Satellite Fault Diagnosis System Based on Quality Problem Knowledge Graph
By constructing a satellite fault diagnosis system based on a knowledge graph of quality issues, the problems of data loss and knowledge sharing in satellite fault diagnosis were solved, multi-dimensional semantic retrieval and fault analysis were realized, and the efficiency and accuracy of fault diagnosis were improved.
Patent Information
- Application Number
- CN202311617451.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-29
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-11-29
AI Technical Summary
Existing satellite fault diagnosis methods suffer from problems such as missing monitoring data, few fault samples, and unclear case rules, which prevent the effective use of historical satellite fault data and the sharing of quality knowledge, thus affecting the efficiency of fault diagnosis.
A satellite fault diagnosis system based on a knowledge graph of quality issues is adopted. By constructing a knowledge graph of satellite quality issues, combining the experience of satellite experts and case data of quality issues, semantic links are established to achieve multi-dimensional semantic retrieval and fault description analysis, and recommend possible fault causes and handling measures.
It enables multi-dimensional correlation analysis of historical satellite fault cases, improving the efficiency and accuracy of fault diagnosis and ensuring the accuracy and efficiency of fault handling decisions.
Smart Images

Figure CN117687824B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a satellite fault diagnosis system based on a knowledge graph of quality problems, and belongs to the field of satellite fault diagnosis technology. Background Technology
[0002] With the deep integration of next-generation artificial intelligence and advanced manufacturing technologies, digital, networked, and intelligent manufacturing management methods will reshape the entire lifecycle of satellites, including design, manufacturing, and operation. Satellite fault diagnosis is a crucial aspect of satellite health management.
[0003] Satellite fault diagnosis methods include model-based methods and model-free methods.
[0004] Model-based fault diagnosis methods require establishing accurate mathematical models for the satellite being diagnosed. Based on the method of parameter acquisition, they are divided into state estimation methods and parameter estimation methods. For complex satellite systems, it is difficult to establish accurate mathematical models; therefore, they are only suitable for fault diagnosis at the satellite system level.
[0005] Model-free methods include signal processing-based methods, data-driven methods, and knowledge-based fault diagnosis.
[0006] Signal processing-based methods employ threshold models to extract features and fuse information from monitoring data. They also extract fault information from signals using methods such as time-domain analysis and wavelet analysis to improve fault identification. Signal processing methods that combine traditional methods with deep learning are gradually becoming the main approach. However, in actual satellite fault diagnosis applications, telemetry signals are diverse and there are redundancies and missing information from some monitoring points, which causes uncertainty in satellite fault characteristic information and limits the application of signal processing methods.
[0007] Data-driven approaches treat fault diagnosis as a pattern recognition problem. They use data analysis and mining techniques such as neural networks, support vector machines, and decision trees to analyze the satellite's operational status based on historical satellite operational data, extract rules to judge operational trends, and address issues such as insufficient fault sample data, lack of domain knowledge, and uninterpretable black-box models in satellite fault diagnosis.
[0008] Knowledge-based fault diagnosis methods include expert system-based methods and case-based methods. Expert system-based methods convert diagnostic knowledge into reasoning rules, and then use these rules to filter or match possible fault modes. Expert system-based methods rely on expert knowledge to organize the input, requiring significant time and effort from experts and making implementation difficult. Case-based methods typically organize case data in a structured manner, making it difficult to analyze the relationships between case data.
[0009] At present, satellite quality problem information is often stored in the form of documents such as zeroing reports. Experience knowledge such as product failure modes and handling measures has not been effectively summarized and refined. There are data silos between various research and development units and subsystems. The above methods cannot solve the problem of collaborative quality knowledge sharing and reference, which is not conducive to the effective use of satellite historical failure data. Summary of the Invention
[0010] The technical problem solved by this invention is to overcome the problems of missing monitoring data, few fault samples, and unclear case rules in existing satellite fault diagnosis methods. It proposes a satellite fault diagnosis system based on a quality problem knowledge graph to solve the problems of sharing and referencing quality knowledge and effectively utilizing historical satellite fault data, thereby improving the efficiency of satellite fault diagnosis.
[0011] The solution of the present invention is: a satellite fault diagnosis system based on a quality problem knowledge graph, which includes a satellite quality problem knowledge graph construction module and a satellite fault diagnosis module based on the quality problem knowledge graph;
[0012] The satellite quality issue knowledge graph construction module consists of two parts: a pattern layer construction module and a data layer construction module. The pattern layer construction module, combining the experience and knowledge of satellite domain experts with quality issue case data, uses an incremental approach to determine the scope of knowledge in the satellite quality issue domain, clarify the elements of the satellite quality issue domain ontology knowledge model, and construct the satellite quality issue domain ontology knowledge model. Semantic links between different satellite quality issue domain ontology knowledge models are established through concept alignment. The data layer construction module performs categorized knowledge extraction from quality issue zeroing reports, existing satellite product failure pattern libraries, and historical satellite failure case data. This instantiation of the satellite quality issue domain ontology knowledge model yields instances, relationships, and attributes. The extracted and merged instances and their attributes are used as key-value pairs as nodes and relationships as edges to form a satellite quality issue domain attribute graph data model. This model is then stored in a graph database, forming the satellite quality issue knowledge graph.
[0013] The satellite fault diagnosis module based on a quality problem knowledge graph employs text preprocessing and text convolutional neural network methods to analyze the fault phenomena input from external sources, identifying fault description keywords, standardized descriptive sentences, and category information. The standardized descriptive keywords and sentences of the fault to be diagnosed are used as query terms. A multi-dimensional retrieval interaction paradigm combining keywords, representation learning, and meta-paths is used to search the satellite quality problem knowledge graph. Finally, a troubleshooting information recommendation method based on multi-dimensional similarity calculation is employed. According to the weighted quantity aggregation principle, a set of similar cases is calculated by comprehensively considering the similarity weights of local semantics, global semantics, and structural semantics. Referring to these similar cases, lists of possible problem causes, possible remedial measures, and possible faulty individual units or components are recommended to engineers for troubleshooting decisions.
[0014] Preferably, the ontology knowledge model for the satellite quality problem domain consists of classes, attributes, relations, functions, and axioms, specifically:
[0015] Classes are used to describe the actual concepts in the field of satellite quality issues. In the ontology knowledge model of satellite quality issues, classes contain two layers of concepts. The first layer of concepts includes two major categories: quality problem cases and faulty satellite information. In the second layer of concepts, the concept of quality problem cases includes quality problems, causes of problems, phenomena of problems, consequences and impacts, responsible units, and handling measures. Faulty satellite information includes faulty satellite models, subsystems to which they belong, faulty units, faulty components, and interruption status.
[0016] Attributes are characteristic descriptions of concepts. Attributes for quality problems include: quality problem identifier, quality problem name, whether it's a repeating attribute, and its current segment; attributes for problem causes include: cause type, cause description, and keywords; attributes for problem phenomena include: time of occurrence, domestic or international location, current stage, and phenomenon description; attributes for consequences and impacts include: whether it affects service, whether it causes single-unit failure, whether it's persistent, whether it's a software or hardware failure, whether it creates a single point of failure, and severity; attributes for responsible units include: name of the responsible unit; attributes for handling measures include: description of the measures and the implementing entity; attributes for faulty satellite models include: satellite launch time, model name, and development unit; attributes for the subsystem to which it belongs include: system name; attributes for faulty individual units include: unit name and unit failure mode; attributes for faulty components include: component name and component failure mode; and attributes for interruption status include: interruption start time and interruption end time.
