Fault Diagnosis Inference Method and System for Avionics Equipment Based on Graph Neural Network

Abstract

An avionics equipment fault diagnosis and reasoning method and system based on graph neural network, belonging to the technical field of fault prediction. The method includes: Step 1: Construct an initial fault knowledge graph based on expert experience; Step 2: Use graph neural network to vectorize the relationships and edges in the initial fault knowledge graph to obtain a fault vector knowledge graph; Step 3: Obtain the fault phenomenon to be detected and query the vector value from the mapping table; Step 4: According to the vector value of the fault phenomenon to be detected, calculate the similarity between the fault phenomenon to be detected and the entities in the fault vector knowledge graph. The entities in the fault vector knowledge graph that are directly connected to the entity with the highest similarity to the fault phenomenon to be detected through relationships are regarded as the entities connected to the fault phenomenon to be detected through relationships. Through the present invention, comprehensive fault causes for fault phenomena can be diagnosed.

Description

Technical Field

[0001] The present invention relates to a fault diagnosis and reasoning method and system for avionics equipment based on graph neural network, and belongs to the technical field of fault prediction and health management. Background Art

[0002] In the field of aerospace, with the continuous improvement of equipment functions and efficacy, the structure and functions of avionics equipment have become increasingly complex, and the difficulty of maintaining and repairing avionics equipment has also increased significantly. Therefore, at present, in order to ensure the normal operation of aviation equipment, improve the reliability of aviation equipment, be able to accurately locate and solve faults in a timely manner when faults occur, and be able to predict faults in a timely manner before faults occur has become one of the important research topics.

[0003] Combining faults with knowledge graph technology and creatively constructing a fault knowledge graph is actually taking advantage of artificial intelligence technology to overcome the limitations of traditional fault analysis. This integration is not only an innovation of the fault analysis method, but also an innovative attempt for knowledge management and intelligent decision-making. The fault knowledge graph integrates various data sources such as existing fault data, expert knowledge, and equipment information, and presents them visually in the form of a graph. Through this structured way of the knowledge graph, the knowledge graph can assist operators to quickly and accurately retrieve and obtain relevant information, and can also discover the potential connections and common features between different fault cases. Because the existing information is incomplete and insufficient, there will be a lack or omission of the relationship between two entities in the construction of the knowledge graph. Therefore, the use of knowledge reasoning is crucial in the knowledge graph. Through knowledge reasoning, the existing knowledge graph relationships can be made richer, making the establishment process of the fault knowledge graph more intelligent and efficient. However, due to the high difficulty and cost of obtaining knowledge of avionics equipment, only a small amount of knowledge can be obtained for some electronic equipment. The knowledge graph constructed based on the existing information always has situations such as less information volume and weaker correlation relationships (that is, the "edges" shown in the graph are not rich enough). Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention provides a fault diagnosis and reasoning method and system for avionics equipment based on graph neural network, which decomposes the fault mode of each fault phenomenon into multiple fault elements, and queries multiple fault causes for each fault element, so as to be able to diagnose comprehensive fault causes for the fault phenomenon.

[0005] To achieve the above-mentioned invention purpose, the present invention provides a fault diagnosis and reasoning method for avionics equipment based on graph neural network, which includes the following steps:

[0006] Step 1: Based on expert experience, sort out multiple fault elements for the fault phenomena of avionics equipment, sort out the fault causes for each fault element, use the fault phenomena, fault elements, and fault causes as entities, and use the relationships between the fault phenomena and fault elements, and between the fault elements and fault causes as edges to construct an initial fault knowledge graph;

[0007] Step 2: Use a graph neural network to vectorize the entities and edges in the initial fault knowledge graph to obtain a fault vector knowledge graph;

[0008] Step 3: Obtain the fault phenomenon to be detected and query the vector value from the mapping table;

[0009] Step 4: According to the vector value of the fault phenomenon to be detected, calculate the similarity between the fault phenomenon to be detected and the entities in the fault vector knowledge graph. The set of entities directly connected by the relationship to the entity in the fault vector knowledge graph with the highest similarity to the fault phenomenon to be detected is regarded as the one-hop entity set connected by the relationship of the fault phenomenon to be detected, denoted as 。

[0010] To achieve the above-mentioned invention purpose, the present invention also provides a system, which includes a storage medium and one or more processors. The storage medium stores a computer program, and the computer program is called by one or more processors to implement the above-mentioned avionics equipment fault diagnosis and reasoning method based on a graph neural network.

