Aero-engine intelligent fault diagnosis and maintenance system based on domain knowledge graph and large language model

Abstract

The invention belongs to the technical field of aero-engine maintenance, and discloses an aero-engine intelligent fault diagnosis and maintenance system based on a domain knowledge graph and a large language model. The system comprises four core modules, wherein a data processing module adopts an AemCasRel model to extract a fault entity and relation triple from maintenance data; the knowledge graph management module stores the triple into a Neo4j database, constructs a structured knowledge graph and provides a management interface; the knowledge visualization module realizes visual display and interactive retrieval of the atlas based on D3. Js; the intelligent question and answer module depends on LangChain and a stellar fire big model and combines a multi-hop path reasoning technology to generate an interpretable diagnosis report. According to the system, the whole process integration of fault knowledge from extraction and management to intelligent diagnosis is realized, the precision, efficiency and intelligent level of maintenance diagnosis are remarkably improved, and the system has a wide application prospect in the fields of aviation, aerospace, machinery and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of aero-engine maintenance, and particularly relates to an intelligent fault diagnosis and maintenance system for an aero-engine based on a domain knowledge graph and a large language model. BACKGROUND

[0002] An aero-engine is a highly complex and technology-intensive core system of aviation equipment, and its operation reliability and safety are crucial. With the increasing complexity of aero-engine structure design and control logic, fault diagnosis and maintenance support work is facing unprecedented challenges. Traditional fault diagnosis relies heavily on the personal experience and knowledge of maintenance experts, and analyzes and judges non-structured or semi-structured data such as maintenance logs and manual documents. This artificially dominated mode is not only inefficient in dealing with massive, multi-source and heterogeneous data, but also limited by the expertise and mobility of experts, which can easily lead to inconsistent diagnostic conclusions, knowledge transfer difficulties and delayed maintenance responses.

[0003] In recent years, as a powerful tool for representing and organizing large-scale domain knowledge, knowledge graphs have shown potential applications in the field of aviation fault diagnosis. By constructing a structured fault knowledge network, a knowledge graph can integrate and correlate fault knowledge to some extent. However, existing methods still have significant shortcomings when constructing a knowledge graph in the field of aero-engine maintenance: first, there is a lack of efficient and accurate relationship extraction methods tailored to the characteristics of aviation maintenance field texts, making it difficult to automatically extract fault entities and their semantic relationships from complex maintenance logs, manuals and other multi-source heterogeneous texts. The construction of a knowledge graph still requires a lot of manual intervention, which is costly and inefficient; second, the constructed knowledge graph is mostly stored statically, with weak visualization and interaction capabilities, making it difficult for maintenance personnel to intuitively understand the associated path and reasoning process of the fault, limiting the depth of application of the knowledge graph in decision support; finally, existing knowledge graph systems lack deep integration with cutting-edge artificial intelligence technologies, and the knowledge retrieval method is rigid, unable to handle complex and ambiguous natural language queries, and unable to provide intelligent reasoning answers with explanations.

[0004] Therefore, the existing technology has not effectively solved the core technical problems faced by the aero-engine maintenance field: how to automatically construct a high-quality fault knowledge graph from multi-source heterogeneous data, how to realize the structured storage, dynamic visualization and interactive reasoning of knowledge, and how to deeply integrate domain knowledge with large language models to generate reliable and interpretable intelligent diagnostic conclusions. SUMMARY

[0005] Aiming at the problems existing in the prior art, the present application aims to provide an aero-engine intelligent fault diagnosis and maintenance system based on a domain knowledge graph and a large language model, which adopts a relation triple extraction method, a knowledge graph front-end visualization, and a large language model intelligent question and answer technology, realizes intelligent fault diagnosis and maintenance of an aero-engine, and meets the intelligent maintenance needs of the aero-engine maintenance field in multiple scenarios.

[0006] In order to realize the above technical purpose, the present application specifically adopts the following technical scheme:

[0007] In one aspect of the present application, an aero-engine intelligent fault diagnosis and maintenance system based on a domain knowledge graph and a large language model is provided, comprising:

[0008] A data processing module is configured to extract fault entities and relation triples from aero-engine maintenance data by an AemCasRel model based on cascaded binary tags;

[0009] A knowledge graph management module is configured to store the extracted triple data into a Neo4J graph database, construct a structured fault knowledge graph, and provide an adding, deleting, searching, and modifying interface;

[0010] A knowledge visualization module is configured to obtain fault knowledge graph data from the knowledge graph management module, and realize visualization display of the knowledge graph by a D3.js tool library, and support node interaction and multi-dimensional search functions;

[0011] An intelligent question and answer module is configured to construct an intelligent agent by a LangChain tool chain, combine multi-hop path reasoning technology and a starfire 4.0 Ultra large language model, access the fault knowledge graph data in the knowledge graph management module, and generate a diagnosis report with an interpretable reasoning chain.

