CN115858807A - Question-answering system based on aviation equipment fault knowledge map - Google Patents

Abstract

The invention discloses a question-answering system based on a fault knowledge map of aviation equipment, which comprises a fault knowledge map module and a fault knowledge question-answering module, wherein the fault knowledge map module comprises modules for fault knowledge modeling, fault knowledge extraction, fault knowledge map storage and the like, the fault knowledge map module establishes a knowledge map mode layer facing the requirement of fault diagnosis service, extracts fault knowledge from massive and multi-source heterogeneous fault text data to establish a fault knowledge map, and stores the fault knowledge map into an ArangoDB database. The fault knowledge question-answering module comprises a question preprocessing module, a question analyzing module, an atlas retrieving module and an answer generating module, the fault knowledge question-answering module processes and analyzes questions provided by a user, matches the questions with corresponding intention templates, maps the corresponding intention templates into image query sentences, retrieves relevant fault knowledge atlases, generates answers to the questions and returns the answers to the questions.

Description

Question-answering system based on aviation equipment fault knowledge map

Technical Field

The invention belongs to the technical field of knowledge maps, and particularly relates to a question-answering system based on a fault knowledge map of aviation equipment.

Background

Currently, the front of aviation equipment maintenance still relies primarily on the personal knowledge reserve and experience accumulation of maintenance engineers to perform fault diagnosis. However, due to the complexity of the aircraft, fault diagnosis is a typical knowledge-intensive activity, and a maintenance engineer needs to spend a lot of time on browsing relevant text data, grasp the knowledge of the functional structure, the working principle and the like of a relevant system to analyze possible fault causes, and then gradually position a fault unit by adopting test means such as observation, measurement, string of parts and the like according to self understanding. The mode has low diagnosis efficiency, often causes long-term fault shutdown of equipment, seriously restricts the equipment perfectness rate, and is a big pain point of a basic unit. In fact, through information management measures, a very large-scale fault document database is accumulated, including equipment maintenance support teaching materials, equipment test diagnosis records, equipment quality control data, equipment maintenance support records, fault analysis and research reports, and the like. The resources are valuable knowledge sources for aviation equipment fault diagnosis, but due to the characteristics of multi-source isomerism and unstructured, the fault data are stored together in a form, but actually are 'information islands' split from each other, knowledge sharing is not really realized, and therefore the resources are wasted.

The knowledge graph is a concept proposed by google in 2012, and basically comprises < entity-relation-entity > triples, wherein entities are represented as nodes, relations are represented as edges, and the nodes and the edges are connected with each other to form a semantic network with a huge scale. The concept of knowledge graph proposed by google aims to dig out the relation between entities from massive webpage information, connect fragmented information on the webpage with an organic semantic network, improve the intelligent capability of a search engine and enhance the search quality and experience of users. By means of the knowledge graph technology, fault knowledge can be mined from massive, multi-source heterogeneous and unstructured fault document data and integrated into a structured and interconnected fault knowledge graph, and a way is provided for sharing and applying the fault knowledge.

The traditional knowledge graph is a semantic network taking noun entities as nodes, can well describe the functional structure, signal parameters and the like of equipment, but has insufficient description capacity on faults, detection and the like.

Disclosure of Invention

Aiming at the existing problems, the invention introduces events on the traditional knowledge graph technology, constructs the fault knowledge graph, and provides a question-answering system based on the fault knowledge graph of the aviation equipment based on the fault knowledge graph so as to describe faults, detection and the like more completely.

The technical solution for realizing the invention is as follows:

a question-answering system based on a failure knowledge graph of aviation equipment is characterized by comprising a failure knowledge graph module and a failure knowledge question-answering module; the fault knowledge map module is used for extracting fault knowledge from fault text data and constructing a knowledge map, and comprises a fault knowledge modeling unit, a knowledge extracting unit, a knowledge map storage unit and a knowledge map human-computer interaction unit;

the fault knowledge question-answering module comprises a question preprocessing unit, a question analyzing unit, a map retrieving unit, an answer generating unit and a man-machine interaction unit, and is used for processing the questions input by the user, retrieving the corresponding knowledge map, generating question answers and returning the question answers to the user.