[0011] Compared with the prior art, the avionics equipment fault diagnosis and reasoning method and system based on a graph neural network provided by the present invention can decompose the fault mode of each fault phenomenon into multiple fault elements, and query multiple fault causes for each fault element, so as to be able to diagnose comprehensive fault causes for the fault phenomenon. Brief Description of the Drawings

[0012] Figure 1 is a flowchart of the avionics equipment fault diagnosis and reasoning method based on a graph neural network provided by the present invention. Detailed Embodiments

[0013] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0014] Figure 1 is a flowchart of the avionics equipment fault diagnosis and reasoning method based on a graph neural network provided by the present invention, as shown in Figure 1As shown, the avionics equipment fault diagnosis reasoning method based on graph neural network provided by the present invention comprises the following steps:

[0015] Step 1: Sort out the avionics equipment failure phenomena based on expert experience Fault element , sort out the fault elements Cause of failure , the fault phenomenon , Fault element and cause of failure As an entity, the fault phenomenon and fault element relationship, fault element and cause of failure The relationships between are used as edges to construct the initial fault knowledge graph;

[0016] Step 2: Use graph neural network to vectorize the entities and edges in the initial fault knowledge graph to obtain the fault vector knowledge graph;

[0017] Step 3: Obtain the fault phenomenon to be detected and query the vector value from the mapping table;

[0018] Step 4: According to the vector value of the fault phenomenon to be detected, the similarity between the fault phenomenon to be detected and the entities in the fault vector knowledge graph is calculated. The entity set directly connected to the entity in the fault vector knowledge graph with the highest similarity to the fault phenomenon to be detected through a relationship is regarded as the one-hop entity set connected to the fault phenomenon to be detected through a relationship, and is recorded as , I is a positive integer.

[0019] The avionics equipment fault diagnosis reasoning method based on graph neural network provided by the present invention also includes the following steps:

[0020] Step 5: Query the fault vector knowledge graph for the one-hop entity set Each entity in The entities directly connected by the edge minus the return flow entity is used as the two-hop entity set of the fault phenomenon to be detected. ,

[0021] In the formula, Representation and Entity The set of entities obtained by subtracting the backflow entities from the entities directly connected by the edge.

[0022] By analogy, query the fault vector knowledge graph with entity set E that is 1 hop away from a a Each entity in the is directly connected to the entities through the edge, and the entities in the reflux are cut as the a-hop entity set of the fault phenomenon to be detected. After multiple reasonings, the comprehensive fault causes for the fault phenomenon can be diagnosed through the quantified knowledge graph of the graph neural network.

[0023] In the present invention, according to the vector value of the fault phenomenon to be detected , calculating the similarity between the fault phenomenon to be detected and the entities in the fault vector knowledge graph specifically includes:

[0024] S1-1: Calculate the embedding matrix after the (k + 1)-layer diffusion of the vector of the fault phenomenon to be detected;

[0025] S1-2: Obtain the entity with the highest similarity to the fault phenomenon to be detected in the fault vector knowledge graph through the following formula: ,

[0026] wherein, is the second norm; represents the vector of any entity in the fault vector knowledge graph after the (k + 1)-layer diffusion embedding matrix; D is the set of actually existing vector triples.

