Spacecraft overall design large model data acquisition method and equipment based on knowledge graph

Abstract

The invention discloses a knowledge graph-based spacecraft overall design large model data acquisition method and equipment, belongs to the technical field of spacecraft overall design, and solves the problems of difficulty in knowledge acquisition, low knowledge reuse rate and weak knowledge intelligent application capability in the existing spaceflight field. The method comprises the following steps: acquiring a data source in the field of overall design of the spacecraft, wherein the data source comprises unstructured data; marking the data source in the step 1; a Bi-LSTM + CRF network is used for named entity recognition, and a Bi-LSTM + CRF model comprises an embedded layer, a Bi-LSTM and a conditional random field CRF; establishing a top layer structure of the knowledge graph, wherein the top layer structure comprises an entity type, a relation type and an attribute value; a database is selected according to the data volume and the query requirement for triple storage; the application of intelligent questioning and answering, a recommendation system and decision-making assistance is realized based on the knowledge graph. The method is suitable for obtaining the fine adjustment data of the spacecraft overall design large model.

Description

Technical Field

[0001] This application relates to the technical field of overall spacecraft design, and particularly to the acquisition of large model data for overall spacecraft design based on a knowledge graph. Background Art

[0002] The current sources for acquiring overall spacecraft design knowledge are divided into two major categories: unstructured knowledge and structured knowledge. Unstructured knowledge can be further divided into explicit knowledge and implicit knowledge. Among them, explicit knowledge mainly includes texts such as engineering manuals, papers, and mission reports formed during the development of models, and implicit knowledge mainly refers to the knowledge contained in expert experience. Structured knowledge mainly consists of the design parameters, test data, and mission logs of spacecraft stored in a MySQL database.

[0003] The main defect of storing unstructured knowledge is that it is difficult to structure, which in turn leads to difficulties in automated analysis and intelligent applications, seriously affecting the reuse of knowledge. The main defect of storing structured knowledge is that it is difficult to store the simulation experiment results mainly in the form of pictures and texts, and it does not support the functions of fast query and inference query. At the same time, such databases of major research institutions are all closed-source, and data is difficult to obtain. In addition to the storage structure, the problem of knowledge base isolation also seriously affects the reuse of knowledge. Traditional overall spacecraft design knowledge bases are mostly stored independently according to the R & D situation of subsystems. Due to the differences in the design teams of each subsystem, the storage logic of the knowledge base will be different, thus forming the problem of knowledge base isolation, seriously affecting the knowledge sharing efficiency and design collaboration efficiency in the development of new models. In addition, when each team constructs a knowledge base, there is a lack of a consistent knowledge quality evaluation system and standards, which will lead to problems of knowledge errors when applying design knowledge across subsystems. Seriously, it will lead to the failure of spacecraft launch and development, causing significant property losses and even safety problems. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems of difficult knowledge acquisition, low knowledge reuse rate, and weak knowledge intelligent application ability in the existing aerospace field, and provides a method and device for acquiring large model data for overall spacecraft design based on a knowledge graph.

[0005] The present invention is realized through the following technical solutions. On the one hand, the present invention provides a method for acquiring large model data for overall spacecraft design based on a knowledge graph, and the method includes:

[0006] Step 1: Obtain the data sources in the field of overall spacecraft design, and the data sources include unstructured data;

[0007] Step 2: Label the data sources in Step 1;

[0008] Step 3: Use the Bi-LSTM+CRF network for named entity recognition. The Bi-LSTM+CRF model includes an embedding layer, Bi-LSTM, and conditional random field CRF.

[0009] Step 4: Establish the top-level structure of the knowledge graph, including entity types, relationship types, and attribute values.

[0010] Step 5: Select a database for storing triples according to the data volume and query requirements.

[0011] Step 6: Develop applications based on the knowledge graph to implement intelligent question answering, recommendation systems, and decision-making assistance.

[0012] Further, step 2 is specifically: Obtain semi-structured data using BIO annotation.

[0013] Further, the Bi-LSTM+CRF model in step 3 is based on the TensorFlow deep learning framework in Keras. Build a Bi-LSTM classification model and use the Python 3.5 environment that supports Chinese processing to compile and run the algorithm.

[0014] The word vector dimension is set to 200 dimensions; the Bi-LSTM model uses 64 neurons, and the batch size is set to 10; Use the callback function to obtain the optimal model parameters, select the Adadelta function as the optimizer, and the learning rate is the default value; The activation function uses softmax, and the loss function uses categorical_corssentropy.

