A method and apparatus for quality assessment and traceability of nodes in aerospace data circulation network

Abstract

This invention discloses a method and apparatus for quality assessment and tracing of nodes in aerospace data circulation networks, relating to the fields of artificial intelligence and data processing technology. First, a standardized aerospace data tracing knowledge graph is constructed, unifying the modeling of three types of entities—data, processing, and subjects—and three types of relationships—production, execution, and input. Second, multimodal feature encoding and heterogeneous graph reconstruction are carried out, fusing quality attributes and topological structure to form a computational graph. Then, a quality regression model is trained based on the heterogeneous graph to achieve continuous numerical prediction of node quality. Finally, combining upstream path constraints and multi-dimensional attribution mechanisms, the root causes of quality problems are located and an interpretable diagnostic report is output. This invention transforms quality analysis from discrete discrimination to continuous quantitative assessment, accurately adapting to heterogeneous network semantics, achieving integrated assessment and tracing, significantly improving the accuracy, efficiency, and interpretability of aerospace data quality analysis, and is applicable to aerospace data end-to-end quality monitoring and reliable circulation scenarios.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and more particularly to a method and apparatus for quality assessment and traceability of nodes in aerospace data circulation networks. Background Technology

[0002] Regarding the implementation methods and technical support means of data circulation networks, the following two representative technical routes have gradually emerged in the existing technologies.

[0003] The first type of method is data traceability implementation based on traditional databases. This type of method is typically based on data traceability models (such as W3C PROV) and uses metadata, system logs, and other means to structurally describe the dependencies and processing procedures formed during data circulation. Relevant traceability information is stored in a centralized database system to support querying and tracing the data generation and circulation process. This approach allows for a relatively complete record of the data processing chain, providing a basis for tracing the data source and processing process afterward. However, this type of method usually relies on a storage and maintenance system centrally managed by a single institution. The traceability records are theoretically still at risk of malicious tampering or incorrect modification due to management errors. Their authenticity and reliability largely depend on the operational capabilities and management reputation of the data management party.

[0004] The second type of method is the trusted traceability method based on blockchain technology, which can effectively alleviate the credibility issues caused by the risk of tampering in traceability records. This type of method introduces distributed ledger technology, utilizing its tamper-proof, decentralized, and traceable characteristics to write key operations and related information in the data processing process into the blockchain, thereby significantly improving the integrity, auditability, and credibility of traceability records. This technical approach has a clear advantage in preventing unauthorized tampering of traceability data. However, although the trusted traceability method based on blockchain improves the credibility of the traceability records themselves, its technical focus remains on the credible storage of traceability information and does not change the way the traceability network is analyzed. In practical applications, when traceability analysis is needed for complex data quality issues, blockchain-based traceability systems usually still require off-chain analysis methods, with technical personnel querying, filtering, and logically reasoning the traceability information. This analysis method has certain limitations in terms of efficiency and intelligent analysis capabilities, making it difficult to systematically model, quantitatively evaluate, and interpretably analyze the quality status of data products or processing stages, thus failing to support practical application needs such as quality verification and problem tracing.

[0005] At present, in order to achieve the goals of data quality monitoring and reliable data flow, there is an urgent need to use data analysis to intelligently assess the quality and trace the source of data flow networks. This is mainly reflected in the following two scenarios:

[0006] (1) Scenario 1 (Verification-based Assessment): In the aerospace data circulation network, various data elements and their algorithmic models are usually accompanied by quality indicators provided by their producers or implementers to describe the accuracy, completeness, and other quality attributes of the data or processing results. However, in the complex multi-stage circulation and processing process, data quality may change due to improper processing methods, unreasonable parameter settings, or mismatched application scenarios. Related quality information fails to synchronously and accurately reflect these changes, leading to inconsistencies between the original quality claims and the actual quality status of the data. Therefore, in practical applications, there is an urgent need for an analytical method that can comprehensively utilize the structural relationships and multi-source information throughout the entire data circulation network to objectively assess and predict the quality status of data products and key processing stages, in order to verify the credibility of existing quality claims and promptly identify potential quality risks.

[0007] (2) Scenario Two (Source-Based Diagnosis): During the flow of aerospace data, when a quality problem is found in a data product (such as a final thematic product or a key intermediate product), practical applications often require a systematic quality source tracing and root cause analysis of its formation process. Since aerospace data typically undergoes multi-source acquisition, multi-level processing, and multi-stage fusion, its quality problems may be caused by the combined effects of multiple upstream data sources or processing stages. Simply identifying the data generation path is insufficient to clarify the actual impact of each influencing factor on the quality defect. Therefore, there is an urgent need for an automated and interpretable method for tracing the source and root cause of quality problems. This method can not only identify the key data sources or processing stages leading to the quality problem, but also quantitatively characterize the relative impact of each upstream factor on the current quality defect, thus providing a reliable basis for the technical analysis of quality problems, process optimization, and responsibility determination.

[0008] By analyzing the two types of demand scenarios mentioned above, it is easy to see that existing technologies (data traceability implementation methods based on traditional databases and trusted traceability methods based on blockchain technology) are still unable to accurately quantify and assess the quality status of data products or processing links based on an established data circulation network, nor can they proactively identify potential quality risks. Furthermore, when quality problems occur, they cannot achieve automated and explainable localization of their causes and key influencing factors. Therefore, a technical method capable of intelligently analyzing the quality status of the entire data circulation process is still lacking. This invention addresses these technical deficiencies by proposing an intelligent quality assessment and traceability diagnosis method for data circulation networks.

[0009] Existing data traceability technologies can record dependencies and processing steps in data circulation and ensure the integrity and reliability of traceability information through database systems or blockchain technologies. However, the technical focus of these approaches is mainly on the acquisition, storage, and reliable preservation of traceability information, lacking intelligent analysis and processing capabilities for quality issues within the established traceability network itself. In application scenarios with multiple links and complex connections, such as aerospace data element circulation networks, relying solely on recording and querying traceability information is insufficient to support quantitative assessment of the quality status of data products or processing stages, proactive identification of potential quality risks, and in-depth analysis of the mechanisms underlying quality problems. Therefore, it is difficult to meet the practical needs of refined quality management and reliable circulation.

[0010] To achieve intelligent analysis of data element circulation networks, the following core scientific questions urgently need to be addressed: In a complex heterogeneous data circulation network composed of multiple types of entities (including data elements, algorithm processing, and participating entities) and multiple types of relationships (including production, use, and execution relationships), how can we, on the one hand, accurately predict continuous numerical indicators such as the quality status of data products and key processing links based on the comprehensive utilization of network structure information and node attribute information, and on the other hand, further reveal the relative influence of different upstream data sources and processing links on the current quality status based on the prediction results, thereby forming interpretable quality assessment and diagnostic conclusions?

