Aircraft semantic knowledge structuring method and system

Abstract

The invention belongs to the technical field of aircraft design, and particularly relates to an aircraft semantic knowledge structuring method and system, and the method comprises the steps: 1, carrying out the paragraph splitting of an aircraft design document; the aircraft design document comprises a task description document, an overall scheme document, a system analysis document and an experimental data document; 2, predicting a design stage corresponding to the split paragraph, and constructing a design stage mark alignment structure; the design stage comprises a task definition stage, an overall scheme stage, a pneumatic design stage, a structure design stage, a system analysis stage and a test verification stage; step 3, carrying out semantic and design parameter extraction on the design stage mark alignment structure; and 4, based on the extracted semantics, designing parameters, and constructing a knowledge graph.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of aircraft design, and particularly relates to an aircraft semantic knowledge structuring method and system. BACKGROUND

[0002] With the deep integration of artificial intelligence technology and aerospace engineering, knowledge-driven intelligent design is becoming an important trend in aircraft system research and development. Especially in the process of developing modern aircrafts with strong task adaptability, fast design iteration and high structural complexity, how to fully tap the large amount of engineering experience and semantic knowledge contained in the design stage, and build a knowledge structure system to support automatic reasoning and optimization decision-making, has become an important foundation for the intelligentization of aircraft design.

[0003] Aircraft design knowledge is scattered in unstructured documents, including design specifications, simulation reports, test records and other text forms. The content is redundant, the structure is loose, and the semantics is inconsistent, which seriously restricts the automatic acquisition, structured management and reuse efficiency of knowledge. Traditional design knowledge modeling methods mostly rely on manual annotation or expert rules, which have problems such as low extraction efficiency, insufficient semantic coverage, and missing knowledge evolution path, and are difficult to meet the efficient modeling and intelligent collaboration needs of complex design tasks.

[0004] Some studies attempt to introduce natural language processing methods to extract keywords and entities from design texts, but in the context of aircraft design, there are problems such as ambiguous professional term expression, ambiguous parameter description, and semantic distribution across paragraphs, which often lead to rough knowledge extraction granularity and insufficient structure expression capability. At the same time, existing methods often ignore the phased characteristics of the design process and the evolution path formed by the iteration of knowledge with the scheme, and cannot truly reflect the structural logic and evolution relationship of aircraft design knowledge.

[0005] In view of the above technical defects, the present application is proposed. SUMMARY

[0006] The purpose of the present application is to provide an aircraft semantic knowledge structuring method and system to overcome or alleviate at least one aspect of the known technical defects.

[0007] The technical solution of the present application is:

[0008] An aircraft semantic knowledge structuring method, comprising:

[0009] Step 1, paragraph splitting of aircraft design documents;

[0010] The aircraft design documents include task description, overall scheme, system analysis and experimental data documents;

[0011] Step two, predict the design phase corresponding to the split paragraph, and construct a design phase label alignment structure;

[0012] Design phase, including task definition, overall scheme, aerodynamic design, structural design, system analysis, test verification phase;

[0013] Step three, semantic and design parameter extraction is performed on the design phase label alignment structure;

[0014] Step four, based on the extracted semantic and design parameters, a knowledge graph is constructed.

[0015] Optionally, in the above aircraft semantic knowledge structuring method, in step two, the split paragraph set is input into the semantic phase structure alignment model, the semantic phase structure alignment model adopts the RoBERTa variant structure, the design phase corresponding to the split paragraph is predicted, and common design phase vocabulary and key prompt words in the aircraft development process are introduced for supervision.

[0016] Optionally, in the above aircraft semantic knowledge structuring method, in step two, in the semantic phase structure alignment model, the split paragraph is BERT encoded, and then a classifier with Softmax added on the basis of RoBERTa is used to predict the corresponding design phase, and a design phase label alignment structure is constructed.

[0017] Optionally, in the above aircraft semantic knowledge structuring method, in step two, in the semantic phase structure alignment model, a context semantic boundary detection mechanism is introduced, combined with syntactic dependency analysis and BERT embedded local consistency scoring, to improve the accuracy of boundary recognition of different design phases of the split paragraph.

[0018] Optionally, in the above aircraft semantic knowledge structuring method, in step three, the design phase label alignment structure is input into the semantic processing model for semantic and design parameter extraction;

[0019] The semantic processing model performs semantic extraction on the design phase label alignment structure, identifies the content conforming to the "design premise-design entity-response scheme" template, and extracts it as a structured triple;

[0020] When the semantic processing model performs semantic extraction on the design phase label alignment structure, a BERT encoding+BiLSTM+CRF structure is adopted, and a structure guiding token is introduced to improve the boundary recognition ability.

[0021] Optionally, in the aircraft semantic knowledge structuring method described above, in step three, the semantic processing model extracts design parameters from the design stage mark alignment structure, identifies the content of the design parameter value expression, extracts the minimum value vmin and the maximum value vmax of the design parameter, and generates a confidence p to model the design parameter interval;

[0022] When the semantic processing model extracts design parameters from the design stage mark alignment structure, for a single-value description of the design parameter value, the minimum value vmin and the maximum value vmax of the design parameter are equal, and for a fuzzy description of the design parameter value, the minimum value vmin and the maximum value vmax of the design parameter are calculated according to the typical margin coefficient learned from the context.

[0023] When the semantic processing model extracts design parameters from the design stage mark alignment structure, for the text of the design parameter value expression content, BERT encoding is performed to obtain an embedding vector H, which is input into two independent fully connected networks to predict the minimum value vmin and the maximum value vmax, and is input into a Sigmoid network to predict the confidence p.