[0015] Further, the embedding layer includes: For an input sequence X=(x1,x2,…,x n ), where n represents the length of the sequence. For each word x∈X in the sentence, construct an embeddingi = w; c h ; c a , where w represents the semantic-level feature, c h represents the character set feature, and c a is used to represent the case feature, and the dimensions are d w , d o and d c .

[0016] Further, the Bi-LSTM is used to extract potential features in the statement; In the embedding layer, use the extraction model to obtain the embedding of each word and use it as a sequence I=(i1,…,i t ,…,i n ) and input it into the Bi-LSTM, where

[0017] There are two LSTM models in the Bi-LSTM, one extracts forward hidden features and the other extracts backward hidden features Specifically, the Bi-LSTM is as follows:

[0018] i t = σ(w xi x t + W hi h t-1 + W ci c t-1 + b i )

[0019] f t = σ(W fi x t + W hf h t-1 + W cf c t-1 + b f )

[0020] c t = f t c t-1 + i t tanh(W xc x t + W hc h t-1 + b c )

[0021] o t = σ(W xo x t + W ho h t-1 + W co c t-1 + b o )

[0022] h t = o t tanh(c t )

[0023] Among them, i t , f t , c t , o t respectively represent the input gate, forget gate, output gate and memory cell at time t, σ represents the activation function, W x , W h , W c represent the weight matrices, b i , b f , b c , b o represent the bias terms;

[0024]

[0025] Combine the forward and backward hidden layer features to obtain the output Use an activation function to extract and features, and then map them to a d l -dimensional space;

[0026] l = tanh(W l tanh(W t h t + b t ) + b l )

[0027] where d l is the number of entity types to be recognized, W l and W t are feature matrices, and b t and b l are bias terms.

[0028] Furthermore, for the features output by each Bi-LSTM, the conditional random field CRF is used to identify the specific label y t at each position of the sequence. The outputs l of the Bi-LSTM layer are combined as a sequence L = [l1, l2,..., l N T and input into the CRF model, where the size of L is n × d l , representing the probability that the i-th word in the sentence is recognized as the j-th label. For a predicted sequence y = (y1, y2,..., y n ), the scoring function of the CRF is defined as follows:

[0029]

[0030] where T represents a transition matrix, and T i,j represents the probability of transitioning from the i-th label to the j-th label. y0 and y n+1 represent the labels at the start and end positions respectively; then the joint extraction model uses a normalization function softmax to calculate the probability of the sequence y;

[0031]

[0032] where Y represents all sequences that X may generate.

[0033] Furthermore, step 4 includes:

[0034] ​Step 4.1: Define the acquisition scope of the ontology: Specifically, determine which step of the overall design the knowledge is acquired up to;

[0035] Step 4.2: Define the knowledge items required in the ontology: Compile the acquisition rules, acquire structured and unstructured data, evaluate the acquired knowledge items, conduct manual verification, form the acquisition criteria to support the completion of the overall design, and form a

[0036] knowledge base;

[0037] Define the acquired knowledge categories: Classify the knowledge items in the knowledge base into categories and fill in the knowledge items according to the defined categories;

[0038] Step 4.3: Define the attributes of the knowledge categories: Define entity attributes, data attributes, domain and range, and entity attribute characteristics;

[0039] Step 4.4: Establish knowledge entities: Create corresponding knowledge entities in the ontology according to the data in the knowledge base.

[0040] Furthermore, step 5 includes:

[0041] Import the semi-structured data in json format into the Neo4j graph database through the Py2neo library; Call the write interface of the py2neo library in the Python environment. The py2neo library is used to create entity nodes, the edges of entity relationships, and execute Cypher query statements;

[0042] Adopt two import methods, the py2neo library and the load csv command, to import the semi-structured data in json format formed after triple extraction into the Neo4j graph data browser.

[0043] In a second aspect, the present invention provides a computer device, including a memory and a processor. A computer program is stored in the memory. When the processor runs the computer program stored in the memory, it executes the steps of a method for acquiring data of a large model for spacecraft overall design based on a knowledge graph as described above.

[0044] In a third aspect, the present invention provides a computer-readable storage medium. Multiple computer instructions are stored in the computer-readable storage medium. The multiple computer instructions are used to cause a computer to execute a method for acquiring data of a large model for spacecraft overall design based on a knowledge graph as described above.