[0012] In one embodiment, the data processing module comprises:

[0013] A semantic encoding unit is configured to encode input text by a MacBERT model to obtain a hidden layer vector representation;

[0014] An attention enhancement unit is configured to strengthen features of the hidden layer vector by a multi-head self-attention mechanism;

[0015] An entity recognition unit is configured to receive the strengthened hidden layer vector, predict the start and end positions of entities in the text by a binary tag marker, and determine the entity boundaries by adopting a Sigmoid activation function in the marker output layer and binarizing the Sigmoid activation function with 0.5 as a threshold value;

[0016] a relation identification unit configured to receive the head entity boundary information from the entity identification unit and the sentence encoding from the semantic encoding unit, generate a context semantic representation fused with the head entity position information by embedding the start and end positions of the head entity into the sentence encoding and performing weighted splicing, and predict the tail entity and its boundary that may exist under each predefined relation type based on the fused representation.

[0017] In an embodiment, the AemCasRel model is trained by a joint loss function composed of a weighted sum of an entity identification cross-entropy loss and a relation extraction cross-entropy loss.

[0018] In an embodiment, the fault knowledge graph defined in the knowledge graph management module includes:

[0019] Entity types: model type, system, fault cause, fault phenomenon, fault impact, fault location, inspection method, and repair method.

[0020] Relation types: contains, causes fault, possible cause, causes phenomenon, causes impact, specific location, needs inspection, and needs repair.

[0021] In an embodiment, the knowledge visualization module includes:

[0022] A data conversion unit configured to convert the subgraph data retrieved from the Neo4J graph database into JSON format through an APOC component;

[0023] A visualization rendering unit configured to receive the JSON format data, parse and realize dynamic rendering of nodes and relations through a D3.js tool library;

[0024] An interactive control unit configured to bidirectionally interact with the visualization rendering unit, receive operation instructions of a user, and control the visualization rendering unit to realize corresponding interactive effects;

[0025] A real-time updating unit having an input end connected to a monitoring interface of the knowledge graph management module and configured to listen to change events of the fault knowledge graph through a multi-thread mechanism, and having an output end connected to a trigger end of the data conversion unit, so that the data conversion unit automatically triggers the data extraction and conversion process again once a data update is detected, thereby realizing real-time refreshing of the front-end visualization content.

[0026] In an embodiment, the intelligent question answering module includes:

[0027] A semantic analysis unit configured to analyze a natural language question of a user through a LangChain tool chain;

[0028] An entity linking unit receives the entity information parsed by the semantic parsing unit, calculates the embedding similarity between the entity information and the nodes in the fault knowledge graph through an Hnswlib vector library, and outputs a Top-N candidate entity set;

[0029] A path reasoning unit receives the candidate entity set, predicts a relationship path through an attention mechanism, introduces a greedy factor exploration strategy for multi-hop reasoning, and obtains a path triple sequence;

[0030] A report generation unit receives the path triple sequence, embeds it into a preset prompt engineering template, and generates a final diagnostic report containing an interpretable reasoning chain by calling a Spark 4.0 Ultra large language model.

[0031] In an embodiment, the path reasoning unit comprises:

[0032] A path prediction sub-unit predicts the relationship type required for the next hop search based on the candidate entity set or the entity result of the previous hop reasoning through an attention weight matrix;

[0033] A query construction sub-unit constructs a corresponding Cypher query statement based on the relationship type prediction result of the path prediction sub-unit, executes the search in a Neo4J graph database, and obtains the associated next node entity;

[0034] A termination judgment sub-unit receives the search result of the query construction sub-unit and the generated path triple sequence in real time, monitors the reasoning process through a maximum hop threshold and a loop detection mechanism, and sends a control signal to the path prediction sub-unit and the query construction sub-unit to forcibly terminate invalid reasoning when the termination condition is met.

[0035] In an embodiment, the aero-engine intelligent fault diagnosis and maintenance system further comprises a feedback optimization module, the input end of which is connected to the output end of the intelligent question and answer module, for collecting the credibility feedback of the user on each hop triple in the reasoning path generated by the intelligent question and answer module; the output end of the feedback optimization module is connected to the path prediction sub-unit in the path reasoning unit, for converting the user feedback into a reward signal, and optimizing and fine-tuning the attention weight matrix in the path prediction sub-unit for predicting the relationship type of the next hop through the gradient descent method, so as to improve the accuracy of subsequent relationship path prediction.