[0027] In the fault inference method of the graph neural network provided by the present invention, vectorizing the entities and edges in the initial fault knowledge graph by using the graph neural network to obtain the initial fault vector knowledge graph specifically includes:

[0028] S2-1: Two related entities and a relationship in the initial fault knowledge graph form a fault triple, and its fault vector triple is , wherein, h represents the head entity vector, r represents the relationship vector, and t represents the tail entity vector value;

[0029] S2-2: Extract all the head entities and tail entities of the initial fault knowledge graph, extract all the relationships of the initial fault knowledge graph, the size of the entity set is , and the size of the relationship set is ;

[0030] S2-3: Randomly initialize the embedding vectors of the entities to obtain the entity embedding matrix H; randomly initialize the embedding vectors of the relationships to obtain the relationship embedding matrix R; at the same time, according to the fault vector triple set, obtain the adjacency matrix A; for each relationship vector r, extract the corresponding subgraph adjacency matrix such that only contains the edges related to the vector r; randomly initialize the change matrix of the relationship vector r;

[0031] S2-4: Calculate the message for the relationship vector r, and further combine the message of the relationship vector with the relationship embedding vector ; Then aggregate all relationships: , and then update the entity embedding vectors to obtain the (k + 1)-th layer embedding matrix of the relationship vector r: , is the activation function; at the same time, obtain the degree matrix of each entity from the adjacency matrix A ;

[0032] S2-5: Calculate the normalized diffusion matrix: , and perform further diffusion on the entity nodes to obtain the diffused embedding matrix: ;

[0033] S2-6: Traverse the fault vector triples in the initial fault knowledge graph set, and calculate the positive loss function according to the following formula : , where represents the diffused embedding matrix of the (k + 1)-th layer of the head entity vector h, represents the diffused embedding matrix of the (k + 1)-th layer of the tail entity vector t.

[0034] In the present invention, using a graph neural network to vectorize the entities and edges in the initial fault knowledge graph to obtain a fault vector knowledge graph includes:

[0035] S2-7: Construct negative samples, and generate false fault vector triples by randomly replacing the head entity vector or the tail entity vector in the fault vector triple , , ;

[0036] S2-8: Traverse the false fault vector triples in the initial fault knowledge graph set, and calculate the negative loss function according to the following formula :

[0037] , where represents the (k + 1)-th layer diffused embedding matrix of the vector ; represents the (k + 1)-th layer diffused embedding matrix of the vector ;

[0038] S2-9: Use the following formula to obtain the total loss L:

[0039] , where represents taking the positive part; D is the set of actually existing triples, is the set of negative example triples;

[0040] S2-10: Update the vector values of the head entity, tail entity, and relation using the gradient descent method.

[0041] In the present invention, the use of a graph neural network to vectorize the entities and edges in the initial fault knowledge graph to obtain a fault vector knowledge graph further includes:

[0042] 2-11: Determine whether the total loss L is minimized. If the total loss L reaches the minimum, use the test fault vector triple for verification. If the verification is qualified, the entities and relations in the initial fault knowledge graph are assigned fixed vectors, and an optimized entity embedding matrix , an optimized relation embedding matrix and an optimized change matrix are obtained, and then the initial fault vector knowledge graph is obtained; if the total loss L does not reach the minimum, replace the updated vector values of the head entity h, tail entity t, and relation r with the vector values before the update, and then return to step S2-4.

[0043] Through the above technical solutions, the present invention quantifies all the entities and relations in the initial fault knowledge graph, assigns fixed vectors to each entity and relation in the initial fault knowledge graph, thereby laying a foundation for diagnosing comprehensive fault causes for fault phenomena.

[0044] Optionally, generate a false fault vector triple for the head entity vector or tail entity vector in the fault vector triple for perturbation, and calculate the perturbation loss through the following formula:

[0045] ,

[0046] In the formula, is the perturbation momentum, T represents the transpose; update the values of the head entity vector h, tail entity vector t, and relation vector r using the gradient descent method, and determine whether the total loss is minimized. If the total loss reaches the minimum, use the direction of the test fault triple for verification. If the verification is qualified, the entities and relations in the initial fault knowledge graph are assigned fixed vectors, and an optimized entity embedding matrix , an optimized relation embedding matrix and an optimized change matrix are obtained, and then the initial fault vector knowledge graph is obtained; if the total loss L does not reach the minimum, replace the updated vector values of the head entity h, tail entity t, and relation r with the vector values before the update, and then return to step S2-4.

[0047] Through the above method, the present invention can make the fault vector knowledge graph have a wider scope of application.