[0045] Advantages of the present invention:

[0046] The present invention proposes to use knowledge graph technology to acquire structured knowledge from unstructured information data in the field of spacecraft overall design and store it in a graph database with query and reasoning functions in an inferable graph structure, such asFigure 1 as shown

[0047] The main content provided by the spacecraft design knowledge graph of the present invention is how to utilize the capabilities of knowledge representation, semantic reasoning, and knowledge fusion of the knowledge graph to solve specific problems such as difficult knowledge acquisition, low knowledge reuse rate, and weak knowledge intelligent application ability in the aerospace field. The present invention applies knowledge graph technology to fuse structured, semi-structured, and unstructured knowledge in the overall spacecraft design field to form an interpretable and inferable high-quality knowledge network under unified expression. However, many problems are encountered in this process:

[0048] First, when defining the knowledge acquisition source, the quality of knowledge acquisition is the key factor in the success or failure of knowledge graph construction. Due to insufficient self-developed data volume, a large amount of expert-level data and general knowledge repositories need to be collected from domestic and foreign sources as supplements, including some closed-source data of research institutions, and professional permissions need to be applied for. The dataset range applicable to the model established by the present invention is wide, including but not limited to: technical documents of institutions such as NASA and JPL, papers related to overall spacecraft design in top international journals in the aerospace discipline, spacecraft-related entries in Wikipedia, domestic textbooks related to overall spacecraft design in the aerospace discipline, and publicly available launch data of domestic and foreign commercial aerospace companies, etc.

[0049] Secondly, in the process of entity, relationship, and attribute triple extraction, a large number of entity categories need to be defined. These entity categories need to be accurate and general, and meet the requirement that the reuse rate reaches more than 80% during the design process. Excessive low-usage and non-usage categories will lead to a reduction in query inference speed and affect the application effect of the knowledge graph. When defining the relationship categories between entities, the design knowledge in the original data does not have relevant information. It is necessary to extract implicit relationship data from the original data according to design experience, and calculate the shortest logical relationship path between entity categories through the path optimization algorithm. Finally, after multiple optimizations, the final entity relationship library is formed. When extracting attribute values, since the numerical units of overall spacecraft design at home and abroad are different in each sub-link, it is necessary to compile an attribute value link library to convert all attribute values in foreign data into domestic standards.

[0050] Finally, when fusing knowledge in multiple heterogeneous data sources, it is necessary to align, simplify entity information, remove redundant information, fill in missing information, and construct a complete high-quality knowledge graph under unified expression. Since the knowledge information in heterogeneous data sources is quite different, it is necessary to complete the knowledge fusion link with high quality in a low-computing-power experimental environment. The present invention adopts a matching algorithm based on string similarity with low computing power requirements to calculate the similarity of word vectors and complete knowledge fusion.

[0051] The present invention is applicable to the acquisition of fine-tuning data for the overall spacecraft design large model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] To more clearly illustrate the technical solutions of the present application, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, other accompanying drawings can also be obtained based on these drawings without creative efforts.

[0053] Figure 1 is the overall technical architecture of the knowledge graph;

[0054] Figure 2 is the overall technical framework of the method of the present invention;

[0055] Figure 3 is the schematic diagram of the BIO annotation experiment;

[0056] Figure 4 is the BIO annotation experiment environment;

[0057] Figure 5 is the schematic diagram of the embedding layer structure of the Bi-LSTM network model;

[0058] Figure 6 is the schematic diagram of the BiLSTM+CRF network model structure;

[0059] Figure 7 is the definition of the relationship between entities in the attitude and orbit control subsystem;

[0060] Figure 8 is the interface code for importing triple data into the Neo4j graph database;

[0061] Figure 9 is the visualization interface diagram of the knowledge graph in the Neo4j graph database;

[0062] Figure 10 is the spacecraft historical design case recommendation system based on the knowledge graph. Specific Embodiments

[0063] The following details the embodiments of the present invention. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present invention, but should not be construed as limiting the present invention.

[0064] Embodiment 1. A method for obtaining data of a large model for overall spacecraft design based on a knowledge graph, the method comprising:

[0065] Step 1: Obtain data sources in the field of overall spacecraft design, the data sources including unstructured data;

[0066] Step 2: Annotate the data sources in Step 1;

[0067] Step 3: Use the Bi-LSTM+CRF network for named entity recognition. The Bi-LSTM+CRF model includes an embedding layer, a Bi-LSTM, and a conditional random field CRF.

[0068] Step 4: Establish the top-level structure of the knowledge graph, including entity types, relationship types, and attribute values.

[0069] Step 5: Select a database for storing triples according to the data volume and query requirements.

[0070] Step 6: Develop applications based on the knowledge graph to implement intelligent question answering, recommendation systems, and decision-making assistance.

