Aviation operation and maintenance knowledge graph construction method and fault solution generation method

Abstract

The invention discloses an aviation operation and maintenance knowledge graph construction method, and the method comprises the steps: constructing a first hierarchy of a knowledge graph, defining an entity word vector as a first hierarchy entity node, and enabling the entity word vector to be obtained through the coding of an aviation logistics equipment name and a part name; constructing a second hierarchy of the knowledge graph, wherein entity nodes and fuzzy nodes of the second hierarchy are defined by the entity nodes of the first hierarchy; constructing a third hierarchy of the knowledge graph, wherein entity nodes of the third hierarchy are defined by nodes of the second hierarchy; constructing a fourth hierarchy of the knowledge graph, wherein nodes of the fourth hierarchy are defined by nodes of the third hierarchy; and constructing a fifth hierarchy of the knowledge graph, wherein nodes of the fifth hierarchy are defined by nodes of the fourth hierarchy. The invention further discloses a fault solution generation method based on the knowledge graph. The speed and accuracy of knowledge query and maintenance scheme matching of an existing knowledge graph can be effectively improved, so that the operation and maintenance efficiency of aviation logistics equipment is improved.

Description

Technical Field

[0001] The present invention relates to the field of aviation logistics equipment operation and maintenance technology, and in particular to an aviation operation and maintenance knowledge graph construction method and a fault solution generation method. Background Art

[0002] With the rapid development of the global aviation logistics industry, logistics equipment such as automated baggage sorting systems and intelligent warehousing systems are increasingly being used in the field of aviation logistics. However, when operating and maintaining these equipment, it is usually necessary to determine the corresponding solutions based on the equipment failure phenomenon. Currently, airports tend to record equipment failure phenomena in text, and query solutions based on the maintenance knowledge and failure cases in the existing maintenance manuals. Currently, there are many sources of operation and maintenance knowledge data in the aviation logistics field, and the operation and maintenance manuals, failure cases and other information of different systems are scattered, making it difficult to form a unified knowledge manual or knowledge base for operation and maintenance personnel to quickly consult.

[0003] In the prior art, knowledge graphs based on entity recognition are used as a knowledge base construction tool. By storing professional knowledge in a structured manner, they can help to efficiently integrate knowledge and make accurate decisions. They have been widely used in many industries such as transportation, financial risk control, and intelligent manufacturing. However, in the field of aviation logistics operation and maintenance, the operation and maintenance knowledge involved is highly professional and complex, which makes it particularly difficult to construct a comprehensive knowledge graph only through entity recognition. At the same time, the existing knowledge graph construction technology is still insufficient in terms of fault-solution matching capabilities, and thus cannot meet the urgent needs for efficient operation and maintenance in the aviation logistics field. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a method for constructing an aviation operation and maintenance knowledge graph and a method for generating a fault solution based on the knowledge graph, which can effectively improve the speed and accuracy of knowledge query and maintenance solution matching of the existing knowledge graph, so as to improve the efficiency of aviation logistics equipment operation and maintenance.

[0005] The technical solution adopted by the present invention is:

[0006] In a first aspect, the present application discloses a method for constructing an aviation operation and maintenance knowledge graph, comprising:

[0007] Constructing the first level of the knowledge graph, the first level includes at least one first-level entity node, defining an entity word vector as a first-level entity node, the entity word vector is obtained by encoding the entity word, and the entity word includes the name of aviation logistics equipment and the name of parts;

[0008] Constructing a second level of the knowledge graph, the second level includes at least one second level entity node and at least two second level fuzzy nodes, defining a first related word vector of the first level entity node as a second level entity node, and the second level fuzzy node is obtained by semantically predicting the second level entity node;

[0009] Constructing a third level of the knowledge graph, the third level comprising a plurality of third level entity nodes, the third level entity nodes being obtained by calculating the second level entity nodes and the second level fuzzy nodes;

[0010] Constructing a fourth level of the knowledge graph, the fourth level comprising a plurality of fourth level entity nodes and a plurality of fourth level fuzzy nodes, the fourth level entity nodes being obtained by calculation according to the third level entity nodes, and the fourth level fuzzy nodes being obtained by semantic prediction according to the fourth level entity nodes;

[0011] A fifth level of the knowledge graph is constructed, wherein the fifth level includes a plurality of fifth level entity nodes, and the fifth level entity nodes are calculated based on the fourth level entity nodes and the fourth level fuzzy nodes.

[0012] As an optional technical solution, before constructing the first level of the knowledge graph, it also includes:

[0013] Acquire fault text data and maintenance text data; wherein the fault text data and maintenance text data are derived from at least one text data in a fault case library, a maintenance manual, and a log record generated during the operation and maintenance of aviation logistics equipment;

[0014] A fault vector group is obtained by using the fault text data, and a maintenance vector group is obtained by using the maintenance text data; wherein the fault vector group includes a fault word vector and a fault sentence vector, and the maintenance vector group includes a maintenance word vector and a maintenance sentence vector.

[0015] As an optional technical solution, the first level of constructing the knowledge graph includes:

[0016] Encoding entity words in the entity word library to obtain entity word vectors; wherein the entity word library contains entity words for aviation logistics equipment names and parts names;

[0017] Extract all entity word vectors, each entity word vector is a node, and all nodes constitute the first level of the knowledge graph.

