Aviation equipment supportability analysis knowledge extraction method based on deep learning

Abstract

The invention provides an aviation equipment supportability analysis knowledge extraction method based on deep learning. The aviation equipment supportability analysis knowledge extraction method comprises the following steps: S1, acquiring aviation equipment life complete cycle data; s2, constructing a supportability analysis database; and S3, constructing a deep learning extraction model oriented to supportability analysis knowledge to complete knowledge extraction to obtain a triple of entities and relationships. The deep learning extraction model comprises a model training unit, a scoring unit and an entity relationship extraction unit, the entity relationship extraction unit is based on a triple of entities and relationships of a three-dimensional matrix, a single-step and single-model combined extraction idea is used, and a bidirectional long-short-term memory network and a self-attention mechanism are introduced, so that the real-time performance of the system is improved; the accuracy of single-step-single-model knowledge extraction for guarantee data is improved, and the knowledge extraction process is simplified. And finally, on the basis of knowledge extraction, constructing an equipment indemnificatory analysis knowledge graph, and realizing structuralization and utilization of indemnificatory analysis history and objective data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aviation equipment knowledge extraction, and particularly relates to an aviation equipment supportability analysis knowledge extraction method based on deep learning. BACKGROUND

[0002] Equipment supportability analysis knowledge is important data accumulated in the process of equipment design, which is generated in various stages of equipment design, test and application, and contains fault analysis data, supportability analysis results and support measures related to equipment support, etc. These data can provide support and reference for the supportability design of new equipment. However, the current supportability analysis lacks demand analysis of historical objective data and lacks corresponding structured means. With the development of data technology, intelligent and intuitive data analysis has become the development direction of data processing technology in various fields, and new technical means such as deep learning and knowledge graph are urgently needed to support the equipment supportability analysis.

[0003] Knowledge graph is an information network that represents real-world entities and relationships using data. It is a commonly used data structuring method and has better scalability and visualization capabilities compared to traditional databases, so it is widely used in various fields. However, knowledge graphs in different professional fields often require different construction methods, especially in highly specialized vertical fields such as equipment support. It is more difficult to construct a graph. In general public fields, many terms in expert knowledge are common in daily life. There are special structured data sets that record entities and relationships in these knowledge. There are some mature technologies, algorithms, and achievements in the construction method and process of knowledge graph. A large part of support-related knowledge is stored in various types of analysis documents in various units, covering support business processes, main activities, input data forms, output data forms, etc. It not only includes related support analysis-related terms, but also includes equipment, devices, parts, and their specific fault mode terms. These terms are highly specialized, require high accuracy, and have large data volumes. It is difficult to efficiently and accurately structure the equipment support analysis knowledge using common knowledge graph construction methods. For example, using manual knowledge extraction methods will face the problem of unacceptable time cost due to large data volume; if using rule-based knowledge extraction, the related knowledge rules are interdisciplinary and cover a wide range, making it difficult to establish and the extraction effect difficult to guarantee; using deep learning-based knowledge extraction can utilize the powerful mapping ability of neural networks to improve the efficiency of knowledge extraction, but traditional deep learning knowledge extraction is divided into two steps: first, use deep learning to realize entity extraction, then use the entity extraction results as input to extract the relationship of the text, and finally get the triplets that can be used to construct the graph. Such knowledge extraction process is called traditional pipeline method, although it is widely used, but it needs to train two deep learning models, and the quality of relationship extraction depends largely on the effect of named entity recognition, making the knowledge extraction process prone to error accumulation. The lack of interaction between the two sub-tasks makes the text information not fully utilized. For the large amount of equipment support analysis knowledge generated in the process of equipment development and use, such knowledge extraction is not real-time, efficient, and of questionable quality.