[0071] In the implementation of this solution, the data source is mainly based on the domestic advanced small satellite design manual as the main data structure, supplemented by authoritative books on the overall design of spacecraft in the domestic and international aerospace fields, forming a large number of expert-level data sources.

[0072] Embodiment 2: This embodiment further limits a method for obtaining data of a large model for the overall design of spacecraft based on a knowledge graph as described above. In this embodiment, step 2 is further limited, specifically including:

[0073] Step 2 is specifically: Use BIO annotation to obtain semi-structured data.

[0074] In this embodiment, BIO annotation is a text annotation method commonly used in natural language processing (NLP) and information extraction tasks, used to identify the boundaries and types of entities or phrases in the text. BIO is an abbreviation for "Beginning, Inside, Outside", which represents three basic states of annotating entities:

[0075] B (Beginning): Represents the beginning part of an entity.

[0076] I (Inside): Represents the internal part of an entity, that is, it is not the first word of the entity, but still belongs to the same entity.

[0077] O (Outside): Represents a word or part that does not belong to any entity.

[0078] A typical application of BIO annotation is named entity recognition (NER), where the task is to identify named entities in the text, such as person names, place names, organization names, etc. Using BIO annotation, each word in the text can be annotated as B, I, or O to indicate whether they belong to a named entity, and if so, whether they are the first word (B) or internal word (I) of the entity.

[0079] In this embodiment, through BIO annotation, named entities in the text and their boundaries can be conveniently identified, which is very useful for NLP tasks such as information extraction and text classification.

[0080] Embodiment 3 further limits a method for obtaining data of a large model for overall spacecraft design based on a knowledge graph as described above. In this embodiment, the Bi-LSTM+CRF model in step 3 is further limited, specifically including:

[0081] The Bi-LSTM+CRF model in step 3 is based on the TensorFlow deep learning framework in Keras to construct a Bi-LSTM classification model, and the Python 3.5 environment that supports Chinese processing is used to compile and run the algorithm;

[0082] The dimension of the word vector is set to 200; the Bi-LSTM model uses 64 neurons, and the batch size is set to 10; the callback function is used to obtain the optimal model parameters, the optimizer selects the Adadelta function, and the learning rate is the default value; the activation function uses softmax, and the loss function uses categorical_corssentropy.

[0083] Embodiment 4 further limits a method for obtaining data of a large model for overall spacecraft design based on a knowledge graph as described above. In this embodiment, the embedding layer is further limited, specifically including:

[0084] The embedding layer includes: for an input sequence X=(x1, x2, …, x n ), where n represents the length of the sequence. For each word x∈X in the sentence, an embeddingi = [w; c h ; c a consisting of three parts is constructed, where w represents the semantic-level feature, c h represents the feature of the character set, and c a is used to represent the case feature, and the dimensions are d w , d o and d c .

[0085] In this embodiment, the purpose of the character set feature c h is to extract the features of the word as comprehensively as possible.

[0086] Embodiment 5 further limits a method for obtaining data of a large model for overall spacecraft design based on a knowledge graph as described above. In this embodiment, the Bi-LSTM is further limited, specifically including:

[0087] The Bi-LSTM is used to extract potential features in a sentence; in the embedding layer, an extraction model is used to obtain the embedding of each word and use it as a sequence I = (i1, …, i t , …, i n ) and input it into the Bi-LSTM, where

[0088] The Bi-LSTM contains two LSTM models, one for extracting forward hidden features and the other for extracting backward hidden features Specifically, the Bi-LSTM is as follows:

[0089] i t = σ(W xi x t + W hi h t-1 + W ci c t-1 + b i )

[0090] f t = σ(W fi x t + W hf h t-1 + W cf c t-1 + b f )

[0091] c t = f t c t-1 + i t tanh(W xc x t + W hc h t-1 + b c )

[0092] o t = σ(W xo x t + W ho h t-1 + W co c t-1 + b o )

[0093] h t = o t tanh(c t )

[0094] where i t , f t , c t , o trespectively represent the input gate, forget gate, output gate, and memory cell at time t, σ represents the activation function, W x , W h , W c represent the weight matrices, b i , b f , b c , b o represent the bias terms;

[0095]

[0096] Combine the forward and backward hidden layer features to obtain the output Use an activation function to extract and features, and then map them to a d l -dimensional space;

[0097] l = tanh(W l tanh(W t h t + b t ) + b l )

[0098] where d l is the number of entity types to be recognized, W l and W t are the feature matrices, b t and b l are the bias terms.