[0018] As an optional technical solution, the second level of constructing the knowledge graph includes:

[0019] Calculate the correlation between each entity word vector and each fault word vector in the fault vector group, extract the fault word vector with the greatest correlation as the first related word vector of the first-level entity node, and all the first related word vectors constitute the second-level entity node of the knowledge graph;

[0020] The bidirectional semantic prediction module is used to perform contextual semantic prediction on the first related word vector corresponding to each second-level entity node, and the two predicted word vectors are used as the two fuzzy nodes of the second-level entity node, and the fuzzy node is used as the second related word vector of the previous level.

[0021] As an optional technical solution, the correlation is calculated according to the following formula:

[0022]

[0023] Among them, cov is the covariance, σ is the standard deviation, w1 is the entity word vector, w2 is the fault word vector, and r(w1|w2) represents the correlation between the entity word vector and the fault word vector.

[0024] As an optional technical solution, the third level of constructing the knowledge graph includes: for all entity nodes and fuzzy nodes of the second level, calculating the probability density of the fault sentence vector in the fault vector group corresponding to each node word vector, and extracting the sentence vector with the largest probability density as the relevant sentence vector of the second level node. All relevant sentence vectors of the second level constitute the third level entity nodes of the knowledge graph.

[0025] As an optional technical solution, the probability density of the fault sentence vector in the fault vector group corresponding to the node word vector is calculated according to the following formula:

[0026]

[0027] Among them, w represents the node word vector, S represents the fault sentence vector, and w o represents the oth row vector of the node word vector matrix, Indicates w o The transpose of s c is the c-th row vector of the fault sentence vector matrix, and N is the window size.

[0028] As an optional technical solution, the fourth level of constructing the knowledge graph includes:

[0029] For the entity nodes at the third level, the probability density of the maintenance word vectors in the maintenance vector group corresponding to each entity node sentence vector is calculated, and the word vector with the largest probability density is extracted as the first relevant word vector of the entity node at the third level. All the first relevant word vectors at the third level constitute the entity nodes at the fourth level of the knowledge graph.

[0030] For the fourth-level entity nodes, the bidirectional semantic prediction module is used to perform contextual semantic prediction on the word vector of each entity node, and the two predicted word vectors are used as the two fuzzy nodes of the entity node and as the second related word vector of the previous level.

[0031] As an optional technical solution, the probability density of the maintenance word vector in the maintenance vector group corresponding to the entity node sentence vector is calculated according to the following formula:

[0032]

[0033] Among them, M represents the entity node sentence vector, K represents the maintenance word vector, and k o represents the oth row vector of the maintenance word vector matrix, Indicates k o The transpose of m o represents the oth row vector of the entity node sentence vector, Indicates m o The transpose of m c represents the c-th row vector of the entity node sentence vector matrix, k c Represents the c-th row vector of the maintenance word vector matrix.

[0034] As an optional technical solution, the fifth level of constructing the knowledge graph includes: for all entity nodes and fuzzy nodes of the fourth level, calculating the probability density of the sentence vector of the maintenance vector group corresponding to the word vector of each node, and extracting the sentence vector with the largest probability density as the relevant sentence vector of the fourth level node. All relevant sentence vectors of the fourth level constitute the fifth level entity nodes of the knowledge graph.

[0035] As an optional technical solution, the bidirectional semantic prediction module includes a forward LSTM unit and a backward LSTM unit, which are used to bidirectionally capture the contextual semantic information of the word vector and splice it to obtain two semantic association information of the word vector.

[0036] In a second aspect, the present application further discloses a method for generating a fault solution based on a knowledge graph, wherein the knowledge graph is constructed by the method in the first aspect above, and the method for generating a fault solution comprises:

[0037] Get the fault word vector to be matched;

[0038] Calculate the correlation between the fault word vector to be matched and the first-level entity node word vector one by one; extract the entity word vector with the greatest correlation with the first-level entity node as the recognition word vector, and select the corresponding first and second associated word vectors as the preferred fault word vector and candidate fault word vector respectively;

[0039] Generate a fault description according to the preferred fault word vector and the candidate fault word vector combined with the knowledge graph;

[0040] Generate a maintenance plan based on the preferred fault word vector and the candidate fault word vector combined with the knowledge graph;

[0041] The fault description and the maintenance plan are combined to generate a fault solution.

[0042] As an optional technical solution, the generation of the fault description includes: based on the node relationships of the knowledge graph, extracting the fault sentence vector corresponding to the preferred fault word vector as the preferred fault sentence vector, extracting the fault sentence vector corresponding to the candidate fault word vector as the candidate fault sentence vector, and decoding the preferred fault sentence vector and the candidate fault sentence vector, and sorting and combining them according to the priority of the preferred and candidate, to obtain the corresponding fault sentence segment as the fault description.