[0024] Optionally, in the aircraft semantic knowledge structuring method described above, in step three, a back-translation model or a paraphrase model is used to perform structure-preserving semantic enhancement on the split paragraphs to generate equivalent but different sentence style domain descriptions, 2-3 semantic equivalent but different sentence style variants are generated for each split paragraph, and a diverse but consistent structure expression is obtained to build an enhanced corpus set.

[0025] When the context diversity of the design stage mark alignment structure is enhanced, a term freezing mechanism is introduced.

[0026] A pre-training language model based on the Transformer architecture is selected, and the semantic processing model is fine-tuned in the aircraft design field using the enhanced corpus set. The goal of fine-tuning is to minimize the comprehensive loss function S:

[0027] S = LMLM + λ × Ldomain + μ × Lstage

[0028] Wherein,

[0029] LMLM is the mask language modeling loss;

[0030] Ldomain is the design entity recognition loss;

[0031] Lstage is the design stage label constraint loss;

[0032] λ and μ are task weight coefficients.

[0033] Optionally, in the aircraft semantic knowledge structuring method described above, in step four, the extracted semantics and design parameters are input into a knowledge graph construction model to construct a knowledge graph.

[0034] The knowledge graph includes an entity set E, a relationship R between entities, and an evolution path graph P, and is stored in the form of a graph database or a nested key-value structure to support retrieval and reasoning.

[0035] Optionally, in the aircraft semantic knowledge structuring method described above, in step four, a semantic embedding mechanism is used to convert all identified design elements into nodes of the knowledge graph.

[0036] The semantic consistency score between the nodes of the knowledge graph is calculated:

[0037] Sim(i,j)=cos(i_vec, j_vec)×δ_structure×γ_context×ω_attribute;

[0038] Sim(i,j) is the semantic consistency score between the nodes i and j of the knowledge graph;

[0039] i_vec and j_vec are embedding vectors generated by BERT encoding of the nodes i and j of the knowledge graph;

[0040] δ_structure is a Boolean item indicating whether the structure categories of the nodes of the knowledge graph are consistent;

[0041] γ_context is the similarity of the nodes of the knowledge graph in the context;

[0042] ω_attribute is the consistency rate of the attribute fields of the nodes of the knowledge graph;

[0043] If Sim(i,j) exceeds the consistency threshold, the nodes i and j of the knowledge graph are fused;

[0044] The evolution path graph P is embedded in the knowledge graph in the form of a directed subgraph.

[0045] Optionally, in the aircraft semantic knowledge structuring method described above, in step four, the quality score of the knowledge graph is calculated to evaluate the quality of the knowledge graph:

[0046] Q(G)=α×C(G)+β×P(G)+γ×Con(G);

[0047] wherein,

[0048] Q (G) is the quality score of the knowledge graph;

[0049] C(G) is the coverage of the knowledge graph;

[0050] P(G) is the precision of the knowledge graph;

[0051] Con(G) is the consistency of the knowledge graph;

[0052] α, β, γ are weights.

[0053] An aircraft semantic knowledge structuring system comprises:

[0054] A paragraph splitting module is configured to split the aircraft design document into paragraphs.

[0055] The aircraft design document comprises a task description, a general scheme, a system analysis, and an experimental data document.

[0056] A design phase label alignment module is configured to predict the design phase corresponding to the split paragraphs and construct a design phase label alignment structure.

[0057] The design phase comprises a task definition, a general scheme, an aerodynamic design, a structural design, a system analysis, and a test verification phase.

[0058] A semantic and design parameter extraction module is configured to extract semantics and design parameters from the design phase label alignment structure.

[0059] A knowledge graph construction module is configured to construct a knowledge graph based on the extracted semantics and design parameters.

[0060] Optionally, in the aircraft semantic knowledge structuring system, the design phase label alignment module inputs the split paragraphs into a semantic phase structure alignment model, the semantic phase structure alignment model adopts a RoBERTa variant structure, predicts the design phase corresponding to the split paragraphs, and introduces common design phase vocabulary and key prompt words in the aircraft development process for supervision.

[0061] Optionally, in the aircraft semantic knowledge structuring system, the design phase label alignment module, in the semantic phase structure alignment model, encodes the split paragraphs using BERT, and then uses a classifier with Softmax added on the basis of RoBERTa to predict the corresponding design phase and construct the design phase label alignment structure.

[0062] Optionally, in the aircraft semantic knowledge structuring system, the design phase label alignment module, in the semantic phase structure alignment model, introduces a context semantic boundary detection mechanism, combines syntactic dependency analysis and local consistency scoring of BERT embedding, and improves the accuracy of boundary recognition of different design phases of the split paragraphs.

[0063] Optionally, in the aircraft semantic knowledge structuring system described above, in the semantic and design parameter extraction module, the design stage mark alignment structure is input into the semantic processing model for semantic and design parameter extraction;

[0064] The semantic processing model performs semantic extraction on the design stage mark alignment structure, identifies content conforming to the "design premise-design entity-response scheme" template, and extracts it as a structured triple;

[0065] When the semantic processing model performs semantic extraction on the design stage mark alignment structure, the BERT encoding+BiLSTM+CRF structure is adopted, and the structure guide token is introduced to improve the boundary recognition capability.

[0066] Optionally, in the aircraft semantic knowledge structuring system described above, in the semantic and design parameter extraction module, the semantic processing model performs design parameter extraction on the design stage mark alignment structure, identifies the content of the design parameter value expression, extracts the minimum value vmin and the maximum value vmax of the design parameter, and generates a confidence p to model the design parameter interval.