[0004] Deep learning-based entity relation joint extraction models effectively address the high cost and low efficiency problems faced by traditional feature engineering-based methods, and also overcome the shortcomings of traditional pipeline methods. These models have shown excellent performance in tasks in multiple fields, which can be mainly classified into three methods: multi-module multi-step, multi-module single-step, and single-module single-step. The multi-module multi-step method divides the extraction process of entities and relations into multiple stages, each responsible for handling different tasks, and integrates these stages through shared parameters, gradually iterating to extract entity relation triples. Although this method is effective in some cases, it still has problems of decoding error accumulation and problem cascade redundancy. The multi-module single-step method constructs a model that requires an optimized joint decoding algorithm to solve the optimal solution and optimal hyperparameters, which undoubtedly increases the complexity of model construction. To overcome the limitations of the above two methods, a single-module single-step joint model appears. This method directly extracts entity relation triples from the text by processing the text once, and shows better performance on public datasets. This method simplifies the model construction process while improving the accuracy and efficiency of extraction. In the specific network structure, bidirectional long short-term memory network and self-attention mechanism both have strong processing power.

[0005] BiLSTM is an advanced RNN architecture that captures bidirectional dependencies in sequences by combining the outputs of both forward and backward LSTM networks. This structure allows BiLSTM to consider both the context information before and after each element in the sequence, making it perform well in processing sequence data such as NLP. BiLSTM's advantage is that it can provide more comprehensive feature representation because it combines forward and backward information, making the model more effective in capturing semantic and dependency relationships in sequence data, and also showing superior performance in tasks such as machine translation, speech recognition, text generation, and question answering systems.

[0006] Self-attention mechanism is a powerful technique that allows the model to dynamically focus on other elements in the input sequence when processing each element, thus better capturing the dependencies within the sequence. Unlike traditional sequence processing models such as recurrent neural networks, which process sequence elements one by one, self-attention mechanism can more efficiently capture long-range dependencies by computing the relevance between each element and other elements while considering the information of the entire sequence. It achieves this through three key steps: first, it computes the relevance scores between each element in the input sequence, then it converts these scores into a probability distribution through a normalization function such as Softmax, and finally it weights the sum of elements in the sequence according to these probabilities to obtain a context vector containing important information. This mechanism not only improves the performance of the model, but also enhances the interpretability of the model, making it more intuitive to show its decision-making process. Therefore, there is an urgent need to provide a knowledge extraction method that can use deep learning methods to analyze equipment supportability knowledge. SUMMARY

[0007] In order to solve the above problems of the prior art, the present application provides an aviation equipment supportability analysis knowledge extraction method based on deep learning, which uses the single-step single-model joint extraction idea and introduces bidirectional long short-term memory network and self-attention mechanism to improve the accuracy of single-step single-model knowledge extraction for support data and simplify the knowledge extraction process. Finally, based on the knowledge extraction, an equipment supportability analysis knowledge graph is constructed to realize the structuring and utilization of support history and objective data.

[0008] Specifically, in one aspect, the present application provides an aviation equipment supportability analysis knowledge extraction method based on deep learning, which includes the following steps: S1, data requirement analysis: collecting aviation equipment life cycle data; S2, database construction: combining supportability data requirements, building a pattern layer, using a top-down method to define entity and relationship information, and completing the construction of the supportability database; S3, constructing a deep learning extraction model for supportability analysis knowledge, the deep learning extraction model includes a model training unit, a scoring unit and an entity relationship extraction unit, the model training unit trains the deep learning extraction model, the scoring unit uses BiLSTM and self-attention layer as a scorer to output a three-dimensional matrix representing the classification result, and the entity relationship extraction unit is based on the entity and relationship triplets of the three-dimensional matrix, specifically including the following sub-steps: S31, construct a BERT pre-training model as a model training unit, the BERT pre-training model receives Input_ids and Attention_mask as input, learns the deep representation of the input sentence through a multi-layer Transformer architecture and outputs a three-dimensional matrix as the input of the scorer, specifically: in a text sample with words, each word is converted into dimensional quantity, the vector representation of the sentence is: , represents the th word in the text converted into a dimensional vector, then the vector is converted into a combined representation , the combined representation is a dimensional vector, represents the combined representation of the th and th words, and finally the text is converted into a three-dimensional matrix as the input of the scorer; , ; S32, score the three-dimensional matrix using a scorer composed of BiLSTM and self-attention layer and output the classification result of each , the calculation formula of the scorer is as follows: ; wherein, is the output of BiLSTM; , are the weights and biases in the nonlinear mapping respectively; SelfAttention is a self-attention layer, and softmax is an activation function, which is used to convert the output of the scorer into a classification of labels; S33, entity relation extraction: based on the classification result of the three-dimensional matrix , add the index and decode to obtain the triple of entity and relation.