[0043] As an optional technical solution, the generation of the maintenance plan includes:

[0044] Based on the knowledge graph, extract the first and second correlation associated word vectors of the preferred fault sentence vector as the first preferred maintenance word vector and the second preferred maintenance word vector; extract the first and second correlation associated word vectors of the first candidate fault sentence vector as the first candidate maintenance word vector and the second candidate maintenance word vector;

[0045] Calculate the probability density of all maintenance word vectors corresponding to their fault sentence vectors, define the probability density as the first matching weight of the corresponding maintenance word vector, and define the second matching weights of the first preferred maintenance word vector, the second preferred maintenance word vector, the first candidate maintenance word vector, and the second candidate maintenance word vector as 0.8, 0.6, 0.4, and 0.2 respectively;

[0046] Determine the product of the first matching weight and the corresponding second matching weight as the matching priority of the corresponding maintenance word vector, match the corresponding maintenance word vector based on the size of the priority, and based on the relationship between the nodes in the knowledge graph, extract the maintenance sentence vector corresponding to the maintenance word vector and sort and combine the maintenance sentence vectors according to the priority of the word vector, decode the maintenance sentence vector to obtain the corresponding maintenance sentence segment, and use it as the maintenance plan.

[0047] In a third aspect, the present application also discloses an electronic device, comprising: one or more processors; one or more memories; the one or more memories storing one or more computer programs, the one or more computer programs comprising instructions, which, when executed by the one or more processors, enable the electronic device to execute the method described in the first or second aspect above.

[0048] In a fourth aspect, the present application further discloses a computer-readable medium, comprising a computer program, which, when executed on a computer, enables the computer to execute the method as described in the first aspect or the second aspect above.

[0049] The beneficial effects of the present invention are:

[0050] 1. Integrate advanced data analysis and entity extraction to identify and integrate multi-source heterogeneous operation and maintenance data.

[0051] 2. Integrate contextual bidirectional semantic prediction technology to expand the scope of semantic recognition and greatly reduce the problem of incorrect or missed recognition.

[0052] 3. Establish a close connection between fault data and solutions, improving the pertinence and effectiveness of solutions.

[0053] 4. Optimize the entity recognition algorithm and matching mechanism to make entity recognition and solution matching faster and more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flowchart of a method for constructing an aviation operation and maintenance knowledge graph in an exemplary embodiment.

[0055] Figure 2 It is a flowchart of a method for generating a fault solution based on a knowledge graph in an exemplary embodiment.

[0056] Figure 3 It is a schematic diagram of an aviation operation and maintenance knowledge graph constructed in an exemplary embodiment.

[0057] Figure 4 is a schematic diagram of the structure of a bidirectional semantic prediction module in an exemplary embodiment.

[0058] Figure 5 It is an architecture diagram of the aviation logistics operation and maintenance knowledge graph management platform in an exemplary embodiment. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components and method steps of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0060] Example

[0061] like Figure 1 As shown, this embodiment discloses a method for constructing an aviation operation and maintenance knowledge graph, including:

[0062] S1. Construct the first level of the knowledge graph, wherein the first level includes at least one first-level entity node, and define an entity word vector as a first-level entity node. The entity word vector is obtained by encoding the entity word, and the entity word includes the name of aviation logistics equipment and the name of parts.

[0063] S2. Construct a second level of the knowledge graph, wherein the second level includes at least one second level entity node and at least two second level fuzzy nodes, define a first related word vector of the first level entity node as a second level entity node, and the second level fuzzy node is obtained by semantic prediction of the second level entity node.

[0064] S3. Construct a third level of the knowledge graph, wherein the third level includes a plurality of third level entity nodes, and the third level entity nodes are calculated from the second level entity nodes and the second level fuzzy nodes.

[0065] S4. Construct a fourth level of the knowledge graph, wherein the fourth level includes multiple fourth level entity nodes and multiple fourth level fuzzy nodes, the fourth level entity nodes are calculated based on the third level entity nodes, and the fourth level fuzzy nodes are obtained by semantically predicting the fourth level entity nodes.

[0066] S5. Construct a fifth level of the knowledge graph, wherein the fifth level includes a plurality of fifth level entity nodes, and the fifth level entity nodes are calculated based on the fourth level entity nodes and the fourth level fuzzy nodes.

[0067] As an optional implementation, before step S1 constructs the first level of the knowledge graph, it also includes:

[0068] Acquire fault text data and maintenance text data; wherein the fault text data and maintenance text data are derived from at least one text data in a fault case library, a maintenance manual, and a log record generated during the operation and maintenance of aviation logistics equipment;

[0069] A fault vector group is obtained by using the fault text data, and a maintenance vector group is obtained by using the maintenance text data; wherein the fault vector group includes a fault word vector and a fault sentence vector, and the maintenance vector group includes a maintenance word vector and a maintenance sentence vector.

[0070] As an optional implementation, the first level of constructing the knowledge graph in step S1 includes:

[0071] Encoding entity words in the entity word library to obtain entity word vectors; wherein the entity word library contains entity words for aviation logistics equipment names and parts names;

[0072] Extract all entity word vectors, each entity word vector is a node, and all nodes constitute the first level of the knowledge graph.

[0073] As an optional implementation, the second level of constructing the knowledge graph in step S2 includes:

[0074] Calculate the correlation between each entity word vector and each fault word vector in the fault vector group, extract the fault word vector with the greatest correlation as the first related word vector of the first-level entity node, and all the first related word vectors constitute the second-level entity node of the knowledge graph;

[0075] The bidirectional semantic prediction module is used to perform contextual semantic prediction on the first related word vector corresponding to each second-level entity node, and the two predicted word vectors are used as the two fuzzy nodes of the second-level entity node, and the fuzzy node is used as the second related word vector of the previous level.