[0099] In this embodiment, the Bi-LSTM bidirectional long short-term memory network can simultaneously obtain the features of the context information of the entity when processing the entity features of the text. For the proper nouns and pronouns in the field, disambiguation processing is realized from the source of acquisition. Compared with the traditional knowledge acquisition method, the accuracy of acquiring knowledge is significantly improved, which is the key technology to ensure the knowledge accuracy and reliability of the expert-level spacecraft overall design knowledge graph constructed by this patent.

[0100] Embodiment 6 further limits a method for obtaining data of a spacecraft overall design large model based on a knowledge graph as described above. In this embodiment, the conditional random field CRF is further limited, specifically including:

[0101] For each feature output by the Bi-LSTM, the conditional random field is used to identify the specific label y t at each position of the sequence, and the output l of the Bi-LSTM layer is combined as a sequence L = [l1, l2,..., l N T ​Input into the CRF model, where the size of l is n×d l , representing the probability that the i-th word in the sentence is recognized as the j-th label. For a predicted sequence y=(y1,y2,…,y n ), the scoring function of the CRF is defined as follows:

[0102]

[0103] where T represents a transition matrix, and T i,j represents the probability of transitioning from the i-th label to the j-th label, and y0 and y n+1 represent the labels at the start and end positions respectively; then the joint extraction model uses a normalization function softmax to calculate the probability of the sequence y;

[0104]

[0105] where Y represents all sequences that X may generate.

[0106] In this embodiment, compared with other models, the conditional random field can better consider the dependency relationships between different labels and thereby improve the accuracy of entity recognition.

[0107] Embodiment Seven. This embodiment further limits a method for obtaining data of a large model for overall spacecraft design based on a knowledge graph as described above. In this embodiment, step 4 is further limited and specifically includes:

[0108] Step 4 includes:

[0109] Step 4.1: Define the acquisition scope of the ontology: Specifically, determine which step of the overall design the knowledge is acquired to;

[0110] Step 4.2: Define the knowledge items required in the ontology: Compile acquisition rules, acquire structured and unstructured data, evaluate the acquired knowledge items, and perform manual compounding to form an acquisition standard to support the completion of the overall design, and form a knowledge

[0111] base;

[0112] Define the acquired knowledge categories: Classify the knowledge items in the knowledge base into categories and fill in the knowledge items according to the defined categories;

[0113] Step 4.3: Define the attributes of the knowledge categories: Define entity attributes, data attributes, domain and range, and entity attribute features;

[0114] Step 4.4: Establish knowledge entities: Create corresponding knowledge entities in the ontology according to the data in the knowledge base.

[0115] This embodiment provides a design method for the top-level structure of a knowledge graph. Combining the technical characteristics of the knowledge graph, it realizes the structured sorting of the complex overall spacecraft design knowledge. Compared with the traditional design mode of one library for one type of large spacecraft, the knowledge graph structure designed in this patent combines the characteristics of generalization, platformization, modularization, and flexibility of small satellites, and accurately defines a general knowledge base for the overall design of micro and small spacecraft that can be highly reused to meet the requirements of rapid design capabilities.

[0116] Embodiment 8. This embodiment further limits a method for obtaining data of a large model for overall spacecraft design based on a knowledge graph as described above. In this embodiment, step 5 is further limited, specifically including:

[0117] Step 5 includes:

[0118] Import the semi-structured data in json format into the Neo4j graph database through the Py2neo library; call the write interface of the py2neo library in the Python environment. The py2neo library is used to create entity nodes and the edges of entity relationships and execute Cypher query statements;

[0119] Use two import methods, the py2neo library and the load csv command, to import the semi-structured data in json format formed after triple extraction into the Neo4j graph data browser.

[0120] This embodiment provides a triple storage scheme for the data of the large model for overall spacecraft design. Among them, JSON format files cannot be directly imported into the Neo4j graph database, so it is necessary to call the write interface of the py2neo library in the Python environment.

[0121] Embodiment 9. This embodiment is an example of a method for obtaining data of a large model for overall spacecraft design based on a knowledge graph as described above. As Figure 2 shown, it specifically includes:

[0122] Step 1: Define the data source: mainly use unstructured data such as the design manuals of self-developed satellites by Harbin Institute of Technology Satellite Technology Research, relevant works on overall spacecraft design, and professional textbooks on overall aircraft design, supplemented by structured data such as databases and web pages.