Furthermore, the fault knowledge map module comprises a fault knowledge modeling unit, a knowledge extraction unit, a knowledge map storage unit and a human-computer interaction unit;

the fault knowledge modeling unit is used for constructing a knowledge graph mode layer facing the fault diagnosis service requirement;

the knowledge extraction unit is used for extracting entities, events, attributes and relations related to faults from massive fault document data and fusing fault knowledge;

the knowledge map storage unit stores the extracted structured fault knowledge by adopting an ArangoDB graph database;

and the knowledge graph man-machine interaction unit is used for displaying the knowledge graph through a visual interface.

Further, the fault knowledge modeling unit for constructing the knowledge graph pattern layer comprises the following steps:

step 1: determining fault diagnosis knowledge elements and ranges;

step 2: defining core concepts, event types, relationship modes and attributes of fault knowledge;

and step 3: normalizing the fault knowledge concept defined in the step 2 according to the standard;

and 4, step 4: and describing entity types, event types, relationship modes and attributes related to the fault, and constructing a fault knowledge model.

Further, the knowledge map mode layer comprises an equipment structure model and an equipment fault model, and the equipment structure model comprises a component unit and static knowledge with the signal parameter name word entity type as the core; the equipment fault model is dynamic knowledge taking fault cases, fault events, event entity types of an inspection method and a removal method as cores; and the equipment structure model and the equipment fault model link the event type entity to the name word type entity for fusion through the argument role of the event.

Further, the argument role is the role an event argument plays in an event, the event argument referring to all participants that constitute the event.

Furthermore, the knowledge extraction unit comprises a name class entity identification subunit, an event class entity identification subunit and a relationship extraction subunit;

a noun entity identification subunit, which adopts a BERT-BilSTM-CRF model to identify a component unit, a signal parameter and a trigger word and argument in an event entity related to the fault from the fault document;

the event type entity identification subunit is used for identifying all entity types and trigger words, event types and entity types of all event types through three steps of entity identification, event confirmation and argument confirmation;

and the relation extraction subunit is used for acquiring the relation between the two fault events through an extraction method based on the rule and extracting the relation by combining a deep learning method.

Further, the question preprocessing unit is configured to perform preprocessing operations of spelling error correction, word segmentation and part-of-speech tagging on the input question, and acquire question information of a substrate granularity;

the question analysis unit comprises a word slot extraction unit and an intention identification unit and is used for carrying out word slot extraction and intention identification on the preprocessed question to obtain an intention template;

the map retrieval unit is used for retrieving related fault information from the knowledge map module according to the intention template obtained by the question analysis module;

the answer generating unit is used for converting, combining and sequencing the fault information obtained by the graph retrieval module and finally returning a question and answer result set;

and the man-machine interaction unit is used for interacting with the user through a visual interface and displaying the question and answer result to the user.

Further, the word slot extraction unit is used for matching the preprocessed question by using a predefined template to obtain a question analysis result, and obtaining slot position information of the entity slot, the mode slot, the attribute slot and the condition slot from the analysis result;

the intention identification unit is used for matching the obtained slot position information to a predefined intention template.

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

firstly, the method is deployed in an aviation equipment big data center based on an aviation equipment fault knowledge map, and fault knowledge is automatically extracted from fault information by acquiring information such as fault phenomena, fault cause analysis, fault elimination experience and the like from the aviation equipment big data center, so that when a maintenance engineer at the maintenance and protection front line encounters an unknown fault, the fault phenomena can be responded to the problem of the maintenance engineer only by inputting the fault phenomena into the system, and the basic level maintenance engineer is assisted to quickly locate a fault unit.

Secondly, the event is introduced into a mode layer of the fault knowledge graph, the event type and the entity type are combined and fused through the argument role of the event type, the defect that the conventional knowledge graph only can construct a knowledge base with nouns as core nodes can be overcome, and dynamic knowledge of faults, detection, repair and the like can be described more completely;

thirdly, the invention designs a fault knowledge extraction method combining entity extraction and event extraction, which can improve the extraction accuracy and recall rate of event entities such as fault events, inspection methods and the like, and is beneficial to improving the quality of fault knowledge maps.