[0048] Using a graph neural network to vectorize entities and edges in the initial fault knowledge graph to obtain a fault vector knowledge graph further includes:

[0049] S2-12: Obtain the vector of another entity in the initial fault vector knowledge graph , and calculate the vector through the following formula and the vector of the difference of the (k + 1)-th layer embedding vectors of ,

[0050] wherein, represents the embedding matrix after diffusion of the (k + 1)-th layer of the vector ; represents the embedding matrix after diffusion of the (k + 1)-th layer of the vector ;

[0051] S2-13: Obtain all entities where is less than the set value, and connect the other entity with all entities less than the set value, and calculate the relationship vector between them;

[0052] S2-14: Repeat step S2-12 and step S2-13, traverse all entities in the initial fault vector knowledge graph to obtain a fault vector knowledge graph.

[0053] Through the above technical solution, the present invention makes the connection of entities in the initial fault vector knowledge graph more comprehensive.

[0054] In the present invention, using a graph neural network to vectorize entities and edges in the initial fault knowledge graph to obtain a fault vector knowledge graph further includes:

[0055] S2-15: Obtain a new fault phenomenon , calculate the similarity between the new fault phenomenon and all entities e in the first fault vector knowledge graph. If the similarity is greater than or equal to the threshold, then use the fault phenomenon as an entity, and connect it through an edge to the entity connected to entity e. Adjust the vector of entity e according to the similarity as the vector of the fault phenomenon , and adjust the vector of the edge connecting entity e and its connected entity according to the similarity as the vector of the edge connecting the fault phenomenon and the entity connected to entity e; if the similarity Less than the threshold value, according to expert experience, sort out new fault modes for this new fault phenomenon, decompose the new fault modes into multiple new fault elements, and sort out multiple new fault causes for the new fault elements; regard the new fault phenomenon, new fault elements and new fault causes as entities, and the relationship between the new fault phenomenon and the new fault elements, and the relationship between the new fault elements and the new fault causes as edges to construct a new initial fault knowledge graph. Then, use a graph neural network to vectorize the entities and edges of the new initial fault knowledge graph to obtain a second fault vector knowledge graph, and fuse the first fault vector knowledge graph and the second fault vector knowledge graph to obtain a fault vector knowledge graph.

[0056] In the present invention, using a graph neural network to vectorize the entities and edges in the initial fault knowledge graph to obtain a fault vector knowledge graph further includes the following steps:

[0057] S2-16: Export the corresponding relationship between the entities, relationships and vectors in the fault vector knowledge graph to obtain a mapping table.

[0058] In the present invention, fusing the first fault vector knowledge graph and the second fault vector knowledge graph to obtain a fault vector knowledge graph includes the following process:

[0059] S2-17: Obtain any entity in the first fault vector knowledge graph , whose vector is , and obtain any entity from the second fault vector knowledge graph , whose vector is Calculate the vector of the vector and the vector of the difference between the embedded vectors after the (k + 1)-th layer of diffusion:

[0060] , where represents the embedded matrix after the diffusion of the vector ; represents the embedded matrix after the diffusion of the vector ;

[0061] S2-18: Obtain all less than the set value, and connect with all entities less than the set value and calculate the relationship vector between them;

[0062] S2-19: Repeat step S2-17 and step S2-18, traverse all entities in the initial fault vector knowledge graph to obtain a fault vector knowledge graph.

[0063] Through the above technical solutions, the present invention can fuse multiple fault vector knowledge graphs with more than or equal to 2, without having to re-quantify the spliced fault knowledge graph again.

[0064] Through the above technical solutions, the present invention expands the fault vector knowledge graph, making the fault phenomena, fault elements and fault causes more comprehensive, which is convenient for obtaining more comprehensive fault causes during fault diagnosis.

[0065] To achieve the above object of the invention, the present invention further provides a system, which includes a storage medium and one or more processors. The storage medium stores a computer program, and the computer program is called by one or more processors to implement the above-mentioned fault diagnosis and reasoning method for avionics equipment based on graph neural network.

