

[0096] Step Six: Learning Subgraphs to Awareness of Design Intent

[0097] In obtaining semantic embeddings of design cases semantic embedding of candidate components Then, it is compared with each design intent query. Learnable coding Perform element-wise multiplication to project the semantic embedding onto the corresponding design intent semantic space, obtaining the node at the [node name missing]. Initial intent-aware embedding in a design intent subgraph:

[0098] ,

[0099] ,

[0100] in, This represents the element-wise multiplication operator. Indicates design intent query The corresponding learnable encoding vector. Subsequently, in each design intent subgraph... Intent-aware graph convolutional propagation with multi-layer symmetric normalization:

[0101] ,

[0102] ,

[0103] in, and These represent design cases. and candidate components In the The first design intent sub-image Intent-aware embedding at the layer; and These respectively represent the design cases and candidate components in the [number]th [year]. The neighbor set in the design intent subgraph. In this embodiment, the number of graph convolution layers in the design intent subgraph is set to L=2. The multi-layer intent-aware embedding is averaged according to the number of layers to obtain the design cases and candidate components in the th design intent subgraph. Ultimate intent perception embedding under each design intent dimension:

[0104] ,

[0105] ,

[0106] By performing the above operations on each of the six design intent sub-graphs, a design case is obtained. and candidate components By embedding six sets of intent perceptions across six design intent dimensions, fine-grained decoupling of design cases and candidate component adaptation features is achieved.

[0107] Step 7: Enhance cross-modal retrieval:

[0108] Considering the sparse matching relationships among candidate components in the design cases, some reusable candidate components may not yet have established matching relationships with the current design cases. This step utilizes design intent-aware embedding to enhance the original matching map across modal neighborhoods. For each target design case... In the Sub-image of design intent In this study, cosine similarity was used to calculate the semantic similarity between the intent-aware embedding of this design case and the intent-aware embeddings of all candidate components that did not establish a positive matching relationship:

[0109] ,

[0110] Where 〈·,·〉 represents cosine similarity.

[0111] From the Among the unmatched candidate parts in each design intent sub-graph, select the one with the highest similarity. One candidate component, as a design case. In the Semantic collaborative neighbor set under each design intent dimension:

[0112] ,

[0113] In this embodiment, the number of semantic cooperative neighbors under each design intent subgraph is set to... The semantic cooperative neighbors retrieved from the six design intent subgraphs are summarized to obtain the target design case. Enhanced neighbor set:

[0114] ,

[0115] The aforementioned semantic cooperative neighbors are incorporated into the original design case—the candidate component matching graph—in the form of weighted edges. In the next step, an enhanced matching graph is constructed, and multi-layer graph convolution propagation is performed on the enhanced matching graph:

[0116] ,

[0117] ,

[0118] in, and These represent design cases. and candidate components In the Layer matching embedding, and These are the learnable initial matching embedding parameters.

[0119] In this embodiment, the number of graph convolutional layers on the enhanced matching graph is set to... The average value of the multi-layer matching embeddings is taken according to the number of layers to obtain the final matching embeddings of the design cases and candidate components:

[0120] ,

[0121] ,

[0122] Step 8: Design Case Study—Construction of Pairwise Comparison Losses for Candidate Components and Joint Optimization of the Model:

[0123] This approach employs pairwise comparisons between design cases and candidate components as the optimization objective, following the InfoNCE loss form, and unifies design intent-aware subgraph learning with cross-modal retrieval enhancement into an end-to-end recommendation learning framework. For the comparison loss from a matching perspective, specific steps are taken for each design case. The candidate parts that match it. As positive samples, other candidate parts in the same small batch and candidate parts with label=0 in pairs.csv are used as negative sample sets. .

[0124] Symmetrically, for each candidate component Design cases that match it As positive samples, other design cases in the same small batch and their corresponding negative matching design cases are used as the negative sample set. The pairwise comparison losses from design case to candidate component and from candidate component to design case are as follows:

[0125] ,

[0126] ,

[0127] in, The temperature parameter is used to control the smoothness of the softmax distribution; in this embodiment, it is preferably set to... .

[0128] The bidirectional contrast loss from the matching perspective is:

[0129] ,

[0130] Similarly, in each design intent sub-graph Bidirectional contrast loss from the perspective of design intent Its structure is consistent with the matching contrast loss described above, only embedding the matches and Replace with the first Intent-aware embedding under a design intent and The final total loss function is composed of the matching-contrast loss. and Contrast loss corresponding to each design intent sub-image Weighted summation yields:

[0131] ,

[0132] in, and These represent the weighting coefficients for the matching contrast loss and the design intent contrast loss, respectively. In this embodiment, they are preferably set to... , .

