Aircraft maintenance knowledge management and intelligent recommendation method based on knowledge graph

Abstract

The invention discloses an aircraft maintenance knowledge management and intelligent recommendation method based on a knowledge graph, and the method comprises the following steps: carrying out the multi-source collection of aircraft maintenance data, and carrying out the classification and coding of the maintenance data, extracting an entity and a type thereof by using a domain self-adaptive entity recognition model according to a splicing vector spliced by the classified coding data; constructing an entity co-occurrence graph according to the entities, performing relation extraction, attribute extraction and logic generation according to the entity co-occurrence graph, and constructing a knowledge graph according to the entities, the entity types, the entity relations, the entity attributes and the entity logic; aircraft fault problems, environmental conditions and maintenance preferences are manually input, and an aircraft fault maintenance scheme is intelligently recommended according to manual input and the knowledge graph; and carrying out resource conflict detection according to the recommended aircraft fault maintenance scheme, and regularly updating the knowledge graph. According to the intelligent maintenance recommendation method provided by the invention, the knowledge graph composed of multi-source maintenance data and multi-factor training are provided, and the accuracy of intelligent recommendation is improved.

Description

Technical Field

[0001] The present invention relates to the field of aircraft, and in particular to an aircraft maintenance knowledge management and intelligent recommendation method based on a knowledge graph. Background Art

[0002] With the advancement of aircraft technology, aircraft systems are becoming increasingly complex, making maintenance more difficult. Traditional knowledge storage methods, such as documents and manuals, are unable to meet the needs of rapid query and decision support. Daily aircraft operations, maintenance records, fault reports, and other data accumulate massive amounts of data. Effectively managing and analyzing this data presents a challenge. Based on this, this paper proposes a knowledge graph-based aircraft maintenance knowledge management and intelligent recommendation method. Summary of the Invention

[0003] The present invention provides an aircraft maintenance knowledge management and intelligent recommendation method based on a knowledge graph, which is characterized by comprising:

[0004] S10, collecting aircraft maintenance data from multiple sources, classifying and encoding the maintenance data, and extracting entities and their types using a domain-adaptive entity recognition model based on a concatenation vector formed by concatenating the classified and encoded data;

[0005] S20. Construct an entity co-occurrence graph based on the entities, perform relationship extraction, attribute extraction, and logic generation based on the entity co-occurrence graph, and construct a knowledge graph based on the entities, entity types, entity relationships, entity attributes, and entity logic;

[0006] S30: Manually input aircraft fault issues, environmental conditions, and maintenance preferences, and intelligently recommend aircraft fault maintenance solutions based on the manual input and the knowledge graph.

[0007] S40. Perform resource conflict detection based on the recommended aircraft fault repair plan and regularly update the knowledge graph.

[0008] As described above, a knowledge graph-based aircraft maintenance knowledge management and intelligent recommendation method is described, in which the collected multi-source aircraft maintenance data is preprocessed, including: for unchangeable files, using a layout analysis tool to extract structured data with coordinate information; for ordinary text, cleaning language tags; generating pseudo-coordinate encoding based on text position; and cutting sensor data into event windows and extracting statistical feature vectors.

[0009] As described above, a knowledge graph-based aircraft maintenance knowledge management and intelligent recommendation method is proposed, in which entities and their types are extracted based on the fused feature vectors and the loss function to perform cross-modal entity recognition. The loss function includes text generation loss, classification loss, and dictionary matching loss.

[0010] As described above, a knowledge graph-based aircraft maintenance knowledge management and intelligent recommendation method is described, in which the entity co-occurrence graph is a preliminary step in constructing the knowledge graph, which can help sort out the preliminary associations between entities. The entity co-occurrence graph consists of a node set and an edge set, where the node is the entity and the edge weight is the number of co-occurrences.

[0011] As described above, a knowledge graph-based aircraft maintenance knowledge management and intelligent recommendation method is described, in which attribute extraction of structured data is performed by traversing each table in the structured table set, screening out cell values that meet the domain constraint value set, using a mapping function to convert column names into attribute names, and merging the attribute name and attribute value pairs that meet the conditions, thereby constructing an entity attribute set.

[0012] As described above, a knowledge graph-based aircraft maintenance knowledge management and intelligent recommendation method is described, in which a conjunction operation is performed on all attributes in the required attribute set of each entity in a certain type of entity set under the limitation of the modal logic necessity operator to ensure that there is a corresponding attribute value so that the entity has the attribute, and then the logical rules that meet the conditions are merged to form the logical axioms of the knowledge graph.

[0013] As described above, a knowledge graph-based aircraft maintenance knowledge management and intelligent recommendation method is proposed, in which when the same entity appears multiple times, entity linking can be performed across modalities, and whether it is a conflicting entity is determined based on semantics. If it is a conflicting entity, manual judgment is required to delete the erroneous entity information.