[0066] To achieve the above object of the invention, the present invention further provides a computer program product, which compiles the above-mentioned fault diagnosis and reasoning method for avionics equipment based on graph neural network into a computer program called and executed by one or more processors using a computer language.

[0067] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation manners described. Obviously, according to the content of this specification, many modifications and changes can be made. These embodiments are selected and specifically described in this specification to better explain the principle and practical application of the present invention, so that those skilled in the art can understand and utilize the present invention well. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A method for fault diagnosis and reasoning of avionics equipment based on graph neural network, characterized in that: The steps include: Step 1: Based on expert experience, sort out multiple fault elements for avionics equipment fault phenomena, sort out the fault causes for each fault element, take fault phenomena, fault elements and fault causes as entities, and use the relationships between fault phenomena and fault elements, and between fault elements and fault causes as edges to construct an initial fault knowledge graph; Step 2: Use the graph neural network to vectorize the entities and edges in the initial fault knowledge graph to obtain the fault vector knowledge graph, including: S2-1: The two entities and relations in the initial fault knowledge graph constitute a fault triple, and its fault vector triple is , where h represents the head entity vector, r represents the relationship vector, and t represents the tail entity vector value; S2-2: Extract all head entities and tail entities of the initial fault knowledge graph, extract all relations of the initial fault knowledge graph, and the size of the entity set is , the size of the relationship set is ; S2-3: Randomly initialize the embedding vectors of entities and relations to obtain the entity and relationship embedding matrices H and R respectively; according to the fault vector triple Set, get the adjacency matrix A; according to the relationship vector r, extract the corresponding subgraph adjacency matrix , so that Only include edges related to vector r; randomly initialize the change matrix of relationship vector r ; S2-4: For the relation vector r, calculate the message , embedding the message and relation vector of the relation vector Combination ; Aggregate all relations: ; Update the entity embedding vector to obtain the k+1th layer embedding matrix of the relationship vector r: , is the activation function; at the same time, the degree matrix of each entity is obtained from the adjacency matrix A ; S2-5: Calculate the normalized diffusion matrix: , further diffusion is performed on the entity nodes to obtain the diffused embedding matrix: ; S2-6: Traverse all fault vector triplets in the initial fault knowledge graph Set, calculate the forward loss function according to the following formula : , where represents the diffused embedding matrix of the head entity vector h at the k+1th layer, represents the diffusion embedding matrix of the k+1th layer of the tail entity vector t; Step 3: Obtain the fault phenomenon to be detected and query the vector value from the mapping table; Step 4: According to the vector value of the fault phenomenon to be detected, the similarity between the fault phenomenon to be detected and the entities in the fault vector knowledge graph is calculated. The entity set directly connected to the entity in the fault vector knowledge graph with the highest similarity to the fault phenomenon to be detected through a relationship is regarded as the one-hop entity set connected to the fault phenomenon to be detected through a relationship, and is recorded as .

2. The avionics equipment fault diagnosis reasoning method based on graph neural network according to claim 1 is characterized in that: The following steps are also included: Step 5: Query the fault vector knowledge graph for the one-hop entity set Each entity in The entities directly connected by the edge minus the return flow entity is used as the two-hop entity set of the fault phenomenon to be detected. , where Representation and Entity The set of entities obtained by subtracting the backflow entities from the entities directly connected by the edge.

3. The avionics equipment fault diagnosis reasoning method based on graph neural network according to claim 2 is characterized in that: According to the vector value of the fault phenomenon to be detected, calculating the similarity between the fault phenomenon to be detected and the entity in the fault vector knowledge graph specifically includes: S1-1: Calculate the vector of the fault phenomenon to be detected The embedding matrix after the k+1th layer diffusion ; S1-2: Obtain the entity with the highest similarity to the fault phenomenon to be detected in the fault vector knowledge graph through the following formula: , In the formula, is the two-norm; A vector representing any entity in the fault vector knowledge graph The k+1-layer diffusion embedding matrix of ; D is the set of real vector triplets.

