Aircraft landing gear fault diagnosis knowledge graph completion method based on meta-learning
Through a meta-learning-based method, the aircraft landing gear fault diagnosis knowledge graph is completed, and the problem of low fault diagnosis efficiency caused by data sparsity is solved, achieving more efficient and accurate fault diagnosis effect.
Patent Information
- Application Number
- CN202510088962.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-20
- Publication Date
- 2025-05-16
AI Technical Summary
Industrial knowledge graphs are prone to data sparsity problems during the construction process, resulting in low efficiency and poor results in equipment fault diagnosis. This problem needs to be solved through knowledge graph completion.
Using a meta-learning-based method, the entities and relationships in the aircraft landing gear fault diagnosis knowledge graph are encoded through the DS-GAT encoder, the representations of entities and relationships are updated using GNN message delivery, and the relationships between unknown entities are predicted through sub-graph construction and meta-training to complete the knowledge graph completion.
It improves the completeness and accuracy of the knowledge graph, thereby improving the efficiency and effectiveness of aircraft landing gear fault diagnosis, and can position and resolve equipment faults more quickly and accurately.
Smart Images
Figure CN120012903A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of fault diagnosis, and specifically relates to a meta-learning-based knowledge graph completion method for aircraft landing gear fault diagnosis. Background Art
[0002] As a structured semantic knowledge base, knowledge graph provides an important tool for information processing, intelligent search, recommendation system, semantic understanding and other aspects in the context of the rapid development of Internet technology. Especially in the industrial field, knowledge graph is used to integrate operation records, expert experience, technical documents, etc. in the production process, thus providing key support for tasks such as fault diagnosis, quality traceability, production scheduling, etc. However, industrial knowledge graph also faces a series of challenges, one of which is data sparsity. The original knowledge in the industrial field is usually multi-source heterogeneous and scattered data, and the process of constructing knowledge graph usually relies on manual and semi-automatic methods, which makes it inevitable that the constructed knowledge graph has incomplete knowledge. Therefore, it is necessary to solve this problem by completing the knowledge graph. Open world knowledge graph completion aims to improve and expand incomplete knowledge graphs by introducing new entities and relationships. With the increase in the scale of knowledge graphs, the method of relying on manual completion of knowledge graphs is inefficient and costly, and is not suitable for handling large-scale completion tasks. It is necessary to study knowledge graph completion methods to achieve intelligent and efficient completion operations.
[0003] The industrial sector involves complex and diverse equipment and process flows, including a large number of mechanical equipment, electrical equipment, and automation systems. These devices usually operate in harsh environments, such as high temperature, high pressure, and high speed, and are susceptible to wear, corrosion, and aging, which can lead to equipment failures. In addition, industrial equipment often requires frequent power-on and power-off operations, which may lead to misoperation or improper use, increasing the risk of equipment failure. Due to the complexity and criticality of industrial equipment, once a failure occurs, it may lead to serious consequences such as production line shutdown and casualties. Therefore, timely and effective fault diagnosis of industrial equipment is crucial.
[0004] Equipment fault diagnosis is a complex task that involves detecting, analyzing, and processing various anomalies and faults, and proposing corresponding solutions. Traditional fault diagnosis methods usually rely on the experience of technicians and direct observation of equipment, and have problems such as low diagnostic accuracy, long time consumption, and reliance on personal experience. With the development of information technology, knowledge graphs, as a structured and semantic knowledge representation method, provide new ideas and solutions for fault diagnosis of industrial equipment. By integrating historical fault data, equipment knowledge, and expert experience, knowledge graphs can help engineers locate and solve equipment failures more quickly and accurately.
[0005] In the industrial field, due to the limitations of graph construction technology and the dispersion of fault knowledge, the constructed knowledge graphs usually have data sparsity problems, resulting in low efficiency and poor results in equipment fault diagnosis. Therefore, completing the industrial knowledge graph is crucial for equipment fault diagnosis. Summary of the invention
[0006] In view of the above-mentioned shortcomings of the prior art, the object of the present invention is to provide a meta-learning-based knowledge graph completion method for aircraft landing gear fault diagnosis to solve the above-mentioned problems.