[0067] When the semantic processing model performs design parameter extraction on the design stage mark alignment structure, for a single-value description of the design parameter value, the minimum value vmin and the maximum value vmax of the design parameter are equal, and for a fuzzy description of the design parameter value, the typical margin coefficient learned from the context is used to calculate the minimum value vmin and the maximum value vmax of the design parameter.

[0068] When the semantic processing model performs design parameter extraction on the design stage mark alignment structure, for the text of the design parameter value expression content, BERT encoding is performed to obtain an embedding vector H, which is input into two independent fully connected networks to predict the minimum value vmin and the maximum value vmax, and is input into a Sigmoid network to predict the confidence p.

[0069] Optionally, in the aircraft semantic knowledge structuring system described above, in the semantic and design parameter extraction module, a back-translation model or a paraphrase model is used to perform structure-preserving semantic enhancement on the split paragraphs to generate equivalent but different sentence descriptions. For each split paragraph, 2-3 semantic equivalent but different sentence variants are generated to obtain diverse but consistent expressions and build an enhanced corpus set.

[0070] When the context diversity of the design stage mark alignment structure is enhanced, a term freezing mechanism is introduced.

[0071] A pre-training language model based on the Transformer architecture is selected, and the enhanced corpus set is used to perform aircraft design field multi-task fine-tuning on the semantic processing model. The goal of fine-tuning is to minimize the comprehensive loss function S:

[0072] S = LMLM + λ × Ldomain + μ × Lstage;

[0073] wherein,

[0074] LMLM is a masked language modeling loss;

[0075] Ldomain is a design entity recognition loss;

[0076] Lstage is a design stage label constraint loss;

[0077] λ and μ are task weight coefficients.

[0078] Optionally, in the aircraft semantic knowledge structuring system, the extracted semantics and design parameters are input into a knowledge graph construction model in the knowledge graph construction module to construct a knowledge graph.

[0079] The knowledge graph includes an entity set E, a relationship R between entities, and an evolution path graph P, and is stored in the form of a graph database or a nested key-value structure to support retrieval and reasoning.

[0080] Optionally, in the aircraft semantic knowledge structuring system, a semantic embedding mechanism is used in the knowledge graph construction module to convert all recognized design elements into nodes of the knowledge graph.

[0081] The semantic consistency score between knowledge graph nodes is calculated:

[0082] Sim(i,j) = cos(i_vec, j_vec) × δ_structure × γ_context × ω_attribute;

[0083] Sim(i,j) is the semantic consistency score between i and j knowledge graph nodes;

[0084] i_vec and j_vec are embedding vectors generated by BERT encoding of i and j knowledge graph nodes;

[0085] δ_structure is a Boolean item indicating whether the structure categories of the knowledge graph nodes are consistent;

[0086] γ_context is the similarity of the knowledge graph nodes in the context;

[0087] ω_attribute is the consistency rate of the attribute fields of the knowledge graph nodes;

[0088] If Sim(i,j) exceeds the consistency threshold, i and j knowledge graph nodes are fused;

[0089] The evolution path graph P is embedded into the knowledge graph in the form of a directed subgraph.

[0090] Optionally, in the aircraft semantic knowledge structuring system, the knowledge graph construction module calculates a quality score of the knowledge graph to evaluate the quality of the knowledge graph.

[0091] Q(G) = a x C(G) + b x P(G) + g x Con(G);

[0092] wherein,

[0093] Q(G) is the quality score of the knowledge graph;

[0094] C(G) is the coverage of the knowledge graph;

[0095] P(G) is the precision of the knowledge graph;

[0096] Con(G) is the consistency of the knowledge graph;

[0097] a, b, and g are weights.

[0098] The present application has at least the following beneficial technical effects:

[0099] The aircraft semantic knowledge structuring method and system provided by the present application integrate technical means such as semantic segmentation, semantic template extraction, parameter interval modeling, evolution path construction, and knowledge graph construction, realize automatic identification, structured expression, and dynamic evolution modeling of implicit semantic knowledge in aircraft design documents, improve the organization efficiency, expression accuracy, and reasoning support capability of design knowledge, significantly improve the accuracy, maintainability, and intelligent reasoning capability of aircraft design knowledge modeling, and can be widely applied to aircraft design process management, intelligent design auxiliary systems, engineering knowledge base construction, and task-oriented design scheme deduction scenarios, improve the design knowledge engineering expression capability, provide intelligent semantic support and knowledge reuse capability for aircraft complex system design, and have important engineering application value and promotion potential. BRIEF DESCRIPTION OF DRAWINGS

[0100] Figure 1 is a schematic diagram of the aircraft semantic knowledge structuring method provided by the present application;

[0101] Figure 2 is a schematic diagram of the aircraft semantic knowledge structuring system provided by the present application.

[0102] In order to better illustrate the present embodiment, some contents in the drawings may be omitted, enlarged, or reduced, and are only used for exemplary illustration, and cannot be understood as a limitation on the present application. DETAILED DESCRIPTION

[0103] In order to make the technical solutions of the present application and their advantages clearer, the technical solutions of the present application will be further clearly and completely described below in conjunction with the drawings. It should be understood that the specific embodiments described herein are only part of the embodiments of the present application, and are only used to explain the present application, but not to limit the present application. It should be noted that, for the purpose of description, only parts related to the present application are shown in the drawings, and other related parts can be referred to the general design.

[0104] In addition, unless otherwise defined, the technical terms or scientific terms used in the description of the present application should be the general meanings understood by the general skilled person in the field to which the present application belongs. In the description of the present application, “comprising” indicates that the concept appearing before the word encompasses the concepts listed after the word and their equivalents, and does not exclude other associated concepts.

[0105] A method for structuring semantic knowledge of an aircraft, as shown in Figure 1 .