4. The avionics equipment fault diagnosis reasoning method based on graph neural network according to claim 3 is characterized in that: The fault vector knowledge graph obtained by vectorizing the entities and edges in the initial fault knowledge graph using the graph neural network also includes: S2-7: Construct negative samples by randomly replacing the fault vector triples Generate a false fault vector triplet from the head or tail entity vector in , , ; S2-8: Traverse the fault vector triples in the initial fault knowledge graph Set, calculate the negative loss function according to the following formula : , where Representation vector The volume embedding matrix after the k+1th layer diffusion; Representation vector The entity embedding matrix after the k+1th layer diffusion; S2-9: Use the following formula to calculate the total loss L: , where represents the positive part; D is the set of real triples, is the set of negative triples; S2-10: Use gradient descent method to update the vector values ​​of head entity, tail entity and relationship.

5. The avionics equipment fault diagnosis reasoning method based on graph neural network according to claim 1 is characterized in that: The fault vector knowledge graph obtained by vectorizing the entities and edges in the initial fault knowledge graph using the graph neural network also includes: 2-11: Determine whether the total loss L is the minimum. If the total loss L is the minimum, use the test fault vector triplet Verification is performed and the entities and relationships in the initial fault knowledge graph are assigned fixed vectors to obtain the optimized entity embedding matrix. , optimized relation embedding matrix And the optimized change matrix , and then obtain the initial fault vector knowledge graph.

6. The avionics equipment fault diagnosis reasoning method based on graph neural network according to claim 5 is characterized in that: The fault vector knowledge graph obtained by vectorizing the entities and edges in the initial fault knowledge graph using the graph neural network also includes: S2-12: Obtaining the vector of another entity in the initial fault vector knowledge graph , calculate the vector by the following formula and vector The second norm of the difference of the k+1th layer embedding vector: , In the formula, Representation vector The embedding matrix after the k+1th layer diffusion; Representation vector The embedding matrix after the k+1th layer diffusion; S2-13: Acquisition less than the set value, and compare the other entity with All entities smaller than the set value are connected, and the relationship vector between them is calculated; S2-14: Repeat steps S2-12 and S2-13 to traverse all entities in the initial fault vector knowledge graph to obtain the fault vector knowledge graph.

7. The avionics equipment fault diagnosis reasoning method based on graph neural network according to claim 6 is characterized in that: The fault vector knowledge graph obtained by vectorizing the entities and edges in the initial fault knowledge graph using the graph neural network also includes: S2-15: Get new fault symptoms , calculate the new fault phenomenon Similarity with all entities e in the knowledge graph of the first fault vector , if the similarity If it is greater than or equal to the threshold, the fault phenomenon As an entity, and connected to the entity connected to entity e through an edge, the vector of entity e is adjusted according to the similarity as the fault phenomenon The vector of the edge connecting entity e and its connected entities is adjusted according to the similarity as the connection failure phenomenon and the edge vector of the entity connected to entity e; if the similarity Less than a threshold, based on expert experience, a new fault mode for the new fault phenomenon is sorted out, the new fault mode is decomposed into multiple new fault elements, and multiple new fault causes are sorted out for the new fault elements; the new fault phenomenon, the new fault element and the new fault cause are taken as entities, and the relationship between the new fault phenomenon and the new fault element, and the relationship between the new fault element and the new fault cause are taken as edges to construct a new initial fault knowledge graph, and then the graph neural network is used to vectorize the entities and edges of the new initial fault knowledge graph to obtain a second fault vector knowledge graph, and the first fault vector knowledge graph and the second fault vector knowledge graph are fused to obtain a fault vector knowledge graph.

8. The avionics equipment fault diagnosis reasoning method based on graph neural network according to claim 7 is characterized in that: Using the graph neural network to vectorize the entities and edges in the initial fault knowledge graph to obtain the fault vector knowledge graph also includes the following steps: S2-16: Derive the correspondence between entities, relations and vectors in the fault vector knowledge graph to obtain a mapping table.

9. A fault diagnosis and reasoning system for avionics equipment based on graph neural network, characterized in that: It comprises a storage medium and one or more processors, wherein the storage medium stores a computer program, and the computer program is called by one or more processors to implement the avionics equipment fault diagnosis reasoning method based on graph neural network as described in any one of claims 1-8.

Images