[0007] According to the present invention, a method for completing a knowledge graph for aircraft landing gear fault diagnosis based on meta-learning is provided, comprising the following steps:
[0008] S1. Use the DS-GAT encoder to encode the entities and relations in the aircraft landing gear fault diagnosis knowledge graph;
[0009] S2, using GNN message passing to obtain the representation of entities and relationships in the aircraft landing gear fault diagnosis knowledge graph;
[0010] S3. Predict the relationship between unknown entities through subgraph construction and meta-training to complete the knowledge graph of aircraft landing gear fault diagnosis.
[0011] Furthermore, in S1, the calculation expressions of entities and relations are as follows:
[0012] Among them, (H, R, T) is the triplet embedding representation obtained by encoding, (h BERT ,r BERT ,t BERT ) and (h R-GCN ,r R-GCN ,t R-GCN ) are the embedding representations obtained by BERT and R-GCN respectively, It is a vector concatenation operation, and CNN() represents the aggregation of two embedding representations through a convolutional neural network.
[0013] Furthermore, in S2, GNN is used to update the representation of entities and relations, where the entity update formula is as follows:
[0014]
[0015]
[0016] Where h, r and t are the head entity, relation and tail entity respectively, σ is the activation function, o(e) is the set of outgoing edges and tail entities of node e, T(e) is the set of incoming edges and head entities of node e, and are the outgoing transformation matrix and incoming transformation matrix of entity e at layer l, respectively. is the linear transformation matrix of the self-loop information at the lth layer, and are the representations of entity e at layer l and layer l+1 respectively; the update formula of the relationship is as follows
[0017]
[0018] in, is the linear transformation matrix of relation r at level l, and are the representations of relation r at the lth layer and the l+1th layer respectively.
[0019] Furthermore, in S3, subgraph construction includes:
[0020] Convert the original training graph into an undirected graph;
[0021] Select nodes of the subgraph by random walk and construct the subgraph;
[0022] Randomly sort the triplets in the subgraph, and count the frequency of each entity and relationship when looping through the triplets in the subgraph;
[0023] After obtaining the sorted triples, a portion of the triples are randomly selected as query triples based on the frequency of entities and relations and the size of the subgraph, and the remaining triples are used as support triples;
[0024] Generate two dictionaries hr2t and rt2h, which are used to store the query triples and the related information of the supporting triples respectively.
[0025] Furthermore, triples whose entity and relation frequencies are greater than 2 are selected as query triples, and triples whose entity and relation frequencies are less than or equal to 2 are selected as support triples.
[0026] Furthermore, in S3, meta-training is implemented through two nested loops: the outer loop iterates the number of meta-training rounds, and each round of meta-training traverses all tasks in the dataset; the inner loop iterates each subgraph data batch.
[0027] Furthermore, in S3, the objective function of meta-training is as follows
[0028]
[0029] Among them, R is the embedding matrix of all relations, G i is the i-th subgraph, θ, are all learning parameters, Q i is the i-th query triple, and m represents the number of samples;
[0030] A self-adversarial loss function is used on the negative sampling task of each query triple
[0031]
[0032] Among them, k is the number of negative samples, (h' i ,r' i ,t' i ) is the i-th negative sample, p(h' i ,r' i ,t' i ) is the self-adversarial weight of the i-th negative sample, which is calculated as follows
[0033]
[0034] Where β is the temperature parameter of the sample. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is a flow chart of a method for completing a knowledge graph for aircraft landing gear fault diagnosis based on meta-learning according to an embodiment of the present invention;
[0036] Figure 2 is a schematic diagram of the DSMeta model structure according to an embodiment of the present invention;
[0037] Figure 3 A local graph in an aircraft landing gear fault diagnosis knowledge graph according to an embodiment of the present invention is shown;
[0038] Figure 4 The curves of the change of indicators MRR, Hits@1, Hits@3 and Hits@10 of the DSMeta model in the ALGFD-Open dataset under different score functions during the iteration process are shown;
[0039] Figure 5 It is the completed knowledge graph subgraph. DETAILED DESCRIPTION
[0040] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.