[0017] Relationships are semantic descriptions of the connections between concepts, including: "quality problem" responsibility lies with "responsible unit", "quality problem" manifestations "problem phenomenon", "quality problem" attributed to "problem cause", "satellite model" causes "quality problem", "quality problem" is mitigated by "responsible measures", "responsible unit" takes "responsible measures", "quality problem" fault location "faulty unit", "quality problem" leads to "consequence impact", and "quality problem" interruption state is "interruption state". The relationships are: "responsibility lies with", "mitigated by", "manifestation", "attributed to", "occurrence", "taken", "fault location", "leads to consequences", and "interruption state".
[0018] Functions are a special type of relation where the first N-1 elements determine the Nth element. Examples include the "subordinate" relationship where "faulty single unit" belongs to "faulty component", the "include" relationship where "fault impact" includes "interruption status", and the "originating from" relationship where "problem cause" originates from "faulty single unit".
[0019] Axioms are facts existing within an ontology that constrain classes and relationships within the ontology. These constraints include value constraints and common-sense constraints. Value constraints include: satellite service interruption start time < satellite interruption end time, and problem occurrence time < interruption start time. Common-sense constraints include the severity of consequences. When the attributes "whether it affects the service," "whether it causes a single machine failure," "whether it forms a single point of failure," and "whether it is persistent" are all true, the severity is higher than when only one or a few of these attributes are true.
[0020] Preferably, the concept alignment method is as follows: taking a satellite quality problem domain ontology knowledge model as a benchmark, traversing the concepts in the benchmark satellite quality problem domain ontology knowledge model, calculating the corresponding similarity of all concepts in other satellite quality problem domain ontology knowledge models, and if the similarity is less than or equal to the concept similarity threshold, then merging the two concepts into the same concept, and establishing a semantic link corresponding to the concept between the two ontology knowledge models;
[0021] The similarity prox(M,N) between the concepts M and N of an ontology is defined as follows:
[0022]
[0023] Among them, P MN It is the common parent class of MN, depth(x) is the depth of x in the class structure, and dom(x) is the attribute domain of x;
[0024] When M and N are discrete values, dom(M) is the number of elements in the domain of concept M, dom(N) is the number of elements in the domain of concept N, and dom(M)∩dom(N) is the number of elements in the union of the domains of concepts M and N.
[0025] When M and N are strings, dom(M) is the number of characters in the string of concept M, dom(N) is the number of characters in the string of concept N, and dom(M)∩dom(N) means that concepts M and N have the same number of characters.
[0026] Preferably, the classification knowledge extraction method is a structured knowledge extraction method based on R2RML or DM mapping, a semi-structured knowledge extraction method based on syntactic rule trees, or an unstructured text extraction method based on a pre-trained neural network model.
[0027] Preferably, entity alignment refers to instance-level fusion by pointing different entity representations to the same objective knowledge element; the specific method is as follows:
[0028] An entity mapping dictionary is established for the satellite quality issue domain. Entity links are established through dictionary matching, which includes both exact matching and fuzzy matching. Fuzzy matching is defined as follows: if the minimum edit distance required to convert string A to string B is less than N / 2, then strings A and B are considered to belong to the same entity; otherwise, strings A and B are considered not to belong to the same entity. The edit distance between strings A and B is lev. A,B (|a|,|b|), where |a|,|b| are the lengths of the strings currently being compared.
[0029]
[0030] Among them, lev A,B (|a|,|b|) represents the edit distance between the first i characters of a and the first j characters of b. Here, i and j are indices starting from 1.
[0031] Preferably, the method for analyzing the input fault description is as follows:
[0032] Based on the thesaurus and stop word list, the fault descriptions input from external sources are standardized, colloquial expressions are removed, and synonyms are replaced.
[0033] Based on the single-machine dictionary and model dictionary, the keyword information that may be involved in the standardized fault description is identified. The keyword information includes fault model, single-machine information, and fault component location.
[0034] A text convolutional neural network is used to classify the output of the first step, namely the normalized fault description sentences, and to map the external fault description sentences to three categories: fault phenomenon, execution operation, and immediate impact, thereby obtaining the category information of key description sentences.
[0035] Preferably, the specific steps of the keyword-based knowledge graph semantic search method are as follows:
[0036] The keywords obtained after analyzing the fault description are used as search inputs.
[0037] Map keywords onto the satellite quality knowledge graph to generate extended entities and relationships;
[0038] Based on the generated entities and relationships, a local subgraph is generated on the satellite quality problem knowledge graph to obtain the graph structure required for structured queries;
[0039] Based on the fault diagnosis objectives, the entities and relationships to be queried in the local subgraph are replaced with variables to generate a structured query subgraph;
[0040] For cases where multiple entities and relationships are mapped to keywords to generate multiple local subgraphs, the set of generated structured query subgraphs is sorted and scored based on keyword search similarity.
[0041] The generated structured query subgraph is submitted to the query engine to obtain query results. The TF-IDF method is used to rank the query results by relevance, resulting in semantic query results {a1, a2, a3, ..., a n}
[0042] Preferably, the specific steps of the knowledge graph semantic search method based on representation learning include:
[0043] By learning operators, the knowledge graph of satellite quality issues is mapped to a low-dimensional continuous vector space;
[0044] The semantic vectors of the classified key descriptive sentences are learned through a text vector model.
[0045] Combining sentence category (entity type) information, the sentence semantic vector and satellite quality problem map are used to perform nearest neighbor calculation in the vector space. Similarity is calculated based on vector distance, yielding similarity results {b1, b2, b3, ..., b...} n}
[0046] Preferably, the specific steps of the knowledge graph structure semantic retrieval method based on meta-paths include:
[0047] Define the entity nodes and meta-paths of extended entities corresponding to keyword items and key sentences;
[0048] On the satellite quality issue knowledge graph, semantic structural similarity between entities is calculated based on meta-paths, and a nearest neighbor query algorithm is executed to obtain structural similarity query results {c1,c2,c3,.....,c n};
[0049] The method for calculating semantic structural similarity between entities is as follows:
[0050]
[0051] p x→y Refers to a knowledge graph metapath instance from object x to object y under metapath P.
[0052] Preferably, a method for recommending obstacle removal information based on multidimensional similarity calculation is adopted, as follows:
[0053] The overall query result set is obtained by taking the union of the three query result sets obtained from considering keyword terms, sentence representation vectors, and meta-paths. Different weights w1, w2, and w3 are assigned to local semantic similarity, global semantic similarity, and structural similarity, respectively, resulting in the query result set R.
[0054] R = w1{a1,a2,a3,....,a n}∨w2{b1,b2,b3,....,b n}∨w3{c1,c2,c3,.....,c n}
[0055] Here, ∨ represents the union operation.
[0056] The query result set is mapped to quality problem cases, and the cases are sorted by similarity based on the semantic similarity of each object contained in each case, local and global semantic similarity, structural similarity, and ranking results. The K cases with the highest similarity scores are obtained as a set of similar fault cases according to the weighted aggregation principle, where K is greater than or equal to 1.
[0057] Output a list of problem causes, a list of solutions, and a list of possible individual units or components from a set of similar failure cases.
[0058] The advantages of this invention compared to the prior art are:
[0059] (1) This invention uses knowledge graphs to manage satellite quality problem data, realizing multi-dimensional correlation analysis of historical satellite failure cases. Compared with the existing structured data storage and management methods, it establishes a shared analysis foundation for satellite quality problem knowledge system and semantic network from top to bottom, improving the efficiency of satellite quality problem case data management and application.
[0060] (2) The fault diagnosis process of this invention adopts a multi-dimensional semantic retrieval interaction paradigm based on the knowledge graph of quality problems, keyword items, representation learning and meta-path. Starting from local semantic similarity, global semantic similarity and structural semantic similarity, the case set with the highest weighted dimension similarity value is recommended to engineers for handling reference, which improves the fault handling decision-making ability while ensuring the accuracy of the solution provided. Attached Figure Description
[0061] Figure 1 This is the architecture of a satellite fault diagnosis system based on a quality problem knowledge graph, as described in this embodiment of the invention.