[0106] Step one, paragraph splitting of the aircraft design document.

[0107] The aircraft design document is subjected to document-level syntax cleaning and paragraph splitting. The split paragraph set , wherein, is the th split paragraph.

[0108] The aircraft design document includes task description, overall scheme, system analysis and experimental data, and related documents, without explicit design stage label, and without using structured syntax, containing a large amount of engineering language and professional terms.

[0109] Step two, predicting the design stage corresponding to the split paragraph, and constructing the design stage label alignment structure.

[0110] The split paragraph set is input into the semantic stage structure alignment model. The semantic stage structure alignment model uses the RoBERTa variant structure to predict the design stage corresponding to the split paragraph, and introduces the design stage common vocabulary and key prompt words in the aircraft development process for supervision.

[0111] In the semantic stage structure alignment model, the split paragraph is subjected to BERT encoding, and then a classifier with Softmax added on the basis of RoBERTa is used to predict the corresponding design stage, and the design stage label alignment structure is constructed, so as to ensure that the subsequent semantic processing task is in the same context window, wherein, is the design stage label alignment structure of the th split paragraph, is the th split paragraph, Design phase, including task definition, overall scheme, aerodynamic design, structural design, system analysis, test verification phase.

[0112] The semantic phase structure alignment model introduces a context semantic boundary detection mechanism, combines syntactic dependency analysis and local consistency scoring of BERT embedding, improves the accuracy of identifying the boundaries of different design phases of the split paragraphs, and avoids semantic processing deviation caused by design behavior across phase interference.

[0113] Step three, semantic and design parameter extraction of design phase mark alignment structure.

[0114] The design phase mark alignment structure is input into the semantic processing model for semantic and design parameter extraction.

[0115] The semantic processing model performs semantic extraction on the design phase mark alignment structure, identifies the content that meets the "design premise-design entity-response scheme" template, and extracts it as a structured triple. For example, for the content "in order to reduce resistance, the main wing adopts a sweepback angle configuration", identify "reduce resistance" as the design premise, "main wing" as the design entity, and "sweepback angle configuration" as the response scheme, and then build a triple.

[0116] When the semantic processing model performs semantic extraction on the design phase mark alignment structure, it uses BERT encoding + BiLSTM + CRF structure and introduces structure-guided tokens to improve boundary recognition capability.

[0117] The semantic processing model performs design parameter extraction on the design phase mark alignment structure, identifies the content of design parameter value expression, such as "the main wing span length should be controlled between 10.2 to 10.8 meters", extracts the minimum value vmin and maximum value vmax of the design parameter, and generates the confidence p to quantify the confidence of the design parameter extraction accuracy, and performs design parameter interval modeling.

[0118] When the semantic processing model performs design parameter extraction on the design phase mark alignment structure, for single-value description of design parameter value, such as "web thickness is 5cm", the minimum value vmin and maximum value vmax of the design parameter are equal, and for fuzzy description of design parameter value, such as "the length of the partition plate is not less than 7.5m", according to the typical margin coefficient learned from the context, the minimum value vmin and maximum value vmax of the design parameter are calculated.

[0119] When the semantic processing model designs the design parameter extraction of the design stage mark alignment structure, for the text of the numerical value expression content of the design parameter, BERT encoding is performed to obtain an embedding vector H, which is input into two independent fully connected networks to predict the minimum value vmin and the maximum value vmax, and is input into a Sigmoid network to predict the confidence p.

[0120] The semantic processing model performs training of the design parameter extraction of the design stage mark alignment structure, and a Huber function is used as a loss function to reduce the sensitivity to the design stage annotation error.

[0121] To improve the robustness of the semantic processing model, context diversity enhancement is performed on the design stage mark alignment structure, a back-translation model or a paraphrase model is used to perform structure-preserving semantic enhancement on the split paragraphs, equivalent but different in sentence style are generated, 2-3 semantic equivalent but different in sentence style variants are generated for each split paragraph, and multiple but consistent structure expressions are obtained to construct an enhanced corpus set.

[0122] When the context diversity enhancement is performed on the design stage mark alignment structure, a term freezing mechanism is introduced to avoid modification of the specific expressions of key parameters such as lift-drag ratio and wingspan when generating equivalent but different in sentence style domain descriptions, and to ensure the consistency of semantics.

[0123] A pre-training language model based on the Transformer architecture is selected, and the enhanced corpus set is used to perform aircraft design field multi-task fine-tuning on the semantic processing model, and the goal of fine-tuning is to minimize the comprehensive loss function S.

[0124] S = LMLM + λ × Ldomain + μ × Lstage;

[0125] wherein,

[0126] LMLM is a mask language modeling loss;

[0127] Ldomain is a design entity recognition loss;

[0128] Lstage is a design stage label constraint loss;

[0129] λ and μ are task weight coefficients.

[0130] Step four, based on the extracted semantics and design parameters, a knowledge graph is constructed.

[0131] The extracted semantics and design parameters are input into a knowledge graph construction model to construct a knowledge graph.

[0132] The knowledge graph comprises an entity set E, a relationship R between entities, and an evolution path graph P, and is stored in a graph database or a nested key-value structure, and supports retrieval and reasoning.

[0133] All identified design elements are converted into nodes of the knowledge graph using a semantic embedding mechanism, and the knowledge graph nodes comprise attribute fields, including entity source document number, context paragraph number, confidence, and version label.

[0134] It is determined whether there is a semantically similar node in the knowledge graph, that is, whether the meanings, structures, context, and the like of two entities are similar, and whether they belong to the same type of parameter, and when similar, the two nodes are fused to avoid duplication and reduce redundancy, and the specific process can be referred to as follows:

[0135] The semantic consistency score between knowledge graph nodes is calculated:

[0136] Sim(i,j) = cos(i_vec, j_vec) x δ_structure x γ_context x ω_attribute.