[0036] In an embodiment, the diagnostic report generated by the aero-engine intelligent fault diagnosis and maintenance system contains fault root cause, maintenance steps, risk warning, and similar fault trend prediction.

[0037] In another aspect of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored on the memory, wherein the processor implements the functions of the intelligent fault diagnosis and maintenance system for aero-engine described above when executing the program.

[0038] The beneficial effects of the present application are:

[0039] 1) The automatic, high-precision extraction and structured storage of aero-engine fault knowledge from multi-source heterogeneous data are realized, overcoming the drawbacks of low efficiency and poor consistency of traditional manual methods.

[0040] 2) Through the visual interaction of the knowledge graph and the dynamic reasoning path display, the transparency and explainability of the fault diagnosis process are significantly improved, and the decision confidence of maintenance personnel is enhanced.

[0041] 3) The field knowledge graph and large language model are fused, and through retrieval-enhanced generation technology, the accuracy and reliability of intelligent question and answer answers are ensured, effectively avoiding the fact errors and illusion problems of traditional large models.

[0042] 4) An intelligent closed-loop system integrating knowledge extraction, management, reasoning, and application is constructed, which significantly improves the precision, efficiency, and intelligence level of aero-engine fault diagnosis, and provides an effective way for the accumulation and inheritance of maintenance knowledge. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is the construction flowchart of the intelligent fault diagnosis and maintenance system for aero-engine of the embodiment of the present application;

[0044] Figure 2 is the construction flowchart of the visual interface of the field knowledge graph of the embodiment of the present application;

[0045] Figure 3 is the retrieval-enhanced generation flowchart based on multi-hop path reasoning technology of the embodiment of the present application;

[0046] Figure 4 is the application interface of the fault knowledge graph in the field of aero-engine maintenance of the embodiment of the present application;

[0047] Figure 5 is the intelligent question and answer application interface of fault diagnosis of the embodiment of the present application. DETAILED DESCRIPTION

[0048] The technical solutions of the present application will be described clearly and completely below in conjunction with specific embodiments, but those skilled in the art will understand that the following described embodiments are part of the embodiments of the present application, not all the embodiments, and are only used to illustrate the present application, and should not be regarded as limiting the scope of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0049] In one embodiment, an intelligent fault diagnosis and maintenance system for aero-engine based on domain knowledge graph and large language model is provided, and the construction process of the system is shown in Figure 1 The data processing module is used to extract fault entity and relationship triplets from aero-engine maintenance data based on the AemCasRel model based on cascaded binary tags. The knowledge graph management module is used to store the extracted triplet data to the Neo4J graph database, construct a structured fault knowledge graph, and provide an add, delete, query and modify interface. The knowledge visualization module is used to obtain fault knowledge graph data from the knowledge graph management module, and realize the visualization display of the knowledge graph through the D3.js tool library, supporting node interaction and multi-dimensional retrieval function. The intelligent question and answer module is used to build an agent through the LangChain tool chain, combine multi-hop path reasoning technology and starfire 4.0 Ultra large language model, access the fault knowledge graph data in the knowledge graph management module, and generate a diagnosis report with an interpretable reasoning chain.

[0050] The data processing module includes a semantic encoding unit, an attention enhancement unit, an entity recognition unit and a relationship recognition unit. The semantic encoding unit is used to encode the input text through the MacBERT model to obtain a hidden layer vector representation. The attention enhancement unit is used to strengthen the features of the hidden layer vector through the multi-head self-attention mechanism. The entity recognition unit is used to receive the strengthened hidden layer vector, and predict the start and end positions of the entity in the text through a binary tag marker; the marker output layer adopts a Sigmoid activation function and is binary with 0.5 as the threshold to determine the entity boundary. The relationship recognition unit is used to receive the head entity boundary information from the entity recognition unit and the sentence encoding from the semantic encoding unit, to generate a context semantic representation fused with the head entity position information by embedding the start and end positions of the head entity and the sentence encoding for weighted splicing; based on the fused representation, the tail entity and its boundary under each predefined relationship type are predicted.