Drawings

FIG. 1 is a correlation diagram of fault events and constituent units;

FIG. 2 is a schema layer of the failure knowledge graph of the present invention;

FIG. 3 is a block diagram of the BERT-BilSTM-CRF model;

FIG. 4 is an architecture diagram of a fault knowledge question and answer;

FIG. 5 is a process diagram of fault knowledge graph modeling;

FIG. 6 is a problem retrieval interface of the present system;

fig. 7 shows the failure analysis/failure attribution interface of the present system.

Detailed Description

In order to make those skilled in the art better understand the technical solution of the present invention, the following further describes the technical solution of the present invention with reference to the drawings and the embodiments.

1. Aviation equipment fault knowledge map construction

The construction of the fault knowledge graph comprises five levels, namely a fault data layer, a knowledge extraction layer, a knowledge graph layer and a human-computer interaction layer.

The failure data layer mainly refers to massive, multi-source heterogeneous and unstructured failure document data accumulated in the whole life cycle class of the equipment, such as equipment maintenance guarantee teaching materials, equipment maintenance guarantee histories, failure analysis and research reports and the like. The knowledge extraction layer mainly refers to identifying entities, events and relations related to faults from massive unstructured fault text data by using a machine learning method, and fusing the knowledge to construct a structured and interconnected fault knowledge map. The knowledge map layer is used for storing the extracted structured fault knowledge by adopting an ArangoDB graph database. The application technology layer is used for matching similar fault cases according to fault scenes by utilizing technologies such as semantic retrieval, knowledge reasoning and problem understanding, reasoning out possible fault reasons, generating problem answers and the like. The man-machine interaction layer is a visual interface directly facing a maintenance engineer, and comprises a visual display of a fault knowledge map, an interactive interface of intelligent question answering and the like.

The construction of knowledge graph in industry field generally adopts top-down construction mode, i.e. firstly, a mode layer is designed facing to the field application requirement, entity type and relation type which may exist in the graph are defined, and the process is generally called knowledge modeling. Only under the constraint guidance of the mode layer, the knowledge can be automatically extracted from the unstructured document data and a knowledge graph can be constructed. The aviation equipment fault knowledge graph is a typical industry field knowledge graph, so a top-down construction mode is adopted. Therefore, firstly, fault knowledge modeling is performed facing to the service requirement of intelligent auxiliary diagnosis, namely, a mode layer of a fault knowledge graph is constructed, namely, fault knowledge modeling is performed. Since the fault knowledge model is an upper-layer mode of the fault knowledge map, the fault knowledge model determines the form of the fault knowledge map, and the quality of the fault knowledge map and the application effect of the fault knowledge model in auxiliary diagnosis can be directly influenced. Therefore, the construction of the fault knowledge model is very important and is a key for determining whether the knowledge graph technology can be successfully applied to the field of aviation equipment fault diagnosis.

Furthermore, fault knowledge modeling is realized by adopting an evaluation iteration method, a fault knowledge model is continuously iterated and perfected by means of measures such as expert evaluation, and finally a fault knowledge model which is commonly approved by experts in the field and faces the requirements of fault diagnosis services is constructed.

The specific process of fault knowledge graph modeling comprises the following steps:

(1) at the beginning of the construction of a fault knowledge model, discussing knowledge elements contained in the field of equipment fault diagnosis together with experts and equipment support personnel in the field, and analyzing the types and characteristics of the knowledge;

(2) constructing a concept system in the field of fault diagnosis, and defining entity types, event types, relationship modes and attributes of fault knowledge;

(3) in order to standardize and standardize fault knowledge, the definitions of professional terms such as national standards, national military standards and industrial standards related to the fault field are collected and used for reference, and relevant concepts in a fault knowledge model are subjected to standardized processing;

(4) describing entity types, event types, relationship modes and attributes related to faults, and constructing a fault knowledge model;

(5) collecting evaluation and improvement opinions of field experts and basic guarantee personnel on the fault knowledge model by adopting a questionnaire survey mode;

(6) and (6) returning to the step (2), and correcting and improving the fault knowledge model. And finally constructing a fault knowledge model which is commonly recognized by experts in the field and has practicability through multiple iterations, and guiding and standardizing the automatic construction of the fault knowledge map.