[0076] As an optional implementation, the correlation is calculated according to the following formula:

[0077]

[0078] Among them, cov is the covariance, σ is the standard deviation, w1 is the entity word vector, w2 is the fault word vector, and r(w1|w2) represents the correlation between the entity word vector and the fault word vector.

[0079] As an optional implementation, constructing the third level of the knowledge graph in step S3 includes: for all entity nodes and fuzzy nodes of the second level, calculating the probability density of the fault sentence vector in the fault vector group corresponding to each node word vector, and extracting the sentence vector with the largest probability density as the relevant sentence vector of the second level node. All relevant sentence vectors of the second level constitute the third level entity nodes of the knowledge graph.

[0080] As an optional implementation, the probability density of the fault sentence vector in the fault vector group corresponding to the node word vector in step S3 is calculated according to the following formula:

[0081]

[0082] Among them, W represents the node word vector, S represents the fault sentence vector, and w o represents the oth row vector of the node word vector matrix, Indicates w o The transpose of s c is the c-th row vector of the fault sentence vector matrix, and N is the window size.

[0083] As an optional implementation, the fourth level of constructing the knowledge graph in step S4 includes:

[0084] For the entity nodes at the third level, the probability density of the maintenance word vectors in the maintenance vector group corresponding to each entity node sentence vector is calculated, and the word vector with the largest probability density is extracted as the first relevant word vector of the entity node at the third level. All the first relevant word vectors at the third level constitute the entity nodes at the fourth level of the knowledge graph.

[0085] For the fourth-level entity nodes, the bidirectional semantic prediction module is used to perform contextual semantic prediction on the word vector of each entity node, and the two predicted word vectors are used as the two fuzzy nodes of the entity node and as the second related word vector of the previous level.

[0086] As an optional implementation, the probability density of the maintenance word vector in the maintenance vector group corresponding to the entity node sentence vector is calculated according to the following formula:

[0087]

[0088] Among them, M represents the entity node sentence vector, K represents the maintenance word vector, and k o represents the oth row vector of the maintenance word vector matrix, Indicates k o The transpose of m o represents the oth row vector of the entity node sentence vector, Indicates m o The transpose of m crepresents the c-th row vector of the entity node sentence vector matrix, k c Represents the c-th row vector of the maintenance word vector matrix.

[0089] As an optional implementation, constructing the fifth level of the knowledge graph in step S5 includes: for all entity nodes and fuzzy nodes of the fourth level, calculating the probability density of the sentence vector of the maintenance vector group corresponding to the word vector of each node, and extracting the sentence vector with the largest probability density as the relevant sentence vector of the fourth level node. All relevant sentence vectors of the fourth level constitute the fifth level entity nodes of the knowledge graph.

[0090] As an optional implementation, the bidirectional semantic prediction module includes a forward LSTM unit and a backward LSTM unit, which are used to bidirectionally capture the contextual semantic information of the word vector and concatenate them to obtain two semantic association information of the word vector.

[0091] like Figure 2 As shown, in another embodiment, the present application further discloses a method for generating a fault solution based on a knowledge graph, wherein the knowledge graph is constructed by the method in the above embodiment, and the method for generating a fault solution includes:

[0092] a. Obtain the fault word vector to be matched;

[0093] b. Calculate the correlation between the fault word vector to be matched and the first-level entity node word vector one by one; extract the entity word vector with the greatest correlation with the first-level entity node as the recognition word vector, and select the corresponding first and second associated word vectors as the preferred fault word vector and candidate fault word vector respectively;

[0094] c. Generate a fault description based on the preferred fault word vector and the candidate fault word vector combined with the knowledge graph;

[0095] d. Generate a maintenance plan based on the preferred fault word vector and the candidate fault word vector in combination with the knowledge graph;

[0096] e. Combine the fault description and maintenance plan to generate a fault solution.

[0097] As an optional implementation, the generation of the fault description in step c includes: based on the node relationships of the knowledge graph, extracting the fault sentence vector corresponding to the preferred fault word vector as the preferred fault sentence vector, extracting the fault sentence vector corresponding to the candidate fault word vector as the candidate fault sentence vector, and decoding the preferred fault sentence vector and the candidate fault sentence vector, and sorting and combining them according to the priority of the preferred and candidate, to obtain the corresponding fault sentence segment as the fault description.

[0098] As an optional implementation, the generation of the maintenance plan in step d includes:

[0099] Based on the knowledge graph, extract the first and second correlation associated word vectors of the preferred fault sentence vector as the first preferred maintenance word vector and the second preferred maintenance word vector; extract the first and second correlation associated word vectors of the first candidate fault sentence vector as the first candidate maintenance word vector and the second candidate maintenance word vector;

[0100] Calculate the probability density of all maintenance word vectors corresponding to their fault sentence vectors, define the probability density as the first matching weight of the corresponding maintenance word vector, and define the second matching weights of the first preferred maintenance word vector, the second preferred maintenance word vector, the first candidate maintenance word vector, and the second candidate maintenance word vector as 0.8, 0.6, 0.4, and 0.2 respectively;

[0101] Determine the product of the first matching weight and the corresponding second matching weight as the matching priority of the corresponding maintenance word vector, match the corresponding maintenance word vector based on the size of the priority, and based on the relationship between the nodes in the knowledge graph, extract the maintenance sentence vector corresponding to the maintenance word vector and sort and combine the maintenance sentence vectors according to the priority of the word vector, decode the maintenance sentence vector to obtain the corresponding maintenance sentence segment, and use it as the maintenance plan.