[0051] Specifically, the data processing module is constructed by the following ways:

[0052] 1.1 Constructing the knowledge graph ontology and the fault entity relationship pattern, the entity types include: aircraft model, system, fault cause, fault phenomenon, fault influence, fault location, inspection method, maintenance method, and the relationship types include: contains, causes fault, possible cause, causes phenomenon, causes influence, specific location, needs inspection, and needs maintenance; secondly, preprocessing the original text data in the maintenance records, equipment manuals, and fault diagnosis documents of the aero-engine, including sentence reorganization and word segmentation; then, labeling the entities and relationships in the text, and arranging the samples into the standardized <sentence, head entity, relationship, tail entity> data format; in the subsequent steps, the input layer, multi-head attention mechanism enhancement layer, and output layer of the binary label marker will be designed respectively.

[0053] 1.2 Semantic encoding unit: design the input layer of the binary label marker, use the pre-trained MacBERT as the semantic encoder to perform context semantic encoding representation on the fault text sentence; given the input text sequence with length N Use the embedding encoding strategy of the semantic encoder to convert the text into a d e dimensional encoding feature vector Get the hidden layer vector through the linear layer:

[0054] F=EW f +b f (1)

[0055] wherein, is the hidden layer matrix, is the linear layer trainable matrix, is the bias matrix, d f is the linear layer output dimension.

[0056] 1.3 Attention enhancement unit: for the hidden layer vector F, further input the multi-head attention layer to calculate and distribute the weight; first, calculate the query matrix Q i , key matrix K i , and value matrix V i of the i-th attention head through linear transformation, as shown in formula 2, then calculate the output weight H i of the i-th attention head, as shown in formula 3:

[0057]

[0058] wherein, is the trainable weight matrix, d kFor attention sub-space dimension, Softmax is the activation function. Then, the multi-head attention results are fused by vector concatenation, as shown in equation 4:

[0059] Z = (Concat(H1,...,H h )W z )(4)

[0060] where Z is the multi-head attention fusion matrix, is the trainable weight matrix, h is the number of attention heads, d p is the output dimension.

[0061] 1.4 Entity Recognition Unit: The attention fusion matrix is input into a linear layer, and the output matrix is calculated, as shown in equation 5:

[0062]

[0063] where, is the output probability score matrix, is the trainable weight matrix, b o is the bias value, and Sigmoid is the activation function; after obtaining the start and end position probabilities of the entity, the final division is performed by setting the score threshold value 0.5 to 0 or 1, and the calculation rule is shown in equation 6:

[0064]

[0065] where P i is the original probability score of the i-th position in the sentence, is the final output score of the i-th position in the sentence.

[0066] 1.5 Relationship Recognition Unit: In the relationship recognition unit, for each recognized entity, it is predicted whether it contains a tail entity under each relationship, i.e. for each relationship type, two identical binary label markers consistent with the entity recognition unit are set, respectively used to predict the start and end positions of the tail entity. It should be noted that, in the entity pointer network, the input is the MacBERT word embedding encoding of the original sentence, and in the relationship pointer network, the input is the MacBERT word embedding encoding of the original sentence plus the encoding of the selected head entity position, which can be determined by the rule of equation 7:

[0067]

[0068] where start and end are the start and end positions of the head entity, respectively, e i is the N-dimensional vector of the i-th token after MacBERT word embedding encoding.

[0069] 1.6 Design of joint training and loss function: the training loss function is composed of entity recognition loss and relation extraction loss; in the training process, the model optimizes the target through the cross-entropy function, and sets a proportion parameter to balance the training loss of entity and relation, further optimizes the performance of the model on fault entity relation extraction, and the calculation formula is shown as formula 8:

[0070] L = L e + βL r (8)

[0071] Wherein, L is the total training loss, L e is the cross loss predicted by the entity pointer network, L r is the cross loss predicted by the relation pointer network, and β is an adjustable parameter.

[0072] The knowledge graph management module is specifically constructed by the following ways:

[0073] 2.1 Data collection and arrangement: through the relation triple extraction method, the entity text and relation information in the maintenance record, equipment manual and fault diagnosis document of the aero-engine are extracted, and arranged into a standardized <head entity, relation, tail entity> data format.

[0074] 2.2 Design of graph database model: the types and attributes of entity nodes are designed, including model, system, fault cause, fault phenomenon, fault influence, fault position, inspection method and maintenance method; and the types and directions of relations are defined, including containing, fault generation, possible cause, phenomenon generation, influence generation, specific position, need to check and need to repair.

[0075] 2.3 Import Neo4J graph database: through the construction of Cypher statement, the arranged standardized data is imported into the graph database, which is converted into graph data suitable for Neo4J format, each entity is stored as a node, and the relation is converted into the edge in the graph; by accurately mapping the attributes of entity and relation, it is ensured that the graph database can accurately express the fault knowledge.