[0009] Preferably, in the classification process of step S32, four types of relationship labels are first defined, the four types of relationship labels are head entity head-tail entity head HB-TB, head entity head-tail entity tail HB-TE, head entity tail-tail entity tail HE-TE and non-entity position none, and relationships defined in the graph pattern layer are encoded , each specific relationship corresponds to a three-dimensional matrix The dimension of this data sample is expanded to The scorer evaluates the three-dimensional matrix under each relation. Each one is scored once, and each of the three-dimensional matrix is ​​scored. The labels are divided into four categories to form a new matrix to be decoded. The new matrix can represent the relationship After scoring all relationships, the information of the head and tail entities is used to expand the new matrix to be decoded into the following form: .

[0010] Preferably, the decoding process in step S33 is as follows: in the i-th relation Lower three-dimensional matrix The matrix to be decoded is formed by scoring. In the text, the tail entity is positioned between the words with the corresponding tags HB-TB and HB-TE, the head entity is positioned between the words with the corresponding tags HB-TE and HE-TE, and other irrelevant positions are filled with the none tag.

[0011] Preferably, the method further includes step S4, which involves using a graph database to construct a knowledge graph based on triples of entities and relations. Specifically, the Neo4j graph database is used as the knowledge storage carrier, and a knowledge graph construction script is deployed on the server. The extracted results are used as input to automatically read entities and relations and generate Cypher statements to build a local knowledge graph.

[0012] Preferably, in step S32 The specific calculation is as follows: ; ; ; ; ; ; ; in, The input vector at the current time step t, This is the hidden state of the previous time step t-1. This represents the cell state at the current time step. This represents the cell state at the previous time step. , , , These are the input gate, forget gate, output gate, and candidate memory content for the current time step. These are the corresponding weights. These are the corresponding biases. and and represent the forward and backward hidden states, respectively.

[0013] Preferably, in step S1, specifically, the life cycle data requirements of the aviation equipment are determined, and the data items in the massive data that affect the equipment support design and actual support work are sorted out, involving data in five fields of support task analysis, support resource analysis, support requirement trade-off analysis and determination, maintenance optimization analysis, and support strategy analysis.

[0014] Preferably, in step S1, the life cycle data of the aviation equipment includes data in the demonstration phase, data in the engineering development phase, data in the test and identification phase, data in the production phase, and data in the use and support phase.

[0015] Preferably, the data in the demonstration phase includes data of an equipment requirement analysis system and data of an equipment requirement management system; the data in the engineering development phase includes data of new equipment development requirements, data of preliminary support schemes, data of equipment support design results, and data of support analysis results; the data in the test and identification phase includes data of new equipment tests and data of support resource matching schemes; the data in the production phase includes data of equipment production and data of support resource development; and the data in the use and support phase includes data of support systems and resources and data of equipment use.

[0016] Preferably, in step S2, the method of constructing the database specifically includes the following sub-steps: S21, defining the boundary and range of support data, and determining that the business fields contained in the support analysis original data from the life cycle data of the aviation equipment are: support requirement trade-off analysis and determination, support task analysis, support resource analysis, support strategy analysis, maintenance optimization analysis, and support data verification; S22, combining the data requirements of the above data to display constraints on related concepts, entities, or attributes, to obtain the entity categories that should be included in the graph and their relationships; S23, the determined entity categories include equipment, systems, subsystems, products, analysis businesses, analysis steps, failure modes, failure causes, maintenance measures, maintenance task categories, maintenance frequency and interval, support resource information, and support model information; the relationship types include belong to, represent, cause, and correspond to.