[0011] To solve the above-mentioned scientific problems, the relevant technical methods need to possess the following capabilities simultaneously:

[0012] (1) Graph structure representation capability: It can effectively model the complex topological structure and its relationships in the data element circulation network, providing a structured expression basis for subsequent analysis;

[0013] (2) Heterogeneous information modeling capability: able to distinguish and utilize the semantic differences of different types of nodes and relationships in the network to characterize their different roles in the process of quality formation and transmission;

[0014] (3) Continuous prediction and impact attribution capability: It can continuously predict the quality status of nodes and provide a structured interpretation of the prediction results, quantitatively characterizing the degree of influence of different upstream factors on the quality status of the target node.

[0015] In existing technologies, analysis methods based on artificial rules and traditional graph theory algorithms are insufficient to simultaneously meet the comprehensive requirements of complex network structure modeling, heterogeneous information processing, and continuous numerical prediction and attribution analysis. Graph Neural Networks (GNNs), which have emerged in recent years, offer a feasible technical approach to solving these scientific problems. GNNs are a type of deep learning model designed for graph-structured data. Through a core "message passing" mechanism, they iteratively aggregate information about node neighborhoods, enabling them to learn high-dimensional representations of nodes by fusing network structure features and node attribute features, and supporting the modeling of complex nonlinear relationships. Therefore, from a technical capability perspective, GNNs possess a certain degree of adaptability in handling prediction and analysis problems in complex heterogeneous networks.

[0016] However, existing graph neural network applications typically model the analysis task as a binary classification problem of nodes and process it based on homogeneous graph neural networks when dealing with related problems. When this type of general technical solution is directly applied to the quality analysis scenario of aerospace data element circulation networks, it is difficult to meet practical needs in terms of continuous quantitative assessment, heterogeneous semantic modeling, and the interpretability of diagnostic results. The main limitations are as follows:

[0017] 1. Existing GNN methods based on classification modeling are difficult to support continuous quantitative quality assessment.

[0018] In existing graph neural networks (GNNs), node analysis tasks are typically modeled as binary or multi-class classification problems. The model outputs discrete class labels to distinguish whether a node is abnormal or meets certain conditions. This modeling approach focuses on class discrimination rather than precise characterization of continuous numerical values. However, in the quality verification scenario of data element circulation networks, the core requirement is to continuously and quantitatively evaluate the quality status of various data products and processing stages within the network to characterize the degree of quality differences and support comprehensive comparison and ranking. Existing GNN methods based on classification modeling struggle to directly output continuous quality values, fail to reflect the severity of quality problems, and cannot support quantitative analysis of the overall quality level. Therefore, they fail to meet the core requirement of continuous quantitative quality assessment.

[0019] 2. GNNs based on homogeneous modeling have difficulty distinguishing the differences in mass transfer in heterogeneous source tracing networks.

[0020] Many commonly used graph neural network models (such as GCN) are designed under the isomorphic graph assumption, employing a unified processing mechanism for different node types and relationship types during information propagation and feature aggregation. This type of model has certain applicability when dealing with graph structures containing only a single node type and a single relation semantic. However, aerospace data flow networks consist of multiple types of nodes and multiple semantic relationships, with different entities and relationships having different mechanisms of action in quality formation and transmission. GNN methods based on isomorphic modeling struggle to distinguish these differences, easily leading to semantic information mixing and thus affecting the accuracy of quality status prediction and diagnosis.

[0021] 3. Existing GNN attribution analysis methods have failed to effectively integrate with the needs of quality traceability and diagnosis.

[0022] Existing graph neural network (GNN) methods typically use the prediction results as the main output after completing node state prediction or anomaly detection. Their related attribution analysis methods mostly exist as independent interpretation modules or post-processing methods, lacking a systematic design and integration centered around the application goal of quality traceability and diagnosis. However, in the traceability-based diagnosis scenario of data element circulation networks, the core requirement is not only to determine whether a data product has quality anomalies, but also to clearly identify the key sources causing quality problems based on the traceability network structure, and quantitatively characterize the relative impact of different upstream data elements or processing links on the current quality status. Therefore, there is an urgent need for an analysis method that can combine GNN prediction results with the traceability structure and transform attribution information into an analysis method that can be directly used for quality diagnosis, to meet the requirements of interpretability and accountability in high-reliability quality traceability scenarios. Summary of the Invention

[0023] Based on the above analysis of the specific limitations of existing graph neural network methods in the quality analysis of aerospace data element circulation networks, this invention proposes corresponding technical concepts focusing on key issues such as continuous quantitative quality assessment, heterogeneous traceability network modeling, and interpretability of quality traceability diagnosis. These concepts specifically include the following three aspects:

[0024] (1) To address the problem that “GNN methods based on classification modeling are difficult to support continuous quantitative quality assessment”, this invention transforms the quality analysis task from the traditional node classification modeling method to the regression modeling method of node quality status. By constructing a model that can predict the continuous quality values ​​of nodes, it realizes the quantitative characterization of the quality status of data products and processing links, thereby supporting application needs such as quality difference analysis, comparison and ranking.

[0025] (2) To address the problem that “the GNN method with homogeneous modeling is difficult to characterize the differences in quality transmission in heterogeneous traceability networks”, this invention uses native heterogeneous graph neural networks as the core analysis model. By explicitly distinguishing different types of nodes and relationships in the model structure, the model can differentiate the different roles of data elements, algorithm processing, and participating entities in the quality formation and transmission process, thereby improving the accuracy of quality status prediction and diagnosis results.

[0026] (3) Addressing the problem that "existing GNN attribution analysis methods fail to effectively integrate with the needs of quality traceability diagnosis," this invention introduces an attribution analysis mechanism oriented towards traceability network structures based on the quality status prediction model. This mechanism models the correlation between the model prediction results and the influence of upstream data elements and processing links, enabling the attribution results to be directly transformed into analytical conclusions with diagnostic significance. This supports the identification of key sources of quality problems and the quantitative analysis of their impact.

[0027] In summary, this invention proposes a method and apparatus for quality assessment and tracing of nodes in aerospace data circulation networks. Through continuous quantitative quality prediction, heterogeneous tracing network modeling, and diagnostic-oriented attribution analysis, it achieves systematic assessment and tracing diagnosis of data quality status. The specific technical solution is as follows:

[0028] A method for quality assessment and traceability of nodes in aerospace data circulation network includes the following steps performed sequentially:

[0029] Step 1: Construct and standardize the aerospace data traceability knowledge graph, receive and parse the aerospace data element circulation network file, and transform the original traceability record into a standardized knowledge graph containing three types of nodes: data elements, algorithm processing elements, and agent subject elements, as well as three types of directed relationships: production, execution, and input.

