CN117687824A

CN117687824A - Satellite fault diagnosis system based on quality problem knowledge graph

Info

Publication number: CN117687824A
Application number: CN202311617451.5A
Authority: CN
Inventors: 张蕊; 张兴超; 周波; 谢宗晟; 郑恒; 应志恒
Original assignee: CHINA AEROSPACE STANDARDIZATION INSTITUTE
Current assignee: CHINA AEROSPACE STANDARDIZATION INSTITUTE
Priority date: 2023-11-29
Filing date: 2023-11-29
Publication date: 2024-03-12

Abstract

The invention relates to a satellite fault diagnosis system based on a quality problem knowledge graph, which comprises two parts of satellite quality problem knowledge graph construction and a satellite fault diagnosis method based on the quality problem knowledge graph, wherein the satellite quality problem knowledge graph construction is used for establishing a multidimensional association relation between satellite historical fault case data, the knowledge graph-based multidimensional retrieval interaction paradigm is adopted, input processing based on fault key description element recognition of a text convolutional neural network is included, and fault possible reasons and maintenance measures are recommended based on a keyword, a similarity calculation method for representing learning and meta-path combination to assist engineers in satellite health management decision, so that intelligent diagnosis of faults at a satellite system level and a single machine level is realized.

Description

Satellite fault diagnosis system based on quality problem knowledge graph

Technical Field

The invention relates to a satellite fault diagnosis system based on a quality problem knowledge graph, and belongs to the technical field of satellite fault diagnosis.

Background

With the deep integration of new generation artificial intelligence technology and advanced manufacturing technology, the digital network intelligent manufacturing management mode will remodel the full life cycle of satellite design, manufacturing, operation and maintenance and the like. Satellite fault diagnosis is an important task for satellite health management.

Satellite fault diagnosis methods include model-based methods and model-free methods.

The fault diagnosis method based on the model needs to establish an accurate mathematical model aiming at the diagnosed satellite, and is divided into a state estimation method and a parameter estimation method according to a parameter acquisition mode. It is difficult for complex satellite systems to build accurate mathematical models that are only suitable for fault diagnosis at the satellite system level.

Model-free methods include signal processing-based methods, data-driven-based methods, and knowledge-based fault diagnostics.

The method based on signal processing adopts a threshold model to perform feature extraction and information fusion on monitoring data, extracts fault information in signals through methods such as time domain analysis, wavelet analysis and the like, improves fault recognition, gradually becomes a main mode based on the traditional method and combined with a deep learning signal processing method, and has the problems of diversity of telemetry signals and redundancy and missing of part of monitoring point information in satellite actual fault diagnosis application, thereby causing uncertainty of satellite fault feature information and limiting the application of the signal processing method.

The method based on data driving uses fault diagnosis as a mode identification problem, uses a neural network, a support vector machine, a decision tree and other data analysis mining technologies to analyze the working state of the satellite according to the historical operation data of the satellite, extracts rules to judge the operation trend, and has the problems of insufficient fault sample data, lack of field knowledge, unexplained model black boxes and the like in the satellite fault diagnosis.

The knowledge-based fault diagnosis method comprises an expert system-based method and a case-based method, wherein the expert system-based method converts diagnosis knowledge into reasoning rules, and possible fault modes are screened or matched through rule reasoning. The expert system-based method relies on expert knowledge to arrange input, so that more time and energy cost are needed to be input by an expert, and the implementation difficulty is high; case-based methods typically organize case data in a structured manner, failing to mine correlations between analysis case data.

At present, satellite quality problem information is often stored in the form of documents such as return-to-zero reports, and experience knowledge such as product fault modes and processing measures are not effectively summarized and refined; the problem of data island exists among development units and subsystems, and the method can not solve the problem of collaborative quality knowledge sharing reference, so that the method is not beneficial to effective utilization of satellite historical fault data.

Disclosure of Invention

The invention solves the technical problems that: the satellite fault diagnosis system based on the quality problem knowledge graph is provided, the problems of quality knowledge sharing reference and effective utilization of satellite historical fault data are solved, and the satellite fault diagnosis efficiency is improved.

The solution of the invention is as follows: the satellite fault diagnosis system based on the quality problem knowledge graph comprises a satellite quality problem knowledge graph construction module and a satellite fault diagnosis module based on the quality problem knowledge graph;

the satellite quality problem knowledge graph construction module consists of a mode layer construction module and a data layer construction module; the model layer construction module is used for combining expert experience knowledge in the satellite field with quality problem case data, determining the knowledge range in the satellite quality problem field by adopting an incremental method, determining the ontology knowledge model elements in the satellite quality problem field, constructing an ontology knowledge model in the satellite quality problem field, and establishing semantic links among ontology knowledge models in different satellite quality problem fields by adopting a concept alignment method; the data layer construction module is used for carrying out classified knowledge extraction on quality problem return-to-zero reports, an existing satellite product fault mode library and satellite historical fault case data, realizing instantiation of a satellite quality problem domain ontology knowledge model, obtaining an instance, a relation and an attribute of the satellite quality problem domain ontology knowledge model, taking a key value pair consisting of the instance and the instance attribute obtained after extraction and fusion as a point and a relation as an edge, forming a satellite quality problem domain attribute map data model, and storing the satellite quality problem domain attribute map data model into a map database to form a satellite quality problem knowledge map;

The satellite fault diagnosis module based on the quality problem knowledge graph carries out fault description analysis on externally input fault phenomena by adopting a text preprocessing method and a text convolution neural network method, identifies fault description keyword term information and normalized description sentences and category information, takes the obtained fault normalization description keyword terms and sentences to be diagnosed as query terms, adopts a multi-dimensional retrieval interaction paradigm based on combination of keywords, representation learning and element paths to retrieve the satellite quality problem knowledge graph, finally adopts a fault removal information recommendation method based on multi-dimensional similarity calculation, calculates a similar case set by comprehensively considering three dimensional similarity weights of local semantics, global semantics and structural semantics according to a weighted quantity aggregation principle, and recommends a possible problem reason list, a possible disposal measure list and a possible fault or single machine part list to an engineer for fault removal decision by referring to the similar cases.