[0076] 2.4 Design of adding, deleting, querying and modifying interface: through the controls of PyQT button, text input box and drop-down menu, new fault nodes or relations can be conveniently added, unnecessary node data can be deleted, error node information can be modified, and specific fault node relation data can be queried.

[0077] 2.5 Update of knowledge graph: through the multi-threading design of automatic update mechanism, when new maintenance record data is imported, the fault entity and relation can be recognized by the triple extraction model AemCasRel and imported into the fault knowledge graph, so as to further enrich and perfect the content of the knowledge graph.

[0078] The knowledge visualization module includes a data conversion unit, a visualization rendering unit, an interaction control unit, and a real-time updating unit. The data conversion unit converts the subgraph data retrieved from the Neo4J graph database into JSON format through the APOC component. The visualization rendering unit receives the JSON format data, parses and implements dynamic rendering of nodes and relationships through the D3.js tool library. The interaction control unit interacts with the visualization rendering unit bidirectionally, receives user operation instructions, and controls the visualization rendering unit to achieve corresponding interactive effects. The input end of the real-time updating unit is connected to the monitoring interface of the knowledge graph management module, which is used to listen to the change events of the faulty knowledge graph through a multi-thread mechanism. The output end is connected to the trigger end of the data conversion unit. Once the data update is monitored, the data conversion unit automatically triggers the data extraction and conversion process, thereby realizing real-time refreshing of the front-end visualization content. Referring to Figure 2 The knowledge visualization module includes a data conversion unit, a visualization rendering unit, an interaction control unit, and a real-time updating unit. The data conversion unit converts the subgraph data retrieved from the Neo4J graph database into JSON format through the APOC component. The visualization rendering unit receives the JSON format data, parses and implements dynamic rendering of nodes and relationships through the D3.js tool library. The interaction control unit interacts with the visualization rendering unit bidirectionally, receives user operation instructions, and controls the visualization rendering unit to achieve corresponding interactive effects. The input end of the real-time updating unit is connected to the monitoring interface of the knowledge graph management module, which is used to listen to the change events of the faulty knowledge graph through a multi-thread mechanism. The output end is connected to the trigger end of the data conversion unit. Once the data update is monitored, the data conversion unit automatically triggers the data extraction and conversion process, thereby realizing real-time refreshing of the front-end visualization content. Referring to

[0079] 3.1 Data conversion unit: Collect user search requests from front-end controls, use the APOC (Awesome Procedures on Cypher) component of Neo4J to convert the retrieved subgraph data in the faulty knowledge graph into JSON file format and export. Support multi-dimensional query methods, such as precise search by inputting fault location and fault phenomenon, and query results are dynamically updated in the front-end interface in real time.

[0080] 3.2 Visualization rendering unit: Parse the exported JSON file through HTML, select the D3.js visualization toolkit for front-end visualization design, customize node and edge styles, including node color, size, shape, to clearly display information of different types of nodes and edges, and enable dynamic graph updating and node interaction. Graph visualization supports dynamic display of reasoning process, which can clearly display the reasoning path from fault phenomenon to fault cause, and highlight relevant nodes and paths.

[0081] 3.3 Interaction control unit: Embed web components with PyQT interface design, which can interact with graph content through clicking, dragging, zooming, expanding, and collapsing operations, and automatically update the clicked node information to the add, delete, search, and modify controls of PyQT; At the same time, it can realize the functions of node click highlighting, dynamic tracking of adjacent paths, and detailed information display of relationships.

[0082] 3.4 Real-time updating unit: adopt multi-thread timing to retrieve the content of the backend database, and automatically refresh the atlas content displayed in the front end when the backend database changes. At the same time, through the design of a notification mechanism, when important changes occur in the atlas data, the system pushes a reminder to the user, ensuring that the user can obtain the latest fault warning and diagnosis result information in a timely manner.

[0083] The intelligent question and answer module includes a semantic analysis unit, an entity linking unit, a path reasoning unit, and a report generation unit. The semantic analysis unit is configured to analyze a user's natural language question by using a LangChain tool chain. The entity linking unit receives entity information analyzed by the semantic analysis unit, calculates embedding similarity between the entity information and nodes in a fault knowledge graph by using a Hnswlib vector library, and outputs a Top-N candidate entity set. The path reasoning unit receives the candidate entity set, predicts a relationship path by using an attention mechanism, and obtains a path triple sequence by using a multi-hop reasoning and a greedy factor exploration strategy. The report generation unit receives the path triple sequence, embeds the path triple sequence into a preset prompt engineering template, and generates a final diagnosis report including an explainable reasoning chain by calling a Starfire 4.0 Ultra large language model.