In order to be able to more fully describe knowledge of faults, checks and the like, event types are introduced into the knowledge graph pattern layer.

(1) Concept of events

An event refers to a state change in a specific environment, such as an action, an activity, or a state being an event. The natural language processing field specifies the concepts related to the event, including event description, event type, time trigger, event argument and argument role, etc. because the definition of the event is abstract and fuzzy, and is not convenient for the common discussion and cooperation of the fields. The content and the related examples are shown in Table 1.

TABLE 1

(2) Fusion of noun entity and event entity

The traditional knowledge graph taking the noun entity as the core is very suitable for describing static knowledge such as system structures, signal parameters and the like of equipment, and dynamic knowledge such as faults, detection, maintenance and the like can be better described by utilizing events. In order to fuse the two types of knowledge, the invention provides an event entity-noun entity fusion mechanism based on the argument role, namely, the event entity is linked to the noun entity through the argument role of the event type, the event type and the argument role thereof are defined firstly, then the entity type is defined, and finally the event type is associated with the entity type through the argument role thereof. For example, an event type of "failure event" is defined, an argument role of "failure location" is defined in the event type, an entity type of "component unit" is defined in the equipment structure knowledge model, and the "failure event" is associated with the "component unit" through the argument of "failure location", as shown in fig. 1.

(3) Mode layer of failure map

As shown in fig. 2, the mode layer of the failure knowledge graph is divided into two large parts, one part is an equipment structure model and is static knowledge taking entities such as component units, signal parameters and the like as cores, the other part is an equipment failure model and is dynamic knowledge taking events such as failure cases, failure events, inspection methods and the like as cores, and the two parts are only fused through argument roles of the models;

the entity types in the mode layer include "constituent units" and "signal parameters". Where a "component unit" refers to a physical element belonging to a particular level of a system (e.g., a subsystem, component, part, etc.), a unit may be composed of several sub-units, or may be a sub-unit of a larger unit. "Signal parameters" refer to parameters that can be measured, and many faults may be characterized as anomalies in the signal. The definitions of all event types and relationship types in the schema layer are shown in tables 2 and 3, respectively. It is noted that the pattern does not distinguish between fault phenomena, fault patterns, fault causes, but rather defines them collectively as "fault events". This is because in unstructured text, the three types of events are very similar in expression, such as the fault phenomenon "slow right car speed is too high" and the fault cause "FRV right cannot open". Therefore, machine learning cannot distinguish the three based on contextual features. In addition, it is not meaningful to distinguish the three in the fault diagnosis because if there is a "cause" relationship between the fault event a and the fault event B, it can be determined that the fault event a is the cause of the fault event B. And sometimes, a fault forms a propagation chain, a fault can be the cause of a fault B, and the fault B can be the cause of a fault C, and at the moment, the unit where the fault is located needs to be found by following a clue along a fault cause and effect chain. If "failure cause" is defined as an event type, the failure propagation chain cannot be described. In summary, a "failure event" and a "cause" relationship to itself are defined.

TABLE 2

TABLE 3

(4) Fault knowledge extraction

The key to the construction of the fault knowledge graph is fault knowledge extraction, including identification of entities and their relationships. The basic methods mainly include two types: heuristic rule-based methods and deep learning-based methods. The heuristic rule based method requires enough prior knowledge to design a reasonable extraction rule, but the extraction accuracy is very high. The deep learning-based method has the capability of processing mass data, can automatically extract features from unlabeled text data under the condition of little or no prior knowledge, describes rich internal information of the text data, does not need to specially design the features, and can avoid incompleteness caused by manually designing the features. In order to fully utilize the advantages of the two methods, a fault extraction method which takes deep learning as a main method and heuristic rules as an auxiliary method is comprehensively adopted, namely, a deep learning method is adopted for extracting most unstructured fault documents, and heuristic rules are adopted for extracting semi-structured data, tables or normative documents with certain structural characteristics;