Preferably, the satellite quality problem domain ontology knowledge model is composed of classes, attributes, relationships, functions and axioms, and specifically:

the system comprises a satellite quality problem domain ontology knowledge model, a satellite quality problem domain ontology knowledge model and a satellite information model, wherein the satellite quality problem domain ontology knowledge model is used for describing actual concepts of the satellite quality problem domain, the satellite quality problem domain ontology knowledge model comprises two layers of concepts, the first layer of concepts comprise quality problem cases and fault satellite information, and the second layer of concepts comprises quality problems, problem reasons, problem phenomena, effect influences, responsibility units and disposal measures, and the fault satellite information comprises fault satellite models, affiliated subsystems, fault single machines, fault machine components and interrupt states;

The attributes are feature descriptions of the concept, and the attributes of the quality problem include: quality problem identification, quality problem name, whether the attribute is repeated or not, and the arc section is located; the attributes of the cause of the problem include: the reason type, reason description, keywords; the attributes of the problem phenomenon comprise the occurrence time, the overseas and overseas, the stage where the problem occurs and the phenomenon description; the attributes affected by the consequences include whether to affect the service, whether to cause a single machine failure, whether to reside, software or hardware failure, whether to form a single point failure, severity; the responsibility unit attribute comprises a responsibility unit name; the attribute of the treatment measure comprises a measure description and an execution subject; the attributes of the fault satellite model comprise satellite transmitting time, model name and development unit attributes; the subsystem comprises a system name; the fault single machine attribute comprises a single machine name and a single machine fault mode attribute; the failure component attributes include a component name, a component failure mode attribute; the interrupt state attribute comprises interrupt start time and interrupt end time;

the relation is a semantic description of the connection relation between concepts, and comprises that the responsibility of the quality problem is in a responsibility unit, the quality problem represents a problem phenomenon, the quality problem is due to a problem reason, the satellite model generates a quality problem, the quality problem is relieved in a disposal measure, the responsibility unit takes the disposal measure, the quality problem fault part faults single machine, the quality problem causes a result effect influence, the quality problem interrupt state is in an interrupt state, and the responsibility is in a release state, the performance is due to the occurrence, the fault part is adopted, the result effect is caused, and the interrupt state is caused;

The function is a special relation, the former N-1 element determines the N element, and the N element comprises a ' subordinate ' relation in a ' subordinate ' fault unit part ', an ' containing ' relation in a ' fault influence ' containing ' interrupt state ' and a ' problem reason ' originating from a ' originating ' relation in a ' single fault unit ';

axioms are facts that exist within a body, constraining classes and relationships within a body: including value constraints and common sense constraints, wherein the value constraints include: satellite service interruption start time < satellite interruption end time, problem occurrence time < interruption start time; common sense constraints include that, in the consequences, the attributes "whether to affect a service", "whether to cause a stand-alone failure", "whether to form a single point failure", "whether to reside" are all more severe than if only one or a few of them are all present.

Preferably, the concept alignment method is as follows: traversing concepts in the ontology knowledge model of the domain of the satellite quality problem by taking the ontology knowledge model of the domain of the satellite quality problem as a reference, calculating corresponding similarity with all concepts of the ontology knowledge model of the domain of other satellite quality problems, merging the two concepts into the same concept if the similarity is smaller than or equal to a concept similarity threshold value, and establishing semantic links corresponding to the concepts between the two ontology knowledge models;

The similarity prox (M, N) of the concepts M and N of the ontology is defined as:

wherein P is _MN For a parent class shared by MNs, depth (x) is the depth of x in the class structure, and dom (x) is the attribute definition field of x;

when M and N are discrete values, dom (M) is the number of elements in the concept M value domain, dom (N) is the number of elements in the concept N value domain, and dom (M) N is the number of elements in the collection of the concept M and the concept N value domain;

when M and N are character strings, dom (M) is the number of characters in the character string of the concept M, dom (N) is the number of characters in the character string of the concept N, and dom (M) ≡dom (N) represents the same number of characters of the concept M and the concept N.

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 a syntax rule tree or an unstructured text extraction method based on a pre-training neural network model.

Preferably, entity alignment refers to instance layer fusion of different entity expressions to the same objective knowledge element at the instance layer; the specific method comprises the following steps:

establishing a satellite quality problem field entity mapping dictionary library, and carrying out entity linking through dictionary matching, wherein the dictionary matching comprises two modes of complete matching and fuzzy matching, and the fuzzy matching mode is as follows: when the minimum editing distance required for converting the character string A into the character string B is less than N/2, the character strings A and B are considered to belong to the same entity; otherwise, the character strings A and B are not considered to belong to the same entity; the edit distance of the character strings A and B is lev _A,B (|a|, |b|), and|a|, |b| are the string lengths of the current comparison:

wherein lev is _A,B (|a|, |b|) represents the edit distance between the first i characters of a and the first j characters of b. Where i, j are indices starting from 1.

Preferably, the fault description analysis method for the input is:

based on the synonym list and the stop word list, carrying out standardization processing on externally input fault description, removing spoken language expression and replacing synonyms;

based on the single machine dictionary and the model dictionary, identifying keyword term information possibly related to the fault description after normalization processing, wherein the keyword term information comprises fault models, single machine information and fault component parts;

and classifying the fault description sentences subjected to the normalized processing output in the first step by adopting a text convolutional neural network, and corresponding the external fault description sentences to three categories of fault phenomenon, execution operation and instant influence to obtain the key description sentence category information.

Preferably, the keyword-based knowledge graph semantic search method comprises the following specific steps of:

the keyword term is obtained after the fault description analysis and is used as search input;

mapping the keyword term on a satellite quality knowledge graph to generate an extended entity and relationship;

Generating a local subgraph on a satellite quality problem knowledge graph based on the generated entity and relation to obtain a graph structure required by structural query;

according to the fault diagnosis target, replacing the entity and the relation to be queried in the local subgraph with variables to generate a structured query subgraph;

for the situation that a plurality of entities and relations are mapped by keywords to generate a plurality of local subgraphs, ranking and scoring the generated set of structured query subgraphs according to the search similarity of the keywords;

submitting the generated structured query subgraph to a query engine to obtain a query result, and performing relevance ranking on the query result by adopting a TF-IDF method to obtain a semantic query result { a ] based on keywords ₁ ,a ₂ ,a ₃ ,....,a _n }。

Preferably, the semantic searching method based on the knowledge graph representing learning comprises the following specific steps:

mapping the satellite quality problem knowledge graph to a low-dimensional continuous vector space through a learning operator;

the classified key description sentences are learned through a text vector model to obtain semantic vectors;

combining sentence category, namely entity type information, executing nearest neighbor calculation on sentence semantic vector and satellite quality problem map in vector space, and based on vectorDistance calculation similarity, obtaining a similar result { b } ₁ ,b ₂ ,b ₃ ,....,b _n }。

Preferably, the knowledge graph structure semantic retrieval method based on the meta-path comprises the following specific steps:

defining entity nodes corresponding to the key terms and the key sentences and meta paths of the extended entities;

on the satellite quality problem knowledge graph, according to the semantic structure similarity among the meta-path calculation entities, executing a nearest neighbor query algorithm to obtain a structure similarity query result { c } ₁ ,c ₂ ,c ₃ ,.....,c _n }；

The method for calculating the similarity of the semantic structures among the entities comprises the following steps:

p _x→y refers to a knowledge-graph meta-path instance from object x to object y under meta-path P.