[0014] The present invention also provides an aircraft maintenance knowledge management and intelligent recommendation system based on knowledge graph, including:

[0015] Entity Extraction Module: This module collects aircraft maintenance data from multiple sources, classifies and encodes the maintenance data, and uses a domain-adaptive entity recognition model to extract entities and their types based on the concatenated vectors formed by the classified and coded data.

[0016] Build a knowledge graph module: Build an entity co-occurrence graph based on entities, extract relationships, extract attributes, and generate logic based on the entity co-occurrence graph, and build a knowledge graph based on entities, entity types, entity relationships, entity attributes, and entity logic;

[0017] Maintenance plan module: Manually input aircraft fault problems, environmental conditions, and maintenance preferences, and intelligently recommend aircraft fault maintenance plans based on manual input and knowledge graph;

[0018] Detection and update module: Perform resource conflict detection based on recommended aircraft fault repair plans and regularly update the knowledge graph.

[0019] The beneficial effects achieved by the present invention are as follows: The present invention provides an intelligent maintenance recommendation method based on a knowledge graph composed of multi-source maintenance data and training with multiple factors, thereby improving the accuracy of intelligent recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0021] Figure 1 This is a flow chart of an aircraft maintenance knowledge management and intelligent recommendation method based on a knowledge graph provided in Example 1 of the present application.

[0022] Figure 2 This is a schematic diagram of an aircraft maintenance knowledge management and intelligent recommendation system based on a knowledge graph provided in Example 2 of the present application. DETAILED DESCRIPTION

[0023] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0024] Example 1

[0025] like Figure 1 As shown, the first embodiment of the present application provides an aircraft maintenance knowledge management and intelligent recommendation method based on a knowledge graph, including:

[0026] S10. Collect aircraft maintenance data from multiple sources, classify and encode the maintenance data, and use a domain-adaptive entity recognition model to extract entities and their types based on the spliced vectors formed by splicing the classified and encoded data.

[0027] S11. Collect aircraft maintenance data from multiple sources.

[0028] Aircraft maintenance-related knowledge is collected from multiple heterogeneous data sources such as aircraft maintenance manuals, fault reports, technical documents, sensor log data, expert experience, historical maintenance records, aviation field dictionaries, etc. The data includes but is not limited to aircraft component information, failure modes, maintenance procedures, required tools and materials, etc.

[0029] S12. Preprocess the collected multi-source aircraft maintenance data.

[0030] Preprocess the collected multimodal data. For unchangeable files, including maintenance manuals and technical drawings, use layout analysis tools to extract text, flow charts, and parameter tables from unchangeable files and generate structured data with coordinate information. The formula is: Among them, Block i represents the i-th independent content block parsed from the document, T i is the text information in the i-th block, are block coordinates, is the coordinate of the upper left corner of the i-th block, is the coordinate of the lower right corner of the i-th block, which locates the spatial position of the block in the document page, τ i ∈{text,table,flow} is the block type, text is the text block, table is the table block, and flow is the flow block.

[0031] For ordinary text, clean the Hypertext Markup Language tags and Extensible Markup Language tags and retain the semantic tags; for semi-structured data, extract the key-value pairs, the formula is KeyValue j ={key j ,value j ,parent_tag j}, KeyValue j For the jth key-value pair object, key j is the key in the jth key-value pair, value j is the value in the jth key-value pair, parent_tag j Generate pseudo-coordinate encoding based on text position is the ratio of the offset within the paragraph to the total length of the paragraph, It is the ratio of the paragraph number to the total number of paragraphs.

[0032] The sensor data is cut by event window, and the formula is t2-t1=ΔT, where Represents a sensor data segment from time t1 to t2, consisting of a set of data points, each of which contains a value v k and the corresponding timestamp t,v k Including temperature, pressure, vibration amplitude, etc., ΔT represents the length of the time window. The statistical feature vector is extracted as f sensor =[μ(v k ),σ 2 (v k ),max(v k )] T ,μ(v k ) represents the sensor data vk The mean of 2 (v k ) represents the sensor data v k The variance of max(v k ) represents the sensor data v k The maximum value in , where T is the transposed sign.

[0033] Aviation Dictionary D dict The entity list is D dict ={e1,e2,...,e m}, used to compare words in the text.

[0034] S13. Perform feature encoding on the preprocessed multimodal data.

[0035] Segment the text in the normal text and the text in the unchangeable file into words and obtain the word sequence T = [w1,w2,...,w n ], encode the word sequence, the formula is H text =BERT(T)∈R n×d , BERT is the encoding model, n represents the length of the text word sequence, and d is the hidden layer dimension of the encoding model.