[0030] Step 2: Vectorize and construct a heterogeneous graph for the standardized knowledge graph, analyze the node and relation types, carry out multimodal feature engineering for different entity types, integrate numerical, category and text semantic features to form node feature vectors, extract the comprehensive quality score as a supervision signal, and construct a heterogeneous computation graph that retains causal topological constraints;

[0031] Step 3: Train the heterogeneous graph neural network quality prediction model. Using the heterogeneous computation graph as input, adopt graph-level data partitioning and batch processing, build a regression model based on the heterogeneous graph Transformer, learn the quality transmission law through message passing and attention mechanism, optimize the parameters with mean squared error as the loss function, and obtain the quality prediction model with the best generalization performance.

[0032] Step 4: Conduct interpretable attribution analysis for quality traceability diagnosis. Using the trained model and heterogeneous computation graph as input, extract the upstream influence subgraph of the target node based on the directed acyclic graph constraint. Combine gradient sensitivity, feature perturbation, and the model's internal attention mechanism to calculate the node attribution score and output a quantitative and interpretable quality traceability diagnosis report.

[0033] A device for quality assessment and traceability of nodes in an aerospace data circulation network, comprising:

[0034] The data traceability knowledge graph construction and standardization module constructs and standardizes the aerospace data traceability knowledge graph, receives and parses the aerospace data element circulation network file, and transforms the original traceability records into a standardized knowledge graph containing three types of nodes: data elements, algorithm processing elements, and agent subject elements, as well as three types of directed relationships: production, execution, and input, after structural constraints, semantic standardization, and consistency verification.

[0035] The knowledge graph vectorization and heterogeneous graph construction module vectorizes and constructs heterogeneous graphs for standardized knowledge graphs, analyzes node and relation types, conducts multimodal feature engineering for different entity types, integrates numerical, categorical, and textual semantic features to form node feature vectors, extracts comprehensive quality scores as supervision signals, and constructs heterogeneous computation graphs that retain causal topological constraints.

[0036] The graph neural network quality regression model training module trains a heterogeneous graph neural network quality prediction model. It takes a heterogeneous computation graph as input, adopts graph-level data partitioning and batch processing, builds a regression model based on the heterogeneous graph Transformer, learns the quality transmission law through message passing and attention mechanisms, optimizes parameters with mean squared error as the loss function, and obtains a quality prediction model with the best generalization performance.

[0037] The interpretability attribution analysis module for quality traceability diagnosis conducts interpretable attribution analysis for quality traceability diagnosis. It takes the trained model and heterogeneous computation graph as input, extracts the upstream influence subgraph of the target node based on the constraint of the directed acyclic graph, and calculates the node attribution score by combining gradient sensitivity, feature perturbation and the internal attention mechanism of the model, and outputs a quantitative and interpretable quality traceability diagnosis report.

[0038] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0039] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0040] The present invention has the following beneficial effects:

[0041] 1. It realizes the transformation of quality analysis results from discrete discrimination to numerical expression, thereby improving the accuracy of quality assessment.

[0042] Existing technologies typically model quality analysis tasks as binary or multi-class classification problems of nodes, resulting in discrete categories that fail to reflect the degree of quality differences between different data products or processing stages. This application models the quality analysis task as a numerical prediction problem of node quality status, enabling the analysis results to numerically characterize the quality status of different data products and processing stages. This improves upon existing technologies in terms of the granularity, comparability, and overall analytical capabilities of quality assessment.

[0043] 2. It makes more effective use of the structural and semantic information of heterogeneous traceability networks, improving the applicability of quality analysis in aerospace data circulation scenarios.

[0044] Compared to existing technologies based on isomorphic graph neural networks, this application distinguishes and models different types of nodes and different semantic relationships in the aerospace data circulation network. This enables the model to reflect the differences in the roles of data elements, algorithm processing, and participating entities during information dissemination and feature aggregation, thus better conforming to the structural characteristics of the aerospace data circulation network. This helps to improve the rationality and stability of quality status prediction and analysis results in such complex scenarios.

[0045] 3. It enables the quality analysis results to directly serve the needs of traceability and diagnosis, thereby enhancing the diagnostic value of the analysis conclusions.

[0046] Existing technologies, after completing quality prediction or anomaly identification, often struggle to integrate the analysis results with the traceability network structure to form diagnostically significant conclusions. This application combines the quality status prediction results with attribution analysis of upstream influencing factors, enabling the analysis results to be linked to specific data elements or processing stages. This provides clear technical support for the source analysis and responsibility determination of quality problems, allowing the quality analysis results to go beyond the "status judgment" level and further support traceability diagnostic decisions. Attached Figure Description

[0047] Figure 1 Diagram of the space-air data circulation network;

[0048] Figure 2 This is a diagram showing the flow relationships within the spectral structure.

[0049] Figure 3 A module structure diagram for constructing and standardizing knowledge graphs for data traceability;

[0050] Figure 4 A module structure diagram for knowledge graph vectorization and heterogeneous graph construction;

[0051] Figure 5 Here is the overall architecture diagram of HGT;

[0052] Figure 6 This is a structural diagram of the interpretable attribution analysis module for quality traceability and diagnosis. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0054] This invention proposes a method for quality assessment and tracing of nodes in aerospace data circulation networks, specifically including the following steps:

[0055] Step 1: The first step of the method of this invention is executed by the Aerospace Data Source Tracing Knowledge Graph Construction and Standardization Module. Unlike existing source tracing implementations that only statically record or simply visualize the data flow process, this module does not directly use a general source tracing model for storage. Instead, it aims to support subsequent quality modeling, graph neural network computation, and interpretability analysis by performing an integrated reconstruction of aerospace data source tracing information for computation and analysis. The core function of this module is to receive and parse one or more pre-built data flow network files describing the flow history of aerospace data elements. Through structural constraints, semantic standardization, and consistency verification, it transforms the original source tracing records into a standardized data source tracing knowledge graph representation that is computable, learnable, and interpretable, laying a unified data foundation for subsequent feature engineering and graph model construction.

[0056] In a specific embodiment of this invention, the input data for this module is a data circulation network stored in JSON format. The data model of this network follows the core source semantics of W3C PROV, but is specialized and extended to meet the needs of aerospace data quality analysis and modeling. Its goal is no longer limited to describing "where the data comes from," but rather to finely depict the causal dependencies and quality correlations between various types of data, processing activities, and responsible entities throughout the complete lifecycle of aerospace data products. In this way, the entire process of an aerospace data product (e.g., a land cover classification map), from raw satellite observation and algorithm processing to final application, can be uniformly modeled as a knowledge graph of aerospace data element circulation oriented towards quality analysis. This knowledge graph uses an attribute graph as its underlying data model and abstracts the following three types of core entity nodes and their directed edge relationships around the key components of aerospace data circulation:

[0057] 1) Data Elements: Represent all entities existing in the circulation network in the form of data. Unlike the traditional approach of merely recording data identifiers or storage locations, data elements in this invention are considered core analytical objects with quality attributes and upstream and downstream dependencies. In aerospace application scenarios, specific examples can be raw satellite remote sensing imagery (such as Gaofen-1 L1 level data), standard data products after geometric or radiometric correction (such as surface reflectance products), advanced thematic products obtained from model inversion (such as vegetation index maps), or a set of ground observation station data used for accuracy verification.