Preferably, a barrier removal information recommendation method based on multidimensional similarity calculation is adopted, and the method specifically comprises the following steps:

the three query result sets obtained by considering the keyword term, the sentence expression vector and the meta path are combined to obtain a total query result set, and different weights w are respectively given to the local semantic similarity, the global semantic similarity and the structural similarity ₁ ,w ₂ ,w ₃ Obtaining a query result set R:

R＝w ₁ {a ₁ ,a ₂ ,a ₃ ,....,a _n }∨w ₂ {b ₁ ,b ₂ ,b ₃ ,....,b _n }∨w ₃ {c ₁ ,c ₂ ,c ₃ ,.....,c _n }

wherein V is union operation.

Mapping the query result set to quality problem cases, and sorting cases according to the semantic similarity, local and global semantic similarity, structural similarity and sorting results of each case, wherein each case comprises each object; obtaining K cases with highest similarity scores as a similar fault case set according to a weighted aggregation principle, wherein K is more than or equal to 1;

The list of problem causes, the list of treatment measures, the list of possible standalone or machine components in the similar failure case set are output.

Compared with the prior art, the invention has the beneficial effects that:

(1) According to the invention, the knowledge graph is adopted to manage satellite quality problem data, so that multidimensional association analysis of satellite historical fault cases is realized, and compared with a storage management method of structured data in the prior art, a shared analysis basis of a satellite quality problem knowledge system and a semantic network is established from top to bottom, and the efficiency of satellite quality problem case data management application is improved.

(2) The fault diagnosis process adopts a keyword term based on a quality problem knowledge graph, a representation learning and meta-path multidimensional semantic retrieval interaction paradigm, and based on local semantic similarity, global semantic similarity and structural semantic similarity, a case set with the highest weighted dimension aggregation similarity value is recommended to engineers to be used as a treatment reference, so that the fault treatment decision capability is improved, and meanwhile, the accuracy of a given solution is ensured.

Drawings

FIG. 1 is a satellite fault diagnosis system architecture based on a quality problem knowledge graph according to an embodiment of the present invention;

FIG. 2 is a flow chart of the ontology knowledge model construction in the satellite quality problem domain according to an embodiment of the present invention;

FIG. 3 is a view of an ontology model for satellite quality problem domain according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a method for identifying entities based on Bi-LSTM model in unstructured knowledge extraction according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of a relationship identification method based on convolutional neural network in unstructured knowledge extraction according to an embodiment of the present invention;

FIG. 6 is a knowledge graph visualization of satellite quality problems based on a Neo4j graph database according to an embodiment of the present invention;

fig. 7 is a flowchart of satellite fault diagnosis based on a quality problem knowledge graph according to an embodiment of the invention.

Fig. 8 is an example of a mapping principle of knowledge graph structured data R2RML with satellite quality problem according to an embodiment of the present invention.

Detailed Description

The invention is further illustrated below with reference to examples.

1. Knowledge graph of satellite quality problem

The satellite quality problem knowledge graph technical architecture is shown in fig. 1, logically consists of a mode layer and a data layer, wherein the mode layer is formalized description of concepts and relations in the satellite quality field and is modeled through an ontology organization; the data layer instantiates concepts based on the mode layer logic, and the method is a concrete embodiment of concept relations in the satellite quality problem field. And combing the knowledge in the satellite quality problem field, establishing a satellite quality problem field ontology knowledge model, extracting entities and relations from the structure and unstructured data such as a zeroing report and the like according to the ontology knowledge model, and constructing a satellite quality problem knowledge graph from top to bottom. The mode layer carries out knowledge modeling in the satellite quality problem field based on the ontology, and semantic unification and knowledge sharing are realized among different unit heterogeneous information sources through concept alignment; based on the specification of the ontology knowledge model, guiding the types of specific examples, relations and attribute elements, obtaining basic elements forming a knowledge graph from structured, semi-structured and unstructured data by adopting a knowledge extraction technology; in order to eliminate ambiguity of various knowledge elements, entity alignment technology is adopted to combine similar semantic items, so that element redundancy of the knowledge graph is reduced, and high quality of the constructed knowledge graph is ensured. After the basic construction of the knowledge graph is completed through the three steps, the fault diagnosis application is developed on the knowledge graph based on the knowledge graph multidimensional semantic retrieval technology.

1.1 satellite quality problem ontology knowledge modeling

The ontology modeling method expresses complex concepts special for the field by integrating and constructing attributes, relations, functions and axioms among the concepts, forms a standardized expression mode of the field concepts, and provides a unified description standard for knowledge modeling and information sharing of the satellite quality problem field. Combining expert knowledge in the satellite quality problem field with quality problem case data to establish a satellite quality problem field ontology knowledge model.

The construction flow of the ontology knowledge model in the satellite quality problem field is shown in fig. 2, firstly, the knowledge domain of the satellite quality problem field is determined by combining with the expert advice in the satellite quality problem field, the clear modeling level is the premise that the knowledge graph correctly plays an analysis role, too few knowledge elements limit the space for analyzing and mining the knowledge graph, and knowledge level mixing causes the knowledge graph to lack a semantic main body and generate modeling deviation. The semantic range of the quality problem and the hierarchical relation of the semantic range and relevant quality elements such as satellite structure function information, product supply information and the like are determined, and the expression mode is the problem which needs to be considered in the step. And secondly, acquisition and analysis of quality problem knowledge are data bases of ontology modeling. Besides the engineering data such as return-to-zero reports, the quality of the quality problems of the aerospace products and the reliability information classification standard are important reference contents, and the industry standards such as guidelines for establishing fault mode libraries of the aerospace products are all important. And establishing a knowledge model of the satellite quality problem field based on the determined knowledge range and the data base. The ontology knowledge model in the satellite quality problem field mainly comprises factors such as concepts, attributes, relations, functions, axioms and examples. The method comprises the steps of taking a fault and satellite model equipment object as traction, wherein a core concept comprises two major categories of quality problem cases and fault satellite information, the quality problem cases comprise concepts of quality problems, problem reasons, problem phenomena, consequence states, responsibility units, arc sections and the like, and the fault satellite information comprises four structural layers of a satellite model, a subsystem, a fault single machine and a fault device component and a corresponding fault mode. The attribute is description and supplement to the concept feature, for example, the problem reasons include attributes such as reason type, reason description, keywords, etc., the effect influence includes attributes such as whether to influence service, whether to cause single machine failure, whether to stay, etc., the attributes include three types of object type, data type and annotation type, and the data type attribute types are usually character string type and boolean type. Relationships are descriptions of interactions between concepts, mainly including cause, responsible, performance, attribution, etc. The functional axiom is to describe the constraint and rule of the relation among concepts, build constraint association on the basis of combining the characteristics of autoreactivity, transitivity, symmetry and the like, describe the true or false of the assertion, and is the basis of carrying out logic detection on the ontology. Examples are conceptual representations that can increase the scope of typical examples or examples that a clear concept contains when modeling an ontology. After six major elements of the satellite quality problem ontology are defined, modeling visualization and logic detection are performed based on protein software, and the satellite quality problem ontology knowledge model visualization management is shown in fig. 3.