[0036] For structural data in unchangeable files and pseudo-labeling of semi-structured data in plain text Encode, the structural data encoding formula is Among them, K is the number of Gaussian kernels, R K is a K-dimensional real vector space, is the block center coordinate, is the center of the kth Gaussian kernel in the x-axis and y-axis directions, l k is the bandwidth of the kth Gaussian kernel. The semi-structured data encoding formula is Sinusoidal is the encoding function, R d is a d-dimensional real vector space.

[0037] Mapping sensor features into vectors, the formula is MLP is a multi-layer perceptron used to map input data to output space, d s Embedding dimension for sensors.

[0038] Entity list D according to the domain dictionary dict ={e1,e2,...,e m} matches words in the current text, the formula is Represents the dictionary matching vector of the i-th word, which is used to indicate whether the word in the current text exists in the domain dictionary. OneHot is a matching indicator function, where the position of the corresponding word in the dictionary is 1, and the other positions are 0. dict | is the dictionary size, that is, the number of words in the dictionary.

[0039] S14. Concatenate and fuse the encoded multimodal features.

[0040] The encoded multimodal features are spliced together, and the spliced features are The spliced features are gated and dynamically fused. The gate weight calculation formula is: g (i) represents the gate vector of the i-th element, σ represents the activation function, W g is the weight matrix, b g is the bias term, and the fused output vector is in, is the fusion feature vector of the i-th element, Represents the text embedding vector of the i-th element, 1-g (i) Represents the complement of the gate vector.

[0041] According to the conditional random field, the most appropriate entity label is assigned to each word in the text, the dependency relationship between the labels is modeled, and the logical errors caused by isolated predictions are avoided. It consists of the label emission score and the transfer matrix. The label emission score formula is S i represents the local confidence that the i-th word belongs to different labels, W c is the trainable weight matrix, b c is the bias term, ε is the label. The transfer matrix F∈R |ε|×|ε| , used to constrain the rationality of the label sequence, F jk represents the transfer score from label j to label k.

[0042] S15. Extract entities and their types based on the domain adaptive entity recognition model.

[0043] According to the fused feature vector, cross-modal entity recognition is performed based on the loss function. The specific formula is:

[0044] Among them, α is the weight of text generation loss, E w～T represents the expected value of all words w in the text T, w~T represents a word w randomly extracted from the text T, P(w|T w ) indicates that given context T w Under the condition of predicting the probability of word w, using the text generation loss E w～T [-logP(w|T w)] Evaluate the performance of the model in generating text to ensure that the model can accurately predict the words in the text. β is the weight of the classification loss, y true represents the true label, h fused Represents the fused feature vector, classification loss -logP(y true |h fused ) is used to evaluate the performance of the model in the classification task, ensuring that the model can correctly predict the category of the entity based on the fused features. λ is the regularization coefficient, S(y true ) represents the true label sequence y true The score of , y' is the possible label sequence, It is used to maximize the score of the correct label sequence while suppressing the wrong sequence. γ represents the weight of the dictionary matching loss, e represents the entity in the dictionary, and D dict Represents a domain dictionary set, containing all possible entities and terms. dict (e) represents the prior probability of entity e in the dictionary, P manual (e) represents the probability of entity e appearing in the manually annotated data, and the dictionary matching loss Used to evaluate the performance of the model in the dictionary matching task to ensure that the model can effectively utilize domain dictionary information and is consistent with the actual data distribution.

[0045] Use the decoding algorithm to find the highest-scoring legal label sequence from the model output, the formula is y * =argmax y P(y|H fused ), y is the output label, according to the optimal label sequence y * , extract the entities in the text and their corresponding label types, the entity-type pair is V = {(e1,type1), (e2,type2), ..., (e N ,type N )}, merging consecutive tags into complete entities.

[0046] S20. Construct an entity co-occurrence graph based on the entities, perform relationship extraction, attribute extraction and logic generation based on the entity co-occurrence graph, and construct a knowledge graph based on entities, entity types, entity relationships, entity attributes and entity logic.

[0047] The entity co-occurrence graph is a preliminary step in building a knowledge graph. It helps to sort out the initial associations between entities, traverse the entity set, and determine which entities appear simultaneously in the same text unit. If two entities appear in the same text unit, they are considered to have a co-occurrence relationship. e ,E),V eis a node set, E is an edge set, which represents the co-occurrence relationship between entities. The node is the entity, and the edge weight is the number of co-occurrences. The formula is used. Calculate the entity importance, d is the damping factor, N is the total number of entities in the entity co-occurrence graph, M(e j ) is entity e i The neighbor entity set, e j Represents entity e i Neighbor entity, PR(e j ) is entity e j The importance score, L(e j ) is e j The number of outgoing edges is from entity e j The number of outgoing edges connecting to other entities. Order the entities in the node set according to their importance scores.

[0048] Relationship extraction is performed based on the entity co-occurrence graph. The formula is: a represents the relationship type, e1 and e2 are two entities, c sx is context information, is the embedding vector of entity e1, e2, is the context embedding vector, σ is the activation function, W a is the weight matrix, b a is the bias term.