[0137] Sim(i,j) is the semantic consistency score between i and j knowledge graph nodes;

[0138] i_vec and j_vec are embedding vectors generated by BERT encoding of i and j knowledge graph nodes;

[0139] δ_structure is a Boolean item indicating whether the structure categories of the knowledge graph nodes are consistent;

[0140] γ_context is the similarity of the context of the knowledge graph nodes;

[0141] ω_attribute is the consistency rate of the attribute fields of the knowledge graph nodes.

[0142] If Sim(i,j) exceeds the consistency threshold value 0.85, i and j knowledge graph nodes are fused to improve the simplicity and consistency of the knowledge graph and reduce reasoning conflicts.

[0143] The evolution path graph P represents the ordered change relationship between design schemes, and its construction depends on design change statement recognition. For example, when it is recognized that "the aerodynamic optimization is performed on the scheme Sk in the scheme Sk+1", a directed edge rk from the node scheme Sk to the node scheme Sk+1 is created, and thus the evolution path graph P {(S1, r1, S2), …, (Sk, rk, Sk+1)} is constructed.

[0144] The evolution path graph P is embedded in the knowledge graph in the form of a directed subgraph, and embedding schemes evolve direction, design change description, change cause and result description, etc. Meta information, support path query and visual backtracking based on timestamp and design logic, suitable for large-scale multi-scheme parallel optimization and change control management scene.

[0145] Calculate the quality score of the knowledge graph to evaluate the quality of the knowledge graph:

[0146] Q(G) = a x C(G) + b x P(G) + g x Con(G);

[0147] Wherein,

[0148] Q(G) is the quality score of the knowledge graph;

[0149] C(G) is the coverage of the knowledge graph, which represents the coverage of the knowledge graph to the information in the aircraft design document;

[0150] P(G) is the precision of the knowledge graph, which represents the accuracy of the information in the knowledge graph;

[0151] Con(G) is the consistency of the knowledge graph, which represents the degree of absence of semantic conflicts or structural contradictions in the knowledge graph;

[0152] a, b, g are weights, usually a can be taken as 0.4, b can be taken as 0.4, g can be taken as 0.2, which can be dynamically adjusted according to the feedback of domain experts.

[0153] When the quality score of the knowledge graph is too low, for example, less than 90%, automatically identify the problems concentrated in which nodes or semantics, generate an error analysis log and submit it to the administrator, or trigger the fine tuning of related models for self correction.

[0154] In a specific example, the aircraft semantic knowledge structuring method disclosed in the above embodiment is used to process a 59-page PDF format overall design document of a certain type of high-speed aircraft, and the constructed knowledge graph involves 8267 entities and 20340 edges, with a coverage of 88.6%, a precision of 92.3%, a consistency of 95.1%, and a quality score of 91.4%. Compared with the analysis process without structuring extraction scheme, the model mapping speed is improved by 3.7 times, and the manual revision cost is reduced by 62%.

[0155] The aircraft semantic knowledge structuring method disclosed in the above embodiment has context context limiting ability for aircraft design document knowledge extraction through paragraph splitting and design stage alignment, avoids semantic confusion between different design stages, and can enhance the understanding and expression ability of complex logical relationship in multi-stage design task.

[0156] By identifying and extracting the contents of design premises, design entities, and response scheme modules in aircraft design documents, a complete semantic triple system is established to realize the structured expression and centralized management of knowledge, which can significantly improve the automation level and semantic accuracy of knowledge acquisition.

[0157] By identifying the content of design parameter value expression, predicting and extracting the minimum value v min, maximum value v max and confidence P of the design parameter, the modeling of a large number of parameters with engineering uncertainty or semantic ambiguity in aircraft design is realized, which can effectively improve the quantitative ability and engineering adaptability of knowledge expression.

[0158] By constructing an evolution path diagram to improve the visual traceability of the design scheme change process, the design scheme evolution logic and change relationship can be identified, which can support reasoning analysis based on design causal logic and provide important support for multi-scheme parallel design, comparison and optimization iteration.

[0159] A knowledge graph quality scoring system is constructed to support systematic detection and structural optimization of knowledge redundancy, conflict and missing problems in multiple sources, ensuring the unity and reliability of the knowledge graph under multiple data sources.

[0160] A kind of aircraft semantic knowledge structured system, as shown in Figure 2 , comprising:

[0161] Paragraph splitting module for splitting aircraft design documents into paragraphs;

[0162] The aircraft design documents include task description, overall scheme, system analysis and experimental data documents;

[0163] Design phase label alignment module for predicting the design phase corresponding to the split paragraphs and constructing design phase label alignment structure;

[0164] Design phase, including task definition, overall scheme, aerodynamic design, structural design, system analysis, test verification phase;

[0165] Semantic and design parameter extraction module for semantic and design parameter extraction of design phase label alignment structure;

[0166] Knowledge graph construction module for constructing knowledge graph based on extracted semantics and design parameters.

[0167] In the design phase label alignment module, the split paragraph set is input into the semantic phase structure alignment model, the semantic phase structure alignment model adopts the RoBERTa variant structure, predicts the design phase corresponding to the split paragraph, and introduces common design phase vocabulary and key prompt words in aircraft development process for supervision.

[0168] In the design stage label alignment module, the semantic stage structure alignment model encodes the split paragraphs by BERT, and then predicts the corresponding design stage by a classifier with Softmax added on the basis of RoBERTa to construct the design stage label alignment structure.