[0017] Preferably, in step S31, the training process of the model training unit specifically includes the following sub-steps: S311, data processing: converting the unstructured original text sequence into a standardized input form suitable for entity and relationship joint extraction by a deep learning model; S312, data training: the data is divided into a training set, a validation set and a test set according to the ratio of 8:1:1, the entity relationship joint extraction deep learning model is used in the training stage, the result of the pre-training model BERT is used as the input, the result of the pre-training model BERT is used for relationship representation, the vector features contained in the text are spliced and fused with the vector features of a certain specific relationship, and then the vector is transmitted to the entity relationship extraction unit through the network layer, and finally the iterative training is completed.

[0018] Compared with the prior art, the beneficial effects of the present application are as follows: (1) The present application proposes a set of knowledge extraction deep learning model for equipment supportability analysis knowledge, aiming at the problem of time-consuming and laborious artificial analysis in the current supportability field, realizes data-driven in the mode of knowledge graph, optimizes the analysis efficiency, and improves the data boundary determination, training link and model structure. On the data, combined with the characteristics that the analysis process in the supportability field is complex and there is no previous unified experience, the supportability data is divided into 5 fields for the first time, and the corresponding logical relationship is proposed for each field as the basis of the pattern layer construction, and the data range and boundary of the supportability historical objective data are determined.

[0019] (2) In the training mode of the model, the present application deviates from the traditional knowledge extraction process, uses the single-step single-model joint extraction idea, and obtains the corresponding triplets in one step. Compared with the traditional knowledge extraction method, one model training link is saved, the algorithm cost of model training and the difficulty of business personnel are greatly reduced, and the support field is more willing to accept the use of historical objective data. In the model structure, the scorer composed of bidirectional long short-term memory network and self-attention mechanism is introduced to strengthen the model's ability to capture the context and improve the accuracy of single-step-single-model knowledge extraction for support data. Finally, based on the knowledge extraction, the equipment supportability analysis knowledge graph is constructed to realize the structurization and utilization of support history and objective data. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 It is a method flowchart of the present application; Figure 2 It is a pre-training model processing process schematic diagram of the present application; Figure 3 It is a model network structure diagram of the present application; Figure 4 It is a decoding process and scorer connection logic schematic diagram of the present application; Figure 5 It is a comparison schematic diagram between entity relationship joint extraction and traditional process in the embodiment of the present application; Figure 6 It is a pattern layer schematic diagram created in the embodiment of the present application; Figure 7 A schematic diagram of a stored map in an embodiment of the present application. DETAILED DESCRIPTION

[0021] Embodiments of the present application will be described below with reference to the accompanying drawings.

[0022] Specifically, in one aspect, the present application provides a method for extracting supportability analysis knowledge of aviation equipment based on deep learning, which comprises the following steps: S1, data requirement analysis: collecting aviation equipment life cycle data. In specific embodiments, first determine the aviation equipment life cycle data requirements, sort out the data items in the massive data that affect the design of equipment supportability and actual support work, involving support task analysis, support resource analysis, support requirement trade-off analysis and determination, maintenance optimization analysis and support strategy analysis data in 5 fields.

[0023] Among them, the aviation equipment life cycle data includes demonstration stage data, engineering development stage data, test identification stage data, production stage data and use and support stage data. The demonstration stage data includes equipment demand analysis system data and equipment demand management system data; the engineering development stage data includes new equipment development requirement data, alternative support preliminary scheme data, equipment support design result data and supportability analysis result data; the test identification stage data includes new equipment test data and support resource matching scheme data; the production stage data includes equipment production data and support resource development data; the use and support stage data includes support system and resource data and equipment use data. And according to the actual conditions of knowledge utilization, it is divided into 5 fields of support task analysis, support resource analysis, support requirement trade-off analysis and determination, maintenance optimization analysis and support strategy analysis.