[0049] Attributes are extracted based on the entity co-occurrence graph, and the attributes of structured data are extracted by the formula Extract, where τ table is a structured table set, ta is a single table in the table set, Attr(e) represents the attribute set of entity e, c j is the column of the table, cv ij is the cell value of row i and column j in table ta, V is the mapping function from column name to attribute name, valid is a set of values that conform to domain constraints; unstructured text attributes are defined by the formula Perform attribute extraction, η is the attribute, SX(e) is the context statement set of entity e, is the balance coefficient between semantics and rules, SRL(o) is the confidence of semantic role annotation of sentence o, PM(o,R pattern ) is the statement o and the regular pattern R pattern The matching score of .

[0050] Using the formula The logical axioms that form the knowledge graph, Represents a set of entities of type h Any entity e in is the modal logic necessity operator, Represents all attributes a in the required attribute set ReqAttr(t) for type h shu Perform conjunction operation, which requires all required attributes to satisfy the subsequent conditions. Indicates that there is a value v, HasAttr(e,a shu ,v value ) indicates that entity e has attribute a shu And the attribute value is v value .

[0051] According to the formula Learn the constraints on key attributes of entities of a specific type h and determine the valid value range of key attributes, thereby providing a constraint basis for the correctness of entity attribute values in the knowledge graph. For the key attribute a shu The valid value range of All values v in suo Perform disjunction operation. ValRange(a shu ,v suo ) represents attribute a shu The value range includes the value v suo , CritAttr(h) is the key attribute set of type h.

[0052] An aircraft maintenance knowledge graph is constructed based on entities, entity types, entity relationships, entity attributes, and entity logic axioms. A graph database is used to store the knowledge graph. This storage method facilitates knowledge query and visualization, and can intuitively present the aircraft maintenance knowledge structure. Hierarchical clustering modeling is used to group entities based on their similarities, forming hierarchical clusters. The ontology mapping constraint formula is used to accurately classify the extracted entities into a standard classification system, unifying the classification standards of entities in the knowledge graph. This makes the knowledge graph more standardized and its semantics clearer, enhances the accuracy and consistency of the knowledge graph, and facilitates the management and application of maintenance knowledge.

[0053] When the same entity appears multiple times, the coordinate information can assist in cross-modal entity linking, using the formula Determine the entity e in the table table With the entity e in the text text Whether they point to the same physical entity, if the spatial block intersection and union (IOU) of the two exceeds the threshold φ, they are considered aligned, otherwise they are mismatched. Is the indicator function. According to the semantics, it is determined whether it is a conflict entity. If it is a conflict entity, the statistical support entity e i The amount of multimodal evidence ZJ(e i ), if ZJ(e i)≥2, it is retained, otherwise it is marked as pending verification. The pending verification entity needs to be manually judged and the incorrect entity information is deleted.

[0054] S30: Manually input aircraft fault problems, environmental conditions, and maintenance preferences, and intelligently recommend aircraft fault maintenance solutions based on the manual input and the knowledge graph.

[0055] The user inputs the parts to be repaired or the current problems of the aircraft, environmental conditions and maintenance preferences. Based on the life cycle of the aircraft parts and the dynamic impact of the environment, the maintenance decision is retrieved and generated from the knowledge graph. The formula is:

[0056]

[0057] Where U is manual input, including Q(U) maintenance demand, P(U) maintenance preference, E(U) environmental status, real-time sensor data or manually input environmental parameters. D(U) is the candidate solution, and the candidate solution generation formula is D(U) = KGS(S KG (Q(U)),LJ repair ), represents the knowledge graph retrieval based on the input maintenance demand Q(U), S KG Represents a local subgraph of the knowledge graph related to maintenance needs, LJ repair For the maintenance rule base. is a multi-objective optimization term, ψ i (U) is the dynamic weight of the input i-th optimization objective, which is adjusted according to the input maintenance preference P(U) and environmental state E(U). is the total number of optimization objectives, including cost, time, reliability, safety, environmental protection, etc. i (D) represents the normalized score of candidate maintenance solution D on the i-th optimization objective. represents the knowledge constraint, R D is the set of rules associated with solution D in the knowledge graph, including maintenance procedures, airworthiness clauses, etc., and ξ(r) is the satisfaction of rule r. is the environmental adaptation item, e v Indicates the environmental factors that affect solution D, Env j is the current environmental parameter value from real-time sensor data, Std e is the standard environmental requirement value from the maintenance manual, and κ is the environmental sensitivity coefficient, which is fitted by historical data. User (D, U) is the user preference item, which consists of explicit preference and implicit preference. Explicit preference is the input preference, and implicit preference is the historical maintenance preference. Sim(Q(U), D) is the demand matching calculation, and the formula is Sim(Q(U), D) = Γ1·TextSim(Q text,D)+Γ2·GrapSim(Q graph ,D),TextSim(Q text ,D) is to calculate the user text query Q text Similarity with the solution description, GrapSim(Q graph ,D) is to calculate the query subgraph Q graph The graph edit distance of the subgraph associated with the solution, Γ1, Γ2 are the related weights.