[0102] In order to better understand the present application, the embodiments of the present application are further described below.

[0103] The aviation operation and maintenance knowledge graph constructed by the method of this embodiment is as follows: Figure 3 As shown, Figure 3 The solid circle in the middle represents the entity node, and the hollow circle represents the fuzzy node. The construction method can be roughly divided into the following process.

[0104] (1) Based on the text data knowledge such as fault case library, maintenance manual, log records, etc. generated during the operation and maintenance of aviation logistics equipment, the fault descriptions or maintenance instructions are marked and extracted to form a fault text database and a maintenance text database.

[0105] (2) Word identifiers and sentence identifiers are used to isolate the words and sentences in the two text databases respectively, and a fault vector group consisting of a fault word vector and a fault sentence vector and a maintenance vector group consisting of a maintenance word vector and a maintenance sentence vector are formed through encoding.

[0106] (3) Based on the list of equipment and parts names in the technical manual of aviation logistics equipment, an entity word library of aviation logistics equipment names and parts names is established, and the entity words therein are encoded to obtain entity word vectors.

[0107] (4) Extract all entity word vectors, define each entity word vector as a node, and all nodes constitute the first level of the knowledge graph.

[0108] (5) Calculate the correlation between each entity word vector and each word vector in the fault vector group, extract the word vector with the greatest correlation as the first related word vector of the first-level node, and all the first related word vectors constitute the second-level entity node of the knowledge graph.

[0109] The correlation calculation formula is as follows:

[0110]

[0111] Among them, cov is the covariance, σ is the standard deviation, w1 is the entity word vector, w2 is the fault word vector, and r(w1|w2) represents the correlation between the entity word vector w1 and the fault word vector w2.

[0112] (6) The bidirectional semantic prediction module is used to perform contextual semantic prediction on the word vector of each entity node in the second level, and the two predicted word vectors are used as the two fuzzy nodes of the node respectively, and the fuzzy node is used as the second related word vector of the previous level.

[0113] (7) For all entity nodes and fuzzy nodes at the second level, calculate the probability density of the sentence vectors in the fault vector group corresponding to each node word vector, and extract the sentence vector with the largest probability density as the relevant sentence vector of the second-level node. All relevant sentence vectors at the second level constitute the entity nodes at the third level of the knowledge graph.

[0114] The probability density calculation formula of the word vector corresponding to the sentence vector is as follows:

[0115]

[0116] Among them, W represents the node word vector, S represents the fault sentence vector, and w o represents the oth row vector of the node word vector matrix, Indicates w o The transpose of s c is the c-th row vector of the fault sentence vector matrix, and N is the window size.

[0117] (8) For the entity nodes of the third level, the probability density of the word vectors in the maintenance vector group corresponding to the sentence vector of each entity node is calculated, and the word vector with the largest probability density is extracted as the first relevant word vector of the third level node. All the first relevant word vectors of the third level constitute the fourth level entity node of the knowledge graph.

[0118] The probability density calculation formula of the sentence vector corresponding to the word vector is as follows:

[0119]

[0120] Among them, M represents the entity node sentence vector, K represents the maintenance word vector, and k o represents the oth row vector of the maintenance word vector matrix, Indicates k o The transpose of m o represents the oth row vector of the entity node sentence vector, Indicates m o The transpose of m c represents the c-th row vector of the entity node sentence vector matrix, k c Represents the c-th row vector of the maintenance word vector matrix.

[0121] (9) For the entity nodes of the fourth level, the bidirectional semantic prediction module is used to perform contextual semantic prediction on the word vector of each entity node, and the two predicted word vectors are used as the two fuzzy nodes of the node and as the second related word vector of the previous level.

[0122] (10) For all entity nodes and fuzzy nodes at the fourth level, calculate the probability density of the sentence vectors of the maintenance vector group corresponding to the word vector of each node, and extract the sentence vector with the largest probability density as the relevant sentence vector of the fourth-level node. All relevant sentence vectors of the fourth level constitute the fifth-level entity nodes of the knowledge graph.

[0123] (11) Construct an operation and maintenance knowledge graph based on the five-level entity nodes and the corresponding word and sentence relationships.

[0124] Among them, the structure of the bidirectional semantic prediction module is as follows Figure 4 As shown in the figure, it consists of a forward LSTM unit and a backward LSTM unit, which capture the contextual semantic information in both directions and concatenate them to obtain two semantic association information of the word vector. LSTM stands for Long Short Term Memory Network.

[0125] For the word vector H t =(h1, h2, h3), performing forward LSTM and backward LSTM at the same time.

[0126] For the forward LSTM, input h1, h2, and h3 in sequence, and predict through LSTM to obtain three forward prediction vectors h R1 、h R2 、h R3 .

[0127] For the backward LSTM, input h3, h2, and h1 in sequence, and predict through LSTM to obtain three backward prediction vectors h L3 、h L2、h L1 .

[0128] The LSTM output expression is as follows:

[0129] h t =σ(W[H t-1 ,B])×tanh(C t )

[0130] Where W and B are the weight matrix and bias vector of LSTM, σ is the sigmoid activation function, and tanh is the hyperbolic tangent activation function; C t is the intermediate state of the input, h t is the output vector at time t, H t-1 is the input vector at time t-1.