[0058] 2) Algorithm Processing Element (Activity): This represents the specific operational entity that processes one or more data elements using algorithms or models, generating new data elements. This type of element not only describes the processing flow itself but is also endowed with evaluation attributes such as methodological quality and reproducibility, allowing the algorithm processing process to be explicitly modeled as a crucial intermediate step affecting data quality. In aerospace applications, specific examples could be a remote sensing data preprocessing algorithm (such as atmospheric correction or orthorectification), a geophysical model inversion process (such as sea surface temperature inversion), or a machine learning classification model (such as a random forest model used for land cover classification).

[0059] 3) Agent Element: Represents the agent or organizational entity responsible for algorithm processing. By incorporating agent elements into a unified graph model, the responsibility relationships between different institutions, systems, or personnel in the data flow process and their potential impact on data quality can be further characterized. In aerospace application scenarios, specific examples could be a data distribution agency, a research institute or university laboratory performing data processing, or a software system that automatically executes batch processing tasks.

[0060] Among the above three types of elements, at least three core directed flow relationships should be defined to characterize the causal dependency structure in the flow of aerospace data, as shown in Table 1. Figure 2 As shown:

[0061] Table 1. Circulation Relationships Among the Three Types of Data Elements

[0062]

[0063] By defining the aforementioned element types and relationships, this invention can unify the originally scattered and heterogeneous aerospace data processing flow into a single attribute graph structure that explicitly expresses the three elements of "data-processing-subject" and their causal dependencies. Unlike traditional source graphs used only for tracing and display, this graph structure is designed from the outset for subsequent quality modeling and graph learning tasks, and its nodes and relationships can carry multi-dimensional quality attributes and semantic information.

[0064] To ensure that data circulation networks from different sources and with different construction methods can be consistently used for subsequent calculations and analyses, this module further standardizes the attribute structure of the input data circulation network, and its typical attributes are shown in Table 2.

[0065] Table 2. Core Aerospace Data Elements and Some Key Attributes

[0066]

[0067] Based on this, the specific implementation process of this module includes the following steps, such as... Figure 3 As shown, its purpose is not simply to load data, but to construct a standardized knowledge graph representation that satisfies the constraints of subsequent graph computation at the semantic, structural, and topological levels.

[0068] (1) File Loading and Syntax Parsing. It receives the path to a file in a data circulation network as input, reads the file content, and converts it into a preliminary in-memory object using a standard JSON parser. If the file does not exist or its content does not conform to JSON syntax specifications, the process is aborted and an error is reported.

[0069] (2) Structure and attribute integrity verification. Based on the required attribute specifications defined in Table 2, the structure constraints of the parsed memory objects are verified. This verification process will check whether the top-level structure is complete (e.g., whether it contains the three required keys "data", "activities", and "agents"), and traverse each element object to verify whether each type of element contains the key attributes necessary to support quality modeling and subsequent learning tasks, thereby avoiding distortion of subsequent analysis results due to missing source information.

[0070] (3) Relationship Consistency and Topology Validation. Given the directional nature of aerospace data flow, consistency and topology validation are performed on the relationships in the graph. For example, it checks whether a reverse reference exists in the target_activities list of each Data node referenced in the inputs_data list of an Activity node. Through bidirectional reference consistency checks, it ensures that the dependencies between data elements and processing elements are semantically closed; and through directed acyclicity checks, it guarantees that the entire data flow network is logically free of circular dependencies, providing a structural basis for subsequent causal analysis and attribution calculations.

[0071] (4) Construction of a computation-oriented structured representation of the knowledge graph. All feature data that has passed the above validations and is organized in list form is reconstructed into a standardized in-memory representation suitable for efficient retrieval and graph computation. For example, a hash map or dictionary can be constructed using each feature's ID as the key, allowing for quick access to the complete attributes of any feature by ID. Based on... Figure 1 The graph structure shown extracts all connections from the reference fields of each element and converts them into a list of relationships (edges). Each item in this list explicitly records a triplet of source node ID, target node ID, and relationship type (e.g., IS_INPUT_FOR, PRODUCES, PERFORMED). For example, for each Data ID in the inputs_data list of an Activity node, an IS_INPUT_FOR relationship is created in the relationship list pointing from that Data ID to that Activity ID. This forms a unified representation that is logically isomorphic to the knowledge graph and structurally suitable for subsequent graph learning processing. Figure 1 As shown, organizations / groups A and B participate in data acquisition and processing activities, producing image product 1 and image product 2 respectively. These are then input into a series of intermediate processing activities, including preprocessing, to produce intermediate data product 4, which is then input into algorithm processing activities. Among these, the data acquisition and processing activities in which organization / group A participates produce image product 3, which is also input into algorithm processing activities. Organization / group C participates in a series of intermediate processing activities, including preprocessing, as well as algorithm processing activities. Finally, the algorithm processing activities produce the final thematic product 5.

[0072] Through the above steps, this module ultimately outputs a space-air data tracing knowledge graph memory object that satisfies consistency constraints at the semantic, structural, and topological levels, thereby providing a unified and computable standardized data foundation for subsequent feature engineering, heterogeneous graph construction, and graph neural network model training.

[0073] Step 2, the second step of the method of this invention, is implemented using a knowledge graph vectorization and heterogeneous graph construction module. Unlike the conventional approach of simply mapping knowledge graphs to node embeddings or for structural similarity analysis, this module does not solely aim at "representation learning." Instead, it focuses on regression prediction modeling and subsequent source attribution analysis of aerospace data quality status. It performs diagnostic-oriented vectorization and heterogeneous graph reconstruction on the standardized data source knowledge graph constructed in the previous step. Its core objective is to systematically transform the knowledge graph, originally represented by a mixture of symbols, text, and multi-source attributes, into a heterogeneous computational graph capable of simultaneously carrying quality information, structural dependencies, and multimodal semantic features, while strictly maintaining the semantic structure and causal direction constraints of the source network. This provides directly usable input for the quality prediction and interpretive analysis of graph neural network models.

[0074] In a specific embodiment of the present invention, the input to this module is the standardized knowledge graph memory object output in step 1, which contains a complete entity attribute dictionary and a list of relation triples. Unlike existing methods that only vectorize single node types or isomorphic structures, this module first explicitly parses the set of entity types (including data elements, algorithm processing elements, and agent elements) and the set of relation types (including IS_INPUT_FOR, PRODUCES, PERFORMED) defined in the knowledge graph, and uses these as meta-structural constraints for heterogeneous graph data objects to limit the subsequent feature organization methods and message passing paths.