[0084] The path reasoning unit includes a path prediction subunit, a query construction subunit, and a termination judgment subunit. The path prediction subunit predicts a relationship type required for next-hop retrieval based on the candidate entity set or an entity result of last-hop reasoning by using an attention weight matrix. The query construction subunit constructs a corresponding Cypher query statement based on a relationship type prediction result of the path prediction subunit, executes retrieval in a Neo4J graph database, and obtains an associated next node entity. The termination judgment subunit receives a retrieval result of the query construction subunit and a generated path triple sequence in real time, monitors a reasoning process by using a maximum hop threshold and a loop detection mechanism, and sends a control signal to the path prediction subunit and the query construction subunit to forcibly terminate invalid reasoning when a termination condition is met.

[0085] Referring to Figure 3 The knowledge graph is constructed in the following manner:

[0086] 4.1 Semantic analysis unit: an intelligent agent is constructed by using a LangChain tool chain to analyze the semantics of a user's question and dynamically determine whether the domain knowledge graph needs to be retrieved; if the intelligent agent determines that the domain knowledge graph needs to be retrieved, step 4.2 is entered, otherwise, step 4.4 is directly entered.

[0087] 4.2 Entity linking unit: first, identify fault entities, and identify and extract a fault entity set {e1, e2,..., en} in a user's question by using the AemCasRel model trained in step 1.k Secondly, the initial node is located, the similarity of the fault entity and the embedding of the knowledge graph node is calculated using the Hnswlib vector retrieval library, and the Top-N candidate nodes are extracted as the initial nodes of the reasoning chain. The similarity calculation formula of the entity and the embedding of the knowledge graph node is defined as:

[0088]

[0089] where e is the extracted fault entity embedding, v j is the jth knowledge graph node embedding.

[0090] 4.3 Path reasoning unit: first, for each reasoning chain, the entities of the historical path and the vector of the user question are fused, the attention weight matrix is input, and the relationship type path required for the next hop is predicted, and the relationship prediction is shown in formulas 10-11:

[0091] z t =Attn(h1+···+h t +q) (10)

[0092] P(r t+1 |z t )=Softmax(W r z t +b r ) (11)

[0093] where h t is the entity embedding vector at step t, q is the user question vector, Attn is the attention weight matrix, z t is the hidden layer, W r is the linear layer weight matrix, b r is the linear layer bias, Softmax is the activation function, P(r t+1 |z t ) is the probability distribution of the next hop relationship type r t+1 ; secondly, the greedy factor ζ selection strategy is introduced to randomly explore non-optimal relationship types with probability ζ and select the maximum probability relationship type with probability 1-ζ, to enhance the exploration ability and robustness of the reasoning chain; finally, according to the selected relationship type, a Cypher query statement is constructed and queried to obtain the associated next node e t+1 , and the retrieved triple (e t , r, e t+1 ) is rewritten into a natural language sentence through templating to enhance the explanation ability of the large model.

[0094] Graph path reasoning iteration, integrating user questions and reorganized sentences under the reasoned path, inputting the large language model, and outputting the decision signal:

[0095]

[0096] wherein Q is the user question, Y i is the reorganized natural sentence at the i-th hop, f LLM is the large language model entry, S t is the end inference termination signal, and G is the prompt sentence of the output content specification. When it is determined that the inference is ended, go to step 4.4, otherwise, go back to step 4.3 to continue predicting the relationship path of the next hop; at the same time, set the path complexity constraint, and when the number of hops is greater than 5 or a loop occurs, then forcibly terminate to avoid invalid inference.

[0097] 4.4 Report generation unit: build a prompt engineering template, set the role of the large language model as a fault diagnosis expert, and require the answer content to contain fault root cause, maintenance steps, and risk warning; if the system retrieves the path result from the graph database is not empty, then embed the triples of the path reasoning unit as factual basis into the prompt engineering template; if the retrieval path result is empty, then do not add triples, and add a low confidence explanation mark, prompting the user that the answer may have uncertainty. The construction of the structured prompt engineering template is shown in formula 13:

[0098]

[0099] wherein Q is the user question, is the retrieved triple, C is the expert role setting, E is the output content specification and explanation, and ψ is the confidence parameter.