[0058] S40. Perform resource conflict detection based on the recommended aircraft fault repair plan and regularly update the knowledge graph.

[0059] Perform resource conflict detection on the intelligently recommended maintenance plan. The formula is: In the time window t, the maintenance plan process a jt Resources Demand Exceed the total amount A warning is triggered when n is the total number of maintenance procedures.

[0060] The knowledge timeliness modeling formula is: Maintenance rule weight Exponential decay over time t, the decay rate ρ is determined by the component type. When ι is marked as expired knowledge, the update mechanism is triggered to form a dynamically updateable maintenance knowledge network, where ι is the expiration threshold.

[0061] As aircraft technology develops and maintenance experience accumulates, the knowledge graph needs to be updated promptly. When new fault cases, maintenance techniques, or tools emerge, new knowledge is added to the knowledge graph through a process of knowledge extraction and fusion.

[0062] Example 2

[0063] like Figure 2 As shown, the second embodiment of the present application provides an aircraft maintenance knowledge management and intelligent recommendation system based on a knowledge graph, including:

[0064] Entity extraction module: includes collection submodule, preprocessing submodule, encoding submodule, splicing and fusion submodule, and extraction submodule.

[0065] Collection submodule: used to collect aircraft maintenance data from multiple sources.

[0066] Aircraft maintenance-related knowledge is collected from multiple heterogeneous data sources such as aircraft maintenance manuals, fault reports, technical documents, sensor log data, expert experience, historical maintenance records, aviation field dictionaries, etc. The data includes but is not limited to aircraft component information, failure modes, maintenance procedures, required tools and materials, etc.

[0067] Preprocessing submodule: used to preprocess the collected multi-source aircraft maintenance data.

[0068] Preprocess the collected multimodal data. For unchangeable files, including maintenance manuals and technical drawings, use layout analysis tools to extract text, flow charts, and parameter tables from unchangeable files and generate structured data with coordinate information. The formula is: Among them, Block i represents the i-th independent content block parsed from the document, T i is the text information in the i-th block, are block coordinates, is the coordinate of the upper left corner of the i-th block, is the coordinate of the lower right corner of the i-th block, which locates the spatial position of the block in the document page, τ i ∈{text,table,flow} is the block type, text is the text block, table is the table block, and flow is the flow block.

[0069] For ordinary text, clean the Hypertext Markup Language tags and Extensible Markup Language tags and retain the semantic tags; for semi-structured data, extract the key-value pairs, the formula is KeyValue j ={key j ,value j ,parent_tag j}, KeyValue j For the jth key-value pair object, key j is the key in the jth key-value pair, value j is the value in the jth key-value pair, parent_tag j Generate pseudo-coordinate encoding based on text position is the ratio of the offset within the paragraph to the total length of the paragraph, It is the ratio of the paragraph number to the total number of paragraphs.

[0070] The sensor data is cut by event window, and the formula is t2-t1=ΔT, where Represents a sensor data segment from time t1 to t2, consisting of a set of data points, each of which contains a value v k and the corresponding timestamp t,v k Including temperature, pressure, vibration amplitude, etc., ΔT represents the length of the time window. The statistical feature vector is extracted as f sensor =[μ(v k ),σ 2 (v k ),max(vk )] T ,μ(v k ) represents the sensor data v k The mean of 2 (v k ) represents the sensor data v k The variance of max(v k ) represents the sensor data v k The maximum value in , where T is the transposed sign.

[0071] Aviation Dictionary D dict The entity list is D dict ={e1,e2,...,e m}, used to compare words in the text.

[0072] Encoding submodule: used to perform feature encoding on preprocessed multimodal data.

[0073] Segment the text in the normal text and the text in the unchangeable file into words and obtain the word sequence T = [w1,w2,...,w n ], encode the word sequence, the formula is H text =BERT(T)∈R n×d , BERT is the encoding model, n represents the length of the text word sequence, and d is the hidden layer dimension of the encoding model.

[0074] For structural data in unchangeable files and pseudo-labeling of semi-structured data in plain text Encode, the structural data encoding formula is Among them, K is the number of Gaussian kernels, R K is a K-dimensional real vector space, is the block center coordinate, is the center of the kth Gaussian kernel in the x-axis and y-axis directions, l k is the bandwidth of the kth Gaussian kernel. The semi-structured data encoding formula is Sinusoidal is the encoding function, R d is a d-dimensional real vector space.