[0075] Building upon this foundation, this module performs multimodal entity feature engineering for quality modeling. This process does not employ a uniform encoding strategy for all nodes; instead, it constructs feature representations for each entity type based on the "different roles of different entity types in the quality formation mechanism." Specifically, the module traverses each entity type in the knowledge graph and independently performs feature extraction, encoding, and fusion processes based on the attribute structure of that entity type to form consistent feature representations within the type and distinguishable representations between types. For example... Figure 4 As shown, this multimodal feature engineering includes at least the following processing steps:

[0076] First, extract numerical attribute information from the entity attribute dictionary to construct a numerical feature vector. (Obtained from numerical attributes), and input into the numerical feature coding unit for processing to obtain the numerical feature coding result. The numerical attributes include, but are not limited to, quantitative indicators describing data quality, such as accuracy and completeness, and are mapped to a unified numerical space through normalization, enabling them to participate in model calculations within the same numerical space.

[0077] Secondly, extract discrete fields of category, role, or function attributes (such as data type, processing stage, subject role, etc.) from entity attributes to construct category feature vectors. (Originated from category attributes), and input into the category feature encoding unit for processing. Discrete features are converted into numerical representations through methods such as one-hot encoding to obtain the category feature encoding result. .

[0078] Next, unstructured text information (such as label, description, etc.) contained in the entity attributes is extracted to construct a text feature vector. (Obtained from text attributes), and input into the text semantic feature encoding unit for processing, mapping it to a high-dimensional semantic space that can capture its deep semantics, thus obtaining the text feature encoding result. The preferred text semantic encoding method in this process is to use a pre-trained language model (such as Sentence-BERT) to capture the deep semantic information contained in the algorithm name, processing instructions, or data description, thereby making up for the problem that it is difficult to express complex semantic differences by relying solely on structured fields.

[0079] After obtaining the encoding results of each modality, the numerical features, category features, and text semantic features encoding results are fused along the feature dimension by the feature fusion unit to obtain the final feature vector of a single node:

[0080] ;

[0081] Among them, symbols This represents the eigenvector concatenation operation; it yields the final eigenvector of a single node. Then, the node features are divided according to the entity type of the nodes and organized into different subsets of node features, including: data element node feature subsets. Algorithm for processing node feature subsets and the subset of agent subject node features All node features together constitute the overall node feature matrix. .

[0082] After constructing the node features, this module further performs structured organization of the quality supervision signals. Unlike approaches that treat quality labels as external supplementary information, this module directly extracts the `overall_quality` field for each entity from the entity attributes of the knowledge graph, and constructs a label vector corresponding to each node according to the node index order within the heterogeneous graph. This label vector will serve as a supervisory signal for the subsequent training of the graph neural network regression model, enabling the model to directly learn the mapping relationship between "source structure - multimodal features - quality status".

[0083] Subsequently, this module constructs the heterogeneous graph topology. This process does not simply generate an adjacency matrix, but rather strictly follows the list of relation triples that have passed consistency and directed acyclicity checks in step 1 to construct a restricted set of directed edges. (Obtained from relation triples), and edges are categorized according to relation type to preserve the semantic relationships and causal direction constraints between different types of nodes. Specifically, firstly, a mapping relationship from its unique identifier to a consecutive integer index is established for each type of entity; then, relation triples are grouped according to relation type, and each group of relations is mapped to a set of edge index tensors that connect only a specific source entity type and a target entity type. Each set of edge indices is stored in the heterogeneous graph data object with its corresponding meta-relation triple (e.g., ('Data', 'IS_INPUT_FOR', 'Activity')) as the key, thereby explicitly preserving the causal dependency direction between different types of nodes at the structural level.

[0084] Through the above processing, this module ultimately outputs a complete heterogeneous graph data object, which is represented as follows:

[0085] ,in, Represents the node feature matrix, Represents a set of directed edges. This represents a node quality label vector. This object integrates multimodal quality-related features at the node level, strictly adheres to the causal topological constraints of the data flow network at the structural level, and directly associates the entity's quality state at the label level. Through this module, this invention systematically transforms the knowledge graph originally used to describe data flow relationships into a heterogeneous computational graph that integrates multimodal features, structural dependencies, and quality supervision signals, which can be used for quality regression prediction and source tracing diagnosis. This provides a direct and unified input format for subsequent graph neural network model training and interpretability analysis.

[0086] Step 3, the third step of the method of this invention, is the training step of the graph neural network quality regression model. Unlike conventional approaches that use graph neural networks as general predictors and only learn the correlation between node attributes and labels, this module focuses on characterizing the formation and transmission mechanism of quality status in the aerospace data circulation network. Based on the aforementioned constructed heterogeneous source tracing computation graph, it employs inductive learning to train a heterogeneous graph neural network quality regression model that can be generalized to different data circulation scenarios. This model is not merely used for static quality scoring of individual nodes, but rather establishes a mapping relationship between node quality status and its upstream and downstream structural environments by learning multiple types of dependencies between "data-processing-subject," thereby providing a unified predictive basis for subsequent quality assessment and source tracing diagnosis.

[0087] In a specific embodiment of the present invention, the input to this step is a list of datasets of heterogeneous graph data objects, where each heterogeneous graph corresponds to a relatively independent aerospace data application process or data product generation process. By jointly training on multiple independent but structurally similar tracing networks, this module enables the model to learn universally applicable laws of mass propagation and impact, rather than simply performing a memorized fit on a single process.

[0088] To this end, this module first implements a graph-level data partitioning strategy for cross-process generalization. Specifically, the input list of heterogeneous graph data objects is randomly divided into non-overlapping training, validation, and test set graph lists according to a preset ratio (e.g., 70% for training, 15% for validation, and 15% for test) under a fixed random seed. This partitioning method ensures that the test set graphs used for final performance evaluation are completely invisible to the model during the training phase, thereby enabling an objective evaluation of the model's ability to predict node quality status when faced with new and unseen data tracing processes.

[0089] Subsequently, graph-level data loaders are constructed for the three types of graph lists mentioned above. Unlike node-level batch processing, this module adopts a batch processing approach with "graphs" as the basic unit: in each training or evaluation iteration, the data loader extracts several heterogeneous graphs from the corresponding graph list and dynamically merges them in memory into a batched heterogeneous graph object containing multiple disconnected subgraphs. This approach allows hardware such as GPUs to process multiple data flow processes in parallel computationally, while mathematically maintaining the independence between each subgraph, thereby avoiding information leakage between different traceability processes. For the training set loader, its shuffle parameter is set to True to randomly shuffle the order of the graphs at the beginning of each training round, enhancing the model's robustness to different structural combinations.