[0131] Combine the forward prediction vector and the backward prediction vector with the word vector of the node respectively, and finally get {(h1, h R1 )、(h2、h R2 )、(h3、h R3 )} and {(h1,h L1 )、(h2、h L2 )、(h3、h L3 The predicted word vectors retain the global feature information of the input text sequence and the associated weight features between each word vector and the context, so that the word vector of the node at this level has a richer representation of the contextual semantic words.

[0132] The process of generating fault solutions using the knowledge graph constructed above generally includes:

[0133] (1) Based on the currently input query text data, extract the words therein and encode them to form the fault word vector to be matched;

[0134] (2) Calculate the correlation between the fault word vector to be matched and the first-level node word vector of the knowledge graph one by one. Extract the entity word vector with the greatest correlation with the first-level node as the recognition word vector, and select its corresponding first and second associated word vectors as the preferred fault word vector and candidate fault word vector respectively.

[0135] (3) Based on the node relationships of the knowledge graph, the fault sentence vector corresponding to the preferred fault word vector is extracted as the preferred fault sentence vector, and the fault sentence vector corresponding to the candidate fault word vector is extracted as the candidate fault sentence vector. The preferred fault sentence vector and the candidate fault sentence vector are decoded and sorted and combined according to the priority of the preferred and candidate, and the corresponding fault sentence segment is obtained as the fault description.

[0136] (4) Extracting the first and second correlation-associated word vectors of the preferred fault sentence vector as the first preferred maintenance word vector and the second preferred maintenance word vector; extracting the first and second correlation-associated word vectors of the candidate fault sentence vector as the first candidate maintenance word vector and the second candidate maintenance word vector.

[0137] (5) Calculate the probability density of all maintenance word vectors corresponding to their fault sentence vectors, define the probability density as the first matching weight of the corresponding maintenance word vector, and define the second matching weights of the first preferred maintenance word vector, the second preferred maintenance word vector, the first candidate maintenance word vector, and the second candidate maintenance word vector as 0.8, 0.6, 0.4, and 0.2, respectively.

[0138] (6) Determine the product of the first matching weight and the second matching weight as the matching priority of the maintenance word vector, match the corresponding maintenance word vector based on the size of the priority, and based on the relationship between each node in the knowledge graph, extract the maintenance sentence vector corresponding to the maintenance word vector and sort and combine the maintenance sentence vectors according to the priority of the word vector, decode the maintenance sentence vector to obtain the corresponding maintenance sentence segment, and use it as the maintenance plan.

[0139] (7) Combine the fault description and maintenance plan to generate a matching fault solution.

[0140] like Figure 5 As shown, the embodiment of the present application also provides an aviation logistics operation and maintenance knowledge graph management platform. Based on the current aviation logistics operation and maintenance business form and actual business needs, a management platform for knowledge graphs for words and sentences is designed, which includes a data acquisition layer, a data processing layer, a data storage layer, a knowledge extraction layer, and a functional application layer.

[0141] The data acquisition layer is mainly divided into two parts: structured / semi-structured data acquisition and unstructured data acquisition. First, the structured or semi-structured text information such as the operation manuals, operation and maintenance manuals, technical specifications, process specifications and standards, and fault diagnosis case collections of all hardware equipment and software in aviation logistics, including mechanical systems, electronic control systems, and IT systems, are extracted. Secondly, the web crawler technology is used to crawl the network related to the aviation logistics field. Patents, papers, research reports, daily operation and maintenance log records and other unstructured information, the obtained text information is cleaned and organized to obtain a text data set that can be used for knowledge extraction.

[0142] The data processing layer annotates the obtained text data set, where the data standard mainly annotates three parts: fault entity location, fault relationship type and relationship role. The position of the fault entity in the sentence is annotated with the BIO tag; the fault relationship type is obtained and annotated from the pre-defined fault entity relationship type library; the relationship role is represented in the form of a triple (entity 1, relationship type, entity 2). Considering that a certain fault entity may have multiple relationship types, they are marked with the numbers "1" and "2". "1" indicates that the word belongs to the first entity in the triple, and "2" indicates that the word belongs to the second entity.

[0143] The data storage layer is mainly used to store all the annotated text data in the form of triples in the entity-relationship database, and to store system data such as user information in the system information database.

[0144] The knowledge extraction layer extracts entities from the text data of the database based on the pre-customized knowledge graph architecture in a top-down manner, encodes word and sentence vectors, constructs a graph hierarchy, determines relationships based on the graph hierarchy, decodes the extracted information, and fills in knowledge to form an aviation logistics operation and maintenance knowledge graph.

[0145] The functional application layer deploys the application of the entire system functions based on the aviation logistics operation and maintenance knowledge graph, including entity recognition, solution matching, and graph visualization.