[0062] Figure 2 This is the process for constructing a domain ontology knowledge model for satellite quality issues according to an embodiment of the present invention;
[0063] Figure 3 Visualization of the domain ontology knowledge model for satellite quality issues in this embodiment of the invention;
[0064] Figure 4 This is a schematic diagram of an entity recognition method based on the Bi-LSTM model in unstructured knowledge extraction according to an embodiment of the present invention.
[0065] Figure 5 This is a schematic diagram of a relation recognition method based on a convolutional neural network in unstructured knowledge extraction according to an embodiment of the present invention;
[0066] Figure 6 This invention provides a visualization of a satellite quality problem knowledge graph based on the Neo4j graph database, as an embodiment of the present invention.
[0067] Figure 7 This is a flowchart of satellite fault diagnosis based on a quality problem knowledge graph, according to an embodiment of the present invention.
[0068] Figure 8 This is an example of the R2RML mapping principle for structured data in a satellite quality problem knowledge graph, as described in this invention. Detailed Implementation
[0069] The present invention will be further described below with reference to the embodiments.
[0070] I. Knowledge Graph of Satellite Quality Issues
[0071] Satellite quality problem knowledge graph technology architecture, such as Figure 1As shown, it logically consists of a schema layer and a data layer. The schema layer is a formal description of concepts and relationships in the satellite quality domain, organized and modeled through ontology. The data layer instantiates concepts based on the schema layer logic, representing the concrete manifestation of conceptual relationships in the satellite quality problem domain. Knowledge in the satellite quality problem domain is organized, and an ontology knowledge model is established. Based on the ontology knowledge model, entities and relationships are extracted from structured and unstructured data such as zero-reports, and a satellite quality problem knowledge graph is constructed from top to bottom. The schema layer models satellite quality problem domain knowledge based on the ontology, achieving semantic unification and knowledge sharing among heterogeneous information sources from different units through concept alignment. Guided by the ontology knowledge model, the types of instances, relationships, and attribute elements are clarified. Knowledge extraction techniques are used to obtain the basic elements constituting the knowledge graph from structured, semi-structured, and unstructured data. To eliminate ambiguity among various knowledge elements, entity alignment techniques are used to merge similar semantic items, reducing redundancy in the knowledge graph and ensuring high quality. After completing the basic construction of the knowledge graph through the above three steps, fault diagnosis applications are carried out on it based on multi-dimensional semantic retrieval technology.
[0072] 1.1 Ontology Knowledge Modeling of Satellite Quality Issues
[0073] Ontology modeling methods express domain-specific complex concepts by integrating attributes, relationships, functions, and axioms between constructed concepts, forming a standardized expression of domain concepts and providing a unified descriptive standard for knowledge modeling and information sharing in the field of satellite quality issues. An ontology knowledge model for the satellite quality issue domain is established by combining expert knowledge in the satellite quality issue domain with case data on quality issues.
[0074] The process of building an ontology knowledge model for the satellite quality problem domain is as follows: Figure 2As shown, firstly, based on expert advice in the field of satellite quality issues, the knowledge scope of the satellite quality issue domain is determined. A clear modeling hierarchy is a prerequisite for the correct analytical function of the knowledge graph. Too few knowledge elements will limit the space for knowledge graph analysis and mining, while mixed knowledge levels will cause the knowledge graph to lack a semantic subject, resulting in modeling bias. Clarifying the semantic scope of quality issues and its hierarchical relationship and representation with related quality elements such as satellite structural and functional information and product supply information is a key issue to consider in this step. Secondly, the collection and analysis of quality issue knowledge is the data foundation for ontology modeling. In addition to engineering data such as zeroing reports, industry standards such as the classification standards for quality and reliability information of aerospace product quality issues and the guidelines for establishing a database of aerospace product failure modes are important references. Based on the determined knowledge scope and data foundation, a knowledge model for the satellite quality issue domain is established. The ontology knowledge model for the satellite quality issue domain mainly includes elements such as concepts, attributes, relationships, functions, axioms, and instances. Guided by faults and satellite model equipment objects, the core concepts include two main categories: quality problem cases and faulty satellite information. Quality problem cases include concepts such as quality problem, problem cause, problem phenomenon, consequence status, responsible unit, and location within the arc. Faulty satellite information includes four structural levels: satellite model, subsystem, faulty unit, and faulty component, along with corresponding fault modes. Attributes describe and supplement the characteristics of concepts. For example, problem causes include attributes such as cause type, cause description, and keywords; consequence impacts include attributes such as whether it affects service, whether it causes unit failure, and whether it is persistent. Attributes include three types: object-type, data-type, and annotation-type. Data-type attributes are typically of string and boolean types. Relationships describe the interaction between concepts, mainly including cause, responsibility, manifestation, and attribution. Function axioms describe the constraints and rules governing relationships between concepts. They establish constraint associations based on features such as combinatorial reflexivity, transitivity, and symmetry, describing the truth value of assertions and forming the basis for logical testing of the ontology. Instances are concrete representations of concepts; typical instances can be added or the scope of instances contained within a concept can be clearly defined during ontology modeling. After clarifying the six key elements of the satellite quality issue knowledge ontology, modeling, visualization, and logic testing were performed using Protege software. The visualized management of the satellite quality issue ontology knowledge model is as follows: Figure 3 As shown.
[0075] 1.2 Knowledge Acquisition of Satellite Quality Issues
[0076] The main steps in acquiring knowledge about satellite quality issues include two parts: extensive collection of quality issue information and knowledge extraction. Sources of satellite quality issue information include quality issue zeroing reports, existing satellite product failure mode databases, and relevant domestic and international quality issue cases. From this information, entities and relationships are extracted from the satellite quality issue knowledge graph instances.
[0077] Knowledge extraction for structured data primarily involves converting relational database data into RDF (Resource Description Framework) data or OWL ontology using mapping languages. Mapping languages include DM (Direct Mapping) and R2RML. DM directly converts the relational database table structure and data into an RDF graph, while R2RML allows for flexible customization of views on relational data. Inputting a relational database conforming to the given pattern results in RDF data described using predicates and types from the target vocabulary. Entity and relation extraction methods for unstructured data are mainly categorized into three types: rule-based methods, statistical learning methods, and deep learning methods. Rule-based methods are suitable for unstructured data with fixed formats and standardized expressions, using manually written syntactic rules for pattern matching to identify entities and relational elements. Statistical learning methods train models based on labeled corpora, requiring the definition of input word features before training. Deep learning methods do not rely on feature engineering; they use word vectors as input and learn from the context through neural networks to identify entities and relations end-to-end.
[0078] 1.3 Knowledge Integration of Satellite Quality Issues
[0079] While knowledge acquisition provides the basic elements for constructing a knowledge graph, the diverse sources of this information often lead to semantic variations and information redundancy among knowledge elements from different sources. This hinders the establishment of connections between knowledge graphs and creates barriers to knowledge sharing. Specifically, different research and development units may use different names for functionally identical product components, such as "navigation subsystem" and "payload subsystem," or "baseband processor" and "reference frequency synthesizer." Computers cannot establish connections between elements with different semantic expressions, resulting in seemingly connected "semantic islands" within the knowledge graph. This prevents the development of association analysis and applications based on the knowledge graph. The interconnectivity and commonality of knowledge elements are crucial for the satellite quality problem knowledge graph to function effectively. Knowledge fusion technology includes entity alignment and ontology alignment.