[0090] In terms of model structure design, this module constructs a native heterogeneous graph neural network regression model to adapt to the coexistence of multiple types of nodes and multiple types of relationships in the source tracing network. In a preferred embodiment, a heterogeneous graph transformer (HGT) is used as the core network structure, and its overall architecture is as follows: Figure 5 As shown, the model consists of an input mapping layer, a heterogeneous information transmission layer, and a quality regression output layer. The input mapping layer is composed of multiple sets of parallel linear mapping units, which are used to uniformly map the multimodal feature vectors of different node types (data elements, algorithm processing elements, and agent subject elements) with different dimensions, constructed in step 2, into the same latent space, thereby creating conditions for cross-type information interaction.

[0091] The information passing layer of the model is key to the quality impact mechanism in its learning source structure, and it consists of one or more stacked HGT convolutional layers. Each HGT convolutional layer uses a ternary relationship consisting of "source node-edge-target node" as the basic unit for message passing. For the target node... And its relationships through nodes (edges) The source node of the connection The model calculates the query vector (Q), key vector (K), and value vector (V) respectively, and their calculation forms are shown below:

[0092] ;

[0093] ;

[0094] ;

[0095] in, Represents the target node In the Layer feature representation, Indicates the source node In the Layer feature representation; , and These represent query weight, key weight, and value weight, respectively.

[0096] Then, combine the edges Feature information, model calculation source node For the target node attention weights During message passing, the model is based on the source node. and edge The characteristics, through message functions Construct message representations and use message weights Perform a linear transformation:

[0097] ;

[0098] in, This represents the message weight parameter. This represents the fusion function of node features and edge features.

[0099] Finally, the model performs weighted aggregation of messages from different source nodes based on attention weights to obtain the target node. The aggregate representation, whose computational form is:

[0100] ;

[0101] in, Indicates a relationship Next source node For the target node Attention weights; Indicates that it is from the source node and edge Jointly constructed message representation; Represents the target node In the Layer feature representation.

[0102] The above results characterize the information fusion results received by the target node from its neighboring nodes in the current layer. Building upon this, the model further performs type-based aggregation and residual connection operations to enhance its ability to model the semantics of heterogeneous structures. Specifically, for aggregation results generated by neighboring nodes of different types, the model groups and aggregates them according to node type to preserve the information differences between different semantic relationships in the heterogeneous structure. Subsequently, layer normalization is performed on the aggregation results to improve the stability of feature distribution and accelerate model convergence.

[0103] After normalization, a residual connection mechanism is introduced, which adds the current layer output to the node's feature representation in the previous layer element-wise. This alleviates the vanishing gradient problem during deep network training and preserves the original feature information. The final node is obtained. In the Layer representation Through the multi-layered heterogeneous information transmission and feature update process described above, the final representation of each node can integrate the information from its multi-hop upstream and downstream nodes, implicitly modeling the transmission and cumulative effect of data quality in the source tracing network.

[0104] At the output layer, the model uses a set of parallel linear regression heads to map the high-dimensional representations of each type of node obtained after the last layer of information transmission to a continuous scalar value. This scalar value represents the model's prediction of the overall quality status of that node. By employing regression modeling, this invention can characterize fine-grained quality differences between different data products, processing stages, or subjects, rather than simply providing discrete anomaly detection results.

[0105] During the model training phase, this module performs an iterative parameter optimization process. In each training round, the model is first set to training mode and forward propagation is performed using the batch heterogeneous graph provided by the training set data loader to obtain the predicted quality scores of all training nodes within the batch. Subsequently, a loss function suitable for continuous quality modeling (preferably mean squared error loss, MSE) is used to measure the difference between the predicted results and the true quality labels. The calculation formula is as follows:

[0106] ;

[0107] in, This represents the number of nodes participating in the training in this batch. For the first The true quality score (label) of each node. For the model to the first The model calculates the predicted quality score for each node. Then, it performs backpropagation via automatic differentiation to compute the gradient of the loss with respect to all learnable parameters. Finally, the optimizer (e.g., the Adam optimizer) updates the model parameters based on the computed gradients.

[0108] During the model validation and selection phase, the model is set to evaluation mode and undergoes a complete forward propagation on the validation set graph list. To measure the model's quality prediction performance on source data not used in training, the root mean square error (RMSE) is calculated as the primary evaluation metric, and its formula is as follows:

[0109] ;

[0110] in, To verify the total number of nodes in the set, and These represent the true and predicted values ​​of the nodes in the validation set, respectively. The RMSE results have the same dimensions as the original data, facilitating an intuitive understanding of the model's prediction error. To prevent overfitting and select the model with the best generalization performance, a selection strategy based on validation set performance is introduced during training. This strategy continuously monitors changes in the RMSE values ​​on the validation set and only saves the model weight parameters corresponding to the historical lowest RMSE values. If the validation set RMSE stops decreasing over multiple training iterations, the training cycle can be terminated early.

[0111] Through the above training and model selection process, this step ultimately outputs a weight file for a heterogeneous graph neural network quality regression model that exhibits good generalization performance across various aerospace data transmission networks. This model can not only predict the quality status of nodes in known tracing networks, but also serve as the foundational model for the next step, the "interpretable attribution analysis module," supporting automated tracing and quantitative diagnosis of the causes of quality problems.

[0112] Step 4, the fourth step of the method of this invention, is an interpretable attribution analysis module for quality traceability and diagnosis. Unlike general interpretability analysis methods that only provide post-hoc visual interpretations of model prediction results, this module does not interpret a single model output in isolation. Instead, it aims to support the traceability and root cause localization of aerospace data quality issues. Combining the directional structural constraints of the data flow network, it performs structurally constrained and path-aware attribution analysis on the prediction results of the graph neural network, thereby transforming the numerical calculation results within the model into analytical conclusions with clear traceability semantics and diagnostic significance.

[0113] The input to this module includes: the weight file of the heterogeneous graph neural network model, which has been trained and has good generalization performance, output from the previous module, and the corresponding complete heterogeneous data source graph object. The core output of this module is not a single attribution score, but a set of quality influencing factors that can directly point to specific data elements, processing links, or responsible entities, used to support the location, interpretation, and decision-making of quality problems.

[0114] The core methodology of this module is built upon the physical and logical constraints of the aerospace data circulation process. Since the aerospace data circulation network semantically describes the process of data evolution from source collection through multiple levels of processing to the target product, this process logically constitutes a directed acyclic graph (DAG). In this structure, the quality state of any node can only be influenced by its upstream data elements, processing activities, and related entities, and cannot be affected by the reverse effects of downstream nodes.

[0115] Based on the above facts, this invention proposes a strict upstream path constraint-based attribution analysis mechanism: before performing any attribution calculations, the attribution space is first constrained at the structural level, allowing only the prediction results of the target node to be interpreted as the combined effect of its upstream "ancestor" nodes. This approach fundamentally avoids the spurious attribution problem that may occur in traditional interpretability methods, which is inconsistent with the actual data flow, ensuring that the attribution results logically meet the causal consistency requirements of source tracing and diagnosis.