[0075] Mapping sensor features into vectors, the formula is MLP is a multi-layer perceptron used to map input data to output space, d s Embedding dimension for sensors.

[0076] Entity list D according to the domain dictionary dict ={e1,e2,...,e m} matches words in the current text, the formula is Represents the dictionary matching vector of the i-th word, which is used to indicate whether the word in the current text exists in the domain dictionary. OneHot is a matching indicator function, where the position of the corresponding word in the dictionary is 1, and the other positions are 0. dict | is the dictionary size, that is, the number of words in the dictionary.

[0077] Splicing and fusion submodule: used to splice and fuse the encoded multimodal features.

[0078] The encoded multimodal features are spliced together, and the spliced features are The spliced features are gated and dynamically fused. The gate weight calculation formula is: g (i) represents the gate vector of the i-th element, σ represents the activation function, W g is the weight matrix, b g is the bias term, and the fused output vector is in, is the fusion feature vector of the i-th element, Represents the text embedding vector of the i-th element, 1-g (i) Represents the complement of the gate vector.

[0079] According to the conditional random field, the most appropriate entity label is assigned to each word in the text, the dependency relationship between the labels is modeled, and the logical errors caused by isolated predictions are avoided. It consists of the label emission score and the transfer matrix. The label emission score formula is S i represents the local confidence that the i-th word belongs to different labels, W c is the trainable weight matrix, b c is the bias term, ε is the label. The transfer matrix F∈R |ε|×|ε| , used to constrain the rationality of the label sequence, F jk represents the transfer score from label j to label k.

[0080] Extraction submodule: used to extract entities and their types based on the domain adaptive entity recognition model.

[0081] According to the fused feature vector, cross-modal entity recognition is performed based on the loss function. The specific formula is:

[0082] Among them, α is the weight of text generation loss, E w～T represents the expected value of all words w in the text T, w~T represents a word w randomly extracted from the text T, P(w|T w ) indicates that given context T w Under the condition of predicting the probability of word w, using the text generation loss E w～T[-logP(w|T w )] Evaluate the performance of the model in generating text to ensure that the model can accurately predict the words in the text. β is the weight of the classification loss, y true represents the true label, h fused Represents the fused feature vector, classification loss -logP(y true |h fused ) is used to evaluate the performance of the model in the classification task, ensuring that the model can correctly predict the category of the entity based on the fused features. λ is the regularization coefficient, S(y true ) represents the true label sequence y true The score of , y' is the possible label sequence, It is used to maximize the score of the correct label sequence while suppressing the wrong sequence. γ represents the weight of the dictionary matching loss, e represents the entity in the dictionary, and D dict Represents a domain dictionary set, containing all possible entities and terms. dict (e) represents the prior probability of entity e in the dictionary, P manual (e) represents the probability of entity e appearing in the manually annotated data, and the dictionary matching loss Used to evaluate the performance of the model in the dictionary matching task to ensure that the model can effectively utilize domain dictionary information and is consistent with the actual data distribution.

[0083] Use the decoding algorithm to find the highest-scoring legal label sequence from the model output, the formula is y * =argmax y P(y|H fused ), y is the output label, according to the optimal label sequence y * , extract the entities in the text and their corresponding label types, the entity-type pair is V = {(e1,type1), (e2,type2), ..., (e N ,type N )}, merging consecutive tags into complete entities.

[0084] Knowledge graph construction module: used to construct entity co-occurrence graphs based on entities, perform relationship extraction, attribute extraction and logic generation based on entity co-occurrence graphs, and construct knowledge graphs based on entities, entity types, entity relationships, entity attributes, and entity logic.

[0085] The entity co-occurrence graph is a preliminary step in building a knowledge graph. It helps to sort out the initial associations between entities, traverse the entity set, and determine which entities appear simultaneously in the same text unit. If two entities appear in the same text unit, they are considered to have a co-occurrence relationship. e ,E),V eis a node set, E is an edge set, which represents the co-occurrence relationship between entities. The node is the entity, and the edge weight is the number of co-occurrences. The formula is used. Calculate the entity importance, d is the damping factor, N is the total number of entities in the entity co-occurrence graph, M(e j ) is entity e i The neighbor entity set, e j Represents entity e i Neighbor entity, PR(e j ) is entity e j The importance score, L(e j ) is e j The number of outgoing edges is from entity e j The number of outgoing edges connecting to other entities. Order the entities in the node set according to their importance scores.

[0086] Relationship extraction is performed based on the entity co-occurrence graph. The formula is: a represents the relationship type, e1 and e2 are two entities, c sx is context information, is the embedding vector of entity e1, e2, is the context embedding vector, σ is the activation function, W a is the weight matrix, b a is the bias term.