[0080] Entity alignment technology is used to fuse different entity representations to a single objective knowledge element at the instance level. Considering the characteristics of satellite quality engineering, where differences in entity designations are often clearly reflected in engineering manuals and other documents, an entity mapping dictionary is established. Entity linking is performed through dictionary matching, using both exact and fuzzy matching methods. Fuzzy matching can employ methods such as string similarity calculation based on edit distance. The edit distance between strings A and B is lev. A,B (|a|,|b|), where |a|,|b| are the lengths of the strings currently being compared.
[0081]
[0082] Among them, lev A,B(|a|,|b|) represents the edit distance between the first i characters of a and the first j characters of b. Here, i and j are indices starting from 1.
[0083] When integrating knowledge graphs from different units, ontology alignment is also required for concept-level fusion. Besides considering expert experience and engineering knowledge, concept alignment similarity can be measured from two aspects: class structure similarity and attribute similarity. The higher the level of similarity and attribute overlap between concepts, the greater the probability that the ontology is similar. The similarity prox(M,N) between concepts M and N of an ontology is defined as follows:
[0084]
[0085] Among them, P MN Let x be the common parent class of MN, depth(x) be the depth of x in the class structure, and dom(x) be the attribute domain of x. In a specific embodiment of the present invention, the concept similarity threshold is 0.75.
[0086] When M and N are discrete values, dom(M) is the number of elements in the domain of concept M, dom(N) is the number of elements in the domain of concept N, and dom(M)∩dom(N) is the number of elements in the union of the domains of concepts M and N.
[0087] When M and N are strings, dom(M) is the number of characters in the string of concept M, dom(N) is the number of characters in the string of concept N, and dom(M)∩dom(N) means that concepts M and N have the same number of characters.
[0088] 1.4 Knowledge Storage and Retrieval of Satellite Quality Issues
[0089] From a data modeling perspective, knowledge graphs are essentially graph data, including RDF graphs and attribute graphs. RDF is a standard data model that represents and exchanges machine-understandable information. It is defined as a finite set of triples (s, p, o), where s and o have a relationship p or the attribute p of s has a value of o. Each resource has an HTTP URI as its unique identifier, making it suitable for knowledge-sharing scenarios. Attribute graphs are the most widely used graph data model. Nodes and edges each have a set of key-value pairs as attributes, which can be used to represent all data in an RDF graph. Edge weights can also be added to support quantitative analysis, making it suitable for knowledge mining and analysis scenarios. The query language for RDF graphs is the declarative query language SPARQL, supporting various query and retrieval mechanisms such as triple schema, subgraph schema, and attribute path. Common query languages for attribute graphs include Cypher and Gremlin, both path-oriented query languages. Cypher is simple and fast, while Gremlin has advantages in advanced traversal. Databases for RDF triples include RDF4J and GraphDB, while databases for attribute graphs include Neo4j and JanusGraph. To support comprehensive analysis of quality issues, key-value pairs consisting of extracted and merged instances and their attributes are used as nodes, and relationships are used as edges to form a domain attribute graph data model for satellite quality issues. This model is stored using Neo4j and supports the Cypher query language. The Neo4j database features ACID transaction processing to prevent data loss in the event of hardware failure or system crash, ensuring the security of satellite data.
[0090] II. Satellite Fault Diagnosis Method Based on Quality Problem Knowledge Graph
[0091] This article focuses on fault diagnosis using a knowledge graph application for satellite quality issues, such as... Figure 7 As shown, the process includes: 1) analysis of fault descriptions input by engineers; 2) a multi-dimensional retrieval and interaction paradigm based on keywords, representation learning, and meta-paths; and 3) a troubleshooting information recommendation method based on multi-dimensional similarity calculation. Based on these three steps, the fault description information input by engineers is preprocessed and used as query input to perform semantic and structural retrieval on the satellite quality problem knowledge graph. Finally, combined with the multi-dimensional similarity calculation results, information such as the causes and measures of similar quality problem cases in the graph is recommended to engineers for troubleshooting decisions. Among these, the fault description analysis input by engineers and the multi-dimensional retrieval paradigm are crucial foundations for determining the effectiveness of the returned troubleshooting information.
[0092] 2.1 Fault Description and Analysis Based on Text Convolutional Neural Network
[0093] Fault description analysis aims to comprehensively and accurately analyze user-input fault information, improving engineers' ability to express their queries. Fault descriptions often suffer from colloquialisms, diverse vocabulary, and complex descriptive elements. This paper proposes a fault description analysis method, which includes text preprocessing and sentence element classification based on a Text Convolutional Neural Network (TextCNN). First, the fault description is standardized by identifying fault models and unit information using three types of thesaurus: a domain-specific thesaurus, a synonym list, and a stop word list, thus recognizing domain-specific vocabulary. Second, a Text Convolutional Neural Network is used to identify and classify key descriptive elements. These two steps reduce the requirements for engineers' input descriptions while standardizing fault descriptions.
[0094] The text preprocessing method processes the colloquial descriptions input by engineers and matches keywords based on the domain dictionary, including: 1) standardizing the fault descriptions input by engineers based on the thesaurus and stop word list, removing colloquial expressions, and replacing synonyms; 2) identifying keywords such as fault model, unit information, and fault component location that may be involved in the standardized fault description based on the unit dictionary and model dictionary.
[0095] A text convolutional neural network (TextCNN) is used to classify fault description sentences to obtain element category information. The TextCNN model is trained based on case data from a knowledge graph. After training, the normalized fault description content is input into the TextCNN model to classify the fault description elements, categorizing the input fault description sentences into element categories such as phenomenon description, execution operation, immediate impact, and other information. Through the classification and recognition of fault descriptions, a correspondence between fault description elements and entity types in the quality problem knowledge graph is established.
[0096] 2.2 A Multi-Dimensional Retrieval Interaction Paradigm Based on Keyword, Representation Learning, and Meta-Path Combination
[0097] After text preprocessing and text convolutional neural network classification, query terms are constructed, including keywords (which may involve fault models, unit information, fault component locations, etc.) and normalized fault description sentences and corresponding entity node types. These query terms need to be used as input to perform retrieval in the satellite quality problem knowledge graph. The search calculation is performed from three dimensions: term semantic features, local semantic features, and global semantic features and structural features. Specifically, it is divided into a keyword-based knowledge graph semantic search method, a representation learning-based knowledge graph semantic search method, and a meta-path-based knowledge graph structure retrieval method.
[0098] 2.2.1 Keyword-based knowledge graph semantic search method
[0099] The keyword-based knowledge graph semantic search method uses keyword terms as query input to intuitively express engineers' information query needs. Subgraphs generated based on keyword terms are used as candidate structured queries for retrieval interaction.
[0100] Using keywords such as fault model, unit information, and fault component location obtained from fault description analysis as retrieval input, candidate structured queries composed of corresponding entities and relationships are constructed based on the mapped keywords. Local subgraphs are generated on the satellite quality problem knowledge graph based on the generated entities and relationships to obtain the graph structure required for structured queries. Some entities and relationships in the local subgraphs are replaced with variables to generate structured query subgraphs. For cases where multiple local subgraphs are generated by mapping multiple entities and relationships to keywords, the generated structured query sets are sorted and scored according to keyword search similarity. The generated structured queries are submitted to the query engine to obtain query results. The TF-IDF method is used to sort the query results to obtain semantic query results based on keywords.
[0101] 2.2.2 Semantic Search Method for Knowledge Graphs Based on Representation Learning
[0102] The knowledge graph semantic search method based on representation learning considers both local and global semantic features of the semantic network. It obtains semantic vectors from the classified key descriptive sentences through a text vector representation algorithm, and performs K-nearest neighbor retrieval and ranking in a low-dimensional continuous vector space by combining sentence category information, i.e. entity type information.