[0169] In the design stage label alignment module, the semantic stage structure alignment model introduces a context semantic boundary detection mechanism to improve the accuracy of identifying the boundaries of different design stages in the split paragraphs by combining syntactic dependency analysis and local consistency scoring of BERT embedding.

[0170] In the semantic and design parameter extraction module, the design stage label alignment structure is input into the semantic processing model for semantic and design parameter extraction.

[0171] The semantic processing model performs semantic extraction on the design stage label alignment structure to identify the content that meets the "design premise-design entity-response scheme" template and extract it as a structured triple.

[0172] When the semantic processing model performs semantic extraction on the design stage label alignment structure, it uses the BERT encoding + BiLSTM + CRF structure and introduces structure-guided tokens to improve boundary recognition capability.

[0173] In the semantic and design parameter extraction module, the semantic processing model performs design parameter extraction on the design stage label alignment structure to identify the content of design parameter numerical expression and extract the minimum value vmin and maximum value vmax of the design parameter, as well as generate a confidence p for design parameter interval modeling.

[0174] When the semantic processing model performs design parameter extraction on the design stage label alignment structure, for single-value description of design parameter value, the minimum value vmin and maximum value vmax of the design parameter are extracted, and for fuzzy description of design parameter value, the minimum value vmin and maximum value vmax of the design parameter are calculated based on the typical margin coefficient learned from the context.

[0175] When the semantic processing model performs design parameter extraction on the design stage label alignment structure, for the text of design parameter value expression content, BERT encoding is performed to obtain embedding vectors H, which are input into two independent fully connected networks for prediction of the minimum value vmin and maximum value vmax, and into a Sigmoid network for prediction of the confidence p.

[0176] In the semantic and design parameter extraction module, a back-translation model or a paraphrase model is used to perform structure-preserving semantic enhancement on the split paragraphs, generate equivalent but different sentence descriptions, generate 2-3 semantic equivalent but different sentence variants for each split paragraph, obtain diverse but consistent structure expressions, and construct an enhanced corpus set;

[0177] When the design stage label alignment structure is enhanced in context diversity, a term freezing mechanism is introduced.

[0178] A pre-trained language model based on the Transformer architecture is selected, and the enhanced corpus set is used to fine-tune the semantic processing model in the aircraft design field. The goal of fine-tuning is to minimize the comprehensive loss function S:

[0179] S = LMLM + λ × Ldomain + μ × Lstage;

[0180] Wherein,

[0181] LMLM is the mask language modeling loss;

[0182] Ldomain is the design entity recognition loss;

[0183] Lstage is the design stage label constraint loss;

[0184] λ and μ are task weight coefficients.

[0185] In the knowledge graph construction module, the extracted semantics and design parameters are input into the knowledge graph construction model to construct the knowledge graph.

[0186] The knowledge graph includes an entity set E, a relationship R between entities, and an evolution path graph P, which is stored in the form of a graph database or a nested key-value structure, supporting subsequent retrieval and reasoning.

[0187] In the knowledge graph construction module, a semantic embedding mechanism is used to convert all recognized design elements into knowledge graph nodes.

[0188] Calculate the semantic consistency score between knowledge graph nodes:

[0189] Sim(i,j) = cos(i_vec, j_vec) × δ_structure × γ_context × ω_attribute;

[0190] Sim(i,j) is the semantic consistency score between i and j knowledge graph nodes;

[0191] i_vec and j_vec are embedding vectors generated by BERT encoding of i and j knowledge graph nodes;

[0192] delta_structure is a Boolean item indicating whether the knowledge graph node structure categories are consistent;

[0193] gamma_context is the similarity of the knowledge graph node in the context;

[0194] omega_attribute is the consistency rate of the knowledge graph node attribute field;

[0195] If Sim(i,j) exceeds the consistency threshold, the i and j knowledge graph nodes are fused;

[0196] The evolution path graph P is embedded in the knowledge graph in the form of a directed subgraph.

[0197] In the knowledge graph construction module, the knowledge graph quality score is calculated, and the knowledge graph quality is evaluated:

[0198] Q(G) = alpha * C(G) + beta * P(G) + gamma * Con(G);

[0199] Wherein,

[0200] Q(G) is the knowledge graph quality score;

[0201] C(G) is the knowledge graph coverage;

[0202] P(G) is the knowledge graph precision;

[0203] Con(G) is the knowledge graph consistency;

[0204] Alpha, beta and gamma are weights.

[0205] For the aircraft semantic knowledge structuring system disclosed in the above embodiment, since it corresponds to the aircraft semantic knowledge structuring method disclosed in the above embodiment, the description is relatively simple, and the specific related parts can be referred to the related description in the aircraft semantic knowledge structuring method part. The technical effects can also be referred to the technical effects of the aircraft semantic knowledge structuring method related part, which will not be repeated here.

[0206] In addition, those skilled in the art should also realize that each module of the aircraft semantic knowledge structuring system disclosed in the present application can be realized by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the present application is generally described in terms of function. Whether the described function is implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can choose different methods to implement the described functions for each specific application and its actual constraints, but such implementation should not be considered beyond the scope of the present application.

[0207] The technical scheme of the present application has been described in combination with the preferred embodiments shown in the drawings, and those skilled in the art should understand that the protection scope of the present application is obviously not limited to these specific embodiments, and those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical schemes after the changes or replacements will fall within the protection scope of the present application.