1.2 satellite quality problem knowledge acquisition

The main step of satellite quality problem knowledge acquisition comprises the steps of widely collecting quality problem information and extracting knowledge. The satellite quality problem information source channel comprises a quality problem return report, an existing satellite product fault mode library, related quality problem cases at home and abroad and the like, and entities and relations in satellite quality problem knowledge graph examples are extracted from the information.

The knowledge extraction for structured data mainly converts relational database data into RDF (resource description framework) data or OWL ontology through a mapping language, the mapping language comprises DM (direct mapping) and R2RML, the DM directly converts the relational database table structure and the data into RDF graphs, the R2RML mapping can flexibly customize views on the relational data, the relational database conforming to the mode is input, and RDF data described by predicates and types in a target vocabulary is output. The entity extraction and relation extraction method for unstructured data is mainly divided into three types, namely a rule-based method, a statistical learning method and a deep learning method, wherein the rule-based method is suitable for unstructured data with fixed formats and standardized expression, and pattern matching is performed by manually writing syntax rules to identify entities and relation elements; the statistical learning method is based on labeling corpus to perform model training, word characteristics needing to be defined before training, the deep learning method does not depend on characteristic engineering, word vectors are used as input, the contexts are learned through a neural network, and entities and relations are identified end to end.

1.3 satellite quality problem knowledge fusion

Basic elements for constructing the knowledge graph are obtained through knowledge acquisition, but the information sources are various, the knowledge elements from different sources often have the problems of various semantic expressions, information redundancy and the like, the establishment of association relations among the knowledge graphs is not facilitated, barriers for knowledge sharing exist, the specific expression is that the names of product composition structures with consistent functions among different development units are different, such as a navigation subsystem and a load subsystem, a fundamental frequency processor and a reference frequency synthesizer, the computer cannot establish the association among the elements with different semantic expressions, so that the existence of individual semantic islands in the seeked associated knowledge graph cannot be realized, the association analysis and application cannot be carried out based on the knowledge graph, and the interconnection commonality of the knowledge elements is the key for playing the application role of the knowledge graph with satellite quality problem. Knowledge fusion techniques include both entity alignment and ontology alignment.

And (3) directing different entity expressions to an objective knowledge element through an entity alignment technology to perform instance layer fusion. Considering the characteristics of the satellite quality engineering field, the differences of entity references are often explicitly embodied in documents such as engineering manuals, and the entity mapping dictionary library is considered to be established, entity links are carried out through dictionary matching, the matching mode comprises two modes of complete matching and fuzzy matching, the fuzzy matching can adopt modes such as character string similarity calculation based on editing distance, and the editing distance of character strings A and B is lev _A,B (|a|, |b|), and|a|, |b| are the string lengths of the current comparison:

The fusion of concept layers is carried out by considering the alignment of the ontology among different unit knowledge graphs, the alignment of the concept can be measured by considering the expert experience and engineering knowledge, the similarity of the concept can be measured by two aspects of similarity of similar structures and similarity of attributes, the higher the hierarchy of the concept is and the higher the attribute overlap ratio is, the greater the probability of the ontology similarity is considered, and the similarity prox (M, N) of the concepts M and N of the ontology is defined as:

wherein P is _MN For a parent class common to MNs, depth (x) is the depth of x in the class structure, and dom (x) is the attribute definition field of x. In one embodiment of the present invention, the concept similarity threshold is 0.75.

1.4 satellite quality problem knowledge storage and retrieval

The knowledge graph is essentially graph data from the perspective of a data model, and comprises an RDF graph and an attribute graph, wherein RDF is a standard data model capable of representing and exchanging machine understandable information, is defined as a limited set of triples (s, p, o), is represented as s and o and has a relation p or the value of the attribute p of s is o, and each resource has an HTTP URI as a unique identifier thereof, so that the knowledge graph is suitable for a knowledge sharing scene; the attribute graph is the most widely adopted graph data model, nodes and edges are provided with a group of key pair value form attributes, all data of the RDF graph can be expressed, edge weight support quantization analysis can be added, and the graph data model is suitable for knowledge mining analysis scenes. The query retrieval language on the RDF graph is a declarative query language SPARQL, and supports various query retrieval mechanisms such as a triplet mode, a sub-graph mode, an attribute path and the like; the query languages commonly used on the attribute graph are Cypher and Gremlin languages, which are both path-oriented query languages, and Cypher is simple and rapid, and Gremlin has more advantages in high-level traversal. RDF-oriented triples databases include RDF4J, graphDB, and attribute map-oriented databases include Neo4j and JanusGraph. For supporting quality problem comprehensive analysis application, key value pairs consisting of the extracted and fused examples and example attributes are used as points, relations are used as edges, a satellite quality problem field attribute map data model is formed, storage is carried out based on Neo4j, and a cypher query language is supported. The Neo4j database has an ACID transaction processing function, prevents data loss under hardware failure or system breakdown, and ensures the safety of satellite data.

2. Satellite fault diagnosis method based on quality problem knowledge graph

The satellite quality problem knowledge graph application is used as a core for fault diagnosis, and as shown in fig. 7, the method comprises 1) analyzing a fault description input by an engineer; 2) Retrieving an interaction paradigm based on keywords, representing a combination of learning and meta-paths; 3) The obstacle-removing information recommendation method based on multidimensional similarity calculation; based on the 3 steps, preprocessing fault description information input by engineering, performing semantic retrieval and structural retrieval on a satellite quality problem knowledge graph as query input, and finally recommending information such as reasons, measures and the like of similar quality problem cases in the graph to engineers to perform obstacle avoidance decision by combining a multi-dimensional similarity calculation result. Wherein the engineer input fault description analysis and multi-dimensional retrieval paradigm are important bases for determining whether the returned troubleshooting information is valid.