[0123] Step 2: Entity recognition and extraction: Complete the triple extraction of the data source through named entity recognition experiments, relationship extraction experiments, and attribute extraction experiments, including:

[0124] Obtaining semi-structured data based on BIO annotation, specifically:

[0125] BIO tagging is a text tagging method commonly used in natural language processing (NLP) and information extraction tasks, which is used to identify the boundaries and types of entities or phrases in the text. BIO is an abbreviation for "Beginning, Inside, Outside", which represents three basic states of tagging entities:

[0126] 1) B (Beginning): Represents the beginning part of an entity.

[0127] 2) I (Inside): Represents the internal part of an entity, that is, it is not the first word of the entity but still belongs to the same entity.

[0128] 3) O (Outside): Represents the words or parts that do not belong to any entity.

[0129] A typical application of BIO tagging is named entity recognition (NER), where the task is to identify named entities in the text, such as person names, place names, organization names, etc. Using BIO tagging, each word in the text can be tagged as B, I, or O to indicate whether they belong to a named entity, and if so, whether they are the first word (B) or the internal word (I) of the entity. As Figure 3 shown, Figure 3 is an example of semi-structured data based on BIO tagging.

[0130] As Figure 4 shown, in the "SMAD" BIO tagging experiment in the spirit annotation software environment, through BIO tagging, it is convenient to identify named entities in the text and their boundaries, which is very useful for NLP tasks such as information extraction and text classification.

[0131] Named entity recognition, specifically:

[0132] The Bi-LSTM+CRF network is a combination of neural network structures commonly used in sequence tagging tasks, especially effective in dealing with NER (named entity recognition tasks). By combining Bi-LSTM and CRF, this model has the ability to simultaneously model context information and label sequence dependencies, bringing higher robustness and accuracy.

[0133] As Figure 6 shown, the Bi-LSTM+CRF model of this embodiment includes three layers, namely the embedding layer, Bi-LSTM, and CRF.

[0134] The embedding layer (Embedding Layer) as Figure 5 shown, for an input sequence X=(x1, x2, …, x n), where n represents the length of the sequence. For each word x ∈ x in the sentence, construct an embedding i = [w; c h ; c a , where w represents the semantic-level feature, c h represents the character set feature, and c a is used to represent the case feature. Their dimensions are d w , d o , and d c , respectively. The purpose of the character-level feature c h is to extract the features of the word as comprehensively as possible. Specifically, the joint extraction model inputs the embedding of each character in the word into a Bi-LSTM, and then takes the output vector of the nth Bi-LSTM model as the overall character set feature c h of the word. Therefore, the total length d i of the input vector = d w + d o + d c . The specific formula of Bi-LSTM is as follows.

[0135] i t = σ(W xi x t + W hi h t-1 + W ci c t-1 + b i ) (1)

[0136] f t = σ(W xf x t + W hf h t-1 + W cf c t-1 + b f ) (2)

[0137] c t = f t c t-1 + i t tanh(W xc x t + W hc h t-1 + b c ) (3)

[0138] o t = σ(W xo x t + W ho h t-1 + W co ct-1 +b o ) (4)

[0139] h t = o t tanh(c t ) (5)

[0140] where i t , f t , c t , o t represent the input gate, forget gate, output gate, and cell memory at time t, respectively. σ represents the activation function, W x , W h , W c represent the weight matrices, b i , b f , b c , b o represent the bias terms, h t represents the final hidden state, which is the processed result of c t through the output gate. All the above parameters need to obtain the final results during training. For the sake of simplicity, in the following text, the LSTM model is briefly represented by LSTM(i t , h t-1 ).

[0141] Bidirectional LSTM layer (Bi-LSTM layer) The main role of Bi-LSTM is to extract the latent features in the sentence. In the embedding layer, the extraction model can obtain the embedding of each word and use it as a sequence I = (i1,…, i t ,…, i n ) input into the Bi-LSTM, where the Bi-LSTM contains two LSTM models, one extracting the forward hidden features and the other extracting the backward hidden features The Bi-LSTM can be represented as follows:

[0142]

[0143] Merging the forward and backward hidden layer features can obtain the output Using an activation function l to extract the and features, and then mapping them into a d l -dimensional space.

[0144] l = tanh(W ltanh(W t h t +b t )+b l ) (7)

[0145] where d l is the number of entity types to be recognized, W l and W t are feature matrices, and b t and b l are bias terms.