[0116] The specific implementation process of this step is as follows: Figure 6 As shown, when the system receives a user-specified target node (e.g., a data product judged to have quality anomalies), it first performs the upstream influence subgraph extraction step. This process starts from the target node and uses a graph traversal algorithm (such as breadth-first search, BFS) to backtrack strictly along the reverse direction of the directed edges representing data dependencies in the source network (e.g., PRODUCES, IS_INPUT_FOR) (i.e., the direction from the target node to its direct predecessor node), extracting layer by layer all upstream ancestor nodes that can influence the target node through legitimate paths. Through this step, the original large-scale source network is pruned into an influence subgraph containing only those nodes that "may have a real impact on the current quality problem." This subgraph retains the complete causal dependency paths structurally while significantly reducing its size, thus providing an analytical space for subsequent attribution calculations that is both semantically consistent with source attribution and computationally controllable.

[0117] After obtaining the influence subgraph, this module further calculates the degree of influence of each ancestor node on the predicted quality score of the target node. Unlike approaches using only a single attribution method, to improve the robustness and reliability of the attribution results, this invention supports introducing multiple complementary attribution mechanisms on the same influence subgraph to perform cross-characterization analysis of quality influence. The implementation methods include, but are not limited to, the following three types:

[0118] (1) Gradient-Sensitive Attribution Methods: These methods quantify the influence of nodes by calculating the gradient of the model output relative to the input features. A preferred implementation is the integral gradient method, which quantifies each "ancestor" node by integrating the gradient of the input feature along the path from a baseline value to its actual value. Features Prediction function for target node The cumulative sensitivity. This method can characterize the trend of the influence of quality changes on the prediction results from the perspective of a continuous function. Its calculation process can be schematically represented as follows:

[0119] ;

[0120] in: Represents ancestor nodes Attribution score for the prediction results of the target node; This represents the actual input feature vector of the model, which is composed of the concatenation or combination of features that affect all nodes in the subgraph. This represents the baseline input feature vector, used as a reference starting point for attribution calculation. Its value can be an all-zero vector or a feature vector representing a "no-information state". Represents the input feature vector Middle and ancestor nodes The corresponding feature components; Represents the baseline input feature vector Middle and ancestor nodes The corresponding feature components; This represents the prediction function of the trained heterogeneous graph neural network model. Its inputs are node features and graph structure information, and its output is the quality prediction score of the target node. Indicates the interpolation coefficients along the integration path; This indicates that the model output corresponds to the input feature components. The partial derivative of the feature is used to characterize the local sensitivity of the prediction result; Indicates to The integration operation from 0 to 1 is used to accumulate gradient information at each point as the input gradually changes from the baseline input to the actual input.

[0121] (2) Attribution methods based on feature perturbation: These methods assess the importance of nodes by simulating hypothetical interventions. A preferred implementation is the feature permutation method. It perturbs each "ancestor" node sequentially. The method assesses the importance of an ancestor node by analyzing its characteristics and measuring the resulting changes in the target node's prediction score. This counterfactual approach evaluates the impact on the target outcome should the node's quality characteristics change, aligning more closely with hypothesis testing in engineering diagnostics. Its attribution score... The calculation method is as follows:

[0122] ;

[0123] in, Represents ancestor nodes Attribution scores; This represents the original input data, including the node features and structural relationships in the subgraph that influence the data. Represents ancestor nodes The input data is perturbed (permuted or randomly rearranged) by the features of the nodes, while the features of the remaining nodes remain unchanged. Indicates the original input data The target node prediction score output by the model is as follows; Indicates at node The predicted score output by the model on input data with perturbed features; : Represents the prediction function of the trained graph neural network model.

[0124] (3) Attribution methods based on the model's internal attention mechanism: These methods directly utilize the interpretability signals within the model. For graph neural network models employing an attention mechanism (such as the preferred embodiment of this invention, HGT), the internally generated attention weights can be used as the attribution basis. This method extracts and aggregates all attribution signals from the "ancestor" nodes. Point to target node (or other intermediate nodes on the path) multi-layered, multi-head attention weights This method calculates the total "attention flow" contributed by each ancestor node to the target node throughout the entire messaging network. It reveals the sources of quality influence from the perspective of the model's internal decision-making logic. A simplified aggregated attribution score is also provided. It can be represented as all those with The weighted sum of the attention weights from the source and within the affected subgraph:

[0125] ;

[0126] in, Represents ancestor nodes Attribution score for the prediction results of the target node; This indicates that an ancestor node in the subgraph is affected. Indicates from ancestor node Nodes on the path to the target node; This indicates that in the influence subgraph, from the ancestor node... The set of nodes traversed by all directed paths to the target node; This represents the number of attention heads in the attention mechanism of a graph neural network; Indicates the first One's attention, ; Indicates the first Under each attention head, from the node Pointing to node Attention weights are used to characterize nodes. For nodes The importance of information transmission; Indicates the first The weighting coefficients of each attention head are used to weight the contributions of different attention heads; This represents a summation operation used to accumulate the contributions of nodes along the path and different attention heads. The score characterizes the model's intrinsic focus on that ancestor node when making decisions.

[0127] The attribution scores mentioned above are used to characterize the influence of ancestor nodes on the prediction results of the target node. The larger the value, the more significant the contribution of the corresponding node to the prediction result. After calculating the attribution scores, this module further performs attribution result output and analysis for diagnostic applications. Specifically, firstly, the attribution scores of all ancestor nodes in the influence subgraph are sorted to identify the set of key influencing nodes that contribute the most to the prediction results of the target node. Subsequently, the system backtracks to the previously constructed data source knowledge graph in-memory representation to extract the semantic attribute information associated with these key nodes, especially their description (process description) and quality (sub-item quality indicators). By combining the quantified attribution scores with structured source semantic information, this module can transform the numerical conclusion of "who the model considers important" into an interpretable diagnostic conclusion of "why this step leads to quality problems". For example, if attribution analysis finds a low-quality data element D10, and its biggest source of influence is the upstream activity element A5, the system will not only point out that A5 has the highest attribution score when generating the report, but will also further extract and display A5's description text (e.g., "The algorithm was executed in an environment that was completely mismatched with its 'spatiotemporal preferences'") or its quality attribute (e.g., "Its 'methodological rigor' score is only 0.6"), thereby deepening the diagnostic conclusion from "the problem lies with A5" to "the problem lies with A5 because it was incorrectly applied in an unsuitable spatiotemporal scenario".

[0128] Ultimately, the system integrates the attribution ranking results, semantic explanation summaries of key nodes, and corresponding visualization results (such as influence bar charts) into a standardized quality traceability diagnostic report, thereby providing a direct and usable decision-making basis for the technical analysis of quality issues, process optimization, and responsibility determination.