[0103] First, translation mechanisms such as TransE and TransH are used to map the satellite quality problem knowledge graph to a low-dimensional continuous vector space. TransE treats entities and relations in the knowledge graph as two matrices. Ideally, after TransE training, the model extracts a vector from each of the entity and relation matrices and performs L1 or L2 computation to obtain the vector of the other entity in the entity matrix, thus representing the triples in the knowledge graph. However, this method does not consider the one-to-many relationship between entities and relations, where a relation pointing to multiple nodes results in vectors with the same value. To address this, TransH abstracts the relations in the triples into a hyperplane in the vector space. Vectors are translated on the hyperplane to calculate the difference between the head and tail nodes, ensuring that the vector representation of the same node differs across hyperplanes of different relations. The scoring function after projection is:
[0104]
[0105] Where h and t represent the vectors of the head and tail entity nodes in the triplet, respectively, and w r The normal vector d represents the hyperplane of relation. rThis represents the translation vector of the relation hyperplane. The smaller the result value is if the triplet relation is correct, and vice versa.
[0106] Use a margin-based sorting function as the loss function:
[0107] L=∑ (h,r,t)∈Δ ∑ (h',r',t')∈Δ' max{0,f r (h,t)+γ-f r' (h',t')}
[0108] Here, Δ represents the set of correct triples, Δ' represents the set of negative examples, and γ represents the margin value used to distinguish between positive and negative examples. The loss function is trained using a mini-SGD optimizer to obtain the vector representation of the satellite quality problem knowledge graph. For the input fault description sentence, a text vector representation method is used. Combining the fault description sentence category information, K-nearest neighbor calculation is performed on the sentence semantic vector and the satellite quality problem graph in the vector space to obtain similarity query results.
[0109] 2.2.3 Knowledge Graph Structure Search Method Based on Meta-Path
[0110] Meta-paths are relational sequences linking object types. They can capture structural semantic information in the satellite quality issue knowledge graph. The meta-path-based knowledge graph structure retrieval method considers the relational structure features of the satellite quality issue knowledge graph, defines meta-paths for entity nodes corresponding to keyword items and key sentences, and performs a nearest neighbor algorithm based on path similarity to obtain similar query results. Specific steps include:
[0111] 1) Define the entity nodes and metapaths of extended entities corresponding to keywords and key sentences, for example...
[0112]
[0113]
[0114] 2) Perform a single-object nearest neighbor query algorithm on the satellite quality problem knowledge graph to obtain the structural similarity query results {c1,c2,c3,.....,c n The method for calculating the similarity between entities is as follows:
[0115]
[0116] p x→yThis refers to instances of knowledge graph metapaths from entity object x to entity object y under metapath P. The more metapaths between two entities, the more similar they are. More connection points indicate higher centrality of entity nodes, corresponding to a higher repetition rate of quality issue cases. Further path mining can be performed using methods such as depth-first traversal and random walks.
[0117] 2.3 A Recommendation Method for Obstacle Clearance Information Based on Multi-Dimensional Similarity Calculation
[0118] The obstacle removal information recommendation method based on multi-dimensional similarity calculation calculates a set of similar cases by comprehensively considering the dimensional similarity weights and the number of neighboring nodes of the matched case nodes according to the weighted quantity aggregation principle. The higher the multi-dimensional similarity and the more neighboring nodes, the higher the case ranks in the recommendation list. The steps of the obstacle removal information recommendation method based on multi-dimensional similarity calculation are as follows:
[0119] 1) The overall query result set is obtained by considering the three query result sets: keyword terms, sentence representation vectors, and meta-paths. Different weights w1, w2, and w3 are assigned to the semantic similarity of terms, local and global semantic similarity, and structural similarity, respectively, to obtain the query result set R:
[0120] R = w1{a1,a2,a3,....,a n}∨w2{b1,b2,b3,....,b n}∨w3{c1,c2,c3,.....,c n}
[0121] Where ∨ represents the union operation, {a1, a2, a3, ..., a...} n} represents the query result set returned by the keyword-based knowledge graph semantic search method, {b1,b2,b3,....,b n} represents the query result set returned by the representation learning-based knowledge graph semantic search method, {c1,c2,c3,.....,c n} represents the query result set returned by the meta-path-based knowledge graph structure search method.
[0122] 2) Map the query result set to quality problem cases, and sort the cases according to the semantic similarity, local and global semantic similarity, structural similarity and ranking results of each object contained in them; obtain the K cases with the highest similarity scores as a set of similar fault cases, where K is greater than or equal to 1;
[0123]
[0124] The entity node a corresponding to the result set o ,b p ,c qCorresponding to the case j (0 < o, p, q <= N, 0 < j <= K)
[0125] Case j The importance weight is composed of the multi-dimensional similarity w of the entity nodes. i (i = 1, 2, 3) and the importance of entity location in the single-dimensional query result set, 1 / o*1 / p*1 / q, determine the result.
[0126] 3) The causes and solutions for similar fault cases are recommended to engineers for troubleshooting reference. After scoring and sorting, a list of candidate causes is obtained. Maintenance personnel perform diagnostic steps according to the list and complete the final cause location based on satellite feedback.
[0127] Repair measures for the same type of fault caused by the same reason in a single unit can be mutually referenced. In addition to the repair measures of fault cases in the quality problem knowledge graph, it is also necessary to refer to the troubleshooting experience of maintenance engineers, handling plans, maintenance manuals and other multi-source information fusion. This process takes into account the differences between multi-source fault repair knowledge, and performs fusion processing on the multi-source maintenance knowledge to generate fused repair measures for engineers to refer to.
[0128] Quality problems occurring during satellite operation are characterized by involving multiple disciplines, a wide range of aspects, and multiple sources and levels of information. It is necessary to summarize and analyze quality problem information of different satellite models using intelligent technologies such as knowledge graphs, so as to realize the construction and application of a knowledge system from quality problem data.
[0129] The knowledge graph-based satellite fault diagnosis method organizes data such as historical cases of quality problems in the form of a knowledge graph. It performs similarity retrieval and calculation between the fault description input by the engineer and the historical case data from three aspects: semantic network local features, global features, and structural features. Finally, it returns the troubleshooting information of the historical fault case with the highest multi-dimensional similarity to the engineer to support troubleshooting decisions, thereby improving the efficiency and accuracy of fault diagnosis and analysis.
[0130] Example 1: Construction and Instantiation of Satellite Quality Issue Ontology Knowledge Model
[0131] A knowledge model on satellite quality issues is established, defining its six key elements. Functional axioms, which constrain and describe the relationships between concepts, establish constraint associations based on features such as combinatorial reflexivity, transitivity, and symmetry, describing the truth value of assertions and forming the basis for logical testing of the ontology. Instances are concrete representations of concepts; typical instances can be added or the scope of instances encompassed by a concept can be clearly defined during ontology modeling. Modeling is performed using Protege software, and logical testing is conducted based on the Protege inference engine.
[0132] Taking the SADA quality problem of solar panel drive structure as an example, a knowledge graph of quality problem is constructed. The data layer of SADA quality problem of solar panel drive structure is organized based on ontology model. The ontology and instance construction information is shown in the table below.
[0133] Table 1. SADA Quality Issues Knowledge Graph Ontology and Corresponding Instance Information
[0134]
[0135]
[0136] Example 2: Knowledge Acquisition of Satellite Quality Issues
[0137] Knowledge extraction for structured data mainly involves converting relational database data into RDF (Resource Description Framework) data using mapping languages. Mapping languages include DM (Direct Mapping) and R2RML. DM directly converts the relational database table structure and data into an RDF graph, while R2RML mapping allows for flexible customization of views on relational data. The input is a relational database that conforms to the schema, and the output is RDF data described using predicates and types from the target vocabulary.