[0041] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0042] See also Figure 1 and Figure 2 ,in, Figure 1 A flow chart of a method for completing a knowledge graph for aircraft landing gear fault diagnosis based on meta-learning according to the present invention is shown. Figure 2 The DSMeta model structure is shown. The method may include the following steps:
[0043] S1. Use the DS-GAT encoder to encode the entities and relations in the aircraft landing gear fault diagnosis knowledge graph.
[0044] Before training, we first need to initialize the entities and relationships in the knowledge graph of aircraft landing gear fault diagnosis. In order to obtain high-quality entity and relationship embedding, the aggregated entity description and neighborhood structure information encoder (DS-GAT encoder) is used. The calculation expressions of entities and relationships are as follows
[0045]
[0046] Among them, (H, R, T) is the triplet embedding representation obtained by encoding, (h BERT ,r BERT ,t BERT ) and (h R-GCN ,r R-GCN ,t R-GCN ) are the embedding representations obtained by BERT and R-GCN respectively, For the vector concatenation operation, CNN represents the aggregation of two embedding representations through a convolutional neural network.
[0047] S2. GNN message passing is used to obtain the representation of entities and relationships in the aircraft landing gear fault diagnosis knowledge graph.
[0048] GNN (Graph Neural Network) can better capture the global information of the graph and consider the global graph features when transmitting and integrating the embedded information. After obtaining the entity relationship embedding representation with entity description information and neighborhood structure information, GNN (Graph Neural Network) is used to update the representation of entities and relationships.
[0049] The update formula for the entity is as follows
[0050]
[0051] Among them, o(e) is the outgoing edge and tail entity set of node e, T(e) is the incoming edge and head entity set of node e, and are the outgoing transformation matrix and incoming transformation matrix of entity e at layer l respectively.
[0052]
[0053] Among them, h, r and t are the head entity, relation and tail entity respectively, σ is the activation function, is the linear transformation matrix of the self-loop information at the lth layer, and are the representations of entity e at layer l and layer l+1 respectively. The update formula of the relationship is as follows
[0054]
[0055] in, is the linear transformation matrix of relation r at level l, and are the representations of relation r at the lth layer and the l+1th layer respectively.
[0056] S3. Predict the relationship between unknown entities through subgraph construction and meta-training to complete the knowledge graph of aircraft landing gear fault diagnosis.
[0057] Subgraph construction
[0058] The local graph neighborhood of a specific triple in the aircraft landing gear fault diagnosis knowledge graph is set, which contains logical information for reasoning about the relationship between target nodes, and the path connecting two target nodes may contain target relationship information. Before model training, it is necessary to first build a subgraph for each source graph, where each subgraph contains supporting triples and query triples. In order to determine the scope of the subgraph, a closed subgraph is defined as a graph derived from all nodes that appear on the path between two target nodes.
[0059] First, the original training graph is converted into an undirected graph. Then, the nodes of the subgraph are selected by random walk, and the subgraph Sub_Graph is constructed. The triples in the subgraph are randomly sorted, and the frequency of each entity and relationship is counted when looping through the triples in the subgraph. After obtaining the sorted triples, according to the frequency of entities and relationships and the size of the subgraph, a part of the triples are randomly selected as query triples Que_Tris, and the remaining triples are selected as support triples Sup_Tris. Specifically, triples with a frequency of entities and relationships greater than 2 are selected as query triples, and triples with a frequency of entities and relationships less than or equal to 2 are selected as support triples. In addition, the subgraph is considered valid only when the number of nodes in the subgraph is greater than or equal to 50. Finally, two dictionaries hr2t and rt2h are generated to store the relevant information of query triples and support triples.
[0060] Meta-training
[0061] Meta-training aims to enable the model to generalize, adapt quickly and perform well when encountering new tasks. The meta-training goal is to produce reasonable entity embedding vectors so that they can be effectively completed when faced with knowledge graphs containing unseen entities. First, a set of subgraphs are randomly sampled from the source graph, and the entities in these subgraphs are regarded as unseen entities to simulate the target graph under the inductive setting. In addition, a portion of the triples in the subgraph are used as query triples, and the remaining triples are used as support triples, where support triples and query triples are mutually exclusive, and support triples contain entities and relations in query triples. Support triples are used to generate entity embedding vectors, while query triples are used to evaluate the rationality of the generated embedding vectors and calculate the training loss.