[0146] The entity recognition function of the aviation logistics operation and maintenance knowledge graph management platform constructed in this embodiment is mainly to mark and identify the entities in the aviation logistics equipment field in the case text or sentence entered by the user. The entities mainly include faulty equipment, fault time, fault mode, fault characteristics, fault causes, etc. After the user enters the text that he wants to recognize in the system front-end entity recognition module page, the system first encodes the input text to be recognized into a word vector, and imports the word vector to be recognized into the background correlation calculation module to calculate the correlation A between the vector and all entity word vectors in the first level of the knowledge graph. If there is an entity word vector with a correlation A greater than 0.6, all entity word vectors with a correlation A greater than 0.6 are extracted and divided into independent entity word vector units. In each unit, the correlation B and correlation C of the first and second correlation fault word vectors of the entity word vector and the word vector to be recognized are calculated respectively. The sum of A, B, and C of each unit is calculated, and the entity word vector corresponding to the maximum value is decoded to obtain the entity word as the entity recognition result. If there is no entity word vector with a correlation A greater than 0.6, it will prompt "Unable to recognize, please re-enter". Finally, the recognition results are fed back to the front-end page for users to view.

[0147] In another embodiment, the present application also discloses an electronic device, comprising: one or more processors; one or more memories; the one or more memories store one or more computer programs, the one or more computer programs include instructions, and when the instructions are executed by the one or more processors, the electronic device executes the methods described in the above embodiments.

[0148] In another embodiment, the present application further discloses a computer-readable medium, including a computer program. When the computer program is executed on a computer, the computer is enabled to execute the method in the above embodiment.

[0149] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the electronic device and computer-readable medium described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0150] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solution according to the embodiment of the present application can be embodied in the form of a software product, which can be stored in a readable medium or on a network, and includes several instructions to enable an electronic device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) to execute the method according to the embodiment of the present application.

[0151] The above description is only a specific implementation mode of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. All technical solutions that fall within the scope defined by the claims of the present invention fall within the protection scope of the present invention.

Claims

1. The aviation operation and maintenance knowledge graph construction method is characterized by: include: Constructing the first level of the knowledge graph, the first level includes at least one first-level entity node, defining an entity word vector as a first-level entity node, the entity word vector is obtained by encoding the entity word, and the entity word includes the name of aviation logistics equipment and the name of parts; Constructing a second level of the knowledge graph, the second level includes at least one second level entity node and at least two second level fuzzy nodes, defining a first related word vector of the first level entity node as a second level entity node, and the second level fuzzy node is obtained by semantically predicting the second level entity node; Constructing a third level of the knowledge graph, the third level comprising a plurality of third level entity nodes, the third level entity nodes being calculated from the second level entity nodes and the second level fuzzy nodes; Constructing a fourth level of the knowledge graph, the fourth level comprising a plurality of fourth level entity nodes and a plurality of fourth level fuzzy nodes, the fourth level entity nodes being obtained by calculation according to the third level entity nodes, and the fourth level fuzzy nodes being obtained by semantic prediction according to the fourth level entity nodes; A fifth level of the knowledge graph is constructed, wherein the fifth level includes a plurality of fifth level entity nodes, and the fifth level entity nodes are calculated based on the fourth level entity nodes and the fourth level fuzzy nodes.

2. The aviation operation and maintenance knowledge graph construction method according to claim 1 is characterized in that: Before constructing the first level of the knowledge graph, it also includes: Acquire fault text data and maintenance text data; wherein the fault text data and maintenance text data are derived from at least one text data in a fault case library, a maintenance manual, and a log record generated during the operation and maintenance of aviation logistics equipment; A fault vector group is obtained by using the fault text data, and a maintenance vector group is obtained by using the maintenance text data; wherein the fault vector group includes a fault word vector and a fault sentence vector, and the maintenance vector group includes a maintenance word vector and a maintenance sentence vector.

3. The aviation operation and maintenance knowledge graph construction method according to claim 2 is characterized in that: The first level of constructing the knowledge graph includes: Encoding entity words in the entity word library to obtain entity word vectors; wherein the entity word library contains entity words for aviation logistics equipment names and parts names; Extract all entity word vectors, each entity word vector is a node, and all nodes constitute the first level of the knowledge graph.

4. The method for constructing an aviation operation and maintenance knowledge graph according to claim 3, characterized in that: The second level of constructing the knowledge graph includes: Calculate the correlation between each entity word vector and each fault word vector in the fault vector group, extract the fault word vector with the greatest correlation as the first related word vector of the first-level entity node, and all the first related word vectors constitute the second-level entity node of the knowledge graph; The bidirectional semantic prediction module is used to perform contextual semantic prediction on the first related word vector corresponding to each second-level entity node, and the two predicted word vectors are used as the two fuzzy nodes of the second-level entity node, and the fuzzy node is used as the second related word vector of the previous level.

5. The method for constructing an aviation operation and maintenance knowledge graph according to claim 4, characterized in that: The correlation is calculated according to the following formula: Among them, cov is the covariance, σ is the standard deviation, w1 is the entity word vector, w2 is the fault word vector, and r(w1|w2) represents the correlation between the entity word vector and the fault word vector.

6. The method for constructing an aviation operation and maintenance knowledge graph according to claim 4, characterized in that: The third level of constructing the knowledge graph includes: for all entity nodes and fuzzy nodes of the second level, calculating the probability density of the fault sentence vector in the fault vector group corresponding to each node word vector, and extracting the sentence vector with the largest probability density as the relevant sentence vector of the second level node. All relevant sentence vectors of the second level constitute the third level entity node of the knowledge graph.