[0138] The structured data table of satellite quality issues is converted into RDF triples. The structured data table includes subsystem name, unit name, faulty component name, fault phenomenon, and fault cause. The unit name is the subject in the structured data table of satellite quality issues. The subsystem, faulty component name, fault phenomenon, and fault cause are listed as objects. The relationship between the subject and object is defined as an attribute. Specifically, the relationship between the unit name and the subsystem is defined as the subsystem to which it belongs; the relationship between the unit and the faulty part is defined as the faulty component; the relationship between the unit name and the phenomenon is defined as the manifestation; and the relationship between the unit name and the cause is defined as attribution. Thus, the example structured data is converted into 8 RDF triples, such as... Figure 8 As shown.
[0139] Entity and relation extraction methods for unstructured data are mainly categorized into three types: rule-based methods, statistical learning methods, and deep learning methods. Rule-based methods are suitable for unstructured data with fixed formats and standardized expressions, using manually written syntactic rules for pattern matching to identify entities and relational elements. Statistical learning methods train models based on labeled corpora, requiring the definition of input word features before training, including word-level, dictionary-level, and document-level features. Entity recognition models primarily employ Hidden Markov Models (HMMs), Maximum Entropy Models (MEMs), and Conditional Random Field Models (CRFs). Deep learning methods do not rely on feature engineering; they use word vectors as input and learn from the context through neural networks to identify entities and relations end-to-end. Entity recognition is often treated as a sequence labeling problem, labeling the beginning, middle, and end of entities using terms like "BIO" or "BIOES." Relation extraction is typically treated as a text classification problem, classifying sentences and their beginning and end entities into corresponding relation categories. Combining the advantages of statistical and deep learning methods, the BiLSTM-CRF model, for example... Figure 4 As shown, entity recognition is achieved using a combination of Long Short-Term Memory (LSTM) networks and Conditional Random Fields (CRF). This includes a vector representation layer, a bidirectional LSM network layer, and a CRF layer. The vector representation layer, learned through a word vector representation model, serves as the input to the bidirectional LSM network, which captures long-distance dependencies in the sentence from both directions. The CRF layer annotates satellite entities in the sentence based on the concatenated vector input from both directions. Supervised relation extraction methods require a large amount of training data; the PCNN long-range supervised relation extraction model, for example... Figure 5 As shown, while enriching the training corpus by aligning texts using the constructed satellite knowledge graph, a segmented pooling strategy based on convolutional neural networks and a multi-instance learning approach are used to reduce noise interference in the existing knowledge graph for accurate relation classification. The knowledge extraction method based on neural network models requires certain implementation costs in terms of training corpus and model debugging, but is suitable for processing massive amounts of unstructured data.
[0140] Example 3: Construction and Storage Management of Satellite Quality Problem Map
[0141] Based on the knowledge graph ontology model of satellite quality issues, triple data is obtained through structured and unstructured knowledge acquisition methods and imported into the Neo4j database in the form of entity and relation triples. Visualization graphics are shown below. Figure 6 As shown, the commands for importing and querying are:
[0142] 1) Create a quality issue entity node and set the quality issue ID as a unique identifier:
[0143] LOAD CSV WITH HEADERS FROM 'file: / / / bdqualityproblem / / / qualityproblemName.csv' AS line CREATE(:qualityproblem{fault_id:line.fault_id,fault_name:line.fault_name,whether_repeat:line.whether_repeat})
[0144] CREATE CONSTRAINT ON (c:Quality Issues)ASSERT c.fault_id IS UNIQUE;
[0145] 2) Create a faulty single-machine entity node:
[0146] LOAD CSV WITH HEADERS FROM 'file: / / / bdqualityproblem / / / fault_standalone.csv' AS line CREATE(:fault_standalone{fault_standalone:line.fault_standalone})
[0147] 3) Create a physical node for the faulty satellite model and set attributes such as "satellite launch time, model name, and development unit":
[0148] LOAD CSV WITH HEADERS FROM 'file: / / / bdqualityproblem / / / satellite_model.csv' ASline CREATE(:satellite_model{lanch_time:line.lanch_time,model_name:line.model_name,research_institution:line.research_institution})
[0149] Create other entity nodes accordingly.
[0150] 4) Create an "occurrence" relationship for the faulty satellite model's "occurrence" quality problem:
[0151] LOAD CSV WITH HEADERS FROM "file: / / / bdqualityproblem / / / Model has a quality problem.csv" AS line MATCH(Satellite Model {model_name:line.model_name}),(Quality Problem {fault_name:line.fault_name}) CREATE(Satellite Model)-[:Occurred {type:line.relation}]->(Quality Problem)
[0152] 5) Create an "attributed to" relationship between the quality problem and its cause:
[0153] LOAD CSV WITH HEADERS FROM "file: / / / bdqualityproblem / / / qualityproblem attributed to the cause of the problem.csv" AS line MATCH(qualityproblem{fault_name:line.fault_name}),(cause of the cause{reason_description:line.reason_description}) CREATE(qualityproblem)-[:attributed to {type:line.relation}]->(cause of the cause of the problem)
[0154] Create other relationship nodes accordingly.
[0155] Query management based on the Cyber language:
[0156] 6) Query the head and tail entity nodes that have a relationship:
[0157] MATCH p = () - [r:`occurred`] -> () RETURN p
[0158] 7) Count the number of quality problems that occurred on each satellite:
[0159] MATCH(n: satellite model)-[r: occurrence]->()RETURN n.title,count(r)asoutgoingDegree ORDER BY outgoingDegree DESC
[0160] Example 4: Fault Description Analysis Based on Text Convolutional Neural Network
[0161] Text preprocessing is based on a thesaurus, a stop word list, a single-machine dictionary, and a model dictionary. It processes colloquial descriptive words and identifies keywords such as fault model, single-machine information, and fault component location that may be involved in the standardized fault description.
[0162] The fault description classification algorithm is implemented based on a convolutional neural network (TextCNN). The accuracy of the TextCNN algorithm is analyzed based on satellite fault description data. The fault description data is derived from 1080 quality problem cases from satellites, resulting in 2790 fault phenomenon descriptions, which are divided into four categories: fault phenomenon, executed operation, immediate impact, and other information. The training dataset contains 2290 data points, and the test dataset contains 500 data points.
[0163] Table 3. Distribution of Fault Description Datasets
[0164]
[0165]
[0166] The TextCNN convolutional neural network model consists of three convolutional layers, one pooling layer, one fully connected layer, a dropout layer, and a ReLU activation layer. The loss function is cross-entropy loss, and the optimization algorithm is the extended Adam optimizer based on stochastic gradient descent.
[0167] Table 4 Fault Description and Analysis Algorithm
[0168]
[0169] By analyzing and comparing the precision, recall, and F1 score of various fault description analysis algorithms, it can be seen that when the fault phenomenon description accounts for 60% of the support set, the recall rate is 77% and the F1 score is 66%; when the execution operation accounts for 18% of the support set, the F1 score is 31%. The accuracy of fault phenomenon identification is higher than that of other elements. Under the same number of support sets, the accuracy of consequence impact is 2% higher than that of execution operation, and the F1 score of execution operation is 10% higher than that of consequence impact. The experiment shows that increasing the number of support sets is an effective way to improve the accuracy and F1 score of fault description element analysis.
[0170] Example 5: Multi-dimensional retrieval based on a combination of keywords, representation learning, and meta-paths
[0171] By performing similarity queries on six types of quality issues—"downlink signal loss," "MERE exceeding limits," "abnormal changes in health indicators," "inter-satellite link transceiver loss," "upload receiver abnormality," and "abnormal remote terminal decryption count"—three similarity retrieval methods were analyzed and verified.
[0172] A multi-dimensional retrieval approach based on keywords, representation learning, and meta-paths compares the similarity between engineer queries and the satellite quality issue knowledge graph from three aspects: local semantics, global semantics, and structural semantics. The keyword-based retrieval method uses keywords matched from single-machine dictionaries and model dictionaries to construct a query subgraph as input. The representation learning-based retrieval method uses text vectors represented by word frequency vectors as query input. The path-based retrieval method uses meta-paths expanded from query terms as query input. The satellite quality issue knowledge graph requires graph vector representation and path mining, represented by TranR, to match retrieval terms and perform similarity comparisons.