[0062] Meta-training is implemented in two nested loops: the outer loop iterates the number of meta-training rounds, each round of meta-training iterates over all tasks in the dataset; the inner loop iterates over each batch of subgraph data. During training, a graph is first built for each task by iterating over each task in the training set. The graph is a subgraph consisting of positive samples that describe the goal of the task and negative samples that are used to train the model to distinguish between positive examples. Then, the parameters are updated by processing the batches of subgraph data, mainly by calculating the loss and using the backpropagation algorithm to update the parameters. At each specified training step, the performance of the model on the validation dataset is evaluated. This is to monitor the generalization ability of the model, and if the performance of the model on the validation set improves, the current model parameters are saved. The strategy of saving the best model parameters helps ensure that the model can achieve good performance on new tasks. Through multiple rounds of iterative, task-oriented subgraph construction and loss optimization, as well as monitoring the model performance and saving the best model parameters, the model can learn generalization ability from a small number of tasks and perform better on new tasks. The objective function of meta-training is as follows
[0063]
[0064] Among them, R is the embedding matrix of all relations, G i is the i-th subgraph, θ, are all learnable parameters, Q i is the i-th query triple, and m represents the number of samples.
[0065] A self-adversarial loss function is used on the negative sampling task of each query triple
[0066]
[0067] Among them, σ is the sigmoid function, k is the number of negative samples, (h' i ,r' i ,t' i ) is the i-th negative sample, p(h' i ,r' i ,t' i ) is the self-adversarial weight of the i-th negative sample, which is calculated as follows
[0068]
[0069] Where β is the temperature parameter of the sample.
[0070] Score function
[0071] The score function is an important indicator for evaluating the credibility of triples in the knowledge graph. This paper considers four different score functions to evaluate triples: TransE, DistMult, ComplEx and RotatE. The following is the calculation method of each score function:
[0072] TransE (based on translation embedding):
[0073]
[0074] DistMult (based on distance multiplication):
[0075]
[0076] ComplEx (based on complex number embedding):
[0077]
[0078] RotatE (rotation based on relationship):
[0079]
[0080] Among them, h, r and t are the head entity, relation and tail entity respectively, f represents the score, and e represents the embedded representation.
[0081] experiment
[0082] Based on the aircraft landing gear risk assessment report and fault investigation report of a large state-owned airline, an aircraft landing gear fault diagnosis knowledge graph was constructed, and the application of the knowledge graph completion method was verified. The graph contains 274 entities and 14 types of relationships, totaling 340 triples. Figure 3 The aircraft landing gear fault diagnosis knowledge graph (Aircraft landing gear fault diagnosis, ALGFD) is shown.
[0083] In order to conduct inductive testing on the aircraft landing gear fault diagnosis knowledge graph, a new inductive dataset ALGFD-Open is created by sampling disjoint subgraphs from the aircraft landing gear fault diagnosis knowledge graph. The information statistics of the aircraft landing gear fault diagnosis dataset ALGFD-Open are shown in Table 1.
[0084] Table 1 ALGFD-Open information statistics
[0085]
[0086] Table 2 shows the experimental results of the comparison model and the four score functions of the present invention on the ALGFD-Open dataset.
[0087] Table 2 Results on the ALGFD-Open dataset
[0088]
[0089]
[0090] From the results in Table 2, it can be seen that the DSMeta model proposed in the present invention is effective on the aircraft landing gear fault diagnosis knowledge graph, and among the four scoring functions, the various indicators of ComplEx perform best, among which the Hits@1 indicator is as high as 0.832, far exceeding the comparison model and other scoring functions, and the Hits@10 indicator is as high as 0.992, indicating that the scoring function ComplEx can best exert the prediction ability of the model on the aircraft landing gear fault diagnosis knowledge graph.
[0091] Figure 4 The curves of the change of the indicators MRR, Hits@1, Hits@3 and Hits@10 of the DSMeta model under different scoring functions in the iteration process on the ALGFD-Open dataset are shown. From the indicators, it can be clearly observed that the effect of the model using different scoring functions is different, from high to low, they are ComplEx, RotatE, TransE, DistMult, from which it can be concluded that the more complex the expression of entities and relationships, the greater the impact on the learning ability of the model, that is, the greater the improvement of the model's prediction ability.