(4.1) noun entity recognition

The target of the recognition of the name word entity is to accurately recognize the component units, the signal parameters, the trigger words and the arguments and the like in the event entity related to the fault from the fault document. The entity identification can be converted into the problem of sentence sequence annotation, the annotation method adopts a BIOE set, B (Begin) represents the beginning of an entity, I (Inside) represents the middle of the entity, O (Other) represents Other parts of a non-entity, and E (End) represents the End of the entity. The letters after B-, I-, O-, E-represent the entity type. For example, the entity type "composition Unit" is represented by the letter U (Unit), and B-U, I-U, E-U represents the beginning, middle and end of "composition Unit", respectively. To achieve this, the classical BERT-BilSTM-CRF model was used, as shown in FIG. 3. The first layer of the model is the BERT layer, which is a language pre-training model responsible for converting characters into vector form. The second layer of the model is a bidirectional long-short term memory neural network (BilSTM) which can well reflect the influence of past and future contents on the current contents, and has the advantage of extracting the sequence characteristics of sentences which are related to the context. The last layer of the model is the Conditional Random Field (CRF) algorithm, which is used for decoding and labeling.

In order to improve the accuracy of entity identification, normalization processing is carried out on data, all numbers are modified into N, chinese character numbers are modified into N, and since the numbers generally have no semantic correlation with the extracted objects, after the numbers are unified, useless semantic information in the context is less.

(4.2) event class entity extraction

Compared with noun entities, event entities have more fuzzy boundaries, larger length change and more complex word composition. For example, the failure mode may be expressed as "XX component + abnormal state" (e.g., "VBV valve stuck"), "XX signal + abnormal state" (e.g., "oil temperature exceeded"), "XX component + XX signal + abnormal state" (e.g., "solenoid valve driving current distortion"), and the like. The events are difficult to be accurately identified by directly adopting a method for identifying noun entities such as BERT-BILSTM-CRF and the like. Therefore, for the extraction of the events, an extraction method combining deep learning and heuristic rules is adopted, and the extraction process is divided into three steps of entity identification, event confirmation and argument confirmation.

Step one, utilizing noun entity recognition technology to recognize all predefined noun entity types and trigger words of all event types in the fault knowledge model. For example, the sentence "the alarm system continuously reflects the overhigh right-sending rotating speed in the descending stage", the composition unit "has the alarm system" and the right-sending ", the signal parameter has the rotating speed", and the trigger word of the fault event has the overhigh trigger word. And step two, determining whether the event type corresponding to the trigger word exists in the sentence. Taking the above sentence as an example, the trigger word "too high" of the fault event is identified, but it cannot be determined whether the sentence includes a fault event only by the trigger word, and further confirmation is required by combining entity information such as a composition unit, a signal parameter, and the like. If there is a case where entities such as constituent elements, signal parameters, etc. and trigger words continuously appear, it is determined that the sentence contains a failure event. Otherwise, vectorizing information obtained by entity identification, fusing the information into sentence sequence vectors through a Conditional Layer Normalization (CLN), extracting sentence-level feature vectors through a text convolution method, realizing event confirmation two-stage classification, and obtaining an event confirmation result. And step three, determining whether the name word entity is the argument role predefined by the event. And if the entity such as the composition unit, the signal parameter and the like and the trigger word continuously appear, directly splicing the composition unit, the signal parameter and the trigger word to be used as a fault event. Otherwise, the sentences, the event types, the entities, the entity types and the relative positions of the entities and the trigger words are vectorized and spliced into a long vector, and then the short characteristic vectors are extracted through a full connection layer to realize argument confirmation two-classification, so that arguments in the events are extracted.

(4.3) relational extraction

The relation extraction adopts a mode of combining rule-based and model-based. The rule-based extraction method is mainly based on understanding of fault data and fault knowledge logic, and some rules, regular expressions, templates and the like are manually written for extraction. For example, in an equipment technical document, the composition relation among fault composition units is directly extracted by relying on the directory relation of each composition unit; and if words of 'causing', 'causing' and 'causing' exist between two fault events in the same sentence, the causal relationship between the two fault events can be determined. By the simple rule, a great number of relations can be accurately extracted, and the method is simple and effective.