[0024] S2, database construction: combining supportability data requirements, building mode layer, using top-down method, defining entity and relationship information. The method for constructing the database specifically includes the following sub-steps: S21, define the boundary and scope of supportability data, determine supportability analysis from life cycle data, and the business fields contained in the original data are: support requirement trade-off analysis and determination, support task analysis, support resource analysis, support strategy analysis, maintenance optimization analysis and support data verification; S22, display constraints on related concepts, entities or attributes in combination with the data requirements of the above data, after detailing the data fields in the above business process, summarize the entity categories and their relationships that should be included in the map; S23, the entity types include equipment, system, subsystem, product, analysis business, analysis step, failure mode, failure cause and maintenance measure, maintenance task category, maintenance frequency and interval, support resource information and support model information; the relationship types include belong, represent, cause and correspond.

[0025] S3, a deep learning extraction model for support analysis knowledge is constructed, the network model includes a model training unit, a scoring unit and an entity relationship extraction unit, the model training unit is used for training the deep learning extraction model, the scoring unit uses BiLSTM and self-attention layer as a scorer to output a three-dimensional matrix representing classification results, and the entity relationship extraction unit is based on the three-dimensional matrix of entity and relationship triplets, and specifically includes the following sub-steps: S31, a BERT pre-training model is constructed as the model training unit, the BERT pre-training model receives Input_ids and Attention_mask as input, learns deep-level representation of the input sentence through a multi-layer Transformer architecture, and specifically: in a text sample with words, each word is converted into a dimensional vector, and the vector representation of the sentence is: , represents the th word in the text converted into a dimensional vector, is a certain word in 1, 2...M. Then the vector is converted into a combined representation , the combined representation is a dimensional vector, represents the combined representation of the th and th words, and finally the text is converted into a three-dimensional matrix as the input of the scorer.

[0026] , .

[0027] The BERT pre-training model construction in step S31 specifically includes the following sub-steps: S311. Data Processing: The data processing workflow aims to transform unstructured raw text sequences into a standardized input format suitable for deep learning models to jointly extract entities and relations. This workflow can be systematically divided into two core stages: data preprocessing and serialization, and model adaptation. In the first stage, the raw corpus is first loaded and cleaned to eliminate irrelevant characters and formatting noise, ensuring the uniformity and cleanliness of the text. Then, relying on punctuation rules or using mature natural language processing tools (such as spaCy), the cleaned text is accurately segmented into independent sentence units, thus forming the basic samples for subsequent analysis. On this basis, annotation experts or automated annotation tools identify the entity boundaries and their types in each sentence, and further determine the semantic relationships between entities, forming structured annotation information. Subsequently, each sentence and its corresponding entity and relation labels are integrated into a structured record and uniformly packaged into a JSON format file to achieve persistent data storage and efficient exchange. In the second stage, the processing flow focuses on integration with the BERT model. Specifically, it requires loading the tokenizer corresponding to a specified pre-trained Chinese model (such as BERT-Base, Chinese). This tokenizer transforms each text sample in the JSON file into an input sequence acceptable to the model. This process includes adding special tags and performing sub-word segmentation, mapping tokens to their corresponding index numbers, and generating attention masks. For sequence labeling tasks, entity and relation labels need to be precisely aligned with the segmented tokens. In joint entity and relation extraction, entity and relation labels include four types: HB-TE, HB-TE, HE-TE, and none. Finally, the encoded index sequence, attention mask, and aligned labels are organized into a tensor and constructed as a data loader, thus providing end-to-end data supply for the model's batch training and inference.