[0133] The Adam optimizer is used to perform end-to-end joint optimization of the model parameters, and the Xavier normal distribution is used to initialize the model parameters.

[0134] Step Nine: Recommendation Score Prediction and Top-N Recommendation List Generation:

[0135] After the model training is completed, it enters the inference phase.

[0136] Design Cases exist The intent-aware embedding and matching embedding under each design intent dimension are concatenated to obtain the final fusion representation of the design case:

[0137] ,

[0138] Similarly, candidate components exist The intent-aware embedding and matching embedding under each design intent dimension are concatenated to obtain the final fused representation of the candidate components:

[0139] ,

[0140] Hypothetical Design Case The upper limit of the mass is Candidate components The quality is First, eliminate candidate components whose quality does not meet the design requirements, that is, only retain candidate components that satisfy the following formula:

[0141] ,

[0142] For candidate components that meet the quality constraints, the inner product is used as the matching scoring function to calculate the design case. With candidate components Predicted match score between:

[0143] ,

[0144] For each spacecraft design case, all candidate components that meet the quality constraints are ranked according to their predicted matching scores. Sort in descending order and remove candidate components that are already a positive match for the design case in the training set.

[0145] The highest-ranked among the remaining candidate parts The 3D models of each component and their parametric design data are used as the final Top-N recommendation list and returned to spacecraft designers or spacecraft digital design systems.

[0146] Step 10: Verify the effectiveness of the method:

[0147] After the model training was completed, in order to verify the effectiveness of the method of the present invention, it was tested on a dataset consisting of 500 spacecraft design cases, 1,000 candidate component design resources, and 10,000 sets of case-component matching samples.

[0148] A full ranking strategy is adopted to filter and rank all candidate components corresponding to each spacecraft design case based on quality constraints.

[0149] We use Recall@N and NDCG@N, which are widely used in recommender systems, as evaluation metrics. Recall@N measures the proportion of positively matching candidate parts in the test set that are covered by the Top-N recommendation list; NDCG@N measures the ranking position of positively matching candidate parts in the Top-N recommendation list, with a larger NDCG@N indicating a higher ranking of the positively matching candidate parts.

[0150] In this embodiment, N=10 and N=20 respectively, Recall@10, Recall@20, NDCG@10 and NDCG@20 are calculated, and the average value of the evaluation results of all spacecraft design cases is taken to evaluate the recommendation effect of the method of the present invention on candidate component design resources.

Claims

1. A method for intelligent recommendation of spacecraft components based on generative intent subgraph enhancement, characterized in that, Includes the following steps: S1. Data preprocessing: Acquire spacecraft design case data, candidate component design resource data, and matching sample data between design cases and candidate components; Spacecraft design cases are used as case nodes, and candidate component 3D models and their parametric design data in the spacecraft standardized design library are used as component nodes. A design case-candidate component matching diagram is constructed based on the confirmed positive matching relationship, and the unmatched relationship is used as a negative sample in the model training process. The engineering attributes and descriptions of candidate components are combined to form the candidate component context text, and the requirement description of the design case is used as the design case requirement text. S2. Design Intent Query Generation Based on a Large Language Model: The context text set of candidate components and design intent generation prompts are input into the large language model. The large language model summarizes the engineering semantics of the candidate components and outputs a preset number of prompts. The design intent query serves as a semantic guide for constructing the design intent subgraph; S3. Evaluation of Intent Response Scores for Candidate Components and Design Cases: Using a large language model, the responsiveness of each candidate component under each design intent query is scored to obtain the intent response vector of the candidate component. For each spacecraft design case, the mean of the intent response scores of the candidate components that are matched to that design case is aggregated to obtain the design case's... Intent response vectors in each design intent dimension; S4. Design Intent Subgraph Construction: Based on the intent response scores of the design case and candidate components, and the preset intent response threshold, the original design case-candidate component matching graph is softly divided into... A design intent subgraph; only when the response scores of the design case and the candidate component under the corresponding design intent are not lower than the intent response threshold, are they included in the corresponding design intent subgraph, and the design case-candidate component matching edge is only retained when both of its two end nodes exist in the corresponding design intent subgraph and are positive matching edges. S5. Semantic Embedding Pre-training: A pre-trained text encoding model is used to encode the context text of candidate components and the requirement text of design cases, respectively, to obtain the initial text embeddings of candidate components and design cases; multi-layer symmetric normalized graph convolution propagation is performed on the original design case-candidate component matching graph, and the multi-layer semantic embeddings are aggregated to obtain the semantic embeddings of design cases and candidate components. S6. Design Intent Aware Subgraph Learning: Perform element-wise multiplication of the semantic embedding with the learnable encoding corresponding to each design intent query, and project the semantic embedding onto the corresponding design intent semantic space to obtain the initial intent awareness embedding of the design case and candidate components. Multi-layer intent-aware graph convolution propagation is performed on each design intent subgraph, and the multi-layer intent-aware embeddings are aggregated to obtain the intent-aware embeddings of design cases and candidate components under each design intent dimension. S7. Enhanced Cross-Modal Retrieval: In each design intent subgraph, calculate the intent-aware embedding similarity between the target design case and candidate components that have not established a positive matching relationship, and select the candidate components with the highest similarity as semantic cooperative neighbors under the corresponding design intent; The semantic cooperative neighbors retrieved from each design intent subgraph are summarized and added to the original design case—candidate component matching graph in the form of weighted edges to construct an enhanced matching graph. Graph convolution propagation is performed on the enhanced matching graph to obtain the matching embeddings of design cases and candidate components; S8. Design Case – Construction of Pairwise Comparison Loss for Candidate Components and Joint Model Optimization: Using positively matched design cases (candidate component pairs) as positive samples, and other nodes and mismatched samples in the same training batch as negative samples, the InfoNCE loss form is used to construct the matching perspective comparison loss from design cases to candidate components and from candidate components to design cases. Simultaneously, design intent perspective comparison losses are constructed on each design intent subgraph. The matching perspective comparison loss is then combined with… The design intent perspective comparison loss corresponding to each design intent subgraph is weighted and summed to obtain the model training objective, and the model parameters are jointly optimized. S9. Recommendation Score Prediction and Top-N Recommendation List Generation: This involves combining design cases and candidate components... The intent-aware embeddings under each design intent dimension are fused with the corresponding matching embeddings to obtain the final fused representation of the design case and candidate components; candidate components that do not meet the quality requirements are filtered according to the quality upper limit of the design case. Calculate the predicted matching score between the design case and the remaining candidate parts, sort them in descending order of predicted matching score, remove candidate parts in the training set that already have a positive match with the design case, and then select the top-ranked parts. A Top-N recommendation list is generated from the 3D models of candidate components and their parametric design data.

2. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In S1, the spacecraft design case data includes at least the case number, spacecraft type, target subsystem, target component family, required functions, required interfaces, required environment, upper quality limit, and design requirement text. Candidate component design resource data should include at least the component number, component name, applicable spacecraft type, subsystem to which it belongs, component family, function label, interface label, environment label, component quality, and descriptive text. The matching samples between design cases and candidate parts include positive matching samples and negative matching samples. When the matching label is 1, the corresponding matching sample is determined as a positive matching sample, and a connection edge is established between the corresponding design case node and the candidate part node. When the matching label is 0, the corresponding matching sample is determined as a negative matching sample, and it is not used as a connection edge in the matching graph. The candidate components are obtained by concatenating their component names, applicable spacecraft types, subsystems, component families, functional labels, interface labels, environmental labels, and descriptive text. Context text This constitutes a candidate component text set: ； For each spacecraft design case, its positive and negative matching samples are randomly divided into training and testing sets in an 8:2 ratio, and each spacecraft design case is guaranteed to retain at least one positive matching sample in the testing set.

3. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In step S2, the following candidate component text set will be selected: ； The design intent-generated prompts are input into a large language model, and the output is... A design intent query, specifically represented as: ； in, This indicates that the large language model inductively summarizes the first result from the contextual text of candidate components. One design intent query; The The value is 6. The six design intent queries are spacecraft type adaptation intent, subsystem consistency intent, component family matching intent, function satisfaction intent, interface compatibility intent, and environment adaptation intent.

4. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In step S3, a large language model is used to analyze each candidate component. In each design intent query The level of response is scored to obtain an intent response score normalized to the interval between 0 and 1: ， in, Indicates candidate components For the first The responsiveness of a design intent query; candidate components exist The response scores under each design intent query form a candidate component intent response vector: ， For each spacecraft design case For its positive matching candidate component set The design case is obtained by aggregating the mean of the intent response scores of each candidate component. In the Response score under each design intent: ， This leads to design case studies. Intent response vector: ， in, Design Cases The corresponding number of positive matching candidate parts, the design case requirement text is used for semantic embedding encoding of the design case, and is not directly used to generate the design case intent response vector.

5. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In step S4, the intention response scores of the design case and candidate components are used as a basis. and Combined with preset intent response thresholds Matching diagrams of candidate components with original design examples Soft division into A design intent sub-diagram; For the A design intent query The design case nodes, candidate component nodes, and matching edges are assigned to the following rules: Sub-image of design intent : ； Among them, only when design cases With candidate components In the The response scores under each design intent are not lower than the intent response threshold. and design cases With candidate components The positive matching edge is only retained in the first position when there is a positive matching edge between them. In the design intent sub-diagram; Through the above threshold control, the original matching map Divided into a set of design intent sub-images: ； Furthermore, a design case node or candidate component node can exist in multiple design intent sub-graphs simultaneously.

6. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In step S5, a pre-trained lightweight BERT model, MiniLM, is used for each candidate component. Context text Encode to obtain candidate parts Initial text embedding ; The MiniLM was used for each design case. Requirement text Code is performed to obtain design examples. Initial text embedding ; In the original design case – candidate component matching diagram Perform multi-layer symmetric normalized graph convolution propagation: ， ， in, and These represent design cases. and candidate components In the Semantic embedding of layers Design Cases The set of positively matched candidate parts Indicates the relationship with candidate components A collection of design examples with positive matching; The semantic embeddings of the multi-layer semantic embeddings are aggregated by mean to obtain the semantic embeddings of the design cases and candidate components: ， ， in, This represents the number of graph convolutional layers in the semantic embedding pre-training process.

7. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In step S6, the semantics of the design case are embedded. semantic embedding of candidate components respectively with the first The learnable code corresponding to each design intent query Perform element-wise multiplication to obtain the design cases and candidate components in the [number]th [stage]. Initial intent-aware embedding in a design intent space: ， ， in, Indicates element-wise product; In the Sub-image of design intent Perform multi-layer symmetric normalized intent-aware graph convolutional propagation: ， ， in, and These represent design cases. and candidate components In the The set of neighbors in a design intent subgraph; By aggregating the mean of the multi-layer intent-aware embeddings, we obtain design cases and candidate components in the [missing information]. Ultimate intent perception embedding under each design intent dimension: ， ， Through The above processing is performed on each design intent sub-drawing to obtain design cases and candidate components. Under the dimension of design intent Group intent-aware embedding.

8. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In S7, for the target design case In the Sub-image of design intent The cosine similarity between the intent-aware embedding and the intent-aware embedding of candidate components that have not established a positive matching relationship is calculated: ， Where ⟨⋅,⋅> represents cosine similarity; Never before designed with a target Among the candidate components that establish a positive matching relationship, the component with the highest cosine similarity is selected. One candidate component, used as a target design case. In the Semantic collaborative neighbor set under each design intent dimension: ， Will The semantic cooperative neighbors retrieved from each design intent subgraph are summarized to obtain the target design case. Enhanced neighbor set: ， The enhanced edges of the semantic cooperative neighbors, weighted by cosine similarity, are connected to the original design case—the candidate component matching graph, and multi-layer graph convolution propagation is performed on the enhanced matching graph: ， ， The mean of the multi-layer matching embeddings is aggregated to obtain the final matching embeddings of the design cases and candidate components: ， ， in, and For learnable initial matching embeddings.

9. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In S8, for positively matched design cases—candidate component pairs Candidate components As a design case Positive samples are used as design cases, along with other candidate components from the same training batch and candidate components with a matching label of 0. negative sample set Design Cases As candidate components Positive samples, along with other design cases and candidate components from the same training batch. Incompatible design cases as candidate components negative sample set ; The matching losses from design case to candidate component direction and from candidate component to design case direction are as follows: ， ， in, For temperature parameters; The bidirectional contrast loss for matching perspectives is: ， In each design intent sub-graph Bidirectional contrast loss from the perspective of design intent And replace the matching embedding with the corresponding first... Intent-aware embedding under each design intent dimension; The total loss function is: ， in, and These represent the weighting coefficients for the matching perspective contrast loss and the design intent perspective contrast loss, respectively. The Adam optimizer is used to perform end-to-end joint optimization of the model parameters, and the Xavier normal distribution is used to initialize the model parameters.

10. The intelligent recommendation method for spacecraft components based on generative intent subgraph enhancement according to claim 1, characterized in that, In S9, design cases will be included. exist The design case is obtained by concatenating the intent-aware embedding and matching embedding under each design intent dimension. Final fusion characterization: ， Candidate components exist The intent-aware embedding and matching embedding under each design intent dimension are concatenated to obtain candidate components. Final fusion characterization: ， Design Case The upper limit of the mass is Candidate components The quality is Candidate parts that meet the following quality constraints will be retained as recommended candidate parts: ， For candidate components that meet quality constraints, the design case uses inner product calculation. With candidate components Predicted match score between: ， Candidate components that meet the quality constraints are ranked according to their predicted matching scores. Sort in descending order, and remove those already in the training set and those already related to the design cases. For positively matched candidate parts, the top-ranked parts from the remaining candidate parts will be selected. The 3D models of candidate components and their parametric design data are used as a Top-N recommendation list and returned to spacecraft designers, spacecraft cloud-based 3D CAD platforms, or spacecraft digital design systems.

Images