Besides the method of enlightening the rules, the method of deep learning is also adopted to extract the relationship. The method converts the relationship extraction into a binary problem. Firstly, finding out paragraphs containing entity pairs or event pairs, extracting the feature vector of each sentence through 'BERT + BilSTM + convolution + pooling', then extracting the feature vector of the paragraph level through 'convolution + pooling', splicing the sentence feature vector containing the entity pairs or the event pairs and the paragraph feature sentences to be used as classification model input for judging whether the predefined relationship exists. For example, in the paragraph "…," the alarm system continuously reflects that the right-hand rotation speed is too high in the descent stage, and then the oil level sensor control component is switched to a fault, …, because … "is caused by the oil level sensor control component failing. The two sentences both contain a fault event, the feature vectors of the two sentences and the feature vector of the paragraph are spliced, and then whether the causal relationship caused by the two sentences exists or not is confirmed through a classification model.

2. Fault question-answering system based on aviation equipment fault knowledge map

As shown in fig. 4, the present invention provides a question answering system based on a failure knowledge graph of aviation equipment, which includes: the system comprises an aviation equipment big data center, a knowledge map module, a question preprocessing module, a question analysis module, a map retrieval module and an answer generation module, wherein the aviation equipment big data center is used for gathering information of basic-level maintenance engineers for online filling of fault phenomena, fault reason analysis and fault elimination experience;

the knowledge map module is used for extracting fault knowledge from fault information of the large data center of the aviation equipment to construct a fault knowledge map;

the question preprocessing module is used for carrying out preprocessing operations of spelling error correction, word segmentation and part of speech tagging on the input question and acquiring question information of substrate granularity;

the question analysis module comprises a word slot extraction unit and an intention identification unit and is used for carrying out word slot extraction and intention identification on the preprocessed question to obtain an intention template; the word slot extracts words, entities, short sentences and the like obtained based on question sentence preprocessing, accurate or fuzzy matching is carried out by utilizing a template, and relevant slot position information including entity slots, mode slots, attribute slots, condition slots (total) and the like is analyzed from the problems. For example, "there are several failure modes for a vibration sensor? "from this question, the entity slot" vibration sensor ", the mode slot" failure mode ", the condition slot" several "can be extracted;

the map retrieval module is used for retrieving related fault information from the knowledge map module according to the intention template obtained by the question analysis module; it maps the intention template information into AQL query statements through rule heuristics and template matching, and concatenates ArangoDB, retrieving entity names, relationship names, attribute values, etc. from the database. Meanwhile, relevant fault documents can be retrieved through entities;

the answer generation module is used for converting, combining and sequencing the fault information obtained by the graph retrieval module and finally returning a question and answer result set; the method comprises the steps of converting and combining results inquired in a database based on an answer template, sequencing answers, and returning an answer set and related documents.

Further, the word slot extraction unit is used for matching the preprocessed question with a template to obtain a question analysis result, and obtaining slot position information of the entity slot, the mode slot, the attribute slot and the condition slot from the analysis result;

the intention identification unit is used for matching the obtained slot position information to a predefined intention template; the method adopts a method of combining template matching and deep learning classification models, wherein the template matching matches slot position information such as entities, modes, attributes, conditions and the like to a predefined intention template through syntactic analysis, and a deep learning discrimination model is adopted when the extracted slot position information cannot be directly matched with a fixed intention template.

After the fault knowledge graph is constructed, a question-answering system based on the fault knowledge graph of the aviation equipment is realized, the relevant fault knowledge graph can be searched for relevant questions proposed by a user based on a human-computer interaction interface, answers to the questions are generated and returned, the user can also be guided to gradually narrow the fault range so as to accurately position the fault unit, and the functional interface of the system is shown in FIGS. 6-7.

Those not described in detail in this specification are within the skill of the art. Although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that various changes in the embodiments and modifications of the invention can be made, and equivalents of some features of the invention can be substituted, and any changes, equivalents, improvements and the like, which fall within the spirit and principle of the invention, are intended to be included within the scope of the invention.

