CN-121980907-A - Spacecraft autonomous health diagnosis method based on graph neural network

CN121980907ACN 121980907 ACN121980907 ACN 121980907ACN-121980907-A

Abstract

The invention discloses a spacecraft autonomous health diagnosis method based on a graph neural network, which comprises the steps of using a spacecraft architecture directional graph to characterize the spacecraft architecture, preprocessing sample telemetry data to obtain training data, using the graph annotation force network to characterize hidden dependency relationships in space among different telemetry variables, constructing and obtaining a graph attention model, using the graph time sequence neural network to characterize dependency in time sequence of different telemetry variables, constructing and obtaining a graph time sequence neural network model, carrying out fusion processing to obtain a spacecraft autonomous health diagnosis model, and realizing real-time prediction of the spacecraft health condition through the spacecraft autonomous health diagnosis model. According to the method, the graph neural network is applied to autonomous health diagnosis of the spacecraft, so that the problem of how to use the graph structure to represent the on-orbit running state of the spacecraft is solved, and on the other hand, the time sequence part with larger influence can be distinguished on the basis of representing key equipment with larger influence on the health state of the spacecraft.

Inventors

MIAO ZHUANG
LI XIAOJUAN
MA JINYUAN
ZHAO WEN
JIANG YU
YE HE
WANG ZHUO
LI WENDONG
HU QIYUAN

Assignees

中国空间技术研究院

Dates

Publication Date: 20260505
Application Date: 20251212

Claims (10)

1. The autonomous spacecraft health diagnosis method based on the graph neural network is characterized by comprising the following steps of: Constructing a model data layer of a spacecraft architecture according to the information flow dependence of the spacecraft and the physical dependence among the single-machine devices of the system, and using a directed graph of the spacecraft architecture to graph the spacecraft architecture; preprocessing sample telemetry data to obtain training data, and completing construction of a model feature layer; Inputting a spacecraft architecture directed graph and training data as a graph attention network, and using the graph attention network to represent hidden dependency relations in space among different telemetry variables to construct a graph attention model; Taking a spacecraft architecture directed graph and training data as inputs of a graph time sequence neural network, using the graph time sequence neural network to represent dependence on different telemetry variable time sequences, and constructing to obtain a graph time sequence neural network model; carrying out fusion processing on the graph annotation force model and the graph time sequence neural network model, and constructing and training to obtain an autonomous health diagnosis model of the spacecraft; And taking the acquired real-time telemetry data of the spacecraft as input of an autonomous health diagnosis model of the spacecraft, outputting a real-time health condition prediction result of the spacecraft through the autonomous health diagnosis model of the spacecraft, and completing real-time prediction of the health condition of the spacecraft.
2. The spacecraft autonomous health diagnosis method based on the graph neural network, which is disclosed by claim 1, is characterized in that a spacecraft architecture directed graph consists of a plurality of nodes and a plurality of edges, wherein the nodes represent single-machine equipment on the spacecraft, and the edges represent information flow directions among the single-machine equipment.
3. The autonomous health diagnosis method for a spacecraft based on the graph neural network according to claim 2, wherein for the spatial position among the single-machine devices of the spacecraft, the single-machine devices which are spatially positioned on the same cabin board or adjacent to the cabin board are placed at the positions of connection or one-hop connection in the spacecraft architecture directed graph, and the single-machine devices which are spatially positioned on the non-adjacent cabin board are placed at the positions of multi-hop connection in the spacecraft architecture directed graph.
4. The spacecraft autonomous health diagnosis method based on the graph neural network according to claim 1, wherein preprocessing sample telemetry data to obtain training data, completing the construction of a model feature layer, comprises the following steps: collecting historical telemetry data before and after a fault period as sample telemetry data; Screening the sample telemetry data, removing the historical telemetry data with the fault time longer than seven days, and obtaining screened historical telemetry data; Marking the screened historical telemetry data according to different single-machine equipment according to the type of the fault, and constructing to obtain a spacecraft fault grade label set Simultaneously, according to the telemetry value of each single machine device in the spacecraft architecture directed graph, each label is determined Corresponding characteristic data Constructing and obtaining a spacecraft characteristic data set Wherein, the label set Each label in (a) The numerical value of the fault is set according to the level of the fault, and the higher the level of the fault is The greater the number of (2); Is one of A vector of the magnitude of the vector, Indicating the number of telemetry values that are to be counted, Representing the length of a telemetry value characterization vector, subscripts Represent the first A single unit device; Label(s) Corresponding characteristic data Randomly dividing according to the proportion of 8:2, and constructing to obtain a training data set And a test dataset Wherein, the method comprises the steps of, A set of feature data for training is represented, A failure level label set for representing training; A set of characteristic data for a test is represented, A set of fault class labels for the test is represented.
5. The autonomous health diagnosis method of a spacecraft based on a graph neural network according to claim 4, wherein the graph semantic force network is based on a spatial neighborhood relation, pair sets The training feature data in the training sequence are aggregated to obtain an aggregated airspace-dependent training feature data set 。
6. The spacecraft autonomous health diagnosis method based on graph neural network according to claim 5, wherein the graph time sequence neural network is based on time sequence neighborhood relation, pair sets The training feature data in the training data are aggregated to obtain an aggregated time-domain dependent training feature data set 。
7. The spacecraft autonomous health diagnosis method based on the graph neural network according to claim 6, wherein the graph time sequence neural network comprises a graph learning module, a graph convolution module and a time sequence convolution module, wherein the graph learning module is used for adaptively capturing spatial relations among variables from time sequence data, the graph convolution module is used for integrating node and neighbor node information, and the time sequence convolution module is used for extracting high-dimensional time sequence features by using one-dimensional expansion convolution kernels of a plurality of standards.
8. The spacecraft autonomous health diagnosis method based on the graph neural network according to claim 6, wherein the fusion processing is performed on the graph annotation force model and the graph time sequence neural network model, and the spacecraft autonomous health diagnosis model is constructed and trained and obtained, and the method comprises the following steps: Outputting graph meaning force network by using vector stitching And graph time sequence neural network output Vector addition is carried out to obtain a spacecraft autonomous health diagnosis model integrating training characteristic data of time domain and space domain dependence, and the input of the spacecraft autonomous health diagnosis model is And Time domain and space domain dependent training characteristic data set obtained by vector addition Output as health diagnosis result ; According to And (3) with The difference value between the two is used as a loss function by using a cross entropy loss function, and gradient descent training is carried out on the autonomous health diagnosis model of the spacecraft by using an Adam optimizer; Using 、 And verifying the prediction result of the trained spacecraft autonomous health diagnosis model to obtain a final optimal spacecraft autonomous health diagnosis model.
9. The spacecraft autonomous health diagnosis method based on the graph neural network according to claim 1, wherein the acquired spacecraft real-time telemetry data is used as input of a spacecraft autonomous health diagnosis model, the spacecraft real-time health prediction result is output through the spacecraft autonomous health diagnosis model, and the real-time prediction of the spacecraft health is completed, and the method comprises the following steps: Constructing a real-time feature data set based on acquired spacecraft real-time telemetry data ; Will be As the input of the spacecraft autonomous health diagnosis model, the spacecraft autonomous health diagnosis model is used for predicting the health condition of the spacecraft in real time, and the predicted health state of the spacecraft is output And 3, the real-time prediction of the health condition of the spacecraft is completed.
10. The method for autonomous health diagnosis of a spacecraft based on a graph neural network of claim 1, further comprising deploying a spacecraft autonomous health diagnosis model and a spacecraft architecture directed graph onto the spacecraft.