[0146] For the features output by each Bi-LSTM, the model uses a conditional random field to identify the specific label y t at each position in the sequence. Compared with other models, the conditional random field can better consider the dependencies between different labels and thus improve the accuracy of entity recognition. Specifically, the model combines the outputs l of the Bi-LSTM layer as a sequence L = [l1, l2, …, l N T and inputs it into the CRF model, where the size of L is n×d l , representing the probability that the i-th word in the sentence is recognized as the j-th label. For a predicted sequence y = (y1, y2, …, y n ), the scoring function of the CRF is defined as follows:

[0147]

[0148] where T represents a transition matrix, T i,j represents the probability of transitioning from the i-th label to the j-th label, and y0 and y n+1 represent the labels at the start and end positions respectively. Then the joint extraction model uses a normalization function softmax to calculate the probability of the sequence y.

[0149]

[0150] where Y represents all possible sequences that X can generate.

[0151] ​The above BiLSTM+CRF model structure is based on the TensorFlow deep learning framework in Keras. A Bi-LSTM classification model is constructed, and the algorithm is compiled and run using a Python 3.5 environment that supports Chinese processing. The dimension of the word vector is set to 200 dimensions. The Bi-LSTM model uses 64 neurons, and the batch size is set to 10. The callback function is used to obtain the optimal model parameters, and the Adadelta function is selected as the optimizer with the learning rate set to the default value. The softmax activation function is used, and the categorical_corssentropy loss function is used.

[0152] (1) Knowledge representation: Design the top-level structure of the knowledge graph, covering entity types, relationship types, attribute values, etc., as Figure 7 shown.

[0153] step1: Define the acquisition scope of the ontology (specifically, which step of the overall design to acquire knowledge to);

[0154] step2: Define the knowledge items required in the ontology (compile the acquisition rules, acquire structured and unstructured data, evaluate the acquired knowledge items, manually composite them to form the acquisition criteria to support the completion of the overall design, and form a knowledge base);

[0155] step3: Define the acquired knowledge categories (classify the knowledge items in the knowledge base. This step can be carried out simultaneously with step2, and fill the knowledge items according to the defined categories);

[0156] step4: Define the attributes of the knowledge categories (as introduced in the above text about the attributes in the ontology, the entity attributes and data attributes, domain and range, and the definition of entity attribute characteristics need to be carried out);

[0157] step5: Create knowledge entities (create corresponding knowledge entities in the ontology according to the data in the knowledge base).

[0158] (2) Knowledge storage: Select the Neo4j graph database for triple storage according to the data volume and query requirements. As Figure 8 and Figure 9 shown.

[0159] The semi-structured data in JSON format is imported into the Neo4j graph database through the Py2neo library. The JSON format file cannot be directly imported into the Neo4j graph database. The py2neo library needs to be called in the Python environment to write the interface. The py2neo library can be used to create entity nodes and the edges of entity relationships, and can also be used to execute Cypher query statements.

[0160] Two import methods, namely using the py2neo library and the load csv command, are adopted to import the semi-structured data in JSON format formed after triple extraction into the Neo4j graph data browser.

[0161] (3) Query and application: Develop applications such as intelligent question answering, recommendation systems, and decision-making assistance based on the knowledge graph, as Figure 10 shown.

[0162] The design of the neural network model of the historical design case recommendation system based on the knowledge graph includes the following steps:

[0163] step1: Convert the design requirement features and multiple knowledge features of historical design cases into feature vectors respectively;

[0164] step2: Further extract features from these feature vectors using fully connected layers or convolutional layers;

[0165] step3: Fuse the feature vectors of the design requirements and multiple knowledge of historical design cases into a vector representation for convenient similarity calculation;

[0166] step4: Calculate the cosine similarity between features, and use the mean square error of this similarity and the design case value as the loss function of this regression model.

Claims

1. A method for acquiring large-scale model data of spacecraft overall design based on knowledge graph, characterized in that: The method comprises: Step 1: Acquire data sources in the field of spacecraft overall design, where the data sources include unstructured data; Step 2: Label the data source in step 1; Step 3: Use the Bi-LSTM+CRF network to perform named entity recognition, where the Bi-LSTM+CRF model includes an embedding layer, a Bi-LSTM, and a conditional random field CRF; Step 4: Establish the top-level structure of the knowledge graph, including entity types, relationship types, and attribute values; Step 5: Select a database to store triples based on the data volume and query requirements; Step 6: Develop applications based on knowledge graphs to implement intelligent question answering, recommendation systems, and decision support.

2. According to claim 1, a method for acquiring large-scale model data of spacecraft overall design based on knowledge graph is characterized in that: Step 2 is specifically: using BIO annotation to obtain semi-structured data.