[0028] S312. Data Training: The labeled data is divided into training, validation, and test sets. During the training phase, a deep learning model for joint entity relation extraction is adopted. The results of the pre-trained model BERT are used as input. The idea of ​​joint entity relation extraction is used to represent the results of the pre-trained model BERT. The vector features contained in the text are concatenated and fused with the vector features of a specific relation. Then, the vectors are passed to the entity relation extraction unit through the network layer. Finally, the accuracy of the model is improved through iterative training, and a deep learning model for joint entity relation extraction that can identify the head and tail entities under a specific relation is obtained.

[0029] S32. Scoring is performed using a scorer composed of a BiLSTM and a self-attention layer. The calculation formula for the scorer is as follows: ; in, wherein, and are the weights and biases in the nonlinear mapping, respectively, SelfAttention is a self-attention layer, and softmax is an activation function used to convert the output of the scorer into a classification over the labels.

[0030] wherein, The specific calculation of is as follows: wherein, is the input vector of the current time step t, is the hidden state of the previous time step t-1, is the cell state of the current time step, is the cell state of the previous time step, are the input gate, forget gate, output gate and candidate memory content of the current time step, respectively, are the corresponding weights, respectively, are the corresponding biases, respectively, and represent the forward and backward hidden states, respectively.

[0031] In the classification process of step S32, four types of relationship labels are first defined, and the four types of labels are head entity head-tail entity head (HB-TB), head entity head-tail entity tail (HB-TE), head entity tail-tail entity tail (HE-TE), and non-entity position (none). At the same time, the ​​​​​​​​​​​​​​​​​​​​are grouped into one of the four types of labels, thus forming a new matrix to be decoded . The new matrix represents the information of the head and tail entities of the relation . After the scoring of all relations is completed, the new matrix to be decoded is expanded to .

[0032] In specific embodiments, in the classification process, four types of labels are first defined, i.e., “head entity head-tail entity head (HB-TB)”, “head entity head-tail entity tail (HB-TE)”, “head entity tail-tail entity tail (HE-TE)”, and “non-entity position (none)”, and the defined relations are encoded . The matrix is scored under each relation to obtain the classification results of each .

[0033] S33, entity relation extraction: based on the classification results of the three-dimensional matrix , the three-tuple of entities and relations is obtained by adding an index and decoding. In the matrix under the relation , the corresponding label of the tail entity is between HB-TB and HB-TE, and the corresponding label of the head entity is between HB-TE and HE-TE. For example, in the matrix under the relation , the position between “HB-TB” and “HB-TE” is the position of the corresponding Chinese character of the tail entity. The specific decoding process is as follows: in the matrix to be decoded formed by scoring the three-dimensional matrix under the i-th relation , the position of the tail entity is between the words with the corresponding labels of HB-TB and HB-TE, the position of the head entity is between the words with the corresponding labels of HB-TE and HE-TE, and the other irrelevant positions are filled with the label none. By locating the positions of the three labels (HB-TB, HB-TE, and HE-TE) in the four types of labels, the corresponding Chinese characters of the head and tail entities can be decoded, and combined with the relation to obtain the three-tuple. By decoding the -dimensional data sample output by the scorer, all the guarantee analysis knowledge three-tuples can be obtained.

[0034] The specific decoding method is shown in Table 1. The connection logic of the decoding process and the scorer is shown in Figure 4 .

[0035] Table 1 Confirming entity of location Corresponding label Tail entity Between HB-TB and HB-TE Head entity Between HB-TE and HE-TE In specific embodiments, the present embodiments take aviation electromechanical products as an example. First, according to the analysis results of the aviation electromechanical product support data demand, relevant text data in 5 fields is collected to form a support analysis knowledge base. The original data in the database is unstructured text, which is cleaned, divided into sentences, and labeled to form a JSON format data set containing entity and relationship annotations, totaling about 50,000 sentences. Subsequently, combined with the aviation electromechanical product support data demand, the mode layer is built, and a top-down method is used to define entity and relationship information. Taking the data in the support task analysis field as an example, the created mode layer includes 43 types of entities and 3 types of relationships, as shown in Figure 6 .