[0129] Another aspect of the present invention provides a device for quality assessment and tracing of nodes in aerospace data circulation network, comprising:

[0130] The data traceability knowledge graph construction and standardization module constructs and standardizes the aerospace data traceability knowledge graph, receives and parses the aerospace data element circulation network file, and transforms the original traceability records into a standardized knowledge graph containing three types of nodes: data elements, algorithm processing elements, and agent subject elements, as well as three types of directed relationships: production, execution, and input, after structural constraints, semantic standardization, and consistency verification.

[0131] The knowledge graph vectorization and heterogeneous graph construction module vectorizes and constructs heterogeneous graphs for standardized knowledge graphs, analyzes node and relation types, conducts multimodal feature engineering for different entity types, integrates numerical, categorical, and textual semantic features to form node feature vectors, extracts comprehensive quality scores as supervision signals, and constructs heterogeneous computation graphs that retain causal topological constraints.

[0132] The graph neural network quality regression model training module trains a heterogeneous graph neural network quality prediction model. It takes a heterogeneous computation graph as input, adopts graph-level data partitioning and batch processing, builds a regression model based on the heterogeneous graph Transformer, learns the quality transmission law through message passing and attention mechanisms, optimizes parameters with mean squared error as the loss function, and obtains a quality prediction model with the best generalization performance.

[0133] The interpretability attribution analysis module for quality traceability diagnosis conducts interpretable attribution analysis for quality traceability diagnosis. It takes the trained model and heterogeneous computation graph as input, extracts the upstream influence subgraph of the target node based on the constraint of the directed acyclic graph, and calculates the node attribution score by combining gradient sensitivity, feature perturbation and the internal attention mechanism of the model, and outputs a quantitative and interpretable quality traceability diagnosis report.

[0134] Another aspect of the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method.

[0135] Another aspect of the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement the method described thereon.

[0136] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0137] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0138] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0139] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

Claims

1. A method for quality assessment and traceability of nodes in a space-air data circulation network, characterized in that, The following steps are performed sequentially: Step 1: Construct and standardize the aerospace data traceability knowledge graph, receive and parse the aerospace data element circulation network file, and transform the original traceability record into a standardized knowledge graph containing three types of nodes: data elements, algorithm processing elements, and agent subject elements, as well as three types of directed relationships: production, execution, and input. Step 2: Vectorize and construct a heterogeneous graph for the standardized knowledge graph, analyze the node and relation types, carry out multimodal feature engineering for different entity types, integrate numerical, category and text semantic features to form node feature vectors, extract the comprehensive quality score as a supervision signal, and construct a heterogeneous computation graph that retains causal topological constraints; Step 3: Train the heterogeneous graph neural network quality prediction model. Using the heterogeneous computation graph as input, adopt graph-level data partitioning and batch processing, build a regression model based on the heterogeneous graph Transformer, learn the quality transmission law through message passing and attention mechanism, optimize the parameters with mean squared error as the loss function, and obtain the quality prediction model with the best generalization performance. Step 4: Conduct interpretable attribution analysis for quality traceability diagnosis. Using the trained model and heterogeneous computation graph as input, extract the upstream influence subgraph of the target node based on the directed acyclic graph constraint. Combine gradient sensitivity, feature perturbation, and the model's internal attention mechanism to calculate the node attribution score and output a quantitative and interpretable quality traceability diagnosis report.

2. The method for quality assessment and traceability of nodes in aerospace data circulation network according to claim 1, characterized in that, In step 1, structural constraints and semantic specifications are checked to verify the unique identifier, name, description, quality attributes, and all necessary attributes of the associated node list for data elements, algorithm processing elements, and proxy subject elements. Consistency checks ensure that bidirectional references are closed and the network is a directed acyclic structure.

3. The method for quality assessment and traceability of nodes in aerospace data circulation network according to claim 1, characterized in that, In step 2, the multimodal feature engineering uses numerical normalization, one-hot encoding, and Sentence-BERT text embedding to process quantitative quality indicators, discrete category attributes, and unstructured text information, respectively, and obtains the final node feature vector by feature concatenation.

4. The method for quality assessment and traceability of nodes in aerospace data circulation network according to claim 1, characterized in that, In step 2, the heterogeneous computation graph stores edge indexes with restricted types using meta-relation triples as keys, explicitly preserving the causal dependency direction and semantic structure of data flow.

5. The method for quality assessment and traceability of nodes in aerospace data circulation network according to claim 1, characterized in that, In step 3, the model training adopts an inductive learning approach, and is jointly trained on heterogeneous graphs of multiple independent aerospace data flow processes. The root mean square error of the verification set is used as an indicator to select the best model and terminate the overfitting training in advance.

6. The method for quality assessment and traceability of nodes in aerospace data circulation network according to claim 1, characterized in that, In step 4, the upstream influence subgraph is extracted by backtracking along the data dependency relationship using breadth-first search, retaining only the ancestor nodes that have an actual causal impact on the target node.

7. The method for quality assessment and traceability of nodes in aerospace data circulation network according to claim 1, characterized in that, In step 4, the attribution results are combined with the semantic attributes of the knowledge graph to generate a diagnostic report, which clarifies the key influencing nodes, the causes of quality defects, and the ranking of the degree of impact, supporting the root cause location of quality problems and the determination of responsibility.

8. A device for quality assessment and traceability of nodes in a space-air data circulation network, characterized in that, include: The data traceability knowledge graph construction and standardization module constructs and standardizes the aerospace data traceability knowledge graph, receives and parses the aerospace data element circulation network file, and transforms the original traceability records into a standardized knowledge graph containing three types of nodes: data elements, algorithm processing elements, and agent subject elements, as well as three types of directed relationships: production, execution, and input, after structural constraints, semantic standardization, and consistency verification. The knowledge graph vectorization and heterogeneous graph construction module vectorizes and constructs heterogeneous graphs for standardized knowledge graphs, analyzes node and relation types, conducts multimodal feature engineering for different entity types, integrates numerical, categorical, and textual semantic features to form node feature vectors, extracts comprehensive quality scores as supervision signals, and constructs heterogeneous computation graphs that retain causal topological constraints. The graph neural network quality regression model training module trains a heterogeneous graph neural network quality prediction model. It takes a heterogeneous computation graph as input, adopts graph-level data partitioning and batch processing, builds a regression model based on the heterogeneous graph Transformer, learns the quality transmission law through message passing and attention mechanisms, optimizes parameters with mean squared error as the loss function, and obtains a quality prediction model with the best generalization performance. The interpretability attribution analysis module for quality traceability diagnosis conducts interpretable attribution analysis for quality traceability diagnosis. It takes the trained model and heterogeneous computation graph as input, extracts the upstream influence subgraph of the target node based on the constraint of the directed acyclic graph, and calculates the node attribution score by combining gradient sensitivity, feature perturbation and the internal attention mechanism of the model, and outputs a quantitative and interpretable quality traceability diagnosis report.

9. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method for quality assessment and tracing of nodes in a space-air data circulation network as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to perform the air-space data circulation network node quality assessment and traceability as described in any one of claims 1 to 7.

Images