2.1 text convolutional neural network based fault description analysis

The fault description analysis aims to comprehensively and accurately analyze fault information input by a user and improve the expression capability of an engineer for inputting inquiry. The input description fault phenomenon often has the problems of spoken language, word diversity, complex description elements and the like, and a fault description analysis method is provided, and comprises text preprocessing and sentence element classification based on a text convolutional neural network (textCNN). Firstly, identifying fault models and single machine information based on three types of word lists of a field professional word list, a synonym word list and a stop word list, carrying out standardization processing on fault description, and identifying field professional words; secondly, identifying and classifying key description elements by using a text convolutional neural network; the requirements for inputting descriptions by engineers are reduced through the two steps, and simultaneously, fault descriptions are normalized.

The text preprocessing method processes spoken language description input by an engineer and matches keyword terms according to a domain dictionary and comprises 1) based on a synonym table and a stop word table, carrying out standardization processing on fault description input by the engineer, removing spoken language expression and replacing synonyms; 2) Based on the single machine dictionary and the model dictionary, identifying keyword item information such as fault models, single machine information, fault component parts and the like which are possibly related to the fault description after the standardized processing;

and classifying the fault description sentences by using a text convolutional neural network to obtain fault description sentence element category information. Training a text convolutional neural network (textCNN) model based on case data in a knowledge graph, inputting normalized fault description contents into the text convolutional neural network model to classify elements of fault description after training, classifying input fault description sentences into element categories such as phenomenon description, execution operation, instant influence and other information, and establishing a corresponding relation between fault description elements and entity types in a quality problem knowledge graph through classifying and identifying the fault description.

2.2 Multi-dimensional retrieval interaction paradigm based on keyword, representation learning and meta-path combination

Through text preprocessing and text convolutional neural network classification, query terms are constructed, including keyword terms (possibly related to fault models, single machine information, fault component parts and other terms), normalized fault description sentences and corresponding entity node types, the query terms are required to be used as input to perform search on satellite quality problem knowledge graphs, search calculation is performed from three dimensions of term semantic features, local semantic features, global semantic features and structural features, and the query terms are specifically divided into a keyword-based knowledge graph semantic search method, a learning-representation-based knowledge graph semantic search method and a meta-path-based knowledge graph structure search.

2.2.1 keyword-based knowledge graph semantic search method

The knowledge graph semantic search method based on the keywords takes the keyword terms as query input, intuitively expresses the query information requirement of engineers, and generates subgraphs based on the keyword terms as candidate structured queries to perform search interaction.

And taking key word items such as fault models, single machine information, fault component parts and the like which are possibly obtained by fault description analysis as search input, constructing candidate structured query composed of corresponding entities and relations according to mapped key words, generating local subgraphs on a satellite quality problem knowledge graph based on the generated entities and relations, obtaining a graph structure required by the structured query, replacing part of the entities and relations in the local subgraphs with variables, generating a structured query subgraph, sequencing and grading a generated structured query set according to the search similarity of the key words when the key words map a plurality of the entities and relations to generate a plurality of the local subgraphs, submitting the generated structured query to a query engine to obtain a query result, and sequencing the query result by adopting a TF-IDF method to obtain a semantic query result based on the key words.

2.2.2 knowledge graph semantic search method based on representation learning

The semantic search method of the knowledge graph based on representation learning simultaneously considers local semantic features and global semantic features of a semantic network, obtains semantic vectors from classified key description sentences through a text vector representation algorithm, and executes K-nearest neighbor search sequencing in a low-dimensional continuous vector space by combining sentence types, namely entity type information.

Firstly, mapping a knowledge graph of a satellite quality problem to a low-dimensional continuous vector space through a TransE, transH translation mechanism and the like to represent a learning operator, taking entities and relations in the knowledge graph as two matrixes by a TransE, respectively extracting a vector from the entity and the relation matrix after training the TransE to calculate L1 or L2 to obtain a vector of the other entity in the entity matrix, and therefore, representing triples in the knowledge graph. The scoring function after projection is:

Wherein h, t respectively represent vectors of a head entity node and a tail entity node in the triplet, and w _r Normal vector representing relational hyperplane, d _r Representing translation vectors of the relational hyperplane. If the triplet relationship is correct, the smaller the result value, and conversely, the larger the result.

Using a margin-based ranking function as a loss function:

L＝∑ _(h,r,t)∈Δ ∑ _{(h'，r',t')∈Δ'} max{0,f _r (h,t)+γ-f _r' (h',t')}

wherein, delta represents a set of correct triples, delta' represents a set of negative examples, gamma represents an edge distance value for distinguishing positive examples from negative examples, and the loss function is trained by a mini-SGD optimizer to obtain satellite quality problem knowledge graph vector representation. And carrying out a text vector representation method on the input fault description sentence, and carrying out K-nearest neighbor calculation on the sentence semantic vector and the satellite quality problem map in a vector space to obtain a similar query result by combining the fault description sentence category information.

2.2.3 method for searching knowledge graph structure based on meta path

The meta path is a relation sequence of link object types, can capture structural semantic information in a satellite quality problem knowledge graph, and is based on a knowledge graph structure retrieval method of the meta path, the relation structure characteristics of the satellite quality problem knowledge graph are considered, the meta path is defined for entity nodes corresponding to key terms and key sentences, and a nearest neighbor algorithm is executed through path similarity to obtain a similar query result, and the method specifically comprises the following steps:

1) Defining entity nodes corresponding to key terms and key sentences and extending the meta-path of the entity, e.g

2) Executing a single-object nearest neighbor query algorithm on the satellite quality problem knowledge graph to obtain a structural similarity query result { c } ₁ ,c ₂ ,c ₃ ,.....,c _n The similarity calculation method between }. entities comprises the following steps:

p _x→y the knowledge graph meta-path example from the entity object x to the entity object y under the meta-path P is that if the meta-paths between two entities are more similar, the more the connection points are, the more centrality of the entity nodes are indicated, and the repetition rate of the corresponding quality problem cases is higher. Path mining may further be performed by depth traversal, random walk, etc.