[0036] Then, joint entity and relationship extraction is performed. The text data is divided into training set, test set and validation set in the ratio of 8:1:1. The training set is input into the model training unit and scoring unit, and the trained scorer can accurately identify the entities and relationships in the support text, with a rate of 91% on the test set. The extraction results on the test set are sent to the entity and relationship extraction unit, and the triple information in the text is obtained after decoding.

[0037] Based on the entity and relationship triplets, the graph database is used to complete the construction of the knowledge graph of the aviation electromechanical product: taking the neo4j graph database as the carrier of knowledge storage, deploying the knowledge graph construction script on the server, taking the extraction results as the input, automatically reading the entities and relationships, generating Cypher statements to realize the construction of the local knowledge graph; the implementation of this script uses the interactive function of Python and Cypher statements, and its processing process is: opening the local neo4j graph database, reading the entity and relationship results decoded in the previous step, identifying and establishing entity types and relationship types, identifying entity and relationship specific content and generating a local graph and saving it. The stored graph provides a data basis for the use of objective support historical data, for example, part of the graph is as shown in Figure 7 .

[0038] The storage script can directly connect with the joint entity and relationship extraction results, and compared with the traditional entity extraction and relationship extraction, the process is greatly optimized, as shown in Figure 5 .

[0039] The above embodiments only describe the preferred embodiments of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A knowledge extraction method for aviation equipment supportability analysis based on deep learning, characterized in that: It includes the following steps: S1. Data Requirements Analysis: Collect data throughout the entire lifecycle of aviation equipment; S2. Database Construction: Based on the data requirements of the assurance analysis, the schema layer is built. A top-down approach is used to define entity and relation information to complete the construction of the assurance analysis database. S3. Construct a deep learning extraction model for knowledge extraction in safeguard analysis to extract entity and relation triples. The deep learning extraction model includes a model training unit, a scoring unit, and an entity and relation extraction unit. The model training unit trains the deep learning extraction model. The scoring unit uses BiLSTM and self-attention layers as scorers to output a three-dimensional matrix representing the classification results. The entity and relation extraction unit is based on the entity and relation triples in the three-dimensional matrix, and specifically includes the following sub-steps: S31. Construct a BERT pre-trained model as the model training unit. The BERT pre-trained model receives Input_ids and Attention_mask as inputs, learns a deep representation of the input sentence through a multi-layer Transformer architecture, and outputs a three-dimensional matrix as the input to the scorer. Specifically, in a text sample, it has... If there are 10 characters, then each character will be converted into 10 characters. The vector representation of this sentence is: , Indicates the first in the text Transformed from individual characters A 1-dimensional vector is then converted into a composite representation. Combination representation for A vector of dimension Indicates the first and the The text is represented by a combination of individual characters, and finally transformed into a three-dimensional matrix. , as input to the scorer; ， ； S32. Using a scorer composed of BiLSTM and self-attention layers to evaluate a three-dimensional matrix. Score and output each The classification results and the scoring formula of the scorer are as follows: ； in, This is the output of the BiLSTM; , These represent the weights and biases in the nonlinear mapping; SelfAttention is the self-attention layer, and softmax is the activation function used to convert the scorer's output into a classification of the label; S33. Entity Relationship Extraction: Based on a 3D Matrix The classification results are labeled and decoded to obtain triples of entities and relations.

2. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 1, characterized in that: In step S32 of the classification process, four types of relationship labels are first defined: head entity head-tail entity head HB-TB, head entity head-tail entity tail HB-TE, head entity tail-tail entity tail HE-TE, and non-entity position none. Simultaneously, in the graph mode layer, the following labels are defined: Encode the relationships Each specific relationship Each corresponds to a three-dimensional matrix. The dimensions of this data sample are expanded to The scorer evaluates the three-dimensional matrix under each relation. Each one is scored once, and each of the three-dimensional matrix is ​​scored. The labels are divided into four categories to form a new matrix to be decoded. The new matrix can represent the relationship After scoring all relationships, the information of the head and tail entities is used to expand the new matrix to be decoded into the following form: .

3. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 1, characterized in that: The decoding process in step S33 is as follows: in the i-th relation Lower three-dimensional matrix The matrix to be decoded is formed by scoring. In the text, the tail entity is positioned between the words with the corresponding tags HB-TB and HB-TE, the head entity is positioned between the words with the corresponding tags HB-TE and HE-TE, and other irrelevant positions are filled with the none tag.

4. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 1, characterized in that: It also includes step S4, which uses a graph database to construct a knowledge graph based on triples of entities and relations. Specifically, the Neo4j graph database is used as the knowledge storage carrier, and a knowledge graph construction script is deployed on the server. The extracted results are used as input to automatically read entities and relations and generate Cypher statements to build a local knowledge graph.

5. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 1, characterized in that: In step S32 The specific calculation is as follows: ； ； ； ； ； ； ； in, The input vector at the current time step t, This is the hidden state of the previous time step t-1. This represents the cell state at the current time step. This represents the cell state at the previous time step. , , , These are the input gate, forget gate, output gate, and candidate memory content for the current time step. These are the corresponding weights. These are the corresponding biases. and These represent the hidden states for the forward and backward directions, respectively.

6. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 1, characterized in that: Step S1 specifically involves: determining the data requirements for the entire life cycle of aviation equipment, sorting out the data items in the massive data that affect the equipment supportability design and actual support work, involving data in five areas: support task analysis, support resource analysis, supportability requirement trade-off analysis and determination, maintenance optimization analysis, and support strategy analysis.

7. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 6, characterized in that: The full life cycle data of aviation equipment in step S1 includes data from the demonstration stage, engineering development stage, testing and evaluation stage, production stage, and usage support stage.

8. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 7, characterized in that: Data during the feasibility study phase includes data from the equipment requirements analysis system and data from the equipment requirements management system; Data during the engineering development phase includes data on new equipment development requirements, data on preliminary alternative support schemes, data on equipment support design results, and data on supportability analysis results. Data in the testing and evaluation phase includes new equipment testing data and support resource matching plan data; data in the production phase includes equipment production data and support resource development data; data in the operation and support phase includes support system and resource data and equipment usage data.

9. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 1, characterized in that: The method for constructing the database in step S2 specifically includes the following sub-steps: S21. Define the boundaries and scope of supportability analysis data. Determine supportability analysis from the full life cycle data of aviation equipment. The business areas included in the raw data are: supportability demand trade-off analysis and determination, support task analysis, support resource analysis, support strategy analysis, maintenance optimization analysis, and support data verification. S22. Based on the data requirements of the above data, explicit constraints are applied to the relevant concepts, entities, or attributes to obtain the entity categories that should be included in the graph and their relationships. S23. The identified entity categories include equipment, systems, subsystems, products, analysis operations, analysis steps, failure modes, failure causes, maintenance measures, maintenance task categories, maintenance frequency and intervals, support resource information, and support model information; the relationship types include belonging, representing, causing, and corresponding.

10. The knowledge extraction method for aviation equipment supportability analysis based on deep learning according to claim 1, characterized in that: The training process of the model training unit in step S31 specifically includes the following sub-steps: S311, Data Processing: Transform unstructured raw text sequences into a standardized input form suitable for deep learning models and perform joint entity and relation extraction; S312. Data Training: The data is divided into training set, validation set and test set in a ratio of 8:1:

1. During the training phase, a deep learning model for joint entity relation extraction is adopted. The results of the pre-trained model BERT are used as input. The results of the pre-trained model BERT are used to represent relations. The vector features contained in the text are concatenated and fused with the vector features of a specific relation. Then, the vectors are passed to the entity relation extraction unit through the network layer, and finally iterative training is completed.

Images