Claims

1. An aircraft semantic knowledge structuring method, characterized in that, The application relates to a method for constructing a knowledge graph of an aircraft design document. Step one, paragraph splitting of an aircraft design document; The aircraft design document comprises a task description, a general scheme, a system analysis and an experimental data document; Step two, prediction of a design stage corresponding to a split paragraph, and construction of a design stage label alignment structure; The design stage comprises a task definition, a general scheme, an aerodynamic design, a structural design, a system analysis and a test verification stage; Step three, semantic and design parameter extraction of the design stage label alignment structure; Step four, construction of a knowledge graph based on the extracted semantics and design parameters.

2. The aircraft semantic knowledge structuring method according to claim 1, characterized in that, In step two, the split paragraph set is input into a semantic stage structure alignment model, the semantic stage structure alignment model adopts a RoBERTa variant structure, the design stage corresponding to the split paragraph is predicted, and common design stage words and key prompt words in the aircraft development process are introduced for supervision.

3. The aircraft semantic knowledge structuring method according to claim 2, characterized in that, In step two, the split paragraph is encoded by BERT in the semantic stage structure alignment model, and then a classifier with a Softmax added on the basis of RoBERTa is used to predict the corresponding design stage, so as to construct the design stage label alignment structure.

4. The aircraft semantic knowledge structuring method according to claim 3, characterized in that, In step two, a context semantic boundary detection mechanism is introduced in the semantic stage structure alignment model, and the accuracy of boundary recognition of different design stages of the split paragraph is improved by combining syntactic dependency analysis and local consistency scores of BERT embedding.

5. The aircraft semantic knowledge structuring method according to claim 4, characterized in that, In step three, the design stage label alignment structure is input into a semantic processing model for semantic and design parameter extraction; The semantic processing model performs semantic extraction on the design stage label alignment structure, identifies the content conforming to the "design premise-design entity-response scheme" template, and extracts the content as a structured triple; When the semantic processing model performs semantic extraction on the design stage label alignment structure, the BERT encoding+BiLSTM+CRF structure is adopted, and the structure guiding token is introduced to improve the boundary recognition capability.

6. The aircraft semantic knowledge structuring method according to claim 5, characterized in that, In step three, the semantic processing model performs design parameter extraction on the design stage label alignment structure, identifies the content of the design parameter numerical expression, extracts the minimum value vmin and the maximum value vmax of the design parameter, and generates a confidence degree p to model the design parameter interval; When the semantic processing model performs design parameter extraction on the design stage label alignment structure, for the single-value description of the design parameter value, the minimum value vmin and the maximum value vmax of the design parameter are equal, for the fuzzy description of the design parameter value, the typical margin coefficient learned from the context is used to calculate the minimum value vmin and the maximum value vmax of the design parameter; When the semantic processing model performs design parameter extraction on the design stage label alignment structure, the text of the design parameter numerical expression content is BERT encoded to obtain an embedding vector H, which is input into two independent fully connected networks to predict the minimum value vmin and the maximum value vmax, and is input into a Sigmoid network to predict the confidence degree p.

7. The aircraft semantic knowledge structuring method according to claim 6, characterized in that, In step three, the split paragraphs are subjected to structure-preserving semantic enhancement using a back-translation model or a paraphrasing model to generate equivalent but differently phrased domain descriptions. For each split paragraph, 2-3 semantically equivalent but differently phrased variants are generated to obtain diverse but structurally consistent expressions, thereby constructing an enhanced corpus. When enhancing the context diversity of the design stage label alignment structure, a terminology freezing mechanism is introduced. A pre-trained language model based on the Transformer architecture is selected, and the semantic processing model is fine-tuned for aircraft design domain multi-tasking using the enhanced corpus. The goal of fine-tuning is to minimize the comprehensive loss function S: S = LMLM + λ × Ldomain + μ × Lstage Where, LMLM is the mask language modeling loss. Ldomain is the design entity recognition loss. Lstage is the design stage label constraint loss. λ and μ are task weight coefficients.

8. The aircraft semantic knowledge structuring method according to claim 7, characterized in that, In step four, the extracted semantics and design parameters are input into the knowledge graph construction model to construct the knowledge graph. The knowledge graph includes an entity set E, entity relationships R, and an evolution path graph P, which are stored in the form of a graph database or nested key-value structure, supporting retrieval and reasoning.

9. The aircraft semantic knowledge structuring method according to claim 8, characterized in that, In step four, a semantic embedding mechanism is used to convert all recognized design elements into knowledge graph nodes. The semantic consistency score between knowledge graph nodes is calculated: Sim(i,j) = cos(i_vec, j_vec) × δ_structure × γ_context × ω_attribute Sim(i,j) is the semantic consistency score between i and j knowledge graph nodes. i_vec and j_vec are the embedding vectors generated by the BERT encoding of i and j knowledge graph nodes. δ_structure is a Boolean item indicating whether the structure categories of the knowledge graph nodes are consistent. γ_context is the similarity of the knowledge graph nodes in the context. ω_attribute is the consistency rate of the knowledge graph node attribute fields. If Sim(i,j) exceeds the consistency threshold, i and j knowledge graph nodes are merged. The evolution path graph P is embedded into the knowledge graph as a directed subgraph.

10. The aircraft semantic knowledge structuring method according to claim 9, characterized in that, In step four, the knowledge graph quality score is calculated to evaluate the quality of the knowledge graph: Q(G) = α × C(G) + β × P(G) + γ × Con(G) Where, Q(G) is the knowledge graph quality score. C(G) is the coverage of the knowledge graph. P(G) is the precision of the knowledge graph. Con(G) is the consistency of the knowledge graph. α, β, and γ are weights.