[0087] Attributes are extracted based on the entity co-occurrence graph, and the attributes of structured data are extracted by the formula Extract, where τ table is a structured table set, ta is a single table in the table set, Attr(e) represents the attribute set of entity e, c j is the column of the table, cv ij is the cell value of row i and column j in table ta, V is the mapping function from column name to attribute name, valid is a set of values that conform to domain constraints; unstructured text attributes are defined by the formula Perform attribute extraction, η is the attribute, SX(e) is the context statement set of entity e, is the balance coefficient between semantics and rules, SRL(o) is the confidence of semantic role annotation of sentence o, PM(o,R pattern ) is the statement o and the regular pattern R pattern The matching score of .

[0088] Using the formula The logical axioms that form the knowledge graph, Denotes the entity set ζ for type h h Any entity e in is the modal logic necessity operator, Represents all attributes a in the required attribute set ReqAttr(t) for type h shu Perform conjunction operation, which requires all required attributes to satisfy the subsequent conditions. Indicates that there is a value v, HasAttr(e,a shu ,v value ) indicates that entity e has attribute a shu And the attribute value is v value .

[0089] According to the formula Learn the constraints on key attributes of entities of a specific type h and determine the valid value range of key attributes, thereby providing a constraint basis for the correctness of entity attribute values in the knowledge graph. For the key attribute a shu The valid value range of All values v in suo Perform disjunction operation. ValRange(a shu ,v suo ) represents attribute a shu The value range includes the value v suo , CritAttr(h) is the key attribute set of type h.

[0090] An aircraft maintenance knowledge graph is constructed based on entities, entity types, entity relationships, entity attributes, and entity logic axioms. A graph database is used to store the knowledge graph. This storage method facilitates knowledge query and visualization, and can intuitively present the aircraft maintenance knowledge structure. Hierarchical clustering modeling is used to group entities based on their similarities, forming hierarchical clusters. The ontology mapping constraint formula is used to accurately classify the extracted entities into a standard classification system, unifying the classification standards of entities in the knowledge graph. This makes the knowledge graph more standardized and its semantics clearer, enhances the accuracy and consistency of the knowledge graph, and facilitates the management and application of maintenance knowledge.

[0091] When the same entity appears multiple times, the coordinate information can assist in cross-modal entity linking, using the formula Determine the entity e in the table table With the entity e in the text text Whether they point to the same physical entity, if the spatial block intersection and union (IOU) of the two exceeds the threshold φ, they are considered aligned, otherwise they are mismatched. Is the indicator function. According to the semantics, it is determined whether it is a conflict entity. If it is a conflict entity, the statistical support entity e i The amount of multimodal evidence ZJ(e i ), if ZJ(e i)≥2, it is retained, otherwise it is marked as pending verification. The pending verification entity needs to be manually judged and the incorrect entity information is deleted.

[0092] Maintenance plan module: Manually input aircraft fault problems, environmental conditions and maintenance preferences, and intelligently recommend aircraft fault maintenance plans based on manual input and knowledge graph.

[0093] The user inputs the parts to be repaired or the current problems of the aircraft, environmental conditions and maintenance preferences. Based on the life cycle of the aircraft parts and the dynamic impact of the environment, the maintenance decision is retrieved and generated from the knowledge graph. The formula is:

[0094]

[0095] Where U is manual input, including Q(U) maintenance demand, P(U) maintenance preference, E(U) environmental status, real-time sensor data or manually input environmental parameters. D(U) is the candidate solution, and the candidate solution generation formula is D(U) = KGS(S KG (Q(U)),LJ repair ), represents the knowledge graph retrieval based on the input maintenance demand Q(U), S KG Represents a local subgraph of the knowledge graph related to maintenance needs, LJ repair For the maintenance rule base.

[0096] is a multi-objective optimization term, ψ i (U) is the dynamic weight of the input i-th optimization objective, which is adjusted according to the input maintenance preference P(U) and environmental state E(U). is the total number of optimization objectives, including cost, time, reliability, safety, environmental protection, etc. i (D) represents the normalized score of candidate maintenance solution D on the i-th optimization objective. represents the knowledge constraint, R D is the set of rules associated with solution D in the knowledge graph, including maintenance procedures, airworthiness clauses, etc., and ξ(r) is the satisfaction of rule r. is the environmental adaptation item, e v Indicates the environmental factors that affect solution D, Env j is the current environmental parameter value from real-time sensor data, Std e is the standard environmental requirement value from the maintenance manual, and κ is the environmental sensitivity coefficient, which is fitted by historical data. User(D, U) is the user preference item, which consists of explicit preference and implicit preference. Explicit preference is the input preference, and implicit preference is the historical maintenance preference. Sim(Q(U), D) is the demand matching calculation, and the formula is Sim(Q(U), D) = Γ1·TextSim(Q text ,D)+Γ2·GrapSim(Q graph ,D),TextSim(Q text ,D) is to calculate the user text query Q text Similarity with the solution description, GrapSim(Q graph ,D) is to calculate the query subgraph Q graph The graph edit distance of the subgraph associated with the solution, Γ1, Γ2 are the related weights.