[0100] Call the large language model Starfire 4.0 Ultra, input the prompt engineering template containing the user question context into the large language model, realize retrieval enhanced generation, and generate a diagnosis report containing an interpretable reasoning chain, maintenance steps, and similar fault trend prediction.

[0101] The aviation engine intelligent fault diagnosis and maintenance system further comprises a feedback optimization module, an input end of which is connected to an output end of the intelligent question and answer module, for collecting user feedback on the credibility of each hop triple in the reasoning path generated by the intelligent question and answer module; an output end of the feedback optimization module is connected to a path prediction subunit in the path reasoning unit, for converting the user feedback into a reward signal, and optimizing and fine-tuning the attention weight matrix for predicting the relationship type of the next hop in the path prediction subunit through the gradient descent method, so as to improve the accuracy of subsequent relationship path prediction.

[0102] The intelligent fault diagnosis and maintenance system for aero-engines also includes a feedback optimization module. Its input is connected to the output of the intelligent question-answering module, used to collect user feedback on the credibility of each hop triple in the reasoning path generated by the intelligent question-answering module. The output of the feedback optimization module is connected to the path prediction subunit in the path reasoning unit, used to convert user feedback into a reward signal, and to optimize and fine-tune the attention weight matrix used to predict the next hop relationship type in the path prediction subunit using gradient descent, thereby improving the accuracy of subsequent relationship path prediction. Specifically, it is constructed in the following way:

[0103] After reasoning, the user can mark each hop in the reasoning chain of the answer as "credible / unreliable." The system converts the feedback into a reward signal R, triggering graph retrieval path optimization to train the attention weight matrix for relation classification prediction in step 4.3. Simultaneously, a path reward decay factor is introduced, causing the reward score to decrease and accumulate with each path. The attention weight matrix is ​​trained using backpropagation with a cross-entropy loss function, achieving adaptive evolution and forming a dynamic optimization loop of "human-machine co-training" to improve the system's long-term performance. The reward score calculation formula is as follows:

[0104]

[0105] Where, r t =1 indicates "trustworthy", r t =-1 indicates "unreliable", and γ∈(0,1) is the decay factor.

[0106] Finally, after implementing the above scheme, refer to Figure 4 As shown, the domain knowledge graph visualization interface can effectively display fault entities and relationship knowledge from multi-source heterogeneous data, and realize various interactive functions and interfaces, intuitively displaying fault associations and reasoning paths, significantly improving the application efficiency of the domain knowledge graph; at the same time, referring to Figure 5 As shown, the system integrates a large language model and combines it with a fault knowledge graph in the field of aero-engine maintenance. This enables enhanced retrieval and generation, allowing for accurate analysis of fault phenomena and generation of solutions in multiple scenarios. It can provide an efficient application tool for fault diagnosis and maintenance in the field of aero-engine maintenance.

[0107] Although the embodiments of this application have been described above in conjunction with the accompanying drawings, this application is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this application, and these are all within the scope of protection of this application.

Claims

1. An aero-engine intelligent fault diagnosis and maintenance system based on a domain knowledge graph and a large language model, characterized in that, The method comprises the following steps: a data processing module for extracting fault entities and relationship triples from aero-engine maintenance data based on a cascaded binary tag-based AemCasRel model; a knowledge graph management module for storing the extracted triple data into a Neo4J graph database, constructing a structured fault knowledge graph, and providing an add, delete, query, and modify interface; a knowledge visualization module for obtaining fault knowledge graph data from the knowledge graph management module and realizing visual display of the knowledge graph through a D3.js tool library, supporting node interaction and multi-dimensional search functions; an intelligent question-answering module for constructing an intelligent agent through a LangChain tool chain, accessing the fault knowledge graph data in the knowledge graph management module by combining multi-hop path reasoning technology and a Starfire 4.0 Ultra large language model, and generating a diagnostic report with an interpretable reasoning chain.

2. The intelligent fault diagnostic maintenance system for aero-engines according to claim 1, characterized in that, The data processing module comprises: a semantic encoding unit for encoding input text to obtain hidden layer vector representation through a MacBERT model; an attention enhancement unit for feature strengthening of the hidden layer vector through a multi-head self-attention mechanism; an entity recognition unit for receiving the strengthened hidden layer vector, predicting the start and end positions of entities in the text through a binary tag marker, and determining the entity boundaries by using a Sigmoid activation function and binarizing the output layer at a threshold of 0.5; a relationship recognition unit for receiving the head entity boundary information from the entity recognition unit and the sentence encoding from the semantic encoding unit, generating a context semantic representation fused with the head entity position information by weighting and splicing the start and end positions of the head entity and the sentence encoding, and predicting the tail entity and its boundaries under each predefined relationship type based on the fused representation.