Claims (8)

1. A question-answering system based on a failure knowledge graph of aviation equipment is characterized by comprising a failure knowledge graph module and a failure knowledge question-answering module; the fault knowledge map module is used for extracting fault knowledge from fault text data and constructing a knowledge map, and comprises a fault knowledge modeling unit, a knowledge extracting unit, a knowledge map storage unit and a knowledge map man-machine interaction unit;

the fault knowledge question-answering module comprises a question preprocessing unit, a question analyzing unit, a map retrieving unit, an answer generating unit and a man-machine interaction unit, and is used for processing the questions input by the user, retrieving the corresponding knowledge map, generating question answers and returning the question answers to the user.

2. The question-answering system based on the aviation equipment fault knowledge graph is characterized in that the fault knowledge graph module comprises a fault knowledge modeling unit, a knowledge extraction unit, a knowledge graph storage unit and a human-computer interaction unit;

the fault knowledge modeling unit is used for constructing a knowledge graph mode layer facing to the fault diagnosis service requirement;

the knowledge extraction unit is used for extracting entities, events, attributes and relations related to faults from massive fault document data and fusing fault knowledge;

the knowledge map storage unit stores the extracted structured fault knowledge by adopting an ArangoDB graph database;

and the knowledge graph man-machine interaction unit is used for displaying the knowledge graph through a visual interface.

3. The question-answering system based on the aviation equipment fault knowledge graph as claimed in claim 2, wherein the fault knowledge modeling unit building the knowledge graph mode layer comprises the following steps:

step 1: determining fault diagnosis knowledge elements and ranges;

step 2: defining a core concept, an event type, a relation mode and attributes of fault knowledge;

and step 3: normalizing the fault knowledge concept defined in the step 2 according to the standard;

and 4, step 4: and describing entity types, event types, relationship modes and attributes related to the fault, and constructing a fault knowledge model.

4. The question-answering system based on the aviation equipment fault knowledge-graph according to claim 3, wherein the knowledge-graph mode layer comprises an equipment structure model and an equipment fault model, and the equipment structure model comprises component units and static knowledge with the signal parameter name word entity type as the core; the equipment fault model is dynamic knowledge which takes fault cases, fault events, an inspection method and a removal method event entity types as cores; and the equipment structure model and the equipment fault model link the event type entity to the name word type entity for fusion through the argument role of the event.

5. The system of claim 4, wherein the argument role is a role that an event argument acts in an event, and the event argument refers to all participants who constitute the event.

6. The question-answering system based on the aviation equipment fault knowledge graph is characterized in that the knowledge extraction unit comprises a name word class entity identification subunit, an event class entity identification subunit and a relation extraction subunit;

a noun entity identification subunit, which adopts a BERT-BilSTM-CRF model to identify a component unit, a signal parameter and a trigger word and argument in an event entity related to the fault from the fault document;

the event type entity identification subunit is used for identifying all entity types and trigger words, event types and entity types of all event types through three steps of entity identification, event confirmation and argument confirmation;

and the relation extraction subunit is used for acquiring the relation between the two fault events through an extraction method based on the rule and extracting the relation by combining a deep learning method.

7. The system of claim 1, wherein the knowledge-graph of the faults of the avionics is based on a statistical analysis of the faults of the avionics,

the question preprocessing unit is used for carrying out preprocessing operations of spelling error correction, word segmentation and part of speech tagging on the input question and acquiring question information of substrate granularity;

the question analyzing unit comprises a word slot extracting unit and an intention identifying unit and is used for carrying out word slot extraction and intention identification on the preprocessed question to obtain an intention template;

the map retrieval unit is used for retrieving related fault information from the knowledge map module according to the intention template obtained by the question analysis module;

the answer generating unit is used for converting, combining and sequencing the fault information obtained by the graph retrieval module and finally returning a question and answer result set;

and the man-machine interaction unit is used for interacting with the user through a visual interface and displaying the question and answer result to the user.

8. The system of claim 7, wherein the word slot extraction unit is configured to perform matching on the preprocessed question by using a predefined template to obtain a question analysis result, and obtain slot position information of the entity slot, the mode slot, the attribute slot, and the condition slot from the analysis result;

the intention identification unit is used for matching the obtained slot position information to a predefined intention template.

Images