3. The method for acquiring data of a large model of spacecraft overall design based on a knowledge graph according to claim 1, characterized in that: The Bi-LSTM+CRF model in step 3 is based on the Tensorflow deep learning framework in Keras to build a Bi-LSTM classification model, and the algorithm is compiled and run using the Python 3.5 environment that supports Chinese processing; The word embedding dimension is set to 200 dimensions; the Bi-LSTM model uses 64 neurons and the batch size is set to 10; the callback function is used to obtain the optimal model parameters, the optimizer selects the Adadelta function, and the learning rate is the default value; the activation function uses softmax, and the loss function uses categorical_corssentropy.

4. The method for acquiring data of a large model of spacecraft overall design based on a knowledge graph according to claim 1 is characterized in that: The embedding layer includes: for an input sequence X = (x1, x2, ..., x n ), where n represents the length of the sequence. For each word x∈X in the sentence, we construct an embedding i=[w; c h ;c a ], where w represents the semantic level feature, c h Indicates the characteristics of the character set, c a Used to represent the features of uppercase and lowercase, the dimensions are d w , d o and d c .

5. The method for acquiring data of a large model of spacecraft overall design based on a knowledge graph according to claim 4 is characterized in that: The Bi-LSTM is used to extract potential features in sentences; In the embedding layer, the extraction model is used to obtain the embedding of each word and used as a sequence I = (i1, ..., i t ,…,i n ) is input into Bi-LSTM, where Bi-LSTM contains two LSTM models, one extracting forward hidden features Another method to extract reverse hidden features Bi-LSTM is specifically: i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i ) f t =σ(W fi x t +W hf h t-1 +W cf c t-1 +b f ) c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c ) o t =σ(W xo x t +W ho h t-1 +W co c t-1 +b o ) h t =o t fishy t ) Among them, i t ,f t ,c t ,o t They represent the input gate, forget gate, output gate and memory unit at time t, σ represents the activation function, W x ,W h ,W c represents the weight matrix, b i ,b f ,b c ,b o represents the bias term; Merge the forward and backward hidden layer features to get the output Use an activation function to extract and The features are then mapped to a d l dimensional space; l=tanh(W l tanh(W t h t +b t )+b l ) Among them, d l is the number of entity types that need to be identified, W l and W t is the feature matrix, b t and b l is the bias term.

6. The method for acquiring data of a large model of a spacecraft overall design based on a knowledge graph according to claim 5, characterized in that: The conditional random field CRF uses the conditional random field to identify the specific label y of each position in the sequence for each feature output by the Bi-LSTM. t , merge the output l of the Bi-LSTM layer into a sequence L = [l1,l2,…,l N ] T Input to the CRF model, where the size of L is n×d l , which represents the probability that the i-th word in the sentence is recognized as the j-th label, for a prediction sequence y = (y1, y2, ..., y n ), the score function of CRF is defined as follows: Where T represents a transfer matrix, T i,j represents the probability of the i-th label being transferred to the j-th label, y0 and y n+1 The labels represent the starting and ending positions respectively; then the joint extraction model uses a normalized function softmax to calculate the probability of the sequence y; Here, Y represents all possible sequences generated by X.

7. The method for acquiring data of a large model of spacecraft overall design based on a knowledge graph according to claim 1, characterized in that: Step 4 includes: Step 4.1: Define the scope of ontology acquisition: the specific step of knowledge acquisition to the overall design; Step 4.2: Define the knowledge items required in the ontology: compile acquisition rules, acquire structured and unstructured data, evaluate the acquired knowledge items, manually compound them, form acquisition standards that support the completion of the overall design, and form a knowledge base; Define and acquire knowledge categories: classify the knowledge items in the knowledge base into categories and fill in the knowledge items according to the defined categories; Step 4.3: Define the attributes of the knowledge category: define entity attributes and data attributes, domains and ranges, and entity attribute characteristics; Step 4.4: Establish knowledge entities: Create corresponding knowledge entities in the ontology based on the data in the knowledge base.

8. The method for acquiring data of a large model of spacecraft overall design based on a knowledge graph according to claim 1, characterized in that: Step 5 includes: Import semi-structured data in json format into Neo4j graph database through Py2neo library; call py2neo library to write interface in Python environment, py2neo library is used to create entity nodes and entity relationship edges and execute Cypher query statements; The py2neo library and the load csv command are used to import the semi-structured data in json format after triple extraction into the Neo4j graph data browser.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor runs the computer program stored in the memory, the steps of the method according to any one of claims 1 to 8 are performed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of computer instructions, and the plurality of computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 8.

Images