11. An aircraft semantic knowledge structuring system, characterized in that, It includes: A paragraph splitting module for splitting aircraft design documents into paragraphs. The aircraft design document includes task description, overall scheme, system analysis, and experimental data document. A design stage label alignment module for predicting the design stage corresponding to the split paragraphs and constructing a design stage label alignment structure. The design stage includes task definition, overall scheme, aerodynamic design, structural design, system analysis, and test verification stage. The semantic and design parameter extraction module is configured to perform semantic and design parameter extraction on the design stage mark alignment structure. The knowledge graph construction module is configured to construct a knowledge graph based on the extracted semantic and design parameters.

12. The aircraft semantic knowledge structuring system of claim 11, wherein, In the design stage mark alignment module, the semantic stage structure alignment model is configured to predict the design stage corresponding to the split paragraph by using a RoBERTa variant structure and introducing common words and key prompt words in the design stage of the aircraft development process for supervision.

13. The aircraft semantic knowledge structuring system of claim 12, wherein, In the design stage mark alignment module, the semantic stage structure alignment model is configured to perform BERT encoding on the split paragraph, and then use a classifier with a Softmax added on the basis of RoBERTa to predict the corresponding design stage, thereby constructing the design stage mark alignment structure.

14. The aircraft semantic knowledge structuring system of claim 13, wherein, In the design stage mark alignment module, the semantic stage structure alignment model is configured to introduce a context semantic boundary detection mechanism, and combine syntactic dependency analysis and local consistency scores of BERT embedding to improve the accuracy of boundary recognition of different design stages of the split paragraph.

15. The aircraft semantic knowledge structuring system of claim 14, wherein, The semantic and design parameter extraction module is configured to input the design stage mark alignment structure into a semantic processing model to perform semantic and design parameter extraction. The semantic processing model is configured to perform semantic extraction on the design stage mark alignment structure, identify content conforming to a "design premise-design entity-response scheme" template, and extract the content as a structured triple. When performing semantic extraction on the design stage mark alignment structure, the semantic processing model is configured to use a BERT encoding+BiLSTM+CRF structure and introduce a structure guide token to improve boundary recognition capability.

16. The aircraft semantic knowledge structuring system of claim 15, wherein, In the semantic and design parameter extraction module, the semantic processing model is configured to perform design parameter extraction on the design stage mark alignment structure, identify content of design parameter numerical expression, extract a minimum value vmin and a maximum value vmax of the design parameter, and generate a confidence p to model the design parameter interval. When performing design parameter extraction on the design stage mark alignment structure, the semantic processing model is configured to, for a single-value description of the design parameter value, extract the minimum value vmin and the maximum value vmax of the design parameter to be equal, and for a fuzzy description of the design parameter value, calculate the minimum value vmin and the maximum value vmax of the design parameter according to a typical margin coefficient learned from the context. When performing design parameter extraction on the design stage mark alignment structure, the semantic processing model is configured to perform BERT encoding on the text of the design parameter numerical expression content to obtain an embedding vector H, input the embedding vector H into two independent fully connected networks to predict the minimum value vmin and the maximum value vmax, and input the embedding vector H into a Sigmoid network to predict the confidence p.

17. The aircraft semantic knowledge structuring system of claim 16, wherein, In the semantic and design parameter extraction module, a back-translation model or a paraphrase model is used to perform structure-preserving semantic enhancement on the split paragraph to generate equivalent but different in sentence style domain descriptions, generate 2-3 semantic equivalent but different in sentence style variants for each split paragraph, obtain diverse but consistent structure expressions, and construct an enhanced corpus set. When the design stage label alignment structure is enhanced in context diversity, a term freezing mechanism is introduced; A pre-training language model based on the Transformer architecture is selected, and a semantic processing model is fine-tuned in the aircraft design field using an enhanced corpus set. The goal of fine-tuning is to minimize the comprehensive loss function S: S=LMLM+λ×Ldomain+μ×Lstage; Where, LMLM is the mask language modeling loss; Ldomain is the design entity recognition loss; Lstage is the design stage label constraint loss; λ and μ are task weight coefficients.

18. The aircraft semantic knowledge structuring system of claim 17, wherein, In the knowledge graph construction module, the extracted semantics and design parameters are input into the knowledge graph construction model to construct the knowledge graph. The knowledge graph includes an entity set E, a relationship R between entities, and an evolution path graph P, which is stored in the form of a graph database or nested key-value structure, supporting retrieval and reasoning.

19. The aircraft semantic knowledge structuring system of claim 18, wherein, In the knowledge graph construction module, semantic embedding mechanism is used to convert all recognized design elements into knowledge graph nodes. Calculate the semantic consistency score between knowledge graph nodes: Sim(i,j)=cos(i_vec, j_vec)×δ_structure×γ_context×ω_attribute; Sim(i,j) is the semantic consistency score between i and j knowledge graph nodes. i_vec and j_vec are embedding vectors generated by BERT encoding of i and j knowledge graph nodes. δ_structure is a Boolean item indicating whether the structure categories of the knowledge graph nodes are consistent. γ_context is the similarity of the knowledge graph nodes in the context. ω_attribute is the consistency rate of the knowledge graph node attribute fields. If Sim(i,j) exceeds the consistency threshold, i and j knowledge graph nodes are merged. The evolution path graph P is embedded in the knowledge graph as a directed subgraph.

20. The aircraft semantic knowledge structuring system of claim 19, wherein, In the knowledge graph construction module, the quality score of the knowledge graph is calculated to evaluate the quality of the knowledge graph: Q(G)=α×C(G)+β×P(G)+γ×Con(G); Where, Q (G) is the quality score of the knowledge graph. C(G) is the coverage of the knowledge graph. P(G) is the precision of the knowledge graph. Con(G) is the consistency of the knowledge graph. α, β, and γ are weights.

Images