3. The intelligent fault diagnostic maintenance system for aero-engines according to claim 2, characterized in that, The AemCasRel model is trained through a joint loss function composed of a weighted sum of entity recognition cross-entropy loss and relationship extraction cross-entropy loss.

4. The intelligent fault diagnostic maintenance system for aero-engines according to claim 1, characterized in that, The fault knowledge graph defined in the knowledge graph management module comprises entity types and relationship types. Entity types: model, system, fault cause, fault phenomenon, fault impact, fault location, inspection method, and repair method; Relationship types: contains, causes fault, possible cause, produces phenomenon, produces impact, specific location, needs inspection, and needs repair.

5. The intelligent fault diagnostic maintenance system for aero-engines according to claim 1, characterized in that, The knowledge visualization module comprises: a data conversion unit for converting subgraph data retrieved from the Neo4J graph database into JSON format through an APOC component; a visualization rendering unit for receiving the JSON format data, parsing and realizing dynamic rendering of nodes and relationships through a D3.js tool library; an interactive control unit for bidirectional interaction with the visualization rendering unit, receiving user operation instructions, and controlling the visualization rendering unit to realize corresponding interactive effects. A real-time updating unit, an input end of which is connected to a monitoring interface of the knowledge graph management module, is configured to listen to a change event of the fault knowledge graph through a multi-thread mechanism, and an output end of which is connected to a trigger end of the data conversion unit, so that the data conversion unit is triggered to re-execute a data extraction and conversion process and refresh the front-end visual content in real time once a data update is monitored.

6. The intelligent fault diagnostic maintenance system for aero-engines according to claim 1, characterized in that, The intelligent question and answer module comprises: A semantic analysis unit configured to analyze a natural language question of a user through a LangChain tool chain; An entity linking unit configured to receive entity information analyzed by the semantic analysis unit, calculate embedding similarity between the entity information and nodes in the fault knowledge graph through an Hnswlib vector library, and output a Top-N candidate entity set; A path reasoning unit configured to receive the candidate entity set, predict a relationship path through an attention mechanism, and introduce a greedy factor exploration strategy for multi-hop reasoning to obtain a path triple sequence; A report generation unit configured to receive the path triple sequence, embed the path triple sequence into a preset prompt engineering template, and generate a final diagnostic report containing an explainable reasoning chain by calling a Starfire 4.0 Ultra large language model.

7. The intelligent fault diagnostic maintenance system for aero-engines according to claim 6, characterized in that, The path reasoning unit comprises: A path prediction subunit configured to predict a relationship type required for searching in a next hop based on the candidate entity set or an entity result of a previous hop reasoning through an attention weight matrix; A query construction subunit configured to construct a corresponding Cypher query statement based on a relationship type prediction result of the path prediction subunit, perform searching in a Neo4J graph database, and obtain an associated next node entity; A termination judgment subunit configured to receive a searching result of the query construction subunit and a generated path triple sequence in real time, monitor the reasoning process through a maximum hop threshold and a loop detection mechanism, and send a control signal to the path prediction subunit and the query construction subunit to forcibly terminate invalid reasoning when a termination condition is met.

8. The intelligent fault diagnostic maintenance system for aero-engines according to claim 1, characterized in that, The intelligent fault diagnosis and maintenance system for an aero-engine further comprises a feedback optimization module, an input end of which is connected to an output end of the intelligent question and answer module, and which is configured to collect a credibility feedback of each hop triple in a reasoning path generated by the intelligent question and answer module; An output end of the feedback optimization module is connected to the path prediction subunit in the path reasoning unit, and is configured to convert the user feedback into a reward signal, and optimize and fine-tune an attention weight matrix in the path prediction subunit for predicting a relationship type in a next hop through a gradient descent method, so as to improve accuracy of subsequent relationship path prediction.

9. The intelligent fault diagnostic maintenance system for aero-engines according to claim 1, characterized in that, The diagnostic report generated by the intelligent fault diagnosis and maintenance system for an aero-engine contains a fault root cause, a maintenance step, a risk warning, and a similar fault trend prediction.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, is arranged to perform the method of any one of claims 1 to 9. The processor implements the functions of the intelligent fault diagnosis and maintenance system for an aero-engine in any one of claims 1 to 9 when executing the program. The processor implements the functions of the intelligent fault diagnosis and maintenance system for an aero-engine in any one of claims 1 to 9 when executing the program.

Images