[0173] The data of six types of duplicate problems were labeled and ranked for similarity with the case data in the satellite quality problem knowledge graph. The evaluation metrics commonly used in recommendation systems, such as Hits Ratio (HR) and Mean Reciprocal Rank (MRR), were then used.
[0174] HR emphasizes the accuracy of model recommendations, i.e., the hit rate, and is concerned with the issue of duplicate cases, i.e., whether the requested item has been recommended. HR@N is the ratio of the number of elements with a predicted ranking less than N.
[0175]
[0176] Where S is the set of all recommendations; hits(i) is used to indicate whether the i-th demand item is included in the list of recommended items by the model.
[0177] The Mean Reciprocal Ranking (MRR) emphasizes the position of user-demanded items in the model's recommendation list, highlighting order and indicating whether the recommended items are placed in a more prominent position.
[0178]
[0179] Where S is the set of all recommendations; p i This indicates the position of the requested item in the recommendation list. If the requested item is not in the list, p i It is infinitely large.
[0180] Taking a knowledge graph retrieval method based on keyword terms and representation learning as an example, this method is compared with a similarity calculation method based solely on character representation.
[0181] Table 5 Sentence Representation Algorithms and Distance Metrics
[0182] Serial Number Preprocessing Text representation Distance metric Algorithm 1 Word segmentation SimHash Haiming Distance Algorithm 2 Word segmentation Word frequency vectors cosine distance
[0183] Algorithm 1 uses Simhash for text feature representation and calculates the similarity between texts based on Hamming distance. Simhash is a local hash sensitive algorithm that obtains feature words through five steps: word segmentation, hashing, weighting, merging, and dimensionality reduction. The Hamming distance is used to calculate the number of different characters at corresponding positions in two strings of equal length; the smaller the Hamming distance, the higher the similarity.
[0184] Algorithm 2 uses word frequency vectors to represent the features of the query text. The word frequency vectors give higher weights to the keyword items. The similarity between the graph representation vector and the word frequency vector is calculated based on the cosine distance. The larger the cosine distance, the higher the similarity.
[0185] The experimental results are compared as follows:
[0186] Table 6 Comparison of the performance of the two algorithms
[0187] serial number HR@5 HR@10 MRR@5 MRR@10 Algorithm 1 0.04% 0.06% 0.02% 0.05% Algorithm 2 36.67% 28.33% 20.22% 69.79%
[0188] Experimental analysis reveals that the knowledge graph similarity calculation method based on keyword terms and representation learning outperforms the simhash calculation method based on Hamming distance, resulting in better recommendation outcomes.
[0189] Example 6: Case Study of Satellite Fault Diagnosis Based on Quality Problem Knowledge Graph
[0190] Using the problem of "satellite inter-satellite ranging values jumping by 12-15ns" as input, and the fault description input as "all 8 navigation signals monitored by the navigation signal monitoring unit experienced time delay ranging value jumps, and the telemetry and control system recovered after sending a reset command to the navigation task processing unit FPGA", the fault input description is preprocessed by word segmentation and then input into a TextCNN convolutional neural network to identify the corresponding execution operation content and fault phenomenon content. Based on the dictionary, the corresponding fault location is "navigation signal monitoring unit" as the keyword. Through similar operation comparison, it is found that "the satellite navigation task processor is designed to achieve accurate backup and recovery capabilities, and the satellite time and the status of each signal after this reset and recovery are as follows". Similar operations such as "restoring all states to their pre-reset state" were performed. Based on the fault location, the same single-unit quality problem was found: "Navigation signal unit monitoring unit B loses lock abnormality after initial on-orbit synchronization." Similar phenomena were described as "B1Ad and B1Ap signals experienced receiver lock abnormality and could not be recovered" and "Satellite telemetry feedback showed that the B1Cp signal received by the on-board navigation signal monitoring unit lost lock." Through these similar quality problem cases, possible fault causes were obtained: "SelectMapSEFI (single-event function interrupt) event occurred in the navigation signal generation FPGA2 or specific code FPGA in the navigation task processing unit" and "FPFA caused by single-event effect." Possible causes of the fault include "I / O interface anomaly (due to FPGA resource limitations, this interface lacks triple-mode redundancy)" and "the 103M counting logic of the FPGA in Monitoring B experienced a flip due to a single-event event in space." Recommended troubleshooting measures are: "First, for satellites unable to autonomously recover from faults, send a specific code reset command from the ground. If the fault cannot be recovered, then send a reset command for the navigation task processing unit FPGA. Second, for satellites capable of autonomous fault recovery, it is recommended to enable the autonomous fault recovery function; if this anomaly occurs, it can recover automatically." "Implement triple-mode redundancy for the local clock module in the FPGA configuration of Monitoring B to improve the software's resistance to single-event events. Subsequently, the on-orbit satellite software will be reconstructed according to the relevant requirements for on-orbit software reconfiguration, and maintenance measures will be implemented for other satellite units under development, drawing inferences from this experience."
[0191] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.
Claims
1. A satellite fault diagnosis system based on a quality problem knowledge graph, characterized by It includes a satellite quality problem knowledge graph construction module and a satellite fault diagnosis module based on the quality problem knowledge graph; The satellite quality issue knowledge graph construction module consists of two parts: a pattern layer construction module and a data layer construction module. The pattern layer construction module combines the experience and knowledge of satellite domain experts with case data of quality issues, and adopts an incremental approach to determine the scope of knowledge in the satellite quality issue domain, clarify the elements of the satellite quality issue domain ontology knowledge model, construct the satellite quality issue domain ontology knowledge model, and establish semantic links between different satellite quality issue domain ontology knowledge models through concept alignment. The data layer construction module performs classified knowledge extraction from the zero-out report of quality problems, the existing satellite product failure mode library, and satellite historical failure case data. It realizes the instantiation of the ontology knowledge model of the satellite quality problem domain, and obtains the instances, relations and attributes of the ontology knowledge model of the satellite quality problem domain. The key-value pairs composed of the extracted and fused instances and instance attributes are used as nodes, and relations are used as edges to form the attribute graph data model of the satellite quality problem domain. The attribute graph data model of the satellite quality problem domain is stored in the graph database to form the satellite quality problem knowledge graph. The satellite fault diagnosis module based on a quality problem knowledge graph employs text preprocessing and text convolutional neural network methods to analyze the fault descriptions of externally inputted fault phenomena. It identifies fault description keywords, standardized descriptive sentences, and category information. The standardized descriptive keywords and sentences of the fault to be diagnosed are used as query terms. A multi-dimensional retrieval interaction paradigm combining keywords, representation learning, and meta-paths is used to search the satellite quality problem knowledge graph. Finally, a troubleshooting information recommendation method based on multi-dimensional similarity calculation is employed. According to the weighted quantity aggregation principle, a set of similar cases is calculated by comprehensively considering the similarity weights of local semantics, global semantics, and structural semantics. Based on these similar cases, lists of possible problem causes, possible remedial measures, and possible faulty individual units or components are recommended to engineers for troubleshooting decisions. The concept alignment method is as follows: taking a satellite quality problem domain ontology knowledge model as a benchmark, traversing the concepts in the benchmark satellite quality problem domain ontology knowledge model, calculating the corresponding similarity of all concepts in other satellite quality problem domain ontology knowledge models, and if the similarity is less than or equal to the concept similarity threshold, then merging the two concepts into the same concept, and establishing the corresponding semantic link between the two ontology knowledge models. The similarity of the concepts M and N of the ontology is defined as: wherein, a parent class common to MNs, is a depth in the class structure, is a property domain of; When M and N are discrete values The number of elements in the domain of concept M. The number of elements in the domain of concept N. The number of elements in the set of the ranges of concepts M and N; When M and N are strings The number of characters in the string M. Let N be the number of characters in the string containing the concept N. This indicates the number of identical characters in concepts M and N; A method for recommending obstacle removal information based on multidimensional similarity calculation is adopted, as follows: The overall query result set is obtained by taking the union of the three query result sets obtained from considering keywords, sentence representation vectors, and meta-paths. Different weights are then assigned to local semantic similarity, global semantic similarity, and structural similarity. Get the query result set : in, This is a union operation; These are semantic query results based on keywords; Based on vectors; The results are from a structural similarity query. The query result set is mapped to quality problem cases, and the cases are sorted by similarity based on the semantic similarity of each object contained in each case, local and global semantic similarity, structural similarity, and ranking results. The K cases with the highest similarity scores are obtained as a set of similar fault cases according to the weighted aggregation principle, where K is greater than or equal to 1. Output the list of causes of similar failures, the list of solutions, and the list of possible individual units or components from the set of similar failure cases.