7. The method for constructing an aviation operation and maintenance knowledge graph according to claim 6, characterized in that: The probability density of the fault sentence vector in the fault vector group corresponding to the node word vector is calculated according to the following formula: Among them, W represents the node word vector, S represents the fault sentence vector, and w o represents the oth row vector of the node word vector matrix, Indicates w o The transpose of s c is the c-th row vector of the fault sentence vector matrix, and N is the window size.

8. The aviation operation and maintenance knowledge graph construction method according to claim 6, characterized in that: The fourth level of constructing the knowledge graph includes: For the entity nodes at the third level, the probability density of the maintenance word vectors in the maintenance vector group corresponding to each entity node sentence vector is calculated, and the word vector with the largest probability density is extracted as the first relevant word vector of the entity node at the third level. All the first relevant word vectors at the third level constitute the entity nodes at the fourth level of the knowledge graph. For the fourth-level entity nodes, the bidirectional semantic prediction module is used to perform contextual semantic prediction on the word vector of each entity node, and the two predicted word vectors are used as the two fuzzy nodes of the entity node and as the second related word vector of the previous level.

9. The aviation operation and maintenance knowledge graph construction method according to claim 8, characterized in that: The probability density of the maintenance word vector in the maintenance vector group corresponding to the entity node sentence vector is calculated according to the following formula: Among them, M represents the entity node sentence vector, K represents the maintenance word vector, and k o represents the oth row vector of the maintenance word vector matrix, Indicates k o The transpose of m o represents the oth row vector of the entity node sentence vector, Indicates m o The transpose of m c represents the c-th row vector of the entity node sentence vector matrix, k c Represents the c-th row vector of the maintenance word vector matrix.

10. The aviation operation and maintenance knowledge graph construction method according to claim 8, characterized in that: The fifth level of constructing the knowledge graph includes: for all entity nodes and fuzzy nodes of the fourth level, calculating the probability density of the sentence vector of the maintenance vector group corresponding to the word vector of each node, and extracting the sentence vector with the largest probability density as the relevant sentence vector of the fourth level node. All relevant sentence vectors of the fourth level constitute the fifth level entity nodes of the knowledge graph.

11. The aviation operation and maintenance knowledge graph construction method according to claim 4 or 8, characterized in that: The bidirectional semantic prediction module includes a forward LSTM unit and a backward LSTM unit, which are used to capture the contextual semantic information of the word vector in a bidirectional manner and concatenate them to obtain two semantic association information of the word vector.

12. A method for generating a fault solution based on a knowledge graph, wherein the knowledge graph is constructed by any one of the methods in claims 1-11, characterized in that: The fault solution generation method comprises: Get the fault word vector to be matched; Calculate the correlation between the fault word vector to be matched and the first-level entity node word vector one by one; extract the entity word vector with the greatest correlation with the first-level entity node as the recognition word vector, and select the corresponding first and second associated word vectors as the preferred fault word vector and candidate fault word vector respectively; Generate a fault description according to the preferred fault word vector and the candidate fault word vector in combination with the knowledge graph; Generate a maintenance plan based on the preferred fault word vector and the candidate fault word vector in combination with the knowledge graph; The fault description and the maintenance plan are combined to generate a fault solution.

13. The method for generating a fault solution based on a knowledge graph according to claim 12, characterized in that: The generation of the fault description includes: based on the node relationships of the knowledge graph, extracting the fault sentence vector corresponding to the preferred fault word vector as the preferred fault sentence vector, extracting the fault sentence vector corresponding to the candidate fault word vector as the candidate fault sentence vector, decoding the preferred fault sentence vector and the candidate fault sentence vector, and sorting and combining them according to the priority of the preferred and candidate, to obtain the corresponding fault sentence segment as the fault description.

14. The method for generating a fault solution based on a knowledge graph according to claim 12, characterized in that: The generation of the maintenance plan includes: Based on the knowledge graph, extract the first and second correlation associated word vectors of the preferred fault sentence vector as the first preferred maintenance word vector and the second preferred maintenance word vector; extract the first and second correlation associated word vectors of the first candidate fault sentence vector as the first candidate maintenance word vector and the second candidate maintenance word vector; Calculate the probability density of all maintenance word vectors corresponding to their fault sentence vectors, define the probability density as the first matching weight of the corresponding maintenance word vector, and define the second matching weights of the first preferred maintenance word vector, the second preferred maintenance word vector, the first candidate maintenance word vector, and the second candidate maintenance word vector as 0.8, 0.6, 0.4, and 0.2 respectively; Determine the product of the first matching weight and the corresponding second matching weight as the matching priority of the corresponding maintenance word vector, match the corresponding maintenance word vector based on the size of the priority, and based on the relationship between the nodes in the knowledge graph, extract the maintenance sentence vector corresponding to the maintenance word vector and sort and combine the maintenance sentence vectors according to the priority of the word vector, decode the maintenance sentence vector to obtain the corresponding maintenance sentence segment, and use it as the maintenance plan.

15. An electronic device, characterized in that: include: one or more processors; One or more memories; the one or more memories store one or more computer programs, the one or more computer programs include instructions, when the instructions are executed by the one or more processors, the electronic device performs the method as described in any one of claims 1-14.

16. A computer readable medium, characterized in that The method comprises a computer program, which, when executed on a computer, causes the computer to execute the method according to any one of claims 1 to 14.

Images