CN-121980279-A - Unmanned aerial vehicle engine fault diagnosis method, system, electronic equipment and storage medium

CN121980279ACN 121980279 ACN121980279 ACN 121980279ACN-121980279-A

Abstract

The invention provides a fault diagnosis method, a fault diagnosis system, electronic equipment and a storage medium for an unmanned aerial vehicle engine, and relates to the field of complex assembly fault diagnosis. The method comprises the steps of firstly realizing intelligent mapping of vibration signal characteristics and failure mechanisms through direct integration of envelope spectrum analysis and a large model, solving the problem that characteristics are extracted and failure is dependent on expert experience in the traditional method, secondly, providing structural knowledge support for a diagnosis process by a constructed multidimensional knowledge graph, enabling a system to have logic reasoning capacity close to expert level while keeping high sensitivity of data driving, and finally, combining the knowledge graph and the large model to effectively inhibit phantom risk of the large model, generating maintenance advice and effectively saving manpower. The generated envelope spectrum image analysis is integrated into a multi-mode large model, and a direct connection is established with an engine fault mechanism, so that the interpretation of the model is enhanced compared with a traditional data driving method.

Inventors

Zhan Xianbiao
WEN LIANG
JIA FAN
LIU ZIXUAN
GAO YUHANG

Assignees

北华航天工业学院

Dates

Publication Date: 20260505
Application Date: 20251206

Claims (10)