2.3 obstacle-removing information recommendation method based on multidimensional similarity calculation

According to the obstacle-removing information recommendation method based on multi-dimensional similarity calculation, according to a weighting quantity aggregation principle, a similar case set is calculated by comprehensively considering the dimension similarity weight and the number of the matched case node neighbors, and the higher the multi-dimensional similarity is, the more the number of the neighboring nodes is, and the more the case is in front in a recommendation list. The obstacle-removing information recommendation method based on multidimensional similarity calculation comprises the following steps:

1) The three query result sets obtained by considering the keyword, the sentence expression vector and the meta path are combined to obtain a total query result set, and different weights w are respectively assigned to the semantic similarity of the term, the local semantic similarity and the global semantic similarity and the structural similarity ₁ ,w ₂ ,w ₃ Obtaining a query result set R:

wherein V is a union operation, { a ₁ ,a ₂ ,a ₃ ,....,a _n The query result set returned by the keyword-based knowledge graph semantic search method is } { b } ₁ ,b ₂ ,b ₃ ,....,b _n The query result set returned by the knowledge graph semantic search method based on representation learning is { c } ₁ ,c ₂ ,c ₃ ,.....,c _n And the query result set returned by the knowledge graph structure searching method based on the meta-path.

2) The query result set is corresponding to the quality problem cases, and the cases are ordered according to the semantic similarity, the local and global semantic similarity, the structural similarity and the ordering result of each object; obtaining K cases with highest similarity scores as a similar fault case set, wherein K is greater than or equal to 1;

entity node a corresponding to result set _o ,b _p ,c _q Corresponding to case _j Upper (0 < o, p, q < =n, 0 < j < =k)

Case _j The importance weight of (2) is the multi-dimensional similarity w of the entity nodes consisting of the importance weights _i (i=1, 2, 3) and entity location importance 1/o 1/p 1/q decisions in the single-dimensional query result set;

3) And recommending the problem reasons and the treatment measures in the similar fault case sets to engineers for fault removal reference. And obtaining a candidate reason list after grading and sorting, sequentially executing diagnosis steps by operation and maintenance personnel according to the list, and completing final reason positioning according to the satellite feedback state.

The maintenance measures among the same faults caused by the same fault cause can be mutually referred to. In addition to the maintenance measures of quality problem knowledge map fault cases, reference is made to multi-source information fusion such as fault removal experience, treatment plans, maintenance manuals and the like of operation and maintenance engineers, and the process considers the difference among multi-source fault maintenance knowledge, carries out fusion treatment on the multi-source maintenance knowledge, and generates the fused maintenance measures for the engineers to refer to.

The quality problem of satellite on-orbit operation has the characteristics of multiple professions, wide involved areas and multiple sources and layers of information, and the quality problem information of satellites of different types needs to be induced and analyzed by intelligent technologies such as a knowledge graph and the like, so that the construction and application from quality problem data to a knowledge system are realized.

The satellite fault diagnosis mode based on the knowledge graph organizes data such as quality problem historical cases and the like in a knowledge graph mode, performs similarity retrieval calculation on fault description input by engineers and historical case data from three aspects of semantic network local features, global features and structural features, and finally returns the calculated historical fault case obstacle removing information with highest multidimensional similarity to the engineers to support obstacle removing decision, thereby improving the efficiency and accuracy of fault diagnosis analysis.

Example 1: satellite quality problem ontology knowledge model construction and instantiation

The method comprises the steps of establishing a satellite quality problem ontology knowledge model, and defining six factors of the satellite quality problem ontology model, wherein functional axiom is constraint and rule description of relations among concepts, constraint association is established on the basis of characteristics such as combination reflexibility, transmissibility and symmetry, and description of true and false of assertion is a basis for logic detection of an ontology. Examples are conceptual representations that can increase the scope of typical examples or examples that a clear concept contains when modeling an ontology. Modeling is performed based on protein software, and logic detection is performed based on a protein inference engine.

Taking the SADA quality problem of the solar array driving structure as an example to construct a quality problem knowledge graph, organizing a SADA quality problem data layer of the solar array driving structure based on an ontology model, and constructing information of the ontology and the example as shown in the following table.

TABLE 1 knowledge graph ontology of SADA quality problem and corresponding instance information

Example 2: knowledge acquisition of satellite quality problems

The knowledge extraction for structured data mainly converts relational database data into RDF (resource description framework) data through a mapping language, wherein the mapping language comprises DM (direct mapping) and R2RML, the DM directly converts the relational database table structure and the data into RDF graphs, the R2RML mapping can flexibly customize views on the relational data, input a relational database conforming to the mode, and output RDF data described by predicates and types in a target vocabulary.

Converting a satellite quality problem structured data table into an RDF triplet, wherein the structured data table comprises a subsystem name, a single machine name, a fault part name, a fault phenomenon and a fault reason; the satellite quality problem structured data table takes a single machine name as a main body; the satellite quality problem structuring data table subsystem, the name of the fault part, the fault phenomenon and the fault reason are listed as objects; the relationship between the subject and the object is defined as an attribute, specifically, the relationship between the name of the single machine and the subsystem is located as the subsystem to which the object belongs, the relationship between the single machine and the failure part is defined as a failure part, the relationship between the name of the single machine and the phenomenon is defined as a representation, and the relationship between the name of the single machine and the cause is defined as a cause. The example structured data is thus converted into 8 RDF triples, as shown in fig. 8.

The entity extraction and relation extraction method for unstructured data is mainly divided into three types, namely a rule-based method, a statistical learning method and a deep learning method, wherein the rule-based method is suitable for unstructured data with fixed formats and standardized expression, and pattern matching is performed by manually writing syntax rules to identify entities and relation elements; the statistical learning method is based on labeling corpus to perform model training, word characteristics which need to be defined and input before training, wherein the word characteristics comprise word level characteristics, dictionary level characteristics and document level characteristics, and the models mainly adopted by entity recognition comprise a hidden Markov model, a maximum entropy model, a conditional random field model and the like; the deep learning method does not depend on feature engineering, takes word vectors as input, learns the context through a neural network, and identifies entities and relations end to end. Entity identification is typically used as a sequence annotation problem, marking the beginning, middle, and ending portions of an entity in the manner of "BIO" or "BIOES", etc. Relationship extraction is typically used as a text classification problem, classifying onto corresponding relationship categories given sentences and head-to-tail entities. The advantages of the comprehensive statistical method and the deep learning method are that a BiLSTM-CRF model is shown in fig. 4, a long-short-time memory network and a conditional random field are combined to perform entity recognition, the method comprises a vector representation layer, a bidirectional long-short-time memory network layer and a conditional random field layer, the vector representation layer is obtained through word vector representation model learning and is used as input of the bidirectional long-short-time memory network, the bidirectional long-short-time memory network captures sentence long-distance dependent information from the forward direction and the reverse direction, and the conditional random field is used for marking satellite entities in sentences based on vector splicing input of the two directions. The relation extraction method based on supervised learning has large demand on the training corpus, a PCNN remote supervised relation extraction model is shown in fig. 5, and the relation classification is accurately performed by reducing noise interference in the existing knowledge graph based on a convolutional neural network segmentation pooling strategy and a multi-instance learning mode while the constructed satellite knowledge graph is utilized to align the text-rich training corpus. The knowledge extraction mode based on the neural network model requires a certain implementation cost on training corpus and model debugging, and is suitable for processing massive unstructured data.