2. The satellite fault diagnosis system based on a quality problem knowledge graph according to claim 1, characterized in that... The ontology knowledge model for the satellite quality problem domain consists of classes, attributes, relations, functions, and axioms, specifically: Classes are used to describe the actual concepts in the field of satellite quality issues. In the ontology knowledge model of satellite quality issues, classes contain two layers of concepts. The first layer of concepts includes two major categories: quality problem cases and faulty satellite information. In the second layer of concepts, the concept of quality problem cases includes quality problems, causes of problems, phenomena of problems, consequences and impacts, responsible units, and handling measures. Faulty satellite information includes faulty satellite models, subsystems to which they belong, faulty units, faulty components, and interruption status. Attributes are characteristic descriptions of concepts. Attributes for quality problems include: quality problem identifier, quality problem name, whether it's a repeating attribute, and its current segment; attributes for problem causes include: cause type, cause description, and keywords; attributes for problem phenomena include: time of occurrence, domestic or international location, current stage, and phenomenon description; attributes for consequences and impacts include: whether it affects service, whether it causes single-unit failure, whether it's persistent, whether it's a software or hardware failure, whether it creates a single point of failure, and severity; attributes for responsible units include: name of the responsible unit; attributes for handling measures include: description of the measures and the implementing entity; attributes for faulty satellite models include: satellite launch time, model name, and development unit; attributes for the subsystem to which it belongs include: system name; attributes for faulty individual units include: unit name and unit failure mode; attributes for faulty components include: component name and component failure mode; and attributes for interruption status include: interruption start time and interruption end time. A relationship is a semantic description of the connection between concepts, including: "quality problem" responsibility lies with "responsible unit", "quality problem" manifestations "problem phenomenon", "quality problem" attributed to "problem cause", "satellite model" experiencing "quality problem", "quality problem" mitigated by "responsible measures", "responsible unit" taking "responsible measures", "quality problem" fault location "faulty unit", "quality problem" leading to "consequence impact", and "quality problem" interruption state is "interruption state". The relationships are: "responsibility lies with", "mitigated by", "manifestation", "attributed to", "occurrence", "taken", "fault location", "leading to consequences", and "interruption state". Functions are a special type of relation where the first N-1 elements determine the Nth element. Examples include the "dependent" relationship in "faulty single unit" belonging to "faulty component", the "include" relationship in "fault impact" containing "interruption status", and the "originating from" relationship in "problem cause" originating from "faulty single unit". Axioms are facts existing within an ontology that constrain classes and relationships within the ontology. These constraints include value constraints and common-sense constraints. Value constraints include: satellite service interruption start time < satellite interruption end time, and problem occurrence time < interruption start time. Common-sense constraints include the severity of consequences. When the attributes "whether it affects the service," "whether it causes a single machine failure," "whether it forms a single point of failure," and "whether it is persistent" are all true, the severity is higher than when only one or a few of these attributes are true.
3. The satellite fault diagnosis system based on a quality problem knowledge graph according to claim 1, characterized in that... The categorized knowledge extraction method is a structured knowledge extraction method based on R2RML or DM mapping, a semi-structured knowledge extraction method based on syntactic rule trees, or an unstructured text extraction method based on pre-trained neural network models.
4. The satellite fault diagnosis system based on a quality problem knowledge graph according to claim 1, characterized in that... Entity alignment refers to the fusion of different entity representations at the instance level, pointing them to the same objective knowledge element; the specific method is as follows: Establish an entity mapping dictionary for the satellite quality problem domain, and link entities through dictionary matching. Dictionary matching includes two methods: exact matching and fuzzy matching. The fuzzy matching method is as follows: if the minimum edit distance required to convert string A into string B is less than N / 2, string A and B are considered to belong to the same entity; otherwise, string A and B are considered not to belong to the same entity.
5. The satellite fault diagnosis system based on a quality problem knowledge graph according to claim 1, characterized in that... The method for analyzing the input fault description is as follows: Based on the thesaurus and stop word list, the fault descriptions input from external sources are standardized, colloquial expressions are removed, and synonyms are replaced. Based on the single-machine dictionary and model dictionary, the keyword information that may be involved in the standardized fault description is identified. The keyword information includes fault model, single-machine information, and fault component location. A text convolutional neural network is used to classify the output of the first step, namely the normalized fault description sentences, and to map the external fault description sentences to three categories: fault phenomenon, execution operation, and immediate impact, thereby obtaining the category information of key description sentences.
6. The satellite fault diagnosis system based on a quality problem knowledge graph according to claim 1, characterized in that... The specific steps of the keyword-based knowledge graph semantic search method are as follows: The keywords obtained after analyzing the fault description are used as search inputs. Map keywords onto the satellite quality knowledge graph to generate extended entities and relationships; Based on the generated entities and relationships, a local subgraph is generated on the satellite quality problem knowledge graph to obtain the graph structure required for structured queries; Based on the fault diagnosis objectives, the entities and relationships to be queried in the local subgraph are replaced with variables to generate a structured query subgraph; For cases where multiple entities and relationships are mapped to keywords to generate multiple local subgraphs, the set of generated structured query subgraphs is sorted and scored based on keyword search similarity. The generated structured query subgraph is submitted to the query engine to obtain query results. The TF-IDF method is used to rank the query results by relevance, resulting in semantic query results based on keywords. .
7. The satellite fault diagnosis system based on a quality problem knowledge graph according to claim 1, characterized in that... The specific steps of the knowledge graph semantic search method based on representation learning include: By learning operators, the knowledge graph of satellite quality issues is mapped to a low-dimensional continuous vector space; The semantic vectors of the classified key descriptive sentences are learned through a text vector model. By combining sentence category (entity type) information with the satellite quality problem map, nearest neighbor calculation is performed on the sentence semantic vector and the satellite quality problem map in the vector space. Similarity is then calculated based on vector distance to obtain similarity results. .
8. The satellite fault diagnosis system based on a quality problem knowledge graph according to claim 1, characterized in that... The specific steps of the knowledge graph structure semantic retrieval method based on meta-paths include: Define the entity nodes and meta-paths of extended entities corresponding to keyword items and key sentences; On the satellite quality issue knowledge graph, semantic structural similarity between entities is calculated based on meta-paths, and a nearest neighbor query algorithm is executed to obtain structural similarity query results. ; The method for calculating semantic structural similarity between entities is as follows: In the metapath From object To object The knowledge graph metapath.
Citation Information
Patent Citations
Satellite electromechanical product life prediction method based on machine deep learning
CN111177916A
Method and device for urban operation knowledge graph construction, and computer equipment
CN111191040A