Description

Spacecraft autonomous health diagnosis method based on graph neural network Technical Field The invention belongs to the technical field of autonomous health diagnosis of spacecrafts, and particularly relates to a autonomous health diagnosis method of a spacecraft based on a graph neural network. Background In recent years, the spacecraft autonomous health management method is concentrated on the health data acquisition judgment and on-orbit treatment aiming at physical telemetry data and a digital model, and along with the continuous increase of the complexity of a spacecraft system and the continuous increase of on-orbit processing capability, the traditional autonomous health management method cannot meet the requirements of intellectualization and autonomy. In order to improve the autonomy and the intelligent level of the spacecraft, reduce the dependence on a ground system, learn the fault rule from the history telemetry data by utilizing the mode recognition capability of the neural network, and realize the automatic detection, diagnosis and prediction of the fault, thus becoming a hotspot problem for the research of the autonomous health field of the spacecraft. However, currently, the neural network is used for monitoring the health of the spacecraft, and network types such as Convolutional Neural Network (CNN), cyclic neural network (RNN) and the like are adopted. Such networks are difficult to fuse multi-modal data of a spacecraft, and have natural disadvantages in terms of the capability of characterizing the complex system structural composition of a spacecraft. Therefore, how to use a neural network for autonomous health diagnosis of a spacecraft is a problem to be solved. Disclosure of Invention The invention solves the technical problems of overcoming the defects of the prior art, providing a spacecraft autonomous health diagnosis method based on a graph neural network, which is applied to spacecraft autonomous health diagnosis, on one hand solving the problem of how to use a graph structure to represent the on-orbit running state of a spacecraft, wherein the graph structure should keep the information dependence, the space dependence and other relations among single-machine equipment of the spacecraft as much as possible, on the other hand, designing a graph neural network structure for graph structure type data to perform pattern learning, and on the basis of being capable of representing key equipment with larger influence on the health state of the spacecraft, distinguishing time sequence parts with larger influence. In order to solve the technical problems, the invention discloses a spacecraft autonomous health diagnosis method based on a graph neural network, which comprises the following steps: Constructing a model data layer of a spacecraft architecture according to the information flow dependence of the spacecraft and the physical dependence among the single-machine devices of the system, and using a directed graph of the spacecraft architecture to graph the spacecraft architecture; preprocessing sample telemetry data to obtain training data, and completing construction of a model feature layer; Inputting a spacecraft architecture directed graph and training data as a graph attention network, and using the graph attention network to represent hidden dependency relations in space among different telemetry variables to construct a graph attention model; Taking a spacecraft architecture directed graph and training data as inputs of a graph time sequence neural network, using the graph time sequence neural network to represent dependence on different telemetry variable time sequences, and constructing to obtain a graph time sequence neural network model; carrying out fusion processing on the graph annotation force model and the graph time sequence neural network model, and constructing and training to obtain an autonomous health diagnosis model of the spacecraft; And taking the acquired real-time telemetry data of the spacecraft as input of an autonomous health diagnosis model of the spacecraft, outputting a real-time health condition prediction result of the spacecraft through the autonomous health diagnosis model of the spacecraft, and completing real-time prediction of the health condition of the spacecraft. In the spacecraft autonomous health diagnosis method based on the graph neural network, the spacecraft architecture directed graph consists of a plurality of nodes and a plurality of edges, wherein the nodes represent single-machine equipment on the spacecraft, and the edges represent information flow directions among the single-machine equipment. In the autonomous health diagnosis method of the spacecraft based on the graph neural network, the single-machine equipment which is positioned on the same cabin board or adjacent cabin boards in the spatial position is placed at a connected or one-hop connected position in the spacecraft architecture directed graph, and the single-machi