Example 3: satellite quality problem map construction and storage management

According to the satellite quality problem knowledge graph ontology knowledge model, triplet data are acquired through a structured knowledge acquisition mode and an unstructured knowledge acquisition mode and are imported into a neo4j database in a form of entity and relation triples, a visual graph is shown in fig. 6, and importing and inquiring execution commands are as follows:

1) Creating a quality problem entity node, and creating a quality problem ID as a unique identification:

LOAD CSV WITH HEADERS FROM 'file:///bdqualityproblem// quality problem name csv' AS line CREATE (: quality problem { fault_id: line. Fault_id, fault_name: line. Fault_name, router_repeat: line. Router_repeat })

CREATE CONSTRAINT ON (c: quality problem) ASSERT c.fault_id IS UNQUE;

2) Creating a fault single entity node:

LOAD CSV WITH HEADERS FROM 'file:///bdqualityproblem///faulty stand-alone csv' AS line CREATE: faulty stand-alone { fault_stand-alone: line. Fault_stand-alone }

3) Creating a fault satellite model entity node, and setting attributes such as satellite transmitting time, model name, development unit and the like:

LOAD CSV WITH HEADERS FROM 'file:///bdqualityplbem// satellite model csv' AS line CREATE (: satellite model { lanch_time: line.lanch_time, model_name: line.model_name, research_inst-tion: line.research_construction })

Other entity nodes are created accordingly.

4) Creating a "generating" relation of the quality problem of the fault satellite model:

LOAD CSV WITH HEADERS FROM "File:///bdqualityproblem// model quality problems occur csv" AS line MATCH (satellite model { model_name: line. Model_name }), (quality problems { fault_name: line. Fault_name }) CREATE (satellite model) - [: types: line. Resolution } ] - >

5) Creating a quality problem "due to" relationship "due to" the cause of the problem:

LOAD CSV WITH HEADERS FROM "file:///bdqualityproblem///quality problem is due to problem cause csv" AS line MATCH (quality problem { fault_name: line. Fault_name }), (problem cause { reflection_description: line. Reflection_description }) CREATE (quality problem) - [ (due to { type: line. Description }) - >

Other relationship nodes are created accordingly.

Query management based on Cyber language:

6) Querying head and tail entity nodes of occurrence relation:

MATCH p= () - [ r:' occurrence ] - > () RETURN p

7) Counting the number of quality problems of each satellite:

MATCH (n: satellite model) - [ r: occurs ] - > -. Count (r) as outgoingDegree ORDER BY outgoing Degree DESC

Example 4: fault description analysis based on text convolutional neural network

Text preprocessing is based on a synonym table, a stop word table, a single-machine dictionary and a model dictionary, spoken descriptive words are processed, and keyword term information such as fault models, single-machine information, fault component parts and the like possibly related to fault description after normalized processing is identified.

The fault description classification algorithm is realized based on a convolutional neural network (TextCNN), and the TextCNN is subjected to algorithm accuracy analysis based on satellite fault description data. The fault description data is divided into 2790 fault phenomenon descriptions based on 1080 quality problem cases of satellites, and the fault description data is divided into four categories of fault phenomenon, execution operation, instant influence and other information, wherein 2290 pieces of training data set and 500 pieces of data of test data set are used.

Table 3 fault description dataset distribution

The convolutional neural network textCNN model consists of three convolutional layers, a pooling layer, a full-connection layer, a dropout layer and a relu activation layer, the loss function selects cross entropy loss, and the optimization algorithm selects an extended Adam optimizer of a random gradient descent method.

Table 4 fault description analysis algorithm

By analyzing and comparing the accuracy rate, recall rate and F1 value of various fault description analysis algorithms, the recall rate is 77% and the F1 value is 66% when the supported concentration accounts for 60%; the F1 value of the executing operation is 31% under the condition that the supporting set accounts for 18%, the identification accuracy of the fault phenomenon is higher than that of other elements, the effect influence is 2% higher than the accuracy of the executing operation under the condition that the number of the supporting sets is the same, the F1 value of the executing operation is 10% higher than the effect influence, and experiments show that increasing the number of the supporting sets is an effective method for improving the analysis accuracy of the fault description elements and the F1 value.

Example 5: multi-dimensional retrieval based on keyword, representation learning and meta-path combination

And carrying out similarity query on 6 quality problems of downlink signal out-of-lock, MERE overrun, health identification abnormal change, inter-satellite link transceiver out-of-lock, uplink receiver abnormality and remote control terminal decryption count abnormality, and analyzing and verifying a 3-type similarity retrieval mode.

And carrying out similarity comparison on the engineer query term and the satellite quality problem knowledge graph from the three sides of local semantics, global semantics and structural semantics based on a multi-dimensional retrieval mode of keywords, representation learning and meta paths. The method is characterized in that a query sub-graph is built by using keywords matched with a single dictionary and a model dictionary as input based on a keyword retrieval mode, a text vector expressed by a word frequency vector and the like based on a retrieval mode for expressing learning is used as query input, a meta path expanded by query terms is used as query input based on a path query mode, and graph vector expression and path mining represented by tranR are required for a satellite quality problem knowledge graph so as to match the retrieval terms and compare the similarity.

And (3) carrying out similarity sorting on the 6-class repeated problem marking repeated data and the case data in the satellite quality problem knowledge graph, and using the common evaluation index hit rate HR (Hits Ratio) and the common average reciprocal rank MRR (Mean Reciprocal Rank) of the recommendation system.

Hit rate HR emphasizes the accuracy of model recommendation, i.e., hit rate, and concerns whether repeat case problems, i.e., demand terms, are recommended, HR@N is the ratio of the number of elements with a predicted rank less than N

S is a set of all recommendations; hits (i) are used to express whether the ith demand item is included in the list of items recommended by the model.

The average reciprocal rank MRR emphasizes the location of the user demand items in the model recommendation list, emphasizes the sequency, and indicates whether the items to be recommended are placed in more prominent locations.

S is a set of all recommendations; p is p _i Representing the position of the demand item in the recommendation list, p if the demand item is not in the list _i Is infinite.