[0092] The application of DSMeta in the knowledge graph of aircraft landing gear fault diagnosis is demonstrated through examples. Table 3 shows the completion result example. The tail entity in the example does not exist in the training graph. The table provides the top three predictions. The triples in bold are the top three predictions. Figure 5 The completed knowledge graph subgraph is shown, where entities with red edges are newly added entities and red edges are predicted edges.
[0093] Table 3. Examples of open world knowledge graph completion results
[0094]
[0095]
[0096] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.
Claims
1. A meta-learning-based knowledge graph completion method for aircraft landing gear fault diagnosis, characterized in that: The following steps are involved: S1. Use the DS-GAT encoder to encode the entities and relations in the aircraft landing gear fault diagnosis knowledge graph; S2, using GNN message passing to obtain the representation of entities and relationships in the aircraft landing gear fault diagnosis knowledge graph; S3. Predict the relationship between unknown entities through subgraph construction and meta-training to complete the knowledge graph of aircraft landing gear fault diagnosis.
2. The method for completing the knowledge graph of aircraft landing gear fault diagnosis based on meta-learning according to claim 1 is characterized in that: In S1, the calculation expressions of entities and relations are as follows: Among them, (H, R, T) is the triplet embedding representation obtained by encoding, (h BERT ,r BERT ,t BERT ) and (h R-GCN ,r R-GCN ,t R-GCN ) are the embedding representations obtained by BERT and R-GCN respectively, It is a vector concatenation operation, and CNN() represents the aggregation of two embedding representations through a convolutional neural network.
3. The method for completing the knowledge graph of aircraft landing gear fault diagnosis based on meta-learning according to claim 2 is characterized in that: In S2, GNN is used to update the representation of entities and relationships, where the entity update formula is as follows: Where h, r and t are the head entity, relation and tail entity respectively, σ is the activation function, o(e) is the set of outgoing edges and tail entities of node e, T(e) is the set of incoming edges and head entities of node e, and are the outgoing transformation matrix and incoming transformation matrix of entity e at layer l, respectively. is the linear transformation matrix of the self-loop information at the lth layer, and are the representations of entity e at layer l and layer l+1 respectively; the update formula of the relationship is as follows in, is the linear transformation matrix of relation r at level l, and are the representations of relation r at the lth layer and the l+1th layer respectively.
4. The method for completing the knowledge graph of aircraft landing gear fault diagnosis based on meta-learning according to claim 3 is characterized in that: In S3, subgraph construction includes: Convert the original training graph into an undirected graph; Select nodes of the subgraph by random walk and construct the subgraph; Randomly sort the triplets in the subgraph, and count the frequency of each entity and relationship when looping through the triplets in the subgraph; After obtaining the sorted triples, a portion of the triples are randomly selected as query triples based on the frequency of entities and relations and the size of the subgraph, and the remaining triples are used as support triples; Generate two dictionaries hr2t and rt2h, which are used to store the query triples and the related information of the supporting triples respectively.
5. The method for completing the knowledge graph of aircraft landing gear fault diagnosis based on meta-learning according to claim 4 is characterized in that: The triplets whose entity and relation frequencies are greater than 2 are selected as query triplets, and the triplets whose entity and relation frequencies are less than or equal to 2 are selected as support triplets.
6. The method for completing knowledge graph of aircraft landing gear fault diagnosis based on meta-learning according to claim 4, characterized in that: In S3, meta-training is implemented through two nested loops: the outer loop iterates the number of meta-training rounds, and each round of meta-training traverses all tasks in the dataset; the inner loop iterates each subgraph data batch.
7. The method for completing the knowledge graph of aircraft landing gear fault diagnosis based on meta-learning according to claim 6, characterized in that: In S3, the objective function of meta-training is as follows Among them, R is the embedding matrix of all relations, G i is the i-th subgraph, θ, are all learning parameters, Q i is the i-th query triple, and m represents the number of samples; A self-adversarial loss function is used on the negative sampling task of each query triple Among them, k is the number of negative samples, (h' i ,r' i ,t' i ) is the i-th negative sample, p(h' i ,r' i ,t' i ) is the self-adversarial weight of the i-th negative sample, which is calculated as follows Where β is the temperature parameter of the sample.