[0097] Detection and update module: Perform resource conflict detection based on recommended aircraft fault repair plans and regularly update the knowledge graph.

[0098] Perform resource conflict detection on the intelligently recommended maintenance plan. The formula is: In the time window t, the maintenance plan process a jt Resources Demand Exceed the total amount A warning is triggered when n is the total number of maintenance procedures.

[0099] The knowledge timeliness modeling formula is: Maintenance rule weight Exponential decay over time t, the decay rate ρ is determined by the component type. When ι is marked as expired knowledge, the update mechanism is triggered to form a dynamically updateable maintenance knowledge network, where ι is the expiration threshold.

[0100] As aircraft technology develops and maintenance experience accumulates, the knowledge graph needs to be updated promptly. When new fault cases, maintenance techniques, or tools emerge, new knowledge is added to the knowledge graph through a process of knowledge extraction and fusion.

[0101] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A knowledge graph-based aircraft maintenance knowledge management and intelligent recommendation method, characterized by: include: S10, collecting aircraft maintenance data from multiple sources, classifying and encoding the maintenance data, and extracting entities and their types using a domain-adaptive entity recognition model based on a concatenation vector formed by concatenating the classified and encoded data; S20. Construct an entity co-occurrence graph based on the entities, perform relationship extraction, attribute extraction, and logic generation based on the entity co-occurrence graph, and construct a knowledge graph based on the entities, entity types, entity relationships, entity attributes, and entity logic; S30: Manually input aircraft fault issues, environmental conditions, and maintenance preferences, and intelligently recommend aircraft fault maintenance solutions based on the manual input and the knowledge graph. S40. Perform resource conflict detection based on the recommended aircraft fault repair plan and regularly update the knowledge graph.

2. The method for aircraft maintenance knowledge management and intelligent recommendation based on knowledge graph according to claim 1, characterized in that: The preprocessing of the collected multi-source aircraft maintenance data includes: for unchangeable files, using layout analysis tools to extract structured data with coordinate information; for ordinary text, cleaning language tags; generating pseudo-coordinate encoding based on text position; cutting sensor data into event windows and extracting statistical feature vectors.

3. The method for aircraft maintenance knowledge management and intelligent recommendation based on knowledge graph according to claim 1, characterized in that: Extracting entities and their types requires cross-modal entity recognition based on the fused feature vector and the loss function. The loss function includes text generation loss, classification loss, and dictionary matching loss.

4. The method for aircraft maintenance knowledge management and intelligent recommendation based on knowledge graph according to claim 1, characterized in that: The entity co-occurrence graph is a preliminary step in building a knowledge graph, which can help sort out the preliminary associations between entities. The entity co-occurrence graph consists of a node set and an edge set. The nodes are entities and the edge weights are the number of co-occurrences.

5. The method for aircraft maintenance knowledge management and intelligent recommendation based on knowledge graph according to claim 1, characterized in that: Attribute extraction of structured data is to traverse each table in the structured table set, filter out the cell values that meet the domain constraint value set, use the mapping function to convert the column name into the attribute name, merge the attribute name and attribute value pairs that meet the conditions, and thus construct the attribute set of the entity.

6. The method for aircraft maintenance knowledge management and intelligent recommendation based on knowledge graph according to claim 1, characterized in that: By performing conjunction operations on all attributes in the required attribute set for each entity in a certain type of entity set under the limitation of the modal logic necessity operator, we ensure that there are corresponding attribute values that enable the entity to have the attribute, and then merge the logical rules that meet the conditions to form the logical axioms of the knowledge graph.

7. The method for aircraft maintenance knowledge management and intelligent recommendation based on knowledge graph according to claim 1, characterized in that: When the same entity appears multiple times, entity linking can be performed across modalities to determine whether it is a conflicting entity based on semantics. If it is a conflicting entity, manual judgment is required to delete the erroneous entity information.

8. An aircraft maintenance knowledge management and intelligent recommendation system based on knowledge graph, characterized by: include: Entity Extraction Module: This module collects aircraft maintenance data from multiple sources, classifies and encodes the maintenance data, and uses a domain-adaptive entity recognition model to extract entities and their types based on the concatenated vectors formed by splicing the classified and coded data. Build a knowledge graph module: Build an entity co-occurrence graph based on entities, extract relationships, attributes, and generate logic based on the entity co-occurrence graph, and build a knowledge graph based on entities, entity types, entity relationships, entity attributes, and entity logic; Maintenance plan module: Manually input aircraft fault problems, environmental conditions, and maintenance preferences, and intelligently recommend aircraft fault maintenance plans based on manual input and knowledge graph; Detection and update module: Perform resource conflict detection based on recommended aircraft fault repair plans and regularly update the knowledge graph.

Images