Taking a knowledge graph retrieval mode based on keyword terms and representation learning as an example, comparing the method with a similarity calculation method only based on character representation,

table 5 sentence representation algorithm and distance measurement

Sequence number	Pretreatment of	Text representation	Distance measurement
				Algorithm 1	Word segmentation	SimHash	Sea distance
Algorithm 2	Word segmentation	Word frequency vector	Cosine distance

The algorithm 1 uses Simhash for text feature representation and calculates similarity between texts based on hamming distance. Simhash is a local hash sensitive algorithm, feature words are obtained through 5 steps of word segmentation, hash, weighting, merging and dimension reduction, and the number of different characters of two character strings with equal length on corresponding positions is calculated through Hamming distances, so that the smaller the Hamming distance is, the higher the similarity is.

The algorithm 2 uses word frequency vectors to perform feature representation on the query text, the word frequency vectors give higher weight to key terms, and similarity between the graph representation vectors and the word frequency vectors is calculated based on cosine distances, and the greater the cosine distance is, the higher the similarity is.

The experimental results were compared as follows:

table 6 comparison of the effects of the two algorithms

Numbering device	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％

Experimental analysis can show that the overall effect of the calculation mode of the similarity of the knowledge graph based on the keyword and the representation learning is better than that of the s im hash calculation mode based on the Hamming distance, and the recommendation result of the first algorithm is better.

Example 6: satellite fault diagnosis case analysis based on quality problem knowledge graph

The problem of 'satellite inter-satellite ranging value jump 12-15 ns' is taken as input, the fault description input is that '8 paths of navigation signals monitored by a navigation signal monitoring unit are in time delay ranging value jump', a measurement and control system sends a navigation task processing unit FPGA reset instruction and then restores ', corresponding execution operation content and fault phenomenon content are identified by preprocessing the fault input description and inputting a TextCNN convolutional neural network, the corresponding fault part of a dictionary is taken as a key word item, the' satellite navigation task processing machine 'is obtained through similar operation comparison, accurate backup and restoration capability is realized, satellite time and signal states of all paths are restored to a state before reset' and the like after the reset is restored, quality problems 'navigation signal unit monitoring unit B is in out-of-lock abnormality' after in-orbit initial synchronization are obtained according to the fault part, the similar phenomena are described as 'B1 Ad and B1Ap two paths of signals, the receiver out-of-lock abnormality is not restored', the satellite return telemetry display satellite navigation signal monitoring unit receives B1Cp signal out-lock is realized, the possible quality problems are obtained through the similar operation items, the possible quality problems are firstly, the possible satellite navigation task processing unit B is enabled to be in a state error is limited by a special fault code, the FPGA (FPGA) is not has a fault-free condition, and the fault is completely-restricted by a specific fault condition, and the fault condition is generated by a special fault interface (FPGA) is firstly, and the fault-free condition is possible, and the fault condition is prevented from being caused by a fault-free fault condition is caused by a fault-caused by a fault condition of a fault code, and a fault-free interface, and a fault condition is firstly has been restricted condition fault condition, if the fault cannot be recovered, a navigation task processing unit FPGA reset instruction is sent; secondly, for the satellite with the autonomous fault recovery, the autonomous fault recovery function is recommended to be started, if the abnormality occurs, the satellite can be recovered automatically, and the three-mode redundancy design is carried out on the local clock module in the FPGA configuration item of the monitoring B so as to improve the single event resistance of the software. The in-orbit satellite software is reconstructed according to the related requirements of the in-orbit software reconstruction, and maintenance measures such as one-to-three and the like are suggested for other satellite single machines under study.

Although the present invention has been described in terms of the preferred embodiments, it is not intended to be limited to the embodiments, and any person skilled in the art can make any possible variations and modifications to the technical solution of the present invention by using the methods and technical matters disclosed above without departing from the spirit and scope of the present invention, so any simple modifications, equivalent variations and modifications to the embodiments described above according to the technical matters of the present invention are within the scope of the technical matters of the present invention.

Claims

1. The satellite fault diagnosis system based on the quality problem knowledge graph is characterized by comprising a satellite quality problem knowledge graph construction module and a satellite fault diagnosis module based on the quality problem knowledge graph;

2. The satellite fault diagnosis system based on the quality problem knowledge graph according to claim 1, wherein the satellite quality problem domain ontology knowledge model is composed of classes, attributes, relationships, functions, axioms, in particular:

3. The satellite fault diagnosis system based on quality problem knowledge graph according to claim 1, wherein the concept alignment method is as follows: traversing concepts in the ontology knowledge model of the domain of the satellite quality problem by taking the ontology knowledge model of the domain of the satellite quality problem as a reference, calculating corresponding similarity with all concepts of the ontology knowledge model of the domain of other satellite quality problems, merging the two concepts into the same concept if the similarity is smaller than or equal to a concept similarity threshold value, and establishing semantic links corresponding to the concepts between the two ontology knowledge models;

4. The satellite fault diagnosis system based on quality problem knowledge graph according to claim 1, wherein 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 a syntax rule tree, or an unstructured text extraction method based on a pre-trained neural network model.

5. The quality problem knowledge graph-based satellite fault diagnosis system according to claim 1, wherein entity alignment means that different entity expressions are directed to the same objective knowledge element at an instance layer for instance layer fusion; the specific method comprises the following steps:

establishing a satellite quality problem field entity mapping dictionary library, and carrying out entity linking through dictionary matching, wherein the dictionary matching comprises two modes of complete matching and fuzzy matching, and the fuzzy matching mode is as follows: when the minimum editing distance required for converting the character string A into the character string B is less than N/2, the character strings A and B are considered to belong to the same entity; otherwise, strings A and B are not considered to belong to the same entity.

6. The satellite fault diagnosis system based on quality problem knowledge graph according to claim 1, wherein the fault description analysis method for the input is:

7. The satellite fault diagnosis system based on quality problem knowledge graph according to claim 1, wherein the keyword-based knowledge graph semantic search method comprises the following specific steps:

8. The satellite fault diagnosis system based on quality problem knowledge graph according to claim 1, wherein the knowledge graph semantic search method based on representation learning comprises the following specific steps:

combining sentence category, namely entity type information, performing nearest neighbor calculation on sentence semantic vectors and satellite quality problem patterns in a vector space, and calculating similarity based on vector distance to obtain a similarity result { b } ₁ ,b ₂ ,b ₃ ,....,b _n }。

9. The satellite fault diagnosis system based on quality problem knowledge graph according to claim 1, wherein the knowledge graph structure semantic retrieval method based on the meta-path comprises the following specific steps:

10. The satellite fault diagnosis module based on the quality problem knowledge graph according to claim 1, wherein the obstacle avoidance information recommendation method based on multi-dimensional similarity calculation is adopted, and specifically comprises the following steps:

wherein V is union operation;