1. The unmanned aerial vehicle engine fault diagnosis method based on the multi-mode large model joint fault knowledge graph is characterized by comprising the following steps of: Step S1, collecting original vibration signals of an unmanned aerial vehicle engine; S2, analyzing the original vibration signal into an envelope spectrum image; S3, extracting feature embedding of the envelope spectrum image through a visual encoder, and performing linear transformation on the feature embedding by using a linear layer as a linear projector to obtain a visual feature vector of the envelope spectrum image; s4, acquiring text embedding of expert knowledge and text embedding of task prompts, wherein the expert knowledge is text information of physical mechanism and diagnosis rules of the engine fault summarized by the expert; step S5, splicing the text embedding of the visual feature vector and the expert knowledge and the text embedding of the task prompt to obtain a spliced vector, and performing position coding on the spliced vector to obtain a coded vector; And S6, carrying out graph reasoning on a pre-constructed fault knowledge graph based on an original vibration signal, generating an optimal relation sub-graph, and generating fault information of the engine through a multi-mode large model based on the optimal relation sub-graph and a coding vector, wherein the fault information comprises a fault reason and corresponding maintenance suggestions, the optimal relation sub-graph comprises a plurality of optimal fault chains, the fault knowledge graph comprises a theory-based sub-knowledge graph and a scene-based sub-knowledge graph, the theory-based sub-knowledge graph is constructed based on a structured theory fault diagnosis guide document, and the scene-based sub-knowledge graph is constructed based on unstructured factory fault work order data.
2. The method of claim 1, wherein the theoretical fault diagnosis instruction document includes an engine fault manual and a device specification, and the factory fault worksheet data includes real engine fault worksheet data from a manufacturer and manual experience data on site; The fault knowledge graph is constructed by the following modes: constructing a sub-knowledge map based on theory based on the theory fault diagnosis guide document; Constructing a scene-based sub-knowledge graph based on the factory fault worksheet data; and integrating the theory-based sub-knowledge graph and the scene-based knowledge graph into a unified enhanced causal fault knowledge graph.
3. The method of claim 2, wherein constructing a theory-based sub-knowledge graph based on a theory fault diagnosis guidance document comprises: converting the theoretical fault diagnosis instruction document into an image format to obtain a document image, extracting text blocks from the document image, and carrying out document layout analysis on the document image so as to keep the connection among the text blocks; identifying text blocks to obtain document text, dividing the document text into articles according to chapters or documents, and further dividing each article into chapters at paragraph level; Extracting entities and relations in a predefined mode from the articles based on the big model, and realizing text semantic mapping for each chapter through the big model; generating descriptive content according to the extracted entities and relations and text semantic mapping; based on descriptive content, a theoretical-based sub-knowledge graph is constructed.
4. The method of claim 2, wherein constructing the scenario-based sub-knowledge graph based on the factory fault worksheet data comprises: extracting target texts in tables in the factory fault worksheets, wherein the target texts comprise fault progress, cause analysis and countermeasure actions; And gradually inquiring and extracting a structured causal chain from the target text based on the causal problem prompt large model, and converting the causal chain into a fault tree format to obtain a knowledge graph based on a scene.
5. The method of claim 1, wherein generating an optimal relational subgraph based on the original vibration signal by graph inference on a pre-constructed fault knowledge graph comprises: extracting the current fault symptom characteristic of the engine from the original vibration signal; Taking the current fault symptom characteristic as an initial node, performing breadth-first search in a theoretical-based sub-knowledge graph, and collecting all associated nodes and side relations thereof to obtain an undirected graph, wherein the traversal depth is set to be 3 layers; Generating a plurality of initial fault chains by enumerating all paths between symptom nodes and potential root cause nodes based on the undirected graph; Invoking a graph neural network model to score the confidence coefficient of each initial fault chain; determining a preset number of initial fault chains with highest reliability in the plurality of initial fault chains as first candidate fault chains; calculating the similarity between the current fault symptom characteristic and each historical fault symptom characteristic in the scene-based sub-knowledge graph; determining a fault chain in which the similarity between the scene-based sub-knowledge graph and the current fault symptom feature is greater than a preset similarity threshold value as a second candidate fault chain; and determining the first candidate fault chain and the second candidate fault chain as optimal relation subgraphs.
6. The method according to claim 1, characterized in that in step S2 the original vibration signal is analyzed into an envelope spectrum image, comprising in particular: Generating an analysis signal from the original vibration signal by using Hilbert transform; calculating the amplitude of the analysis signal to obtain an envelope signal; and performing fast Fourier transform on the envelope signal to generate an envelope spectrum image.
7. The method of claim 1, wherein the multi-modal large model, linear projector, and visual encoder are jointly trained using the LoRA method.
8. Unmanned aerial vehicle engine fault diagnosis system based on multi-mode large model joint fault knowledge graph, which is characterized by comprising: the first processing module is used for collecting original vibration signals of the unmanned aerial vehicle engine; The second processing module is used for analyzing the original vibration signal into an envelope spectrum image; The third processing module is used for extracting feature embedding of the envelope spectrum image through the visual encoder, and performing linear transformation on the feature embedding by using the linear layer as a linear projector to obtain a visual feature vector of the envelope spectrum image; The fourth processing module is used for acquiring text embedding of expert knowledge and text embedding of task prompts, wherein the expert knowledge is text information of physical mechanism and diagnosis rules of the engine fault summarized by the expert; the fifth processing module is used for splicing the text embedding of the visual feature vector and the expert knowledge and the text embedding of the task prompt to obtain a spliced vector, and performing position coding on the spliced vector to obtain a coded vector; The sixth processing module is used for carrying out graph reasoning on a pre-constructed fault knowledge graph based on the original vibration signal, generating an optimal relation sub-graph, generating fault information of the engine through a multi-mode large model based on the optimal relation sub-graph and the coding vector, wherein the fault information comprises fault reasons and corresponding maintenance suggestions, the optimal relation sub-graph comprises a plurality of optimal fault chains, the fault knowledge graph comprises a theoretical sub-knowledge graph and a scene-based sub-knowledge graph, the theoretical sub-knowledge graph is constructed based on a structured theoretical fault diagnosis guide document, and the scene-based sub-knowledge graph is constructed based on unstructured factory fault work order data.
9. An electronic device, characterized in that the electronic device comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps in the unmanned aerial vehicle engine fault diagnosis method based on the multi-mode large model joint fault knowledge graph according to any one of claims 1 to 7.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program, which when executed by a processor, implements the steps in a method for diagnosing an engine failure of an unmanned aerial vehicle based on a multi-modal large model joint failure knowledge graph according to any one of claims 1 to 7.

Description

Unmanned aerial vehicle engine fault diagnosis method, system, electronic equipment and storage medium Technical Field The invention belongs to the technical field of complex assembly fault diagnosis, and particularly relates to an unmanned aerial vehicle engine fault diagnosis method, system, electronic equipment and storage medium. Background With the rapid development of modern manufacturing technology, the reliability and operational stability of mechanical systems are subject to requirements of higher standards. As a core transmission component in mechanical equipment, unmanned aerial vehicles mainly rely on the relative motion of wings and air to generate lift force, and are extremely susceptible to potential faults in long-time and large-range operation. A single engine failure may propagate to the entire system, causing multiple equipment failures, and in some cases, instability of the entire system. Therefore, accurate and timely diagnosis of unmanned aerial vehicle engine faults is critical to ensure normal operation and reliability of mechanical systems. The traditional diagnosis method relies on manual analysis of the characteristic frequency of the vibration signal, but suffers from noise interference and other problems under complex working conditions, so that the diagnosis efficiency is limited. In the data driving method, the data driving method automatically learns fault characteristics from monitoring data such as vibration, temperature and the like through algorithms such as machine learning, deep learning and the like, and the diagnosis automation level is remarkably improved by combining a signal processing technology (such as fast Fourier transform and envelope spectrum analysis) and a deep learning model (such as a convolutional neural network and a long-term and short-term memory network), but the problems of poor model interpretation and the like still exist. Disclosure of Invention In order to solve the technical problems, the invention provides an unmanned aerial vehicle engine fault diagnosis method, an unmanned aerial vehicle engine fault diagnosis system, electronic equipment and a storage medium. The invention discloses an unmanned aerial vehicle engine fault diagnosis method based on a multi-mode large model combined fault knowledge graph, which comprises the following steps: Step S1, collecting original vibration signals of an unmanned aerial vehicle engine; S2, analyzing the original vibration signal into an envelope spectrum image; S3, extracting feature embedding of the envelope spectrum image through a visual encoder, and performing linear transformation on the feature embedding by using a linear layer as a linear projector to obtain a visual feature vector of the envelope spectrum image; s4, acquiring text embedding of expert knowledge and text embedding of task prompts, wherein the expert knowledge is text information of physical mechanism and diagnosis rules of the engine fault summarized by the expert; step S5, splicing the text embedding of the visual feature vector and the expert knowledge and the text embedding of the task prompt to obtain a spliced vector, and performing position coding on the spliced vector to obtain a coded vector; And S6, carrying out graph reasoning on a pre-constructed fault knowledge graph based on an original vibration signal, generating an optimal relation sub-graph, and generating fault information of the engine through a multi-mode large model based on the optimal relation sub-graph and a coding vector, wherein the fault information comprises a fault reason and corresponding maintenance suggestions, the optimal relation sub-graph comprises a plurality of optimal fault chains, the fault knowledge graph comprises a theory-based sub-knowledge graph and a scene-based sub-knowledge graph, the theory-based sub-knowledge graph is constructed based on a structured theory fault diagnosis guide document, and the scene-based sub-knowledge graph is constructed based on unstructured factory fault work order data. According to the method of the first aspect of the invention, the theoretical fault diagnosis instruction document comprises an engine fault manual and equipment specifications, and the factory fault work order data comprises real engine fault work order data from manufacturers and on-site manual experience data; The fault knowledge graph is constructed by the following modes: constructing a sub-knowledge map based on theory based on the theory fault diagnosis guide document; Constructing a scene-based sub-knowledge graph based on the factory fault worksheet data; and integrating the theory-based sub-knowledge graph and the scene-based knowledge graph into a unified enhanced causal fault knowledge graph. According to the method of the first aspect of the invention, a theoretical sub-knowledge graph is constructed based on a theoretical fault diagnosis guide document, comprising: converting the theoretical